Experimental Syntax

:perinienta
jyitax
Experimenta
Syntax
Applying Objective Methock
to Sentence Judgments
Wayne Cowart
I SAGE Publications
I International Educational and Professional Publisher
Thousand Oaks London New Delhi
Copyright © 1997 by Sage Publications, Inc.
All rights reserved. No part of this book may be reproduced or utilized in any form or by
any means, electronic or mechanical, including photocopying, recording, or by any
information storage and retrieval system, without permission in writing from the pubHsher.
For information address:
SAGE Publications, Inc.
2455 Teller Road
Thousand Oaks. California 91320
E-mail: order@sagepub.com
SAGE Publications Ltd
1 Oliver's Yard
55 City Road
London ECIY ISP
SAGE Publications India Pvt Ltd
B-42, Panchsheel Enclave
Post Box 4 1 0 9
New Delhi 110 017
Printed in the United States of America
Library of Congress Cataloging-in-Publication Data
Cowart, Wayne.
Experimental syntax: Applying objective methods to sentence
judgments / author, Wayne Cowart.
p. cm.
Includes bibliographical references and index.
ISBN 0-7619-0042-X (alk. paper).—ISBN 0-7619-0043-8 (pbk.:
alk. paper)
1. Grammar, Comparative and general—Syntax. 2. Linguistics—
Methodology. 3. Acceptability (Linguistics) I. Title.
P291.C68 1996
415—dc20 96-35620
97 98 99 00 01 02 03 10 9 8 7 6 5 4 3 2 1
Acquiring Editor: Alex Schwartz

Editorial Assistant: Eileen Carr
Production Editor: Diana E. Axelsen
Production Assistant: Denise Santoyo
Typesetter ά Designer: Andrea D. Swanson
Indexer: Virgil Diodato
Cover Designer: Ravi Balasuriya
Contents
Preface viii
1. Introduction: Are Judgments Stable? 1

1.1. Doubts About the Stability of Judgments 3
1.2. Grammatical Theory and the Measurement
of Acceptability 7
1.3. Evidence of Stability Within Populations 12
1.4. Stability of Responses to Individual Sentences 22
1.5. Outline of the Book 27
2. Error Variance in Sentence Judgments 31
3. Designing Experiments on Acceptability 39

3.1. Variance and Variance Partitioning 41
3.2. Experiment Design 44
3.3. Designing and Constructing Sentence Materials 46
3.4. Controlling Context 51
4. The Sentence Judgment Task 55
5. Presenting Sentence Materials to Informants 63
6. Response Methods and Scaling Issues 67

6.1. Category Scale Methods 70
6.2. Ratio Scale Methods 72
7. Sampling 79
7.1. Representativeness 79
7.2. Linguist Informants 81
7.3. Sample Size 81
7.4. Comparing Groups 84
8. Settings for Experiments 85

8.1. Survey Experiments 86
8.2. Laboratory Experiments 88
8.3. Field Settings 88
9. The Organization and Construction of
Questionnaires 89
9.1. General Instructions and Demographic Data 90
9.2. Response Training and Practice 90

9.3. Judgment Criteria 91
9.4. Practice and Benchmark Materials 91
9.5. The Main Sentence List 92
9.6. Constructing Questionnaires 96
10. Coding and Decoding the Data 103

10.1. Scarming 104
10.2. Patching, Parsing, Decoding, and Sorting
Scanned Data 105

10.3. Keying and Verifying Unscannable Responses 110
11. Summarizing the Data 111

11.1. By-Ir\formants Summaries 112
11.2. By-Materials Summaries 114
11.3. Summaries for Filler and Benchmark Data 116
12. StatisHcal Issues 119

12.1. Alternative Tests for Category Scale Data 120
12.2. Testing Generalizations Across Informants and
Token Sets 121

12.3. Variation Accounted For 122
12.4. Getting Traiiung or Assistance in Statistics 126
Appendbc A: A Reader's Guide to Statistics 129

A.l Descriptive Tools 129
A.2 hiferential Statistics 131
Appendix B: Statistical Supplement to Chapter 1 137
B.l. Subjacency 137

B.2. "That"-Trace 138
B.3. Coordination and Binding Theory 139
B.4. Stability of Responses to Individual Sentences 139
Appendbc C: Excel as a Syntactician's Workbench 141
C.l. Accessing Software Tools and Sample Files 142
C.2. Building Questionnaires (Chapter 9) 143
C.3. Coding/Decoding Data (Chapter 10) 153
C.4. Summarizing Data (Chapter 11) 158

C.5. Transferring Data to a Statistical Program 161
Appendbc D: Token Set Data From a "That"-Trace
Experiment 163
Appendix E: Sample Questionnaire for Scannable

Line Drawing 167
References 175
Author Index 179
Subject Index 181
About the Author 187

Preface
This book provides an introduction to objective methods by which

linguists and others may describe patterns of sentence acceptability
in speech communities. An investigator with a well-formed question
about a matter of fact relative to sentence acceptability is shown how
to design, execute, and analyze an appropriate experiment. The book
is augmented by software tools relevant to various aspects of the
preparation, execution, and analysis of survey experiments, plus a
number of example files. These materials are provided by way of the
World Wide Web site maintained by the Linguistics Program at the
University of Southern Maine, at http:/www.usm.maine.edu/~lin.
The book is intended primarily for those whose background in
syntactic theory is more extensive than their backgrovmd in experi
mentation and statistical analysis. It should be useful to a range of
people from seasoned syntactic researchers who have not previously
used experimental methods to students who are new to both syntac
tic research and experimental methods. Those who come to the book
with a strong background in experimental work will find many
passages can be skimmed without loss. Although the book will
discuss a variety of statistical issues, it is not intended to be a text in
statistics or general research design. Rather, the aim is to show how
mostly commonplace principles of these disciplines can be brought
to bear on experiments on sentence judgments. The text will point
out some statistical topics and sources along the way but will discuss
statistical issues only insofar as necessary to specify and justify
appropriate methods for experiments on sentence judgments.
The main thread of the discussion will present all the theoretical
and practical issues an investigator needs to be aware of to design
vjji
Preface ix
and execute meaningful experiments on judgments of acceptability.

Two kinds of supplementary material appear in appendixes at the
back of the book. First, although Chapter 1 reports on a number of
experimental findings, the chapter is meant to be approachable for
readers with no prior backgroimd in statistics. Thus only informal
accounts of the results appear in the main text of Chapter 1. More
detailed statistical reporting on the results presented in this chapter
appears in Appendix B. Second, the main discussion in the book will
not deal with software tools and techniques relevant to executing the
methods described. Some readers famiUar with word processing and
spreadsheet software will be able to work their way through the
process on their own. Those who want more specific guidance on
how to use software tools to carry out the various tasks that constitute
the preparation, execution, and analysis of an experiment will find
relevant discussion in Appendix C.
The term objective methods in the title of the book perhaps
deserves brief comment. Sentence judgments appear to arise within
human subjective experience. Although there seems to be a threat of
controversy at every turn in the discussion of consciousness, I will
assume the classical view that consciousness is a domain of private
experience closed to all external observers. How then can there be
objective methods for the study of a subjective phenomenon? Phi
losophers of science noticed some time back that so-called objective
observations were themselves subjective experiences, not fundamen
tally different than any other sort of subjective experience.^ Rather,
what distinguishes those experiences we call objective observations
from others is that they are those on which we can achieve intersub
jective agreement. In other terms, these are the observations where
different observers can reliably agree on what counts as an appropri
ate linguistic expression of the observation (e.g., "The meter is read
ing 3.7." "The reddish-brown layer is below the whitish one." "This
patient is aphasic"). From this vantage point, it requires only a
relatively small step to get to a notion of "objectivity" relevant to the
study of sentence judgments.
Although the immediate evanescent experience of a sentence,
including the sense that it is, or is not, acceptable, seems to be
inescapably private, the communicative integrity of linguistic com
munities rests, by hypothesis, on shared grammars. That is, we
expect that different members of a given linguistic community will
be led to similar experiences of a very wide range of sentences by a
Xii EXPERIMENTA
EXPERIMENTAL SYNTAX
grammar that they hold, more or less, in common. In terms of

Chomsky's notion of I-language (Chomsky, 1986), being a function
ing member of a linguistic commuruty depends upon achieving some
degree of (rather intricate) alignment between one's own imique,
private, I-language and the typical or normal patterns of the I-lan
guages of other members of one's commimity. Indeed, each speaker's
I-language can be seen as an estimate of the grammatical norms of
the community (or communities) within which the individual ac
quired language and currently uses it, although this estimate is
obviously subject to distortion by a variety of factors operating on
the acquisition process and the generally slower and perhaps less
orderly processes of grammatical change in adults.
Thus there is an object of linguistic inquiry that becomes ap
proachable via intersubjective methods. This is the standard or aver
age pattern of judgments across sets of related sentences within a
coherent linguistic community. By asking for judgments from many
individuals, we can uncover typical patterns of judgments within
commimities; these patterns should bear on theories of the grammars
that shape those judgments. Variants of these methods may also be
useful in providing reliable descriptions of the patterns of response
typical of individual speakers, but as we'll see in Chapter 2, the
problem of reUably characterizing individuals is far harder than
detecting differences in judgments across sentence types.
The leap from objective observations of the external world to
intersubjective observations of the inner world is not so great as it first
appears. In the classical objective case, we assimie that there is some
external reality that governs and constrains, and thus accotmts for, the
commonalities in the reported subjective experiences (i.e., the observa
tions) of different observers. When we apply intersubjective methods to
the study of human grammars, we assimê that there is some inner
reality in the brains of the speakers we test that governs and constrains,
and thus accounts for, the commonalities and structure in the reported
experiences of each individual speaker, and also that similarities across
sjjeakers are explained by similarities in this inner reality.
Acknowledgments
The project out of which this work arises has been in develop
ment for some time and has benefited from comments and sugges
Preface xi
tions from many sources. My greatest debt is to my colleague Dana

McDaniel, who has collaborated extensively on some of the experi
mental work reported here and has contributed in innumerable ways
to the project as a whole. Helen Cairns has commented ever so faithfully
and usefully on many drafts of papers and chapters related to this work.
I am also indebted to her for passing along to me a certain "picky-picky"
tradition in the construction of example sentences that arose at the
University of Texas in the 1960s. I hope this work will contribute
something to that tradition by helping to better describe differences
among sentences of the "same" type. Jennifer Hsu, Fritz Newmeyer,
Ken Pugh, and Nancy Richards also read and commented on all or part
of the manuscript, each providing very helpful suggestions. Tom Bever
was instrumental in sparking a certain biological turn in my thinking
that led to heightened interest in syntactic patterns in populations and
the individual members of those populations. Michael Studdert-
Kennedy fostered my interest in biological perspectives on language
before I could figiire out quite how or whether I wanted to address such
issues. Diana Smetters and Ken Williams helped to get this line of research
going by running two of the earlier experiments reported here.
I am very grateful to Carson Schutze for providing a draft of
his recent volume (1996) on grammaticahty judgments. His thor
oughgoing and scholarly review of the literature in this area has been
an invaluable aid.
Ed Battistella at Wayne State College and David Basilico at the
University of Alabama at Birmingham provided indispensable assis
tance in collecting data on their campuses. Although almost none of
the data are used here, I am also very grateful to colleagues at the
University of Arizona and Queens College/CUNY for extraordinary
cooperation with a recent large experiment (as yet unpublished) that
contributed significantly to the project out of which this book has
grown. At Arizona, Rudy Troike and Tom Miller, as well as many of
their colleagues in the Composition Program, helped us collect hun
dreds of responses from undergraduates there. At Queens College,
Helen Cairns and several instructors in the Linguistics Program
helped us collect many responses there as well.
Len Katz provided some helpful suggestions on statistical is
sues. Guy Garden and Hiroshi Nagata provided copies of papers that
proved quite useful at various stages of this project.
I also benefited from comments and questions from members
of the audience at presentations and coUoquia on parts of this work
Xii EXPERIMENTAL SYNTAX
at CLS '89, ESCOL '90, LSA '90, LSA '96, University of Massachu
setts/Amherst, Haskins Laboratories, City University of New York,
and the University of New Hampshire.
Alttiough I have benefited from the assistance, advice, and/or
cooperation of all of these colleagues, they bear no responsibility for
such foolishness as may remain in this text. All defects in the book
are solely due to the tiny green bugs that seemed to rain down on the
manuscript whenever I worked on it in my backyard.
The experimental work reported here was supported in part by
grants from the National Science Foimdation, SBR-9422688 (awarded
to the author and Dana McDaiuel), and the National Institutes of
Health, R01NS22606.
Finally, and mostly, I am grateful to Judy and Cayce for repay
ing my neglect with much more patience and affection than I think
I'm entitled to.
Note
1. See Popper ( 1 9 5 9 / 1 9 3 5 , pp. 44-48) for some comments or\ the role of inter-
subjective agreement in establishing scientific observations.
Web Site for This Book
Some Excel spreadsheets and other tutorial

materials relevant to the topics this book discusses
are available at the following site on the World
Wide Web:
http:/www.usm.maine.edu/~lin
1
Introduction: Are
Judgments Stable?
So far as research practice is concerned, one of the most striking

contrasts between generative linguistics and its immediate predeces
sors lies in the fundamental importance assigned to speakers' intro
spective judgments of sentence acceptability (Chomsky, 1957,1986).
In an earlier era, Bloomfield (1930) had specifically rejected reference
to the speaker's internal states as beyond the reach of science. Chom
sky instead embraced introspective judgments as an essential tool for
the exploration of the mental resources underlying language use.
This move has been extremely fruitful. The new data made available
to linguists via the inclusion of judgments of sentence acceptability
have, in conjunction with other innovations, brought a vast array of
syntactic phenomena within the reach of contemporary theory. What
ever the merits of various theoretical enterprises, there is no doubt
that we now have access to an enormous treasure of descriptive
generalizations about sentence structure in a wide range of languages
1
2 EXPERIMENTAL SYNTAX
that could never have been acquired without reliance on judgment

data.
Nevertheless, there have been continuing doubts about the
empirical reliability and theoretical interpretation of judgment data
as well as qiiistions about what constitutes an appropriate technique
for gathering judgment data. A variety of authors have alleged that
judgments are gravely compromised by instability of several differ
ent kinds. Some have found evidence of extreme disagreements from
informant to informant (or linguist to linguist) while others have
reported evidence that informants often disagree with themselves if
responses given on different occasions are compared. Others have
suggested that the orderly and consistent patterning of judgments
across various theoretically related sentence types occurs, if at all, in
the judgments of linguists, but not in the unstudied responses of
typical speakers. These tmcertainties are complicated by the fact that
there seems to be no consensus on how to gather judgments apart
from a widespread tolerance of informal methods in which the
linguist consults her ovm intuitions and those of the first handy
irvformant (what we might call the "Hey, Sally" Method). Further,
the lack of any consensus on just how judgments are related to
grammars (what, if anything, is the difference between grammatical
ity and acceptability) has sometimes been seen as an insuperable bar
to any scientifically appropriate use of judgments.
This book has a simple purpose that bears on these issues; it
wUl describe a family of practical methods that yield demonstrably
reliable data on patterns of sentence acceptability in populations of
linguistically naive speakers. This chapter will review some evidence
that the methods to be described here do in fact yield reliable results.
However, the bulk of the book is a tutorial intended to allow a researcher
with little or no prior training in experimental work to design and
execute small-scale experiments that yield reliable data on Uie kinds of
judgment issues syntacticians typically care about. Thus, for the most
part, the book ignores prior debates about the stability and interpreta
tion of judgment data, focusing iiwtead on describing a sovmd practical
tedmique whose merits can be independently assessed by interested
investigators. The book will not be much concerned with some notions
related to stability in the psychology and statistics literatures.' The best
source a reader may turn to for a general review of the history and
theory of judgments in linguistics is Carson Schutze's volume. The
Empirical Base of Linguistics (1996).
Introduction 10
Before launching into a practical explication of the method,

there are some background issues that need to be addressed. First,
the next section wUl briefly review some of the kinds of evidence that
have raised doubts about the stability of judgment data, as well as
some potentially relevant data from allied disciplines that suggest
that the problems of stability may not be as bad as is sometimes
alleged. Second, we will ask whether there is any way that grammati
cal theory can or should constrain our approach to measuring accept
ability—especially where a grammatical theory posits only two (or a
very few) degrees of grammaticahty. This section also asks what it
means to claim that a particular kind of data is or is not stable.
In the third and fourth parts of this introductory chapter, we
turn to empirical evidence that stable patterns of judgments occur
among linguistically naive informants. This section lays out a num
ber of experimental findings that come from applications of the
methods described in the rest of the book. These results will demon
strate that the methods do in fact produce reliable data of the kind
that bear on syntactic theory.
The chapter ends with an overview of the rest of the book.
1.1 Doubts About the Stability of Judgments
Linguists were not the first to use introspective judgments in a

scientific context. Introspective judgments of various kinds were
widely applied in psychological investigations in the late nineteenth
and early twentieth centuries. During tt\is era, psychologists studied
a variety of conscious phenomena via introspective methods. Yet,
despite their wide apphcation and acceptance, many varieties of
introspective procedures were generally rejected earlier in this cen
tury in part because of evidence that the results of these procedures
were unstable, seeming to vary across individuals and laboratories.
Boring's (1953) historical review of introspection suggests, however,
that the rejection of introspection was more theoretical than practical.
Many of the conscious phenomena that had been the focus of intro
spective investigation continued to be studied, often by quite similar
means. Verbal reports that were once interpreted as evidence of
private conscious states now became objects of study in themselves;
instead of asking why consciousness took such and such a form, the
psychologist asked instead why the verbal report took such and such
a form. What changed in the behaviorist era had to do more with the
theory of what consciousness is (or isn't) than with a thoroughgoing
rejection of introspective method.
Doubts about the stabiHty of judgment data began to appear in
the linguistic literature very soon after the emergence of generative
linguistics. As early as 1961, Hill reported on a very informal study
of several nonnaive informants in which he found substantial dis
agreement about the status of several of the 10 sentences he examined
(HUl, 1961). In the early 1970s, Garden (1970,1973) found evidence
of dialect differences among naive informants as to how negatives
and quantifiers interact in certain cirounstances. Other participants
in the subsequent discussion (Heringer, 1970; Labov, 1972; Stokes,
1974, among others) also fovmd evidence of substantial disagree
ments among informants.^ Labov (1975) called attention to differ
ences in experimentally derived judgments reported in these and
related papers. Snow and Meijer (1977) collected judgment data from
Dutch informants on 24 sentences via both a rating and a ranking
procedure. Although they found substantial agreement among in
formants on sentence rankings, there was no significant correlation
between the ratings and rankings obtained from individual infor
mants (provided at separate sessions 1 week apart). Snow and Meijer
were primarily impressed by the evidence of instability they foimd.
Ross (1979) describes a study in which 30 informants, including
some linguists, were asked to rate 13 sentences in various ways.
Although Ross is quite candid about the statistical limitations of his
analysis (the work being intended essentially as a pilot study), he
finds much evidence of disagreement among informants about the
status of particular sentences. He found subsets of informants that
agreed on a few sentences, but no two ii\formants provided the same
ratings for even a third of the sentences. Ross also noted that infor
mants seemed to be least likely to agree on more peripheral items
that show generally low acceptability.
A very thorough and careful critique of judgment data appears
in Labov (1975). In a review of numerous published and unpublished
findings, Labov found extensive evidence of disagreements among
informants. He also found many disagreements between published
assessments of sentences in the linguistics literature and the assess
ments of naive informants. He concluded that variation in judgments
was widespread and disorderly enough to call into question much
work based on intuitive data.
Introduction 10
A number of investigators have found evidence that individual

ir\formants can give different assessments on different occasions or
under different conditions (Carroll, Bever, & Pollack, 1981; Nagata,
1987a, 1987b, 1988,1989a, 1989b, 1989c, 1989d, 1992; Snow & Meijer,
1977). Labov (1975) also concludes that informant judgments don't
always agree with the informant's own linguistic behavior, raising
questions about both the stability and the interpretation of those
judgments.
There is also considerable anecdotal evidence of variability in
judgments from linguists who teach syntactic theory. In both gradu
ate and undergraduate courses, it seems to be commonplace for
students to disagree with judgments given in texts and the research
literature. Although these disagreements may sometimes reflect mis
understanding of the issues relevant to a given sentence, such dis
agreements are too commonplace to be dismissed lightly.
Despite these indications of significant instability in judgment
data, there is also evidence that sentence judgments are, at least some
of the time, highly stable. Some debates in syntactic theory turn on
relatively small differences in judged acceptability, but there remain
many contrasts that are quite dramatic. If, for example, one is con
cerned with the standard ordering of the major elements of the
English clause, there will be little or no variation in relevant judg
ments. Personal introspection may be reliable for some of these
syntactic contrasts. Newmeyer (1983) also argues that many dis
agreements over judgments are only apparent, and that, whatever
the limitations of judgment data, their manifest utility in furthering
syntactic research far outweighs whatever limitations attach to them.
Given the skepticism described above, it is perhaps suφrising
that judgments of "grammaticahty" have found increasing use in the
psychological literature. Roughly 60 papers over the last two decades
have used one or another grammaticahty judgment method, about
40 of these in the last decade. Many of these papers report some
variety of judgment procedure that is potentially relevant to linguis
tic issues. In addition to issues in general cognitive psychology and
language processing (e.g., Ferreira & Henderson, 1991), these papers
have addressed questions in neurolinguistics (e.g., Bemdt, Salasoo,
Mitchum, & Blumstein, 1988; Blumstein, Milberg, Dworetzky, Rosen,
& Gershberg, 1991; Linebarger, Schwartz, & Saffran, 1983; Wulfeck
& Bates, 1991), LI and L2 acquisition (e.g., Ellis, 1991; Schachter &
Yip, 1990), reading (e.g.. Fowler, 1988), and the cognitive theory of
10 EXPERIMENTA SYNTAX
EXPERIMENTAL
judgment process (e.g., Carroll et al., 1981; Cowart, 1994; Gerken &
Bever, 1986; Nagata, 1987a, 1987b, 1988,1989a, 1989b, 1989c, 1989d,
1990,1991,1992).^
There is a notable contrast between most psychological uses of
judgments and a number of the linguistic studies that have criticized
judgments on the basis of experimental findings. Many of the experi
mental linguistic studies mentioned above tacitly assume that each
encounter between an informant and a sentence yields (at least in the
great majority of cases) a definitive index of that informant's assess
ment of that sentence. In statistical terms, these studies assimie that
there is Uttle or no "error variance" in sentence judgments (the
concept of error variance will be discussed in the next chapter).
Surprisingly few linguistic attempts to study judgments experimen
tally have given serious consideration to the possibility that an
informant's judgments might be subject to some amoimt of random
variation around a stable mean. Another important feature of some
of the linguistic studies mentioned above is that they effectively
assume that each sentence is a definitive representative of its syntac
tic type, sometimes using only a single representative of each sen
tence type tested. Such practices make no allowance for the widely
recognized fact that lexical differences between different tokens of
the same syntactic type can produce quite substantial differences in
judged acceptability. Many of these studies also ascribe no signifi
cance to order, in effect assuming that the order in which a set of
sentences is presented exerts no effect on judged acceptability.
The contrast to standard practice in psychophysics is particu
larly striking. Like much of contemporary syntactic research, psycho
physics is very much concerned with descriptions of the private
subjective states of human subjects. In the case of psychophysics,
however, the private states in question are generally responses to
relatively simple external physical events such as flashes of light,
brief tones, touches of the subject's skin, and so on. Even with these
far simpler stimuh, it is commonplace to have each subject judge
dozens or even hundreds of iivstances of the same stimulus. This effort
is imdertaken for the simple reason that subjects routinely give
variable responses to different presentations of a single stimulus. By
asking each subject to judge each stimulus many times, psychophysi
cists are able to establish highly reliable descriptions of those re
sponses that will generalize to future sessions with the same subject
and to data collected from other subjects as well. From a linguistic
Introduction 10
point of view, this bears emphasis: Although there are many stimuli
for which each subject's responses will show a highly stable norm, it
is often difficult to detect that norm on the basis of small numbers of
responses. The nature of sentences obviously makes it inappropriate
to present the same sentence to one informant many times,* but there
may be an important hint in standard psychophysical practice and
experience that the normal variability of individual sentence judg
ments could be much higher than linguists have generally assumed.
The next chapter will pursue this issue in more detail.
1.2 Grammatical Theory and the Measurement of Acceptability
Part of the task of devising and justifying sound objective

methods for gathering judgment data—^methods that can put to rest
the doubts and concerns raised above—is to adopt a clear and defen
sible view of the relation between grammatical theory and the pro
cess of forming judgments. I will asstime that theories of grammar
are partial models of the human cognitive system that should, among
other things, help to explain why judgments pattern as they do. The
principal question here is whether there is any aspect of grammatical
theory that can or should constrain the procedures by which judg
ment data are gathered or analyzed. In particular, because grammati
cal theories sometimes countenance only a few different degrees of
grammaticahty, is there any way that this feature ought to affect the
collection or analysis of judgment data?
Chomsky (1965) provides a relatively straightforward answer
to this question in what has become the classical position on the role
of judgments of sentence acceptability in generative theory. In this
view,^ which exploits the distinction between competence and per
formance, the act of forming or expressing a judgment of accept
ability is a kind of linguistic performance. The grammar that a
linguistic theory posits in the head of a speaker does not exercise
exhaustive control of judgments any more than it exercises exhaus
tive control of the processes of uttering or comprehending sentences.
While forming a sentence judgment, a speaker draws on a variety of
cognitive resources to perceive the auditory or visual signals through
which the sentence is expressed, to parse or process the sentence as
its elements are identified, to form an evaluation of the sentence, and
to organize and execute a response. Among the cogrxitive resources
the speaker will draw on in executing a performance of this kind

there is, by hypothesis, a grammar. Even if every component works
optimally, whatever influence the grammar achieves arises in col
laboration, or perhaps competition, with other components of the
cognitive system. Whatever pattern of grammaticahty values a gram
mar might assign to a set of sentences, the grammar relies on the rest
of the cognitive system to instantiate those grammaticahty values in
overt judgments of acceptabiUty. The resulting judgments could
pattern quite differently than the grammaticahty values we might
like them to reflect.
Needless to say, with these assvmiptions, however many de
grees of grammaticahty the grammar might distinguish, the grammar's
interactions with other components could easily yield judgments
that are more (or less) finely differentiated than the grammaticahty
values the grammar provides.^
Within this frame of reference, judgments of acceptability pro
vide evidence that bears on theories of grammar, but that evidence
is not always subject to easy interpretation. As Bever (1974) noted,
this model of acceptability judgments implies that the study of
judgments and their imphcations for the theory of grammar will
proceed in parallel with the study of the full network of cognitive
resources involved in executing judgments. Patterns of results that
emerge in judgment data are potentially explained by reference to
(a) a model of grammar, (b) other components of the cogrutive
system, or (c) interactions between the grammar and those other
components. For example, it is widely accepted that extragrammati
cal components or features of the processing system (e.g., memory
limitations) can make fully grammatical sentences difficult in ways
that show up as low acceptability ratings, and it is at least conceivable
that extragrammatical components of the system could "rescue"
ungrammatical sentences, giving them higher acceptability ratings
than their grammatical status would predict (see Lcmgendoen &
Bever, 1973, and Bever, 1974, for suggestions as to how this situation
might arise). It is worth noting here that the fact that acceptability
data are responsive to many kinds of influence does not entail that
all those influences must be identified and understood before accept
ability data can be put to work in any particular domain, including
grammatical theory; we'll come back to this issue below.
In this context, the measurement of judgments of acceptability
ought to aim at standard goals of measurement throughout the
Introduction 10
sciences; observations on judgments of acceptability ought to be as

reliable as possible and they should capture the subtlest differentia
tions that current theory is able to say something about.
Thus the goal of the procedures described in this book is to
assess the relative acceptabihty of sentences. These procedures will
assume that sentence acceptabihty varies continuously. They will not
assume or attempt to establish any discrete levels of sentence accept
ability. The minimal goal of the procedures is to allow an investigator
to determine whether or not there is a (statistically) reliable difference
in acceptabiUty between two sentences for a particular population,
and, if there is, which of the two is the more acceptable. Further, the
procedures should make it possible to compare differences in accept
abiUty; given two pairs of sentences, these procedures should make
it possible to determine which pair has the greater difference in
acceptability. Because the procedures can be appUed to large groups
of informants, where a sufficiently large and homogeneous popula
tion is available, it should be possible to discriminate the accept
ability values of almost any two sentences (although, as indicated
earlier, this will be worthwhile only where a theory is available that
can make some sense of these discriminations).
In the view adopted here, there is no such thing as an absolutely
acceptable or unacceptable sentence, although there might be such a
thing as an absolutely grammatical or ungrammatical sentence. No
matter how acceptable (or unacceptable) any particular sentence proves
to be, these procedures will allow for the possibihty that some other
sentence will have an even more extreme value. Any attempt to estab
lish criteria for absolute acceptabiUty or imacceptabiUty would cortfront
serious difficulties. Among other things, experience demonstrates that
even highly acceptable and theoreticaUy uncontroversial sentences are
often judged less than perfect by naive informants. There is evidence
that changes of instructions or context as well as other psychological
manipulations can shift acceptabiUty values, evoi for sentences that
seem to Ue at the extremes of the acceptabiUty scale (see CarroU et al.,
1981; Cowart, 1994; Nagata, 1988, 1992, as weU as Schutze, 1996, pp.
129-160). Thus, even if we were to assume an imderlying binary gram
matical mechanism, there are no dear and weU-motivated criteria for
detemtuning what observed acceptabiUty values are suffident to estab
lish that a given sentence is (or is not) absolutely acceptable.
There are other empirical phenomena aside from judgments of
global acceptabiUty that grammatical theories may be relevant to.
such as acceptability under a particular interpretation or scheme of

reference relations, or in a given context. The focus here on the overall
acceptability of isolated sentences is not meant to shght any of these
other kinds of data. Many of the techniques to be described below
will also be relevant to the collection of other kinds of judgments.
However, questions about the overall acceptabiUty of isolated sen
tences have played a predominant role in the development of linguis
tic theory to date. The role of this book is to advocate better methods
for collecting that very central kind of evidence.
Another issue needs to be mentioned concerning the relation
between grammatical theory and judgments of acceptabiUty. There is a
common misconception to this effect: To test grammatical theories by
reference to experimentaUy derived judgment data, it is first necessary
to have a fairly thorough theoretical understanding of the processes by
which informants produce judgments in experimental settings (i.e., to
have answers to the many theoretical questions impUdt in the model
outlined above). There are at least three reasons to reject this view.
First, the sources of judgment data that emerge from formal
experiments are no more obscure than the sources of judgments
gathered by way of traditional informal methods. S)mtacticians may
deploy a thoughtful and thoroughgoing battery of techniques to
protect the sentence judgments they make from a variety of extra
granmiatical factors. There has been, however, no empirical demon
stration of the effectiveness of the training through which most
syntacticians acquire these tools, nor any demonstration of the effec
tiveness of the techniques themselves. Nor is there any tradition of
reporting or acknowledging the ir\formal attempts an investigator
may make to protect reported judgments against known or sus
pected distorting factors; this makes it hard for the field as a whole
to develop any greater insight into what does and does not work.
Ir\formally, it appears that some investigators are good at providing
judgments that are relatively free of unwanted influences and are
representative of general facts about the target language while other
investigators are not. It is up to the good sense and experience of
individual readers to teU which is which. So far as their "purity" and
relevance are concerned, formal experimental results are certairUy no
more problematic than the products of traditional methods, while in
other respects they are clearly superior.^
Second, the important point about judgments of acceptability
(however evidence about them is acquired) is not that all of the
Introduction 11
variation in judgments is controlled by the syntactic properties of the

target sentences but that some of it is. That other factors also play a
role is at worst an inconvenience; it is not a bar to making effective
use of the phenomenon. Thoughtful research design and modem
statistical techniques make it quite practicable to design experiments
that, through their structure, isolate those components of variation
in judgments that are attributable to (what the theory says ought to
be) syntactic manipulations, regardless of what other factors may be
contributing to overall variation in judgments in a particular experi
ment (see Chapter 3). This capacity to reUably associate manipula
tions with results is all most sciences need to make some progress in
theory building.
Third, given that both grammatical theory and psycholinguis
tics have a legitimate interest in the study of judgments, research
would stop if workers in each research tradition were to insist that
workers in the other tradition be the first to provide a theory of
judgments. If an independent referee were to choose which was to
proceedfirst,grammatical theory would make the better bet because,
on the whole, the issues there seem clearer, more tractable, and more
amenable to the construction of a richly structured theory.
In short, no issue of "impurity" or lack of insight blocks exploi
tation of experimentally derived judgment data in the study of
grammatical theory. Whether experimental studies of judgments
contribute to the development of grammatical theory will depend
upon whether investigators find clever and insightful ways to bring
the studies to bear on important theoretical issues, not on any intrin
sic limitations in this kind of evidence.^
One other connection between grammatical theory and the
measurement of acceptability should be noted here. Stabihty is a
theory-relative construct; to describe the stabihty of a phenomenon
within a population, we must know what theory is to be applied to
it. Most psychological phenomena vary in some degree, and that
variability can be quantified in various ways, but whether any par
ticular degree of variability counts as large or small can only be
decided relative to specific questions and purposes that derive from
a theory. What's needed is that the theory provide categories, distinc
tions, and descriptive parameters of some sort. Describingtiiestabil
ity of the phenomenon is a matter of comparing the empirical differ
ences in behavior or performance that can be linked to the theory
with the random scatter arovmd those theory-linked differences.
To detennine how variable syntactic judgments are within a given

population, we first find the average empirical difference in accept
abihty for two sentence types that syntactic theory discriminates and
then compare the variability of those judgments to the observed differ
ence between the two means. More generally, we look at a wide variety
of the kinds of contrasts between sentences that syntactic theory sug
gests are important. If those differences the theory predicts are generally
small relative to the typical variability of the judgments, we might be
justified in concluding that judgments are highly variable. On the
contrary, the evidence to be reviewed below shows that variabifity
within and between samples of native speakers of American English is
generally sntall relative to the observed differences in acceptabihty
associated with a number of theoretically derived distinctions.
Needless to say, there will be no simple or univocal answers to
questions about the stabihty of grammatical phenomena in speech
commimities. It seems likely that stability in the judgments in a
community will vary by gr<mimatical domain, that binding or ECP
effects may show different degrees of stabihty than parasitic gaps or
crossover effects.
We turn now to some direct experimental evidence for the
stability of judgments among nonlinguists.
1.3 Evidence of Stabiiity Within Populations
The data presented in this and the next section are intended to
show that the kinds of methods described in this book are able to
detect stable patterns of judgments in speech conununities. These
findings will not respond to all the issues raised in the critical
literature on judgments. They will show, however, that one popula
tion (native speakers of American English) has stable patterns of
response with respect to some theoretically significant syntactic is
sues, and that those patterns of resporwe can be reliably measured
via the methods to be described in this book. I will treat all other
questions of stability, reUability, and validity as secondary to these
issues (see Endnote 1, this chapter). Because the rest of the book will
describe in detail the methods used here, a very brief summary will
suffice as background to the experiments discussed in this chapter.
The experiments use questionnaires whose materials were con
structed according to constraints that typically apply in on-line ex
Introduction 13
periments using sentence stimuli. The materials for each experiment

consist of a collection of paradigmlike token sets. Within each token
set, all of the sentences (each being a representative of a sentence
type) are, as nearly as the design of the experiment will allow,
identical to each other. Questioruiaires are constructed according to
a counterbalancing scheme that ensures that each informant judges
exactly one member of each token set, judges representatives of all
the sentence types, and judges the same number of sentences in each
category. The counterbalancing scheme also ensures that each mem
ber of each token set is judged by some subset of the informants in
the experiment. Informants judge from three to sbc or more sentences
of each token type. The experimental sentences are randomly min
gled with filler sentences of diverse kinds. Wherever possible, the
order of presentation of the materials is varied across informants.
Informants rate each sentence on some sort of scale, with details
varying across experiments. All of these aspects of experimental
design and procedure will be discussed in detail later in the book.
All of the results reported in this chapter have been subjected
to appropriate statistical tests, although I will report those results
only in nontechnical terms within the text of this chapter. Details of
statistical tests and results are presented in Appendix B.' The main
statistical notion I will use here in the main text is "significant" (or
"reliable"). In statistics, a significant or reliable difference is one that
is unlikely to have occurred by chance. That is, statistical tests such
as those used here estimate how likely it is that an observed differ
ence (say, between two means) in the results of an experiment would
have occurred if in fact there were no consistent difference between
these means in the sampled population. Characterizations such as
"significant" or "reliable" are based on explicit criteria; for example,
in the psychological and social science literature, a result is usually
said to be significant if the probability of it occurring by chance is .05
or less. One or two other statistical notions used later in the chapter
are also discussed in Appendix A.
The quantitative results of these experiments will be reported
in graphs that will show the relative acceptability of the several
sentence types tested. Associated with each data point will be error
bars that indicate how much variability there was around the indi
cated mean (see Figure 1). If the error bars for two means overlap, or
come close to it, the two means may not be reliably different. The
larger the gap between two means, relative to the size of their error
Figure 1. Data points with error bars.
bars, the more likely it is that there is a reliable difference between

those means. The vertical scale of all of the graphs used here is
defined in terms of standard score (z-score) units. Standard score
scales locate each data point relative to the average of all the data
covered by a given graph, which is associated with the zero point on
the scale. They have the advantage that they capture all of the
available information about the relative acceptability of the tested
sentence types, while abstracting away from the specifics of the scale
used to collect the data (see Appendix A for further discussion).
This section will describe three sets of empirical results relating
to different issues in syntactic theory. In each case, the results will
show three things. First, some or all of the syntactic contrasts tested
prove to be highly stable properties of the sampled population.
Second, each experiment corroborates one or more contrasts in ac
ceptability that have been assumed in the syntax literature. Finally,
two of the studies reported in this section reveal further aspects of
the phenomena in question that have been overlooked (or ignored)
in the syntax literature. My purpose in presenting this work here is
not to argue for any particular theoretical conclusions but to demon
strate that the procedures employed here can detect reliable differ
ences among the kinds of sentence types that typically matter to
grammatical theory.
Introduction 15
1.3.1 Subjacency
Sentences such as (lb), (Ic), and (Id) appear to involve extrac

tion from a picture-NP in a way that violates the subjacency constraint
(Chomsky, 1973)."
(1) a. Control'^ Why did the Duchess sell a portrait of Max?

b. Indefinite Who did the Duchess sell a portrait of?
c. Definite Who did the Duchess sell the portrait of?
d. Specified Who did the Duchess sell Max's portrait of?
Subject
Syntactic accounts of relevant cases have generally assumed that

extraction from NPs with specified subjects, as in (Id), is not grammatical,
and that extraction tmm indefinite NPs such as (lb) is grammatical.
Opinions have varied on extraction from definite NPs sudi as (Ic). An
experimental examination of the relevant acceptability facts used token
sets such as (1)." Questiormaires designed along the lines discussed above
were presented to 88 linguistically naive informants. In this experiment
and the next, the same order of materials was used for all informants."
The results are simmiarized in Figure 2.
The small size of the error bars relative to the large differences
between the data points suggests that the pattern is highly reUable, and
this impression is confirmed by the statistical results reported in Ap
pendix B. Each of the three rightmost sentence types is reliably less
acceptable than the sentence type to its left.
The pattern of these results strongly corroborates the standard
assumption that Specified Subject cases such as (Id) are unaccept
able. The most interesting feature of the result, however, is not in the
contrast between the Control and Specified Subject cases but in the
relation of the Indefiiute and Defiiute cases to the two extremes.
Although each of the intermediate cases is sigiuficantly more accept
able than the Specified Subject cases, both the Indefinite and the
Definite cases are reliably less acceptable than the Control case, and
reliably different than each other. Insofar as a theory of grammar
might be held responsible for explaining these results, that theory
must explain why a seemmgly quite normal sentence like (lb) should
be so sharply distinguished from the Control case.^^
The next experiment was designed as a foUow-up to the first. It
provides two new control cases against which to gauge the low
1.0 -ι 1
N 88
0.8 I =
~ 0.6
(0
'c
2 0.4
o
υ
0,2
1 "
f o.o - 5
φ
Iαϊ8 -0.2
- -0.4 - *
-0.6
-0.8 »
.1.0 J 1 1 1
Control Indefinite Definite Specified
Subject
Figure 2. Relative judged acceptability of three cases of extraction from a
picture-NP.
acceptability of the Indefinite case in (lb) above. The experiment
used token sets structured as in (2).
(2) a. Control Why did the Duchess sell a portrait of Max?
b. Indefinite Who did the Duchess sell a portrait of?
with "of"
c. Indefinite Who did the Duchess sell a portrait to?
with "to"
d. Specified Subject Who did the Duchess sell Max's portrait to?
with "to"
Introduction 17
1.0 -, 1
0.8 - Ν =41
χ
δ 0.4-
8
Γ
Ν, 0.2- 0
£ τ
00
IS - - 0 ."2
υ
I -0.4
f-0.6
S -0.8
Έ
-1.0 J 1
I 1 — — 1
Control Indefinite Definite Specified
with "of" with "to" Subject
with "to"
Figure 3. Relative judged acceptability of extractions from picture-NPs compared
with extraction from a sister to NP.
Here (2a,b) replicate the Control and Indefinite conditions of the first
experiment, and (2c) and (2d) provide added relevant points of
comparison. Cases such as (2c) are maximally similar to (2b) on the
surface but do not involve extraction from within a picture-NP. Thus,
if the low relative acceptability of the Indefinite cases in the first
experiment resulted from informants applying a prescriptive rule
against sentences ending in prepositions, (2c) ought to be similarly
compromised. The contrast between (2c) and (2d) allows for a com
parison to the pattern attained with the superficially similar cases in
(1) and provides a third clear case of a class of apparently acceptable
sentences against which to compare cases such as (2b).
The procedure was unchanged except that a smaller group of
16
informants was used (Ν = 41). The results of the second experiment
are summarized in Figure 3.
As is evident, the hidefinite cases ending in "of" were markedly

less acceptable than either the original Control cases or the hidefinite
cases ending in "to." As before, the contrast between the hidefmite
"of" cases and the Control was significant, but the contrast with the
bidefinite "to" cases was also reUable. There was no significant
difference between the two cases ending in "to." The overall pattern
of results was significant.
Thus sentences such as 2b (and lb above) seem to be consis
tently less acceptable than other relevant sentence types. This result
may be related to Chomsky's observation that matrix verbs vary in
their compatibihty with extractions from picture-NPs (Chomsky, 1977,
pp. 114-115). In any event, it seems that (2b) does not have quite the
status usually assumed for it.
These two experiments also provide an empirical basis for
considering the question of whether sentence acceptability (as op
posed to grammaticahty values) varies continuously or by discrete
steps. These results strongly support the claim that acceptability
varies continuously. Strictly speaking, it is possible that there are two
or more distinct patterns in the populations tested here. One might
imagine, for example, that all informants in the first experiment reject
the Specified Subject cases and all accept the Control cases, but that
there is a dialect spht on the Indefinite and Definite cases. Averaging
across this dialect split gives the appearance of gradedness even
though the imderlying phenomenon is essentially discrete—indeed,
binary. There is, however, no support for this analysis here. One
prediction of this analysis is that the variability of informant re
sponses in the Indefinite and Definite cases will be greater than that
in the other two. There is no evidence of such differences in either of
these experiments; variabihty around the means is roughly similar
across all experhnental conditions within each experiment.
1.3.2 "That"-Trace
The third experiment examines the "that"-trace effect described

by Chomsky and Lasnik (1977). The standard observation is that
there is a certain subject-object asymmetry when "that" is present in
cases such as (3). Extraction from either site is acceptable without
"that" (3a,b), but only object extraction (3d) is acceptable with "that"
(compare with 3c).
Introduction 19
1.0
No "that"
0.8
With "that"
c 0.6
ο 0.4
Ν
I
0.2
0.0
& -0.2
< -0.4 +
f -0.6 +
§ -0.8 + Ν = 32
0)
-1.0
Subject Object
Extraction Extraction
Figure 4. The "that"-trace effect, an interaction between extraction site and
the presence of that.
(3) No "that"
a. Subject Extraction I wonder who you think likes John.
b. Object Extraction I wonder who you think John likes.
With "that"
c. Subject Extraction I wonder who you think that likes John.
d. Object Extraction I wonder who you think that John likes.
Data were collected from 32 informants. Eight different orderings of

the materials were used.
The results are summarized in Figure 4. There is no difference
in acceptability between Subject and Object Extraction in the absence
of "that," but there is a reliable difference in the presence of "that."
Note also that there is a reliable "that" effect that co-occurs with the
"that"-trace effect; both the sentence types containing "that" were
less acceptable than the two without "that." This last finding raises
the question of how (or whether) the "that"-trace effect might be
related to the apparent general bias against all extraction over "that."'^
The general pattern here is reliable, as the error bars suggest.
1.3.3 Coordination and Binding Tlieory
The final set of preliminary observations concerns an interac

tion between reflexives and coordinate NPs. An earlier pilot study
indicated that reflexives inside coordinate NPs did not necessarily
conform to the Binding Theory (Chomsky, 1981). The token set in (4)
was constructed to look for Binding Theory effects where a reflexive
appears in first conjunct position inside a coordinate NP.
(4) No Coordination
a. Local Antecedent Cathy's parents require that Paul
support himself.
b. Remote Antecedent Paul requires that Cathy's parents
support himself.
Simple Coordination
c . Local Antecedent Cathy's parents require that Paul
support himself and the child.
d. Remote Antecedent Paul requires that Cathy's parents
support tiimself and the child.
Coordination With "Both"

e. Local Antecedent Cathy's parents require that Paul
support both himself and the child.
f. Remote Antecedent Paul requires that Cathy's parents
support both himself and the child.
In particular, the question of interest was how informants would

evaluate the Remote Antecedent cases, which the Binding Theory
would rule ungrammatical. All of the Local Antecedent cases are
assumed to be grammatical. The cases incorporating "both" were
expected to heighten any effects due to the introduction of coordinate
structure by adding a cue at the leading edge of the coordinate NP.
The 43 informants who participated used 24 different orderings
of the materials. The results are summarized in Figure 5.
These findings strongly confirm the influence of the Binding
Theory in all three pairs of cases. The difference between the two
Introduction 21
1.0
- ο — Local Antecedent
0.8 \
—·— Remote Antecedent
(η
Έ
3
0.4 \
0.2 I
1
I-ο
α>
ο>
τ3
3
-5
C
(Ο
0)
/V = 4 3
-1.0 ι
I 1
I
No Simple Coordination
Coordination Coordination witii "both"
Figure 5. Relative acceptability of sentences containing reflexives with local

and remote antecedents by context of the reflexive.
Antecedent Location conditions (local versus remote) is reUable overaU

and for each of the three coordination conditions.
In addition to the persistent Binding Theory effects, these results
also show that reflexives with remote antecedents are significantly more
acceptable when the reflexive appears within a coordinate NP. Looking
only at the Remote Antecedent cases, there was a sigruficant increase in
acceptabiUty from the No Coordination condition to the Simple Coor
dination condition, and from the Simple Coordination condition to the
Both Coordination condition. The question then becomes why it naight
be that coordinate structures seem to override or partly nullify effects
of the Binding Theory (see Frazier & Clifton, 1996, pp. 35-37, for some
comments on the theoretical issues these findings raise).'*
The picture derived from the several experiments described
above is unmistakable. Each experiment confirms some well-attested
contrast in acceptability, and all of the phenomena are very stable.^'

All of the principal findings are robust enough to be detected with
samples much smaller than those used here. There are also some
results that are not obviously consistent with current theory or that
challenge empirical assumptions that have played some role in gram
matical research (e.g., the increased acceptability of reflexives with
remote antecedents in coordinate NPs). Although it is far from clear
whether any of these findings have important theoretical implica
tions (or what those implications might be), they are reliable aspects
of patterns that have otherwise proven interesting and worthy of
intensive study. Even if none of the findings reported above is
theoretically consequential, it is noteworthy that traditional observa
tional methods have apparently allowed these facts to be misrepre
sented or overlooked. Taken together, results outlined above provide
no support for the suggestion that judgments of acceptability are
compromised by extreme variability. There appear to be highly
stable patterns of relative sentence acceptability.
1.4 Stability of Responses to Individual Sentences
The reliable differences among judgments of sentence types de

scribed above are compatible with different degrees of consistency in
the judged acceptability of the individual sentences that fall within those
types. This section will present evidence on the stability of speakers'
assessments of the relative acceptability of individual sentences.
An indication of the stabihty of responses to individual sen
tences appears in Figure 6. Here, 12 sentences were judged by the
same 54 informants on two occasions separated by a week or more.
The sentences varied widely in acceptability and were presented in
one continuous sequence in the same order at both experimental
sessions. As is apparent from the graph (and the correlation coeffi
cient), this group of informants gave, on average, almost exactly the
same ratings to these 12 sentences in the two sessions.
The procedures to be described in this book are designed to
disentangle, insofar as possible, the characteristic acceptabihty rat
ings of individual sentences from effects of the contexts in which
those sentences occur. Thus, in the ideal case, every informant will
see the sentences they are to judge in a different order than does any
other informant. This raises the question of how much of the obvious
Introduction 23
1.0 τ
Ν = 54
CM r^ = .97
c
ο
'«
0.5 I
to
0)
CO
0.0
I -0.5 +
8
φ -1.0 •
Ο)
•σ
3
C -1.5 1
CO
φ
-2.0
-2.0 -1.0 0.0 1.0
Mean Judged Acceptability (Session 1)
Figure 6. Identical sentences judged on two occasions.
N O T E : Each data point represents the mean of 54 judgments from Session 1 and the mean of 5 4
judgments from Session 2.
Stability evident in Figure 6 is due to the fact that the sentences were
presented in the same order on both occasions. The large differences in
acceptabiUty among these sentences also help in achieving stable re
sults; contrasts of theoretical interest often wiU be much more subtle.
To estimate the stability of responses to individual sentences
where the order of presentation differs in the two sessions completed
by each informant and where, within any one session, various sub
sets of informants see the materials in different orders, we isolated
and analyzed a subset of the data from a "that"-trace experiment (n
= 52) in which each informant participated in two sessions.
In this experiment, the filler sentences were identical in the two
sessions, differing only in the order in which they were presented.
The filler list used was carefully constructed. A preliminary list of
highly acceptable sentences was pretested in an earlier experiment

and all the sentences getting relatively poor ratings were edited to
improve acceptabihty. Further testing showed that this editing suc
ceeded in producing a list of sentences of relatively uniform high
acceptability. This set of filler sentences was then randomly divided
so as to isolate a subset of one third of the sentences. These one third
of the fillers were then systematically deformed so as to make them
conspicuously and dramatically deviant. Further testing showed that
this manipulation was successful, yielding the intended strong con
trast between the two groups of sentences. These mixed fillers, two
thirds highly acceptable, one third very vmacceptable, were used in
both sessions of the experiment.
This makes for a strong test of the stabihty of responses to indi
vidual sentences because, within each subgroup of filler sentences, the
range of differences from sentence to sentence is relatively small. Thus,
if there are any differences in acceptabihty within one of these groups,
they ought to be relatively subtie differences, ones that would be
difficult to detect and vulnerable to change from session to session.
Similarity across sessions was examined by computing a corre
lation coefficient for the degree to which the judged acceptability of
individual filler sentences in Session 1 predicted the judged accept
ability of the same sentences in Session 2. The highly acceptable and
the highly unacceptable items were treated separately for these pur
poses. Figure 7 summarizes the results.
The figure shows that there was a strong relation between the
judged acceptabiUty of these sentences in the two sessions. Correlation
tests within the two subsets confirmed the impression in Figure 7; there
was a strong and reUable correlation between first and second session
ratings for both the High AcceptabiUty items and the Low AcceptabiUty
items. Thus, even v«thin these two rurrow ranges, and with the sentences
appearing in different orders at the two sessions, the rated acceptabiUty
of individual items was quite similar across the two sessions.
Another frame of reference for considering the stabihty of
responses to individual sentences focuses on the relative accept
ability of the several sentences that make up a single token set. That
is, we are particularly interested in knowing whether speakers con
sistently differentiate among the several closely related sentences
that make up individual token sets.
Evidence relevant to this issue comes from another "that"-trace
study. In this experiment, data were collected from 332 university
Introduction 25
3.0 1.0
Low Acceptability High Acceptability
Sentences Sentences
3.2 r » = .72 1.2 Λ» = .70
3.4 1.4
0
3.6 ο <g 1.6
0 ° /
3.8 1.8
Ν =32 Ν »64
4.0 ° , 1 2.0
4.0 3.8 3.6 3.4 3.2 3.0 2.0 1.8 1.6 1.4 1.2 1.0
Mean Judged Acceptability / S e c o n d Session
Figure 7. Plots of Session 1 versus Session 2 results for 96 individual filler

sentences.
N O T E : A sentence that w a s given a rating of " 1" (highest possible) b y all informants in both Session
1 and Session 2 would b e represented b y a point at the intersection of the " 1 " lines in the graph
for the High Acceptability cases. A sentence given the lowest possible rating ("4") b y all informants
in both sessions would be indicated at the intersection of the " 4 " lines in the graph for the L o w
Acceptability cases. If the results from Session 1 perfectly predicted results from Session 2, then
all the data points would lie on the diagonal of one of the graphs. "M" refers to the n u m b e r of
sentences covered b y each plot. See Appendix A for i n t e φ r e t a t i o n of the values.
undergraduates at Wayne State College (NE), the University of Ala

bama at Birmingham, and the Uruversity of Southern Maine. The
materials for this experiment were somewhat different than those used
in the earlier "that"-trace study in that they did not involve indirect
questions. A different main clause verb was used for each of the 20 token
sets constructed for this study. A sample token set is given in (5).
(5) a. Who do you suppose invited Ann to the circus?

b. Who do you suppose Ann invited to the circus?
c. Who do you suppose that invited Ann to the circus?
d. Who do you suppose that Ann invited to the circus?
The overall results of this study were very similar to those reported
in Section 1.3.2 above, except that the gap between Cfcject Extraction cases
with and without "that" was smaller, although still reliable overall. There
was a small statistically reliable difference between the results from the
Alabama cohort and the other two, which resulted from the fact that the
gap between the Object Extraction cases with and without "that" did not
appear in the results for the Alabama cohort.
To get at evidence on the stability of judgments on individual

sentences, the results for each token set were summarized separately
for the three cohorts (Nebraska, Alabama, Maine). This produced a
table of four means (for the four conditions in [5]) for each of 20 token
sets, for each cohort. The 20 token sets themselves and graphs for all
of the token sets appear in Appendix D.
Upon inspection of this evidence, two observations were imme
diately apparent. First, many of the token sets showed distinctive
patterns. Although there was a clear "family resemblance" across all
the token set graphs, several showed noticeably unique patterns.
Second, the patterns for individual token sets were strikingly similar
across the three cohorts. Whatever it was that made the pattern for
any one token set stand out from the rest in, for example, the Ne
braska cohort, it appeared that those same traits were generally
evident in the patterns for the other two cohorts.
These findings are illustrated for 4 of the 20 token sets in Figure 8.
Note that in the graphs in Figure 8 (as in all those in Appendbc E), each
data point represents a different subset of informants than do any of the
other data points in the same graph. Thus the NT/SE data point (No
"that"/Subject Extraction) for the Alabama cohort and the "suppose"
token set represents not only a different group of people than do the
other two NT/SE data points, but (because of the coimterbalandng
scheme; see Section 9.5.1) it also represents a different subset of the
Alabama cohort than is represented by any of the other three Alabama
data points in the same graph. In short, the fact that the patterns are as
similar as they are across the three cohorts for each token set reflects a
very high degree of sinularity across the three subpopulations repre
sented here. The three cohorts do not agree merely by sentence type
(liking NT/SE sentences and rejecting WT/SE sentences); rather, they
agree, for the most part, token set by token set and sentence by sentence.
1.4.10verview of Evidence
Taken together, the results presented in this chapter show

clearly that there is a stable natural phenomenon of sentence accept
ability; we find that for all of the syntactic phenomena considered
here, native speakers of American English exhibit stable, clear-cut
patterns of acceptabiUty differences across sentence types. Although
these patterns are of the kind commonly discussed in the syntax
Introduction 27
1.40
1.40
-.10
-.60
MT/SE NT/OE WT/SE WT/OE MT/SE NT/OE WT/SE WT/DE
Figure 8. Sentence by sentence mean judged acceptability data for four

token sets in a "that"-trace experiment.
N O T E : Each line in each graph represents data collected from iiiformants at a different site
(Nebraska, Alabama, Maine). N T / S E = N o "that," Subject Extraction cases; W T / O E = W i t h "that,"
Object Extraction cases; and so on. Each data point is the standardized mean (see Appendix A ) of
all the responses for that cohort for that sentence. Each of the four token sets w a s based on the
verb indicated in the graph for that token set.
literature, these results also show that there are sometimes details of
these patterns that have gone unnoticed or underestimated in syn
tactic research. The evidence reviewed above also demonstrates that
there are practical experimental methods by which the phenomenon
of sentence acceptability can be measured and assessed.
1.5 Outline of the Book
This chapter has stressed stability; the next will stress variabil
ity. At least where the judgments of naive informants are concerned.
it is the details of variability in those data that provide the strongest

motivation for relying on formal experimental procedures in assess
ing acceptability phenomena. Chapter 3 is the keystone of the rest of
the book. It introduces the notion of variance control and reviews the
major sources of variance that need to be controlled in experiments
on acceptability. It lays out the logic underlying experiments like
those described above. The chapter also considers constraints on the
form that sentence materials may take and some issues related to the
construction of experimental and filler materials. With this material
as backgroimd, investigators will have a better sense of what the
essential features are of the procedures discussed below, and be
better equipped to modify those procedures appropriately when the
need arises. Chapter 4 considers some of the different kinds of
information that sjmtactic investigators may seek and the issues that
arise in getting at that information experimentally. Alternative modes
of presenting sentences to informants and of collecting and measur
ing their respoi\ses are reviewed in Chapters 5 and 6, respectively.
Chapters 7 and 8 discuss sampling and other issues linked to the
actual execution of experimental procedures. Many issues arise in
connection with the organization of questionnaires, and Chapter 9
covers this topic in some detail. Recovering the data from an experi
ment and initial statistical summaries are discussed in Chapters 10
and 9 while Chapter 12 reviews some more specialized issues in
statistics £md measurement. Finally, Appendix C is a tutorial in
tended to introduce spreadsheet novices to the functions in Excel (a
spreadsheet program) that are most useful for creating and analyzing
acceptability experiments.
Notes
1. In the context of psychometrics especially, psychologists are often con

cerned with the reliability of individual measurements, that is, with the consistency
of repeated measurements on the same individual. Needless to say, some degree of
reliability in this individual sense is a prerequisite for obtaining reliable evidence on
populations, but experimental science focusing on the trails of populations can
proceed with cruder instruments than those that are required when the goal of
measurement is to characterize one particular individual (especially where that mea
surement may have practical consequences for the individual).
The concept of reliability is also coimected with questions about whether meas
urements accurately reflect the true state (i.e., true score) of the targeted property of the
individual. Measurements may be reliable (in the sense of consistent) and yet systemati
Introduction 29
cally distort the characteristic they are meant to meastire (giving, for example, values that
are consistently too high or too low in particular parts of the range of the measurement).
Reliability is also connected with validity, questions about what it is that a particular
psychological test actually measures (e.g., to what extent IQ is a measure of a cognitive
trait rather than an inadvertent index of socioeconomic background). Schutze's Chapter
3 (19%) is essentially a review of the literature on validity as it relates to judgments.
All of these issues are obviously relevant to work with judgments in linguistics,
but collecting reliable and scientifically useful data on patterns of sentence accept
ability in speech communities does not require the prior resolution of any questions
of these kinds. Although I will keep to the relatively informal usages of the linguistic
literature when any of these issues come up, I will try to indicate dearly which notion
of stability is relevant wherever we need to corwider matters other than stable
measures of the characteristics of linguistic communities.
2. Newmeyer (1983) argues that most of these disagreements are not about
the facts of acceptability. He suggests that in several cases judgments were constant,
but conflicts among theorists on how various sources of unacceptability should be
accounted for drove changes in the status ascribed to certain sentences.
3. This sample of papers includes those that were identified in a search of the
PsychlNFO database (American Psychological Association). The search target w a s
specified as follows: grammatical? (In) judgment?. This was intended to select all
papers whose database entry included adjacent ii«tances (in either order) of the target
terms or their variants.
4. Nagata (1987a, 1987b, 1988,1989a, 1989b, 1989c, 1989d) notwithstanding,
il appears to be essential to the nature of sentences (and human responses to them)
that they have identities in ways that flashes of light do not. Linguists and psycholo
gists, in taking an interest in the structure of language, naturally view sentences as
categories and see particular utterances as irwtances of those categories. From the
standpoint of day-to-day usage, however, repeating sentences seems to have some
status akin to re-eating a meal; the food wasn't quite right, or the conversation took a
bad turn, so we hit the rewind button and did it again on the spot. Sentences seem to
be ordinarily experienced as unique and unrepeatable events. Although w e can
individuate flashes of light (or sounds, or smells, or the like) by identifying each with
a time, this seems somehow contrary to the ordinary experience of such events. From
a psychological point of view, it seems all the light that flows from a bulb is part of
the same quantity or entity; individuating particular flashes feels a bit like giving
different personal names to a person's left arm, right arm, left ear, and so on. In short,
it does serious violence to the ordinary way of relating to sentences to ask informants
to judge the same sentence many times. This does not mean, of course, that repeated
presentations in experimental settings are unthinkable, but it does raise questions
about how to relate those events to more typical uses of sentences.
5. I am, with some trepidation, attempting here to summarize a view of these
issues that emerges in various of Chomsky's writings over many years. Readers who
wish to consider the issues in more detail should consult Schutze (1996), especially
his Chapter 2, which cites the various relevant works of Chomsky's.
6. Bever and his colleagues (Bever, 1974; Bever & Carroll, 1981; Gerken &
Bever, 1986; Katz & Bever, 1976; see also Schutze, 1996, pp. 62-70) have argued that
there are very good reasons to explain evidence of gradedness in judgments solely by
reference to performance systems while maintaining a discrete model of the grammar.
7. Formal methods have the distinct advantage that they are public and
replicable. They allow for productive ways of resolving data disputes where standard
informal methods do not. With traditional methods, disputes about data can easily
become dead-ends, regardless of the actual quality of the data in question.
8. Even if linguists have wildly misconstrued the origin of judgments (sup

pose, as some would claim, there is simply no distinct cogiutive subsystem that has
properties or roles like those assigned to grammars), evidence of stable patterns of
judgments in speech communities is likely to be scientifically important. If the phenom
ena of acceptability are stable and richly structured in the ways that current evidence
suggests they are, accoimting for these facts will still stand as an important challenge
for cognitive theory.
9. Appendbc A provides a brief introduction to the various statistical concepts
and procedures that are cited or used in Appendix Β or elsewhere in the book.
10. The size of the error bars in the graphs is determined by the standard error
of the mean, a notion that is discussed in Appendix A.
11. Chomsky's account of these cases changed substantially in the "Barriers"
framework (1986).
12. Control is used here in the sense of "control group."
13. Preliminary results of this and the next experiment were reported pre
viously (Cowart, 1989a, 1989b). The design for this experiment called for 24 items but
1 item was deleted by error.
14. The experiment was executed in collaboration with Diana Smetters.
15. It should be noted that the result is averaged over many sentences. The
reported result is not for the particular example sentences in (1). Section 1.4 and
Appendix Ε provide some information on sentence-by-sentence differences in a
"lhat"-trace experiment. In the subjacency experiments described in this section, there
may be important effects associated with individual verbs.
16. The experiment was conducted by a student. Ken Williams, as part of an
M.A. thesis project. No analysis on token sets was done.
17. In subsequent work on this phenomenon, w e have replicated this differ
ence at almost every opportuiuty, although the magnitude of the difference is usually
less than what is seen here.
18. Ross (1970) and Zribi-Hert2 (1989) discuss cases suggestive of the phe
nomenon observed here.
19. All of the phenomena described above emerged in the course of longer-
term investigations that have demonstrated the same or similar effects several times.
In the case of the "that"-trace phenomena, as we'll see, very similar patterns have been
detected in widely separated populations around the United States.
2
Error Variance
in Sentence
Judgments
In countless psychophysical, cognitive, and physical tasks, human

performance is characterized by a random scatter of individual ob
servations around a more or less stable mean. In statistics, this scatter
is termed error variance and it is ubiquitous in observations of human
performance. The purpose of this chapter is to provide a clearer
picture of the magnitude of error variance in sentence judgments and
to consider its significance for the problem of getting reliable infor
mation about patterns of judgments in the general population. Al
though the results described in Chapter 1 demonstrate a high degree
of stability in a number of judgment phenomena, a closer examina
tion of these and other similar studies also reveals much greater error
variance affecting individual responses than traditional linguistic
approaches to judgments have generally allowed for.
31
Standard methods in syntactic research reflect relatively httle

concern with the role of error variance in judgments of sentence
acceptability. This is easily seen in the way that linguists have some
times approached experimental work on judgments. HiU's (1961)
paper reports the responses of individual informants in some detail,
noting considerable divergence around points where it seems Chom
sky (1957) anticipated agreement. Notably, however, even though he
seemed to be probmg for evidence covmter to Chomsky's empirical
assumptions, it does not seem to have occurred to Hill to present the
sample sentences to his informants a second time. Although Ross's
(1979) study mentions in passing the question of "repeatability" in
cormection with asking his informants to rate their confidence in the
judgments they provide, he apparently did not do any explicit checks
of repeatability. Whatever doubts these investigators might have
harbored about the reliability of judgment data, they did not check
to see whether informants would agree with themselves upon doing
the same questionnaire a few weeks or months later. In both these
cases, the investigators behaved as though each response from each
hiformant was (or was close to) a definitive representation of that
informant's attitude toward that sentence and the sentence type it
represents.
In typical psychological studies, error variance is assumed;
investigators routinely collect numbers of observations from each
participant, often ask each participant to respond to each stimulus
many times, and report the procedural and/or statistical methods
used to control error variance. Standard practice recognizes that
individual observations are not reliable, and sophisticated practitio
ners may evaluate findings differently accordmg to their confi
dence in the means employed to control error variance. The fact
that the syntax literature is characterized by consistent inattention
to the methods by which sentence judgments are gathered and
summarized is not compatible with due concern for the role of error
variance.
Nevertheless, error variance is present in judgments of accept
ability. This is easily seen in a more detailed analysis of the "that"
trace study mentioned in Section 1.4 above (although the same point
could be made with data from almost any experimental study of
judgments). In this study, each informant judged five different rep
resentatives of each of the four sentence type categories shown in (6).
Error Variance in Sentence Judgments 33
(6) a. Who do you suppose invited Ann to the circus?

b. Who do you suppose Ann invited to the circus?
c. Who do you suppose that invited Ann to the circus?
d. Who do you suppose that Ann invited to the circus?
A variant of a standard psychophysical scaling method (see

Section 6.2) was used to express the relative acceptability of individ
ual sentences. As usual, the mean of each informant's responses
within each sentence type category was taken as the index of the
informant's attitude toward that category.
A follow-up analysis of this study isolated a subset of 88 in
formants who had all judged the same experimental and filler sen
tences (although in one of two different orders). The follow-up
analysis calculated the range between the informant's highest and
lowest ratings within each sentence type (e.g.. With "that"/Object
Extraction) and the mean rating within each category. This made it
possible to construct a rough index of the size of the "that"-trace
effect for each informant.^ This measure was then compared with the
range of ratings within various sentence type categories.
This analysis reveals what might appear to be an extraordinary
degree of inconsistency in informant responses; for roughly 90% of the
informants, the average range of variation unthin sentence type catego
ries exceeded the difference between the "that"-h-ace violations (6c) and
the average of the other three types. Indeed, on average, for individual
informants the mean range within categories was more than 80% larger
than the size of the "that"-b-ace effect. The "that"-trace effect was smaller
than the average range within categories even for half of the 10 infor
mants who showed the least overall variation within categories. In
short, for most informants a great many of their individual sentence
judgments were uruepresentative of the average acceptability of the
category from which the sentence was drawn.
The consequences of this degree of variability in individual
responses are illustrated by the data from five informants shown in
Figure 9. An informant can be said to show the "thaf'-trace effect if
the average of the five bars in the With "that"/Subject Extraction is
lower than the average of the bars in the other three graphs. Infor
mant 1 is perhaps the single "best" informant in this subset, showing
very little variation within each sentence type category and a clear
difference between the "that"-trace violations (With "that"/Subject
Extraction) and the other three types. Informant 2 is much more like
No 'Itiaf/ No "ttiatV With •Itiaf/ With "thatV

Subj Ext Obi Ext Subj Ext Obj Ext
Informant 1
Max mfm category range = 1

Avg w/in category range = .5
1 2 3 45
luu uuu
1 2 34 5 1 2 3 4 5
Informant 2
Max w/ln category range s s

Avg w/ln category range = 3.3
1 2 3 46 1 2 34 5
uili
1 2 3 4 5
Informant 3
Max w/in category range = 7

Avg wAn category range = 4 . 3
1 2 3 4 S
luh
1 234 5
Ja-A
1 2 34 5 1 2 3 4 5
Informant 4
Max w/!n category range s 8

Avg w/ln category range = β
Li
1 2 34 5
lui tUi
1 234 5 1 2 3 4 5 1 2 3 4 5
Informant 5
Max w^n category range " 9

Avg w/ln category range = 6
Overall:
1 2 34 5 1 234 5 1 2 34 5 12 3 4 5
Avg w/in category
range = 3.5
Figure 9. Within-category variation in response for individual informants.

N O T E : Each row o f graphs represents data from one informant. Each graph s h o w s the informant's
responses to all five of the sentences of a particular type that the informant judged. Each bar in
one o f ttie graphs represents the informant's response on one sentence. T h e sentence types are
identified at the top o f each column of graphs. T h e vertical axis of the graphs shows the rating
scale, running from 0 to 9.
the typical case. Here, although there is a difference in average

response between the "that"-trace violations and the other three
(difference = 1.8), the range of variation within every sentence type
category exceeds this difference (5,2,3,3, respectively). Furthermore,
there were many more variable informants who showed patterns hke
those seen in Informants 3,4, and 5. Although the first two of these
informants shows the "that"-trace effect overall, the range of vari
ation within sentence categories overwhelms average between-cate
gories differences in every case.
This degree of error variance, however, did not obscure the overall
finding. Iriformants gave high ratings to the (6a) and (6b) sentences,
slightly (but reliably) lower ratings to the (6d) sentences, and far lower
ratings to the (6c) sentences. The overall pattern was highly reliable.^
In short, when results were averaged across the several sen
tences to which each informant responded and averaged across
informants, the error variance demonstrated earlier was not suffi
cient to overcome or seriously obscure the underlying evidence of
the "that"-trace effect. Similar patterns of substantial error variance
accompanied by clear and consistent overall patterns have been
found for other syntactic phenomena as well.
A further indication of the magiutude of error variance in
judgments comes from another "that"-trace experiment in which 21
informants (out of a far larger sample) responded to exactly the same
questionnaire on each of two occasions a week or more apart. This
contrasts with the previous study where the informants saw the
target sentences in two different orders. These questionnaires in
cluded only three sentences of each syntactic type. Nevertheless, the
general "that"-trace effect evident in the full experiment and earlier
studies was reliably replicated with this small subset of informants.^
On the other hand, an informant's means for the four experimental
conditions from the first session were generally poor predictors of
that informant's means from the second session (see Figure 10). As
we saw earlier, highly stable patterns may be discernible in data from
a group of informants without those same patterns being readily
detectable in the results of individual informants.
These findings indicate that there is normally more than enough
error variance in judgments of sentence acceptability given by naive
informants to make the control of this variance a proper concern of
any reader of proffered general claims about the relative accept
ability of sentence types. Nevertheless, as the data reviewed in
Chapter 1 indicate, this error variance is easily controlled by standard
tools of experiment design and statistical analysis, so long as these
tools are properly applied.^
It is not clear how the judgment perfonnance of experts com
pares with that of typical naive informants. It may be that particular
Subject Extraction Object Extraction
2 0
rVoe
1.5 •
0 οο ο ο
0 e, 1.0
0 0
0.5
NO ^^^--"""^ ο 0
•tJurt" •— 0.0
° 0
-0 5 ο
0 0
ο 0 -1.0 ο
-IS
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
With
•that"
-1.5
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2 . 0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2 . 0
Mean Judged Acceptability Mean Judged Acceptability
{ζ -score units) / Session 1 (z -score units) / Session 1
Figure 10. T h e s e f o u r g r a p h s p l o t first s e s s i o n v e r s u s s e c o n d s e s s i o n d a t a

for 2 1 i n f o r m a n t s w h o w e r e g i v e n e x a c t l y t h e s a m e v e r s i o n o f a " t h a t " - t r a c e
questionnaire o n e a c h of t w o occasions.
NOTE: A t each session, each informant judged the same three sentences in each experimental
condition (e.g., the N o "that'VSubject Extraction condition). T h e data point representing each
informant in each graph is determined b y the mean o f the informant's three responses at the first
session (X axis) and the mean o f the informant's three responses on the s a m e three sentences at
the second session (Y axis). If informants had given the same or nearly the s a m e responses at both
sessions, all data points would lie at or near the diagonal in each plot. T h e values estimate the
share of variance in the second session that can be explaiiKd b y reference to the data from the first
session.
individuals (through specialized training, experience, exceptional

insight, or sensitivity) are able to reliably estimate general patterns
of sentence acceptability in a community. There is, however, usually
no way to tell whether a particular pattern of judgments offered by
an individual is representative of the target population apart from
applying experimental procedures like those used here.
The procedures described in this book assume the existence of

error variance in sentence judgments and apply various measures to
control that variance. The most important of these measures are the
use of multiple informants and multiple instances of any sentence
type whose acceptability is to be estimated. The general problem of
designing experiments on judgments that properly accotmt for the
existence of error variance and the other constraints peculiar to
studies of sentence materials is considered in the next chapter.
Notes
1. The difference between the informant's average rating for the With "that"/
Subject Extraction cases and the average of the other three categories was taken as the
measure of the "thaf'-trace effect for each informant.
2. From this point on in the text, when needed, statistical results will be
reported in endnotes. Readers unfamiliar with the tests reported should consult
Appendix A.
There was a reliable interaction between the Presence or Absence of "that" and
the Extraction Site factors, f i(l, 331) = 361, ρ < .001. Both the main effects were also
significant, F i ( l , 331) = 690, ρ < .001, for "that," and F i ( l , 331) = 281, ρ < .001, for
Extraction Site. Various formal and informal methods suggest that the effects were
robust enough to be tested with samples as small as eight informants without undue
risk (= .25) of a Type Π error. We'll return to sample size issues in Chapter 9.
3. The "that" versus Extraction Site interaction was reliable, F i ( i , 20) = 11.4, ρ <
.01, as was the main effects of "that," Fi(i, 20) = 11.i, ρ < 01, and Extraction Site, F i ( l ,
20) = 11.0, ρ <.01.
4. Readers unfamiliar with the theory of measurement and reliability may find
it paradoxical that a measurement procedure, such as giving acceptability judgments
for sentences, can give reliable evidence about the relative acceptability of sentences
or sentence types for a population while providing very unstable information about
the individual informants in the sample on which those general findings are based.
There is n o paradox. What is happening is simply that the analysis of the experiment
as a whole is aggregating weak information gleaned from the individual informants
to provide a strong indication of the state of affairs in the target population. If there
were a way to subject each informant to a measurement effort as intensive as that
applied to the full sample (e.g., collecting 50 responses on every target sentence), it is
likely that we could acquire information about each informant roughly as reliable as
that we are able to obtain for a population from a sample.
A corollary is that having achieved reliable measurement of a population via a
sample does not give evidence that we have achieved reliable measurement of
individuals that make up the sample. This raises some concerns because when
linguists collect experimental data, they are often tempted to examine and comment
on the results for individual informants. Such efforts ought to be reserved for cases
where the reliability of the measurement procedure (vis-A-vis individual informants)
has been established. As of now, linguists have available no experimental procedure
for gathering judgment data that meets this criterion.
3
Designing
Experiments
on Acceptability
We do experiments to estimate the properties of a population on the

basis of tests applied to a sample drawn from that population.
The sample is always divided in some fashion so that we can
compare the effects of one experimental condition against another.
In a minimal psychological experiment, the sample is divided into a
test group and a control group, with some "treatment" being applied
to the test group. We expect to leam something about the effects of
the treatment in the population by looking for measurable differences
between the test and control groups after the treatment has been
applied.
Unfortunately, we can expect that there will almost always be
some difference between the two groups—even if we do nothing at
all except select and measure them. Thus the central problem in the
interpretation of experiments is deciding when a difference between
39
groups within a sample is large enough to warrant the inference that

there is also a difference in the population the sample represents.
The statistical approach to this problem aims to provide explicit
rational control of thefrequency of errors in this process. It holds out
no prospect of assurance, except the assurance that, under certain
assumptions, no more than some predetermined proportion of the
inferences we draw in inteφΓeting experiments will lead to error.
This level of control is achieved by using the normal random
scatter of observations that occurs within almost all samples as a kind
of yardstick against which to measure the differences among groups
or conditions. The backgroimd scatter, error variance, will spontane
ously produce differences between test and control groups even
without experimental interventions. Through careful statistical mod
eling of error variance, however, we can determine how often purely
accidental differences of various sizes will arise. On the basis of these
estimates, we can determine how likely it is that a particular observed
difference between two groups arose purely by chance. If the ob
served difference is, by this analysis, sufficiently unlikely to have
come about solely by chance, we may be warranted in drawing the
inference that there is a real difference in the population correspond
ing to the difference between the groups in cin experiment.
Another way to describe the same strategy is to say that we will
use within groups variance (the error variance that we find within
any sample of individuals) as a basis for gauging the size of the
between groups variance (the difference between groups that is,
potentially, the product of our experimental manipulation).
There are two problems for research design that arise in this
context. The tests and activities within an experiment must be struc
tured in some way to produce differences between groups. If success
ful, this activity will produce reliable differences that will support
inferences about the relevant population. However, reliable differ
ences are of little scientific interest unless we can determine how titey
came about in the experiment. Thus the second and more challenging
aspect of research design is the task of ensuring the meaningfulness
of the results achieved in experimental studies.
The fimdamental work that research design must do to achieve
meaningful results is to control variance. In many experiments (in
cluding ones on judgments), every response may be uruque and the
responses may spread over a wide range. These responses are poten
tially useful exactly because they differ from each other, provided
Designing Experiments on Acceptability 41
that some of this variabUity is under the experimenter's control.

Research design aims to impose structure on the variability in experi
mental results so that observed differences can be given some theo
retical inteφretation.
Because the statistical tests that are relevant to the problems
vmder consideration here are defined around the techrucal concept
of "variance," it will be helpful to review this notion in a httle detail.
As we'll see, effective design ahns at exercising thorough experimen
tal (i.e., procedural) and statistical control of three particular kinds
of variance.
3.1 Variance and Variance Partitioning
A statistical summary of a collection of scores will generally

need to indicate both the "location" of the collection of scores (e.g.,
their mean or average) and the degree to which the scores are spread
out arovmd that mean. One fairly obvious way to represent the
"spread-outness" of the scores would be to subtract the mean of the
full set of scores from each individual score and take the average of
these difference scores. Because of the way the mean is defined,
however, the sum and (therefore) the average of the difference scores
will always be zero. This difficulty is overcome by squaring each
difference score (which makes them all positive) before taking their
average. This procedure is illustrated in the two columns to the right
of the coluitm of judgment responses in Table 1. The mean of the
squared difference scores is termed the variance of the full set of
scores.
Designing a useful experiment can be seen as finding a way to
partition this value, the total variance of the set of scores, into two or
more components.
Suppose, for example, that the 20-judgment responses were
obtained in the following way. The first 10 scores represent the
responses of 10 iriformants who judged Sentence A, and the second
10 represent informants who judged a different sentence. Sentence
B. As illustrated in the middle and rightmost columns of Table 1, we
can now use the mean for each group of informants to partition the
total variance calculated earlier.
This partitioning is achieved first by using the mean for each
group, instead of the overall mean, to calculate the difference and
Table 1 Partitioning the variance for a set of scores.

Each Individual Each IndvldiMl'*
Each I Γββροηββ coinpervd group m««n
iMpoHM eompand
to Ih* IndlvMuar* oofflptrad to th*
to Ηι· grand « Μ η grand mMn
group HMan
Squarod Squarod Squand
JudgnMt Mtanooi DIftoranM MhnnM OlffaranM Diflaiwna
OIRaroaea
Soon Seen Soon Soora Soora Seora
2 .2.85 8.70 .1.60 2.56 -1.35 1.82
2 •2.SS e.70 .1.60 2.66 -1.35 1.82
RaapontM β 1.0S 1.10 2.40 5.76 -1.35 1.82
S 0.05 0.00 1.40 1.98 -1.35 1.62
6 1.0S t.io 2.40 5.76 .1.35 1.82
Sum 3 -1.fl5 3.80 •0.60 0.36 -us 1.82
36 5 0.05 0.00 1.40 1.96 -1.35 1.82
Group Mmi^ 3 -i.es 3.80 •0.60 0.36 -1.35 1.62
3.e0 1 .3.95 15.80 .2.60 6.76 -1J5 1.82
3 -i.es 3.80 .0.60 0.38 -1.35 1.82
^ .0.95 0.90 .2.30 6.29 1.35 1.82

Swito Β 4 .0.95 0.90 .2.30 S.2« 1.35 1.82
3 •1.95 3.80 .3.30 10.89 1.35 1.82
η··ροη··· 9 4.05 16.40 2.70 7.29 1.35 1.82
3 .1.95 3.80 .3.30 10.89 1.35 1.82
9 4.05 18.40 2.70 7.29 1.35 1.82
Sum β 1.05 1.10 .0.30 0.09 1.35 1.82
63 9 4.05 18.40 2.70 7.29 1.35 1.82
Group M*a«| 7 2.05 4.20 0.70 0.49 1.35 1.82
9 4.05 18.40 2.70 7.29 1.35 1.82
».»0
All
Raaponaae
Sum
• 9
Grand Maanl
4.9B
Sun of
Dllfaranoa 0.00 0.00 0.00
Soora*
(Sumof
Squmd Sum of
Squara* 126.95 90.50 36.45
OWarwioM
SOOTM)
(Mawiof
Squxd Varlanca 6.35 • 4.53 1.82
Diffannca
Soorai)
Within Batwaan
Tota
Totall
Varianc
Variancaa
Greupa * Group*
Varlanca Varlanca
NOTE: Each row in the main body o f the table represents a single informant. T h e informants are
divided into two groups, each of which includes ten individuals w h o supplied one judgment response
each. The first column of difference scores compares the individual's response to the grand m e a n ( 5 3 ) .
The second CQlumn of difference scores compares each informant's judgment response to die mean
for that informanl's group (3.90 for flie first group and 6.70 for ΰκ second). Finally, rightmost
column of difieieiKX scores shows ttte difference Iwtween the mean for Uie iiifbrmant's group arid ttie
grand mean. Thus, for the first ten informants this difference score is always 3.90 - 5.3 = - 1 . 4 0 , and
for Uie second it is 6.70 - 5 3 = 1 4 0 . Toward the bottom of the table, each value in the Sum of Squares
row shows the sum of the sûiml difference values in Oie same column. The values in the variance
row aie obtained b y dividing the values in Uie Sum of Squares row b y Uve number of individuals (20).
Crucially, tiie sum of Ote Within Groups and Between C i o u p s Variances equals the Total Variance.
This example demonstrates the general logic of variance partitioning, but not its actual application in
statistical tests. In statistical tests, the goal is to csftmote relevant p a r a m e t e r for a population b y way
of data takenfiroma sample drawn from that population. This complicates the logic and the math,
but the underiying concept remains (he same.
squared difference values for each individual score within each

group. This removes the effects of the overall difference between the
two groups from the calculation of the squared difference value for
each score. It is as though every informant is part of the same group
judging the same sentence. Thus the variance that remains is just the
error variance discussed in the previous chapter, the variance that
arises mostly from the ordinary meaningless scatter of scores around
any mean. Another, somewhat more transparent, term for this vari
ance is within groups variance and this is what is used in Table 1.
The term enor variance notwithstanding, it could be that all of
this variance is due to orderly processes and is ultimately explain
able. However, as Kerlinger (1973, p. 78) suggests, our ignorance of
its sources and our inability to control much of it—at least in any one
experiment—forces us to regard it as meaningless. It should be
noted, however, that the role of error variance is quite benign, despite
the need to control and minimize it in experimental design. It is
precisely error variance that most statistical tests use as the yardstick
against which experimental effects are measured and rehability esti
mated. That is, what many common statistical tests do is estimate the
overall stability of the sampled population from the measure of error
variance. This provides an index of how large differences between
groups must be to be unlikely to have occurred by chance.
Tviming to the experimental effects in the example in Table 1, it is
now possible to calculate a further variance component based directly
on the mearw for Sentence A and Sentence B. The mean for each of the
two groups is shown at the left of Table 1. In therightmosttwo columns
of the table, we calculate the difference between each informant's group
mean and the overall mean of the full data set, as well as the square of
this difference value. This, of course, is highly redimdant because the
group mean does not vary within the group. Just as we did with the
squared difference values for the individual scores, we take the mean
of the squared difference values for the group means at the bottom of
the rightmost coliunn. This is our second component of the totcil vari
ance in the experiment. This component reflects the experimental effect,
the differences in judgments due to the differences between Sentence A
and Sentence B, and is usually called between groups variance.
We can demonstrate that this procedure in fact consists in a
partitioning of the variance in the entire data set by adding up the
between groups variance and the within groups variance; the siun of these
variances is equal to the total variance.
The application of variance partitioning in actual statistical

tests involves many complexities that need not concern us here.
Among other things, variances computed in typical statistical tests
are understood as estimates of variances in the population from
which the sample is drawn. This complicates various aspects of the
calculation. Analysis of variance techniques also allows many differ
ent independent comparisons to be made within one experiment.
That is, factorial designs allow simultaneous tests on two or more
different treatments (e.g., the Sentence A/Sentence Β contrast) within
one experiment. Despite the added mathematical complexity these
extensions involve, the underlying logic of variance partitioning
remains essentially the same.
3.2 Experiment Design
To review, the art of experiment design consists in controlling

variance. There are two reasons for this. First, it is the ratio of the
between groups variance and within groups variance that will ulti
mately decide questions about the reliability of the result. If the
experimenter does not provide a manipulation that produces a suf
ficiently large difference between the means of the groups, there will
(from a statistical point of view) be no result at all. Second, it is the
appropriateness of the manipulations that induce differences be
tween the group means and the steps taken to ensvire that only these
differences distinguish the groups that will determine the meaning
fulness and validity of the result.
For the purposes of this discussion, I will look at experimental
design mostly as a problem of manipulating certain categorical vari
ables (e.g.. Sentence Type A versus Sentence Type B) so as to induce
effects on a continuous variable (acceptability). A finer differentia
tion of parameters will be needed eventually, but this distinction
should serve well enough here.
In designing an experiment, it is useful to consider three kinds
of variance that must be controlled. The first two we've already
encountered.
Between groups variance is due to the experimental manipula
tions constructed by the experimenter. In the example in Table 1, it
is the variiince due to the difference between Sentence A and Sen
tence B. The design problem for between groups variance is to ensure
that whatever difference is observed is in fact due to the intended

factor.
Within group variance, the second type relevant here, is gener
ally inevitable, ultimately benign, and not particularly difficult to
manage so long as its existence and approximate magnitude are
appropriately recognized in the design of an experiment.
The third kind of variance might be termed extraneous systematic
variance. This consists of any part of the variance in the experiment
that is orderly, but not the part of the variance that the experimenter
means to manipulate. If, for example, long sentences are judged less
acceptable than short ones, and if the sentences in some experimental
condition. Condition A, are consistently shorter than those in Condi
tion B, then A may be judged better than Β for different reasons than
the experimenter intends. We say that the contrast between Condi
tions A and Β is confounded; any difference we observe in connection
with this contrast may be due to either the syntactic difference be
tween them or to the length difference. Confounding can make it
impossible to derive any interesting or useful conclusion from an
experimental result. Control of extraneous systematic variance is an
extremely important issue in the design of experiments on sentence
materials because there seem to be many different ways that it can
arise. Section 3.3 will address this issue in detail.
Extraneous variance can also be an issue in the control of error
variance. When, for example, each group of infonnants in an experi
ment consists of two different kinds of persons (in equal propor
tions), there may be no issue of confoimding, but ttiere may be a
problem all the same. As noted, detecting statistically reliable differ
ences among experimental conditions depends ultimately on the
ratio of between groups and within groups variance. If there are large
systematic differences among two different types of persons (e.g.,
speakers of two different dialects), it may be worthwhile to control this
factor so as to reduce error variance. If, for example, the population from
which a particular group of irtformants is selected includes members of
two different dialect communities who have differing responses to the
experimental stimuli, the experimenter has two choices. The experi
menter can select each group of informants in such a way as to ensure
that approximately equal numbers of members of the two commimities
appear in each group of iiiformants. This should ensure that the
difference between members of the two conununities will have no
untoward effect on the experiment. Alternatively, the experimenter
might choose to select informants from only one of the two commimi
ties, or run two parallel experiments using informants from just one of
the communities in each of the two procedures. These latter alternatives
have the advantage of reducing the total error variance in the experi
ment, which improves the experimenter's opportunity to detect real
differences associated with the intended manipulatiorw.
One of the key challenges in designing meaningful and inter
esting experiments is to ensure that any systematic variance in the
experiment is either appropriately identified and directly manipu
lated or effectively neutralized and submerged within the error
variance in the experiment. As we'll see, the main focus of this effort
in the design of experiments on acceptability is directed toward the
design of the materials that will represent the various experimental
conditions. In an experiment on judgments, by far the most hkely
source of extraneous systematic variance hes in the opportimity to
inadvertently introduce differences between the materials in various
conditions.
Apart from the sentence materials themselves, the main hkely
sources of extraneous systematic variance in experiments on sen
tence judgments are the ordermg of those materials and differences
among informants. Each of these factors will be considered below.
3.3 Designing and Constructing Sentence IVIaterials
For experiments on syntax or syntactic processing issues, the

design and execution of sentence materials largely determines what,
if anything, the result will mean.
What makes the materials so critical is that there are many
intricately interacting systematic factors affecting judgments of indi
vidual sentences. There is evidence suggesting that an ir\formant's
response to an mdividual sentence may be affected by many different
lexical, syntactic, semantic, and pragmatic factors, together with an
assortment of extralinguistic influences that become haphazardly
associated with linguistic materials or structures. In particular, the
frequency and/or famiUarity of the lexical items, the length of those
items, the semantic properties of each item, and each item's associa
tions with other items and with miscellaneous facets of the infor
mant's personal history can affect informant response. Syntactic
factors such as clausal structure, the specific devices that are used to
implement various syntactic roles (e.g., choice of complementizer or

relative pronoun), the complexity of a structure, the famiUarity or
frequency of a structure, and parsabUity can all influence judgments.
Semantic factors such as the use of negation or the logical complexity
of a sentence (e.g., "I couldn't fail to disagree with you less"),^ or
pragmatic factors having to do with what presuppositions or entail
ments the informant is (or is not) sensitive to may affect judgments.
The interpretation the informant assigns to the sentence may have
unpredictable associations with the informant's life history that can
substantially affect judged acceptabiUty. In short, any one person's
response to any one sentence is usually massively confoimded.
These observations are hardly surprising. Linguists have long
been aware that judgments can be affected by many factors having
little or no relevance to syntax. Although various steps have been
taken to try to control these extrasyntactic influences, the effective
ness of these efforts has never been demonstrated.
The procedures described in this book differ from more typical
Unguistic approaches to this problem. Rather than attemptmg to sup
press extrasyntactic influences, these procedures control these influ
ences by trying to spread them as uniformly as possible across aU the
sentence types to be tested, and by assuming that (because each sentence
may be iiifluenced by a uruque assortment of the factors listed above)
no one sentence of a given syntactic type is an ideal representative of
that type. The procedures described here asstune that sentence judg
ments are subject to a host of factors contributing to extraneous system
atic variance. The factors enimierated above are for the most part,
apparently, principled and consistent in their effects, but many are not
among the kind of systematic factors that are thought to matter to
syntax. Thus the goal of experiment design here is to ensure that these
misceUaneous systematic but theoreticaUy irrelevant influences do not
inadvertently aUgn with whatever factors are of syntactic interest in a
particular experiment. If two sentences are meant to implement a
contrast between, say, violating or not violating the subjacency con
straint, the goal here is to ensure that, as nearly as possible, it is only
subjacency that differentiates the two cases—aU other lexical, syntactic,
semantic, pragmatic, and misceUaneous psychological factors being
uiiiformly distributed across the two sentences.
This ideal, of course, can be very hard to attain in a variety of
circumstances, so it is often helpful to employ what are caUed factorial
designs, as iUustrated with the "that"-trace materials in Table 2. The
Table 2 A factorial presentation of materials for a "thaf'-trace study.
Presence/absence of "that"
Without "that" With "that"
Extraction Site
Subject Extraction Who do you think Who do you think that
Ukesjohn? likes John?
Object Extraction Who d o you think Who do you think that
John likes? John likes?
NOTE: Eadi dimension of the tablerepresentsa "factor" of the design. Thus Uvere is a "ttuit"factor
(Presence/Absence of "that") and an Extraction Site factor. Each factor in this design is said to
have two "levels" (e.g.. Subject Extraction and Object Extraction are each levels of the Extraction
site factor. This is said to tie a 2 χ 2 design because it has two leveb on each of two factors.
experimental ideal would be to take "Who do you think likes John?"

and make it into a "thaf'-trace violation without changing its surface
form at all. That of course can't happen, but by adding "that" to the
sentence, the interpretation of any observed difference in accept
ability between the forms of this sentence with and without "that"
becomes more difficult. The fact that the sentence becomes a word
longer, that the added word is of imknown relevance to many of the
confounding factors listed above, that the clausal structure of the
sentence is somehow highlighted by the added word, and so on, all
raise legitimate questions about how any observed difference in
acceptability is to be interpreted. Factorial designs control for this
kind of confotmding by effecting the very same sort of manipulation
in another set of very closely parallel conditions. Thus, instead of
irwerting "that" only into a subject extraction case, it is also inserted
into an object extraction case. Because this is apparently immune to
the "that"-trace effect, it can serve as a control case. Now the "that"
trace effect is, strictly, the difference between two differences: the
difference between subject extraction sentences with and without
"that," <md the parallel contrast in object extraction sentences. The
"that"-trace effect is now seen as a greater decrement in acceptability
when "that" is added to subject extraction cases than occurs when
"that" is added to object extraction cases.
Another useful way of looking at the tJφe of structure shown
in Table 2 is that it isolates two factors. Extraction Site and the
Presence or Absence of "that," and sets up an interaction between
them. Structuring materials in terms of factorial patterns makes it
possible to use some of the more powerful statistical tools available.
Table 3 A factorial presentation of a coordination paradigm.
Antecedent Location
Local Antecedent Remote Antecedent
Coordination Type
N o Coordination Cattiy's parents require Paul requites that Cathy's
that Paul support himself. parents support himself.
Simple Coordination Cathy's parents require Paul requires that Cathy's
that Paul support himself parents support himself
and the child. and the child.
Coordination with Cathy's parents require Paul requires that Cathy's
"both" that Paul support both parents support both
himself and the child. himself and the child.
N O T E : There are two factors. Antecedent Location and Coordination Type, with two levels of the
first of these and three levels of the second. Thus this is described as a 2 χ 3 design.
Another illustration of a factorially structured set of materials

comes from the coordination experiment discussed in Section 1.3.3
(see Table 3). Here, one of the two intersecting factors, the Coordina
tion Type factor, has three "levels." Indeed, factorial designs may
have, in principle, any number of different factors and each factor
can have any number of levels. However, practical and interpretable
experiments will usually have only very modest numbers of factors
and levels of those factors.
The main virtue of designing materials in terms of factorial
structures is that it facihtates distributing extraneous systematic
factors uniformly across the experimental conditions. In paradigms
such as those shown above, whatever effects are pecuhar to the
particular lexical items and grammatical structures used in any one
sentence are likely to appear in all the others as well.
Unfortunately, the formal symmetry of a well-designed token
set carmot ensure the relevant sort of linguistic uruformity across all
the sentences within that token set. Formally driven variations in the
specific form of some base sentence type can sometimes produce
surprising interactions where one or more sentences in the set be
come strange in a way uruike parallel cases in other token sets. If
properly controlled, such interactions can be a research topic. If not
controlled, they confoimd and can completely imdermine particular
experiments.
In a typical psychophysical experiment that was intended to
compare six very similar stimuli (as in Table 3), all sbc stimuh would
probably be presented to each informant (possibly many times each).

There are, however, further constraints Aat are usually applied in
experimental work on sentence materials; no more than one member
of a token set such as in Table 3 is presented to any one informant,
and each sentence an informant judges is seen orUy once. The single
presentation constraint is motivated by the assumption discussed
earlier that sentences change infonnants. By hypothesis, the infor
mant is never able to confront a sentence in quite the same way twice.
Any future encounter with the same sentence is likely to be influ
enced in some way by the prior encounter. (See Note 4, Chapter 1,
p. 29, this volume.) The one-sentence-per-token-set constraint is
similarly foimded. There is experimental evidence (Bransford &
Franks, 1971) indicating that presentations of different but related
sentences within an apparently random list of sentences can never
theless be integrated by informants. Thus again the processing of the
second sentence seems likely to be affected in some way by the
informant's prior encovmter with a related sentence. In effect, the
constraints against repeated presentations or multiple presentations
from a single token set are aimed at making each encounter between
informant and sentence as much like a natural informative encounter
as possible. The mechanics of constructing sentence lists that respect
these constraints are discussed in Section 9.5.1.
Finding Verbs
Fleshing out an experiment design is mostly a matter of build

ing a list of token sets that correctly implement the design without
inadvertently introducing extraneous systematic variance of some
kind. In many cases, the core of this process consists in finding a set
of verbs appropriate to the needed structures. Indeed, it is often wise
to develop, carefully evaluate, and filter a list of verbs before con
structing any actual sentences.
The best aid in this process is the experience of an investigator who
is very well versed in the issue at hand. Syntactidans can sometimes
readily generate lists of verbs that exhibit specific properties.
Where more formal searches are undertaken, dictionaries that
include explicit grammatical tags (if they are sufficiently detailed)
may be helpful. Often, however, dictionaries generate lists of candi
dates that include many items that are irrelevant to a particular
search, making the process of identifying the target items quite
tedious. A shortcut that sometimes facihtates this sort of search can

be done with a thesaurus. The investigator begins by building the
largest possible set of appropriate candidates from memory or expe
rience and then looks each of these up in a thesaurus (electronic ones
are handy for this purpose). Generally the verbs linked to the those
in the initial hst will include some with similar syntactic properties.
This process can then be repeated for all new items foimd.
3.4 Controlling Context
There is evidence that the context in which sentences are judged

(i.e., other sentences seen previously) can affect judgments. Two
elements of the experiment designs used in the work reported m
Chapter 1 are intended to manage context effects.
Order of presentation clearly affects sentence responses. In
deed, some lines of psycholinguistic investigation have been built on
exactly the influence that a preceding sentence can exert on the
syntactic processing of a following sentence (see Bock, 1987, 1990).
Thus the procedures to be described below will present any one list
of materials in at least two different orders to efface order effects as
much as possible. This of course does not eliminate order effects.
Rather, it merely distributes them across the various experimental
conditions in such a way that they are unlikely to systematically
distort effects attributable to the targeted differences among sentence
types.
Filler sentences, which made up the majority of the sentences
informants judged in all of the experiments described in Chapter 1,
play several important roles in a judgment experiment. They create
the linguistic background against which the experimental sentences
are judged and at the same time provide a kind of global anchor that
can strongly affect both the overall range of responses given to the
experimental sentences and the location of that range relative to the
mean of all the responses in a session. Filler sentences, when well
chosen, can provide a kind of benchmark against which experimental
results from different experiments can become comparable. Fillers
also disguise the pattern of the experimental items, making it harder
for informants to guess which kinds of sentence the experimenter is
interested in. This can be important because when informants are
able to detect repeating patterns in the materials or other aspects of
the experiment, there is a danger that they will adopt response

strategies that are unrepresentative of their general behavior. Finally,
filler sentences can also serve to qualify informants by allowing the
experimenter to determine whether the informant is behaving "nor
mally" without reference to the experimental materials. For example,
sometimes in a session an informant will invert the response scale,
using the high end to represent sentences of low acceptability and
vice versa. Such cases are easy to detect with appropriate analyses
on fillers.
Because fillers create the general background against which the
experimental sentences are judged, they can have a marked effect on
the average rating given to experimental sentences. Cowart (1994)
shows that using a mixed filler list (two thirds highly acceptable
sentences, one third sentences of very low acceptability) as opposed
to a pure list (all highly acceptable) can shift overall ratings on
experimental up substantially (though without exerting any dis
cernible influence on the overall pattern of response).
The best strategy is to include a balanced list of fillers that
includes approximately equal nimibers of sentences at a wide range
of acceptability values.
3.4.1 Overview
In the approach described in this chapter, the core of a sound

design for an experiment on sentence acceptability is a paradigmlike
factorial structure that specifies and integrates one, two, or more
syntactic variables whose effects on acceptability are to be tested. The
heart of the process of executing an experiment designed along these
lines consists in developing a list of token sets that accurately imple
ment the intended syntactic contrasts without simultaneously induc
ing imwanted or vmcontroUed differences among the experimental
conditions. The background for the presentation of these experimen
tal sentences is provided mostly by a set of filler sentences. By careful
control of the distribution of experimental and filler sentences across
lists to be judged by various groups of informants, and by appropri
ately varying the order in which materials are presented, it is possible
to construct experiments that reliably detect a wide range of syntac
tically interesting differences in the relative acceptability of various
sentence types.
All of the designs discussed so far have been what are called
"within subject" designs. That is, they are intended to detect differ
ences across sentence types by assessing each informant's average
response in each of several sentence type categories. By augmenting
a within subjects design with a between groups factor that distin
guishes in some fashion among different types of informants, designs
of this kind can be extended to address many sorts of differences
among groups.
Note
1. This gem of unprocessable English is due to Raymond Smullyan.

4
The Sentence
Judgment Task
At various times, investigators have sought judgments on accept

ability, on acceptability relative to some context, on grammaticality,
on possibility or plausibility assuming some referential interpreta
tion, on truth relative to a picture, and on a variety of other aspects
of speaker responses to sentences. Although the relevance and inter
est of these various kinds of judgments is clear, the ability of infor
mants to differentiate these types has not been carefully investigated.
Most of the issues of experiment design that are discussed
elsewhere in this book are not very sensitive to the kind of question
put to the informant. Most of the general principles relevant to the
design of materials and the overall organization of an experiment
would apply to experiments on, say, judgments relative to scope
possibilities or possible patterns of reference relations, but I will
elsewhere assume that the target is a global judgment of the accept
ability of each sentence in isolation. This sort of judgment seems to
55
reflect the operation of powerful and orderly aspects of the cognitive

system of language, and thus is often seen as centrally important to
linguistic theory, hi this chapter, however, I will briefly consider
some of the issues that arise when informants are asked for different
sorts of judgments.
The puφOse of this chapter is simply to point out that the effects
that an experimenter's instructions have on an informant's perfor
mance can never be assumed. Although there is good reason to
explore effects associated with differing forms of judgment tasks, it
will take a good deal of work to make the contrast between different
sorts of tasks useful. I will illustrate this claim with a brief review of
an experiment that was designed to test naive informants' abihty to
control the criteria they apply in forming judgments. Although the
result is hardly definitive, it suggests something of the kind of work
that is required to make alternative judgment tasks useful.
One of the concerns most commonly expressed about the use
of naive informants is that it is thought to be difficult to provide
mformants with instructions that reliably elicit the "right" sort of
judgments, that is, ones uncontaminated by extralinguistic concerns
(see Bley-Vroman, Felix, & loup, 1988; Garden, 1970; Newmeyer,
1983, p. 61, among others). The experiments described in Chapter 1
generally ignored this issue; they used instructions that asked in
formants for very informal intuitive reactions to the sentences and
specifically emphasized that informants were not to try to assess the
sentences according to any grammatical rules they might have learned
in school. This does not reflect a lack of concern for achieving imcon
taminated results, but arises from the approach to this issue that is
reflected throughout this book; here the emphasis is on experimental
control rather than direct manipulation of the informant's approach
to the task. Thus the intuitive instructions applied in most of the
experiments described earlier should invoke a very inclusive crite
rion for judgment; the informant is invited to respond to each sen
tence on any grounds (other than school grammar) that seem appro
priate to the informant.^ The coordination experhnent described in
Section 1.3.3, however, divided informants into two groups, each of
which was given one of the sets of instructions shown in Table 4. The
"Intuitive Instructions" are representative of those used in most of
the other experiments described in Chapter 1 while the "Prescriptive
Instructions" were designed to invoke a prescriptivist stance toward
presented sentences. This particular contrast was constructed be
The Sentence Judgment Task 57
T a b l e 4 A l t e r n a t i v e i n s t r u c t i o n s for a j u d g m e n t e x p e r i m e n t .
Intuitive Instructions
Please read each of the sentences listed below. For each sentence, w e would like
you to indicate your reaction to the sentence. Mark your response sheet A, B, C, or
D. Use (A) for sentences that seem fully normal, and understandable to you. Use
(D) for sentences that seem very odd, awkward, or difficult for you to understand.
(Note: DO NOT USE "E.") If your feelings about the sentence are somewhere
between these extremes, use one of the middle responses, Β or C. THERE ARE NO
"RIGHT" OR "WRONG" ANSWERS. Please base your responses solely on your
gut reaction, not on rules you may have learned about what is "proper" or
"correct" English.
Prescriptive Instructions
Please read each of the sentences listed below. For each sentence, we would like
you to indicate whether or not you think the sentence is a well-formed,
grammatical sentence of English. Suppose this sentence were included in a term
paper submitted for a 400-level English course that is taken only by English majors;
would you expect the professor to accept this sentence? Mark your response sheet
A, B, C, or D. Use (A) for sentences that seem completely grammatical and
well-formed. Use (D) for sentences that you are sure would not be regarded as
grammatical English by any appropriately trained person. (Note: DO NOT USE
"E.") If your judgment about the sentence is somewhere between these extremes,
use one of the middle respoiwes, Β or C. Use Β for sentences you think probably
would be accepted but you are not completely sure. Use C for sentences you think
probably would not be accepted.
cause various commentators suggested that the rather laissez-faire

intuitive instructions we have generally used might suppress sensi
tivity to grammatical factors relevant to the coordination experiment
described in Section 1.3.3.
For current purposes, the key question is whether informants
actually use any different standards when they are asked to apply
school grammar criteria in making sentence judgments. Note that the
standard view seems to be that all of the Remote Antecedent sen
tences in (4) are unacceptable and presumably ungrammatical; it also
provides no basis for differentiating these sentence types.
The result of this manipulation is displayed in Figure 11. The
difference in instructions did not produce any difference of pattern
that seems to matter to linguistic theory. Apparently, the same fac
tors govern informant responses under both types of instructions.^
This result is not, of course, a definitive showing that instruc
tions cannot affect informant responses, although it does corroborate
- o - Intuitive Instructions / Local Antecedent

-ώ- Intuitive Instructions / Remote Antecedent
- · - Prescriptive Instructions / Local Antecedent
-A— Prescriptive Instructions / Remote Antecedent
1.0
% 0.5;
S
Si 0.0 φ
I
§
<
^ -0.5 f
T3
D
I -1.0 Intuitive Instructions Ν =22

Prescriptive Instructions Ν =21
-1.5
No Simple Coordination
Coordination Coordination with "both"
Figure 11. T h e i n t e r a c t i o n o f c o o r d i n a t i o n a n d b i n d i n g effects u n d e r t w o

different t y p e s o f i n s t r u c t i o n s to i n f o r m a n t s .
a general impression gained over many experiments that informants

have very little abihty to deliberately adjust the criteria they apply in
giving judgments. Rather, the result serves primarily to emphasize
that it is an empirical question whether any particular difference in
instructions in fact affects informant responses, and, if there is such
a difference, what form that difference takes. It seems plausible that
instructions to informants to approach sentences in different ways
can be used to highhght or emphasize different aspects of their
overall response to sentences. The current result shows, however.
that investigators who wish to ask informants to make different kinds

of judgments will need to demonstrate two things. First, they will need
to show that giving informants instructions other than those that ask for
a global, intuitive assessment of the sentence produces consistent, sys
tematic differences of pattern in informant responses.^ Second, they will
need to diagnose those differences that arise. That is, to make those
differences of pattern fully useful, investigators will need to demon
strate that one sort of instruction makes informants comparatively more
(or less) sensitive to specific aspects of the presented sentences (e.g., to
processing difficulty associated with, say, rare vocabulary or known
parsing complexity, or to semantic factors connected with reference or
the empiricaltiruthof the sentence). Insofar as these tests can be met,
alternative instructions may be useful in exploring the cogrutive re
sources that shape judgments.
An Aside on tlie Training ofUngnists
In this context, it is perhaps appropriate to consider the com

monplace suggestion that syntactidans in the course of their training
gain insights into natural grammar that aid them in giving judgments
uncontaminated by "extragrammatical" influences. This too is an
empirical claim, but not one that has, so far as I am aware, been
subjected to appropriate tests. There is nothing inherently wrong
with the suggestion. There are many areas within medicine, the
natural sciences, technology, and the arts where practitioners, after
appropriate formal or informal training, become able to make sophisti
cated judgments about natural phenomena that cannot be replicated by
untrained observers. Within linguistics, acquiring specialized skills of
this sort is a standard part of training in some subdisciplines, such
as phonetics. Nevertheless, in all the cases relevant to science, sophis
ticated observations can be checked against objective procedures.
There are means by which the skills of individual investigators can
be assessed by comparison to some observer-independent standard.
The same standards that apply to the assessment of procedures
used with naive informants could apply to the training procedures
to which linguists are subjected. If it is possible, for example, to train
novice investigators to apply clear alternative criteria to judgments,
it ought to be possible to demonstrate the effectiveness of these
procedures. For example, a relevant experiment would subject a
reasonably homogeneous group of graduate students to a single

clearly specified course of study and training. At the end of this
training, it ought to be possible to demonstrate that members of this
cohort consistently differ from the general population m the pattern
of response they produce with respect to numbers of syntactically
relevant patterns. Furthermore, there ought to be at least as much
consistency across individuals within this cohort as there is in the
general population, and it should also be possible to demonstrate
what specific factors the trained observers are more and less sensitive
to relative to naive iitformants. There are a variety of obvious prac
tical reasons why it would be difficult to organize and evaluate a
training program along these lines.
The fact that no appropriately vaUdated training program for
syntactic informants is available makes data obtained from expert
informants particularly difficult to interpret. Although there is little
doubt that advanced training can lead to useful sophistication in the
rendering of sentence judgments, there is no way to know how
rehable current training regimens are nor any formal way to assess
the skills of individual sophisticated informants. There is also good
reason to believe that advanced training is at best a mixed benefit.
Although it might be that sustained practice can sharpen an individ
ual's abihty to give reliable judgments, there are also reasons to
suspect (as has often been suggested) that training can produce some
theory-motivated bias. It is also worth considermg that syntactic
training leads to some amount of judgment learning; to read a syntax
article intelligently usually requires that the reader hold in mind a
particular pattern of judgments, whether or not those are the ones
the reader finds most natural. Enough practice in this exercise could
seriously distort an individual's ability to render the judgments that
are most natural to his or her own grammar.
These considerations suggest that the relatively objective pro
cedures described in this book (or other objective procedures) can
play a useful role in providing relevant fines of evidence that do not
rely on expert skills of unknown reliability.
Notes
1. It was thought appropriate to ask the informant to ignore school grammar

precisely because this is an iiifluence that seems to be something informants are
consciously aware of. Asking informants to make distinctions among syntactic, se
mantic, pragmatic, and other sorts of influences is quite another matter.
2. The interaction among the instructions variable, antecedent location, and
coordination type did yield a small reliable difference, Fi(2, 82) = 3.71·, F2(2, 46) =
4.18*. Note, however, that there appears to be no difference in pattern of any theoreti
cal interest for the results obtained under the two sets of instructions.
3. Gross differences in overall acceptability will be relatively easy to obtain,
but these are of no particular theoretical interest.
5
Presenting
Sentence
Materials to
Informants
All of the experiments discussed in Chapter 1 used written presen

tations of sentence materials to informants. There is little doubt that
judgments will differ for at least some classes of sentences if the
sentences are presented to informants in some other way. There are,
however, drawbacks associated with any choice of presentation method.
Table 5 presents an overview of some advantages and disadvantages
of various modes of presentation. Although this is hardly an exhaus
tive list of all the possible permutations, it may be useful in suggest
ing some of the considerations that ought to guide the choice of mode
of presentation.
Interesting facets of the psycholinguistics of the judgment pro
cess may be revealed by manipulating mode of presentation, although
63
Table 5 C o m p a r i s o n s o f a l t e r n a t i v e w a y s of p r e s e n t i n g s e n t e n c e m a t e r i a l s
to informants.
Presentation Mode Advantages Disadvantages
Written Easy to prepare and Relies on reading ability,

(printed administer. Facilitates which is both limited and
questionnaires) collecting data from variable in the general
large numbers of population. Speakers may
informants. Permits have different expectations
testing many informants (or tolerances) for the
simultaneously. Multiple syntax of written sentences
versions of the question than they do for spoken
naire can be presented sentences.
simultaneously.
Spoken Does not rely on literacy Experimenter may
(live presentation skills of the informant. inadvertently bias or
by experimenter) Can l>e used to present influence the informant's
materials to groups of responses. Testing groups
informants simul of informants
taneously. About as simultaneously is
easy to prepare as somewhat more awkward
written presentations. than with written
presentations. Potential
for different patterns of
results from different
experimenters. All
informants present must
respond to the same
materials in the same order.
Recorded Similar to above. More Time-consuming to prepare
(computer or tape controllable and more and execute. Somewhat
recording presented consistent across cumbersome equipment
via loudspeaker) multiple presentations may be required. All
of the same sentences informants present must
(e.g., to different groups respond to the same
or individuals). materials in the same order.
Recorded Similar to above. Better Time-consuming to prepare.
(computer or tape, control of experimental Relatively unnatural mode
presented via situation. Better of interaction with
headphones in opportunity to elicit informant. More labor-
laboratory setting) full cooperation and intensive per informant.
effort of informant. Expensive and difficult to
Materials can be n m more than one
presented to each informant at once.
informant in a unique
order.
Synthesized speech Similar to above. Most readily available
(computer generated Potentially provides synthesis technologies
speech) maximum control of produce poor quality
stimulus. Potentially speech. Tweaking improves
easier to prepare than result but at the cost of
recorded presentations. reducing the basic
advantage of easy
preparation. Unnatural
mode of interaction in the
informant session.
Presenting Materials to Informants 65
the constraints discussed in Chapter 4 will apply here as well; if

systematic differences in patterns of response can be demonstrated
for different presentation modes, it will be necessary to diagnose and
identify those differences to exploit them effectively.
6
Response Methods
and Scaling Issues
The simplest method of collecting responses in informal syntactic re

search is an instance of a forced choice procedure—the informant is
given only "yes" and "no" as possible responses. Increasing awareness
of intermediate possibilities (in judgments and in the grammar) and the
frequent discomfort of informants faced with forced choices has led
many investigators to offer informants one or more "maybe" or other
intermediate responses between the extremes of the scale (see Schutze,
1996, pp. 44-48, for a review of current descriptive practice).
Investigators have sometimes been concerned to try to match
the array of responses they offer informants to the array of degrees
of grammaticality they posit in the grammar. This is less important
than sometimes supposed, as the discussion in Section 1.2 in Chapter
1 suggests. If the set of possible responses exceeds the number of
categories provided by the grammar, no information is lost by pro
viding excess categories, and continued investigation ought to reveal
67
whether and how categories can be combined. Even at the opposite

extreme, if grammaticality is assumed to vary continuously, useful
data could still be collected via forced choice procedures. More
grammatical sentences will get a higher percentage of "yes" re
sponses, and the frequency of "yes" responses should change con
tinuously as grammaticality changes. This rather insensitive proce
dure will require larger numbers of informants to reliably detect
small differences in grammaticality, but there is no inherent limit in
the degree of detail it could detect.
When investigators offer informants an array of response op
tions (even if there are only two), they are engaging in an exercise in
measurement; that is, they are attempting to assign labels or numbers
to a phenomenon according to a rule. In such circumstances, it
becomes important to attend to the properties the investigator and
informant are ascribing to those symbols or numbers that represent
the outcome of each judgment event.
An array of possible responses may be either ordered or imor
dered. A set of unordered responses might differentiate among
reasons for deviance, as in a set such as "OK," "Sounds incorrect,"
or "Hard to understand." In such a case, the analysis of the result will
proceed along quite different lines than the analyses of the sample
experiments discussed in Chapter 1. Here the investigator will need
to compare the percentage of responses in the various categories for
various items, but there will be no continuum that unites the catego
ries. This is an instance of what statisticians term nominal data (see
Table 6); it consists in distributing a set of individuals, or a set of
responses, across a set of vmordered, discrete categories.
In more typical judgment studies, the various response options
are seen as an ordered series of points on a scale. In this case, several
different mathematical interpretations of the relation between the
points on the scale are possible and the difference between interpre
tations may have important consequences for the interpretation of
scale values and for the choice of appropriate statistical tests.^
If the response options are merely ordered, the scale yields
what is known as "ordinal data"; here, if two sentences are found to
have different scale values, we can determine from those values
which sentence is the more acceptable, but nothing more. In particu
lar, a difference in scale values on an ordinal scale carries no infor
mation about the size of the acceptability difference between sen
tences. To capture this further kind of information, the scale must
Response Methods and Scaling Issues 69
Table 6 Measurement scales.

Scale Operatims Performed Permissible Transformations
Nominal Identify and classify Substitution of any number

for any other number
Ordinal Rank order Any change that preserves
order
Interval Find distances or Multiplication by a constant
differences Addition of a constant
Ratio Find ratios, fractions. Multiplication by a constant
or multiples
SOURCE: After Stevens (1975).
yield what is called "interval data." An interval scale carries infor

mation about the sizes of the differences between points on the scale.
A 10°C range between daily high and low temperatures reflects the
same size interval whether the day's high is 10°C or 20°C, but we
cannot say (at least from a physicist's point of view) that a 20°C day
is twice as warm as a 10''C day because this scale lacks a true zero
point. When a phenomenon can be measured on an interval scale, a
range of more powerful statistical tools becomes available for exam
ining the resulting data. In the case of sentence acceptability, meas
uring on an interval scale makes it possible to determine which of
two differences in acceptabihty is the larger one. A study that seeks to
compare the impact of a subjacency violation and an ECP violation
would need interval level data.
In principle, there is a still more informative interpretation that
is sometimes available, but its relevance to the measurement of
acceptability is uncertain. If a scale yields "ratio data," it becomes
possible to calculate the relative sizes of individuals. That is, we can
say of a cabbage that it weighs twice as much as another because this
scale does have a true zero point. If in the linguistic domain a set of
subjacency violations and a set of ECP violations could be measured
on a ratio scale, we might be able to describe the results in terms such
as these: Subjacency violations are, on average, twice as acceptable
as ECP violations, or 33% more acceptable, and so on. Whether a ratio
scale for measuring acceptabiUty can be constructed is quite imclear
at present. See the references in Endnote 1 in this chapter and Bard,
Robertson, and Sorace (1996) for further discussion of issues in
scaling.
The problem of measuring acceptability is the problem of con

structing a response mechanism for informants to use that will
reliably deliver interval level data. I will assume that the human
cognitive system constructs judgments in such a way that there is in
principle a relatively smooth and continuous interval scale of accept
ability along which the acceptability values of all sentences lie. It does
not follow from this assumption that any particular response scale
will actually produce interval level data. Even when an investigator
conceives of a set of response options as an interval scale and in
structs informants to use it accordingly, merely ordinal data may
result. The balance of this chapter is concerned with this issue.
6.1 Category Scale Methods
A category scale is simply a sequence of respor\se categories that

are understood to be uniformly spaced along some underlying contin
uum. Thus we could construct category scales with two, five, twenty,
or more discrete steps that would evenly divide up the perceived
continuua of pitch, loudness, or other perceptual phenomena.
Category scales have been in use for a very long time^ and they
continue to be of value. Indeed, all of the experiments described in
Section 1.3 used category scales. Category scales are attractive be
cause they are conceptually simple, easy to explain to informants,
and quick and easy for informants to use. They also produce data
that are relatively easy to collect and to process.
Ur\fortunately, when linguists have constructed scales on which
to rate sentences, they have not always attended carefully to the
distinction between ordinal and interval scales. Occasionally an in
vestigator will clearly intend that a series of resporise categories be
seen and used as a scale, and yet, in addition to describing the
extremes or the scale as a whole, will also attempt to describe the
inner categories. The scale given in Table 7 illustrates this approach.
Even without the associated numerical labels, most informants
could tell from these descriptions which of two response categories
is meant to be the better one. It is very difficult, however, to ensure
that infonnants will inteφret the descriptions of the intermediate
categories such that all adjacent pairs of values represent the same
sized difference in acceptability. Attempts to explicitly describe the
inner scale values are probably as likely to defeat uruform treatment
Table 7 Caricature of an ill-advised technique for constructing a category

scale.
Rating Description
1 Entirely natural and acceptable

2 Acceptable, but possibly a little unnatural
3 Doubtful, although possibly acceptable
4 Worse, but not entirely unacceptable
5 Completely uiuiatural
6 Terrible
Table 8 A definition of a five-point category scale.

. . . Mark your response sheet A, B, C, D, or E. Use "A" for sentences that seem fully
normal, and understandable to you. Use "E" for sentences that seem very odd,
awkward, or difficult for you to understand. If your feelings about the sentence are
somewhere between these extremes, use whichever one of the middle responses
seems most appropriate
of the scale as to facihtate it. If informants don't manage to construct

approximately equal intervals between the vmits of the scale, rating
responses may not meet the mathematical assumptions of statistical
tests such as (-tests and ANOVAs.
It is generally wiser simply to identify a scale for the informant,
and perhaps to describe its end points, without attempting to de
scribe any intermediate scale values. Table 8 illustrates this approach.
This scale description is meant to invite uniform treatment of
the intervals without explicitly discussing the relation between dif
ferent scale categories.
Unfortunately, the intended relation between the steps of a
category scale can be difficult to obtain in practice. Stevens and
Galanter (1957) demonstrated that when category scales are applied
to a variety of continuously varying phenomena, subjects often use
these scales as though there were systematically different sized in
tervals between steps of the scale. This is especially true when the
stimuli to be judged are distributed imevenly across the scale. Ges
cheider (1976, p. 108) suggests one possible cause of this difficulty is
that subjects usually feel pressed to try to use the full scale for
whatever stimuh are presented and will try when possible to use all
categories about equally frequently. Although distortions due to
these problems are not necessarily dramatic, these considerations do

suggest caution when investigators are considering the use of a
category scale.
In general, it seems wisest to strive for the highest level of
measurement that seems theoretically appropriate to the phenome
non in question. This gives the best prospects for the reliabihty of
statistical results and provides the richest description of the phe
nomenon. Nevertheless, there may be circumstances where the sun
phcity and economy of the category scale will outweigh the risk that
it will fail to yield interval level data. Some theoretical questions are
resolvable by evidence that merely determines whether two sen
tences types have the same or different degrees of acceptability (i.e.,
where ordinal results are sufficient). Even where the magnitudes of
differences are crucial, category scales can yield interval level results
if they are used with appropriate caution. The most important step
to take in preparing an experiment that will attempt to derive interval
scale data from a category scale procedure is to pretest and carefully
evaluate all the materials to be presented to informants. The goal here
is to ensiu-e that the materials are more or less uniformly distributed
over tiie range of acceptabihty values that the experiment will tap,
without any noticeable preponderance of relatively very good or
very bad sentences. If, for example, the ratio of target sentences to
fillers is kept low and the fillers are distributed uniformly over a wide
range of acceptabihty values, it seems likely that interval level mea
surement can be achieved for the target sentences. Also, although
there are no practical statistical tools for correcting data from a
category scale that has produced ordinal data, we'll see in Chapter
12 that there are statistical ways to ensure that no inappropriate
conclusioiis are drawn from such data.
6.2 Ratio Scaie IMethods
Psychophysicists originally devised ratio scale procedures not

only in the hope of providing more accurate scaling results but also
in respor\se to this discipline's long-held interest in locating the
thresholds of various perceptual continua. Although the notion of a
threshold of acceptabihty is of doubtful relevance to the study of
language, ratio scale procedures provide the best prospects for achiev
ing reliable interval level measurement of judgments of acceptability.
Stevens (1975) describes a variety of psychophysical applications of

ratio scaling procedures and Bard et al. (1996) discuss the application
of ratio scaling methods (especially magnitude estimation) to accept
ability judgments.
Ratio scaling methods have been shown to provide a better
alignment between scaled responses and a variety of physical con
tinuua than typically arise with category scale procedures (Stevens
& Galanter, 1957). In general, ratio scaling methods are characterized
by an effort to give the subject maximimi freedom to set the magni
tude and range of the responses. If, for example, the subject is asked
to gauge the relative brightness of flashes of light, one possible
response method would be for the subject to draw a line for each flash
of light. The subject's goal is to make the ratios of the line lengths
correspond to the brightness ratios of the flashes of light. In doing
this, the subject is free to make lines of any length (within a prede
fined space) so long as the correct ratios are maintained. There are a
great many response measures of this general type and they are
widely used in psychophysics.
Although the analysis of ratio scale results is somewhat more
labor-intensive than the analysis of category scale results, the extra effort
is manageable for smaller studies (those that collect, say, 100 or fewer
responses from each informant and use perhaps one to two dozen
informants). For larger studies, however, the extra effort required to use
ratio scale procedures can be quite burdensome. Thus this section will
describe two standard ratio scale procedures (magnitude estimation
and line drawing) as well as further variants of these methods that have
been specifically adapted for use in large studies.
6.2.1 Magnitude Estimation
Magrtitude estimation is one of the most commonly used meth

ods of collecting responses in ratio scale procedures. In magnitude
estimation, informants are told to represent their evaluation of each
stimulus by assigning it a number. The number the informant chooses
can be as large or small as he or she likes. It may be any real number
or fraction, so long as it is not negative. The informant's choice of
number is constrained only in this way: The proportional relation
among the numbers assigned to different stimuli should reflect the
proportions of the stimuli themselves; thus a sentence that is judged
twice as good as another should get a number twice as high as that

assigned to the other sentence (Baird & Noma, 1978; Gescheider,
1976; Lodge, 1981; Stevens, 1975).
Bard et al. (19%) report on an extensive series of tests with
magnitude estimation and other ratio scale procedures that provide
compelling evidence of the utiUty of these procedures for judging
acceptabihty. Although most informants will not have encoxmtered this
procedure before, they seem to learn to use it rapidly and reliably.
The principal drawback to using magnitude estimation is that
(although it has been shown to be very widely appUcable) it requires
a very modest degree of mathematical sophistication on the part of
the informant. Not all informants are equally comfortable with num
bers and it may be that some adult informants do not have sufficient
faciUty with numbers to use magnitude estimation. Particularly where
informants are very young, infirm, or ilhterate, magnitude estima
tion may not be appropriate.
6.2.2 Une Drawing
Line drawing as a response is attractive because it demands

nothing of the informant's numerical sophistication and rehes instead
on the abihty to recognize proportions among hnes. Line drawring
responses are constructed by informants drawing lines in clearly de
fined spaces to represent the relative acceptabiUty of a sentence. Residts
obtained with Une drawing procedures aUgn very closely with those
obtained with magnitude estimation (Bard et al., 1996).
6.2.3 Scannable Une Drawing
A major Umitahon of magiutude estimation and line drawing

methods is that coding data obtained via these methods is relatively
time-consimîg and somewhat error-prone. In magnitude estimation,
each nvimerical response must be read and keyed into a computer file.
With line drawing, each line response must be measured by hand.
We have devised several scaimable line drawing procedures in
our laboratory in an attempt to combine the convenience and econ
omy of machine-scored response forms with the effectiveness and
simpUcity of line drawing in obtaining reUable interval level data.
Our procedures use a form that provides a ten-point scale for

each response item. We instruct informants that they should regard
the ten-point scale as a line (see the sample questionnaire in Appen
dix E). By selecting one of the ten items on the scale, they were
marking the end point of a line. Although they are marking only the
end point of the line, irvformants are told that they should use long
lines to represent high acceptability and shorter lines to represent
lower acceptability. Crucially, informants are free to set the range of
those variations in any way they like. They can, for example, vary
their responses between the second and fourth points on the scale, or
between the first and tenth. Their goal is simply to use line lengths
represented in this way to accurately indicate the relative accept
ability of the sentences presented to them. We find that informants
readily grasp the instructions for using this method. By incorporat
ing some elementary psychophysical scaling tasks witiiin various
experiments, we have demoitstrated that informants use this method
quite effectively and accurately to represent perceived magnitudes.
As a test of the effectiveness of our scannable line drawing
procedure, we conducted parallel experiments, one using a magni
tude estimation procedure and the other using a scaimable ratio scale
response. The results are summarized in Figure 12. The results from
the two procedures were reliably similar (p < .001), with the magru
tude estimation results accounting for about 85% of the variance in
the scaimable ratio scale responses.
Appendix Ε presents a set of sample instructions for a scannable
line drawing procedure. The principal disadvantages of this proce
dure are that it takes more time to describe adequately for informants
and, given its relative novelty, it requires greater caution in applica
tion. Until the procedure is applied more widely, it will be prudent
for investigators to collect some data on a simple psychophysical task
along with sentence judgment data to demonstrate the effectiveness
of the procedure within each experiment. An example of a checking
procedure appears at the top of the fourth page of the sample
questionnaire shown in Appendix E.
6.2.4 Laboratory Implementations
When ratio scale methods are used in a laboratory setting, it

becomes possible to use a variety of response mechanisms that are
9.0
r»=.8e
ο oJK
ο
ο ° °° χ " "
ο ο οο
/Ο ο
5.0 ο
° Χ ο
^4.0 4
3.0 4
0 ( 5 ^ °9 ο
^ ο
ο 0
0.0 1 1
0.0 5.0 10.0 15.0 20.0 25.0

Mean Magnitude Estimation Response
N = 19
F i g u r e 12. Results from a scannable ratio scale (SR) p r o c e d u r e c o m p a r e d

with magnitude estimation.
N O T E : T h e graph compares the geometric m e a n s of scale responses from the nonstandard
procedures to results obtained with magnitude estimation on die s a m e sentences. T h e target
sentences w e r e 90 filler sentences that provided the context for a pair o f 'that^'-trace experiments,
each of which used a different scaling procedure. T h e numi>er of informants w h o used each scaling
procedure is shown in the axis labels for each graph.
easier to explain to informants and more natural than scannable line

drawings. Informants can set the positions of dials that control
sounds or Ughts (e.g., louder/brighter indicates higher acceptabihty),
press triggers (higher pressure indicates higher acceptability), or set
sliders on mechanical scalelike devices. Because of their simphcity
from the informant's point of view, these response procedures will
be attractive when they can be conveniently applied.
Overview
One central fact about hiunan judgments about sentences seems

to be that these phenomena are very richly structured. There appear
to be subtle but stable differences among judgments people give
across the whole range of sentences in a language. The problems of
measurement considered in this chapter are the problems of captur
ing reliable information about those phenomena. The techiUques and
procedures discussed above seem to hold out the hope that linguists
can capture a great deal more information about judgment phenom
ena than they have to date. Given the insights that have been uncov
ered via informally gathered judgment results, there is reason to
hope that having more detailed and more reliable irxformation about
this phenomenon can help to imcover significant further insights.
Notes
1. There has been extensive discussion in the psychophysics literature on the

question of which statistics are and are not appropriate with the various levels of
measurement. Standard sources include Stevens (1975), Marks (1974), Baird and
Noma (1978), and Gescheider (1976). A useful collection of papers debating questions
about the propriety of using various statistical tests with various kinds of data appears
in Kirk (1972, pp. 47-80).
2. Gescheider (1976, p. 110) reports on a scale astronomers used successfully
for centuries to gauge the brighmess of stars.
7
Sampling
The experiments described in this book are attempts to estimate the

properties of a population on the strength of data collected from a
sample drawn from that population. The relevant population might
be the speakers of contemporary American English, the speakers of
a particular midwestem variety of contemporary English, elderly
English speakers in the Northeast, and so on. The population of
speakers is always finite at the moment the sample is drawn, but
open and able to include new speakers indefinitely (at least in prin
ciple). Obviously, the reliability of any estimate based on a sample is
sensitive to the representativeness of that sample.
7.1 Representativeness
There are many familiar domains within the social and behav
ioral sciences where exacting control of the sampling process is
critical to the representativeness of the sample. Polling on political
79
and social issues is perhaps the most famihar example. Here very
carehilly drawn samples of 1,000 or more individuals are common. In
general, the sampUng process is critical whenever "typical" individuals
are (or may be) imrepresentative of the population as a whole, or where
the investigator intends to estimate the proportion of infrequent types
of individuals (e.g., generally healthy 35-year-old males who show signs
of subclinical aphasia). Statistical theory also leads to a concern with
sampling because most statistical tests are based on a model that
assumes random samphng from the relevant population.
This is not to say, however, that samphng is always a critical
issue. Many samples used in biological, medical, and psychological
research are controlled more by issues of convenience and economy
than by sampling theory, and this does not generally reflect tmdue
laxity on the part of the researchers who use these samples (although
it can). Many biological phenomena are sufficiently consistent across
individuals within a population to make drawing importantly un
representative samples far harder than drawing representative ones.
For example, the relative length of the bones of the human upper arm
and upper leg is quite consistent across individuals (though, of
course, not invariant). Almost any small, casually drawn sample
from individuals found in a school, hospital, jail, or shopping mall
will provide an estimate of this ratio that wiU clearly distinguish
humans from chimps or other primates. Countless other anatomical,
physiological, and even behavioral traits show similar consistency
across individuals in numerous species.
Notable consistency is also apparent in a variety of human
behavioral traits, especially in language. Small samples are accepted
for a wide variety of articulatory, acoustical, and perceptual studies
of speech because theoretically relevant patterns can be readily and
rehably detected in these small samples.
Thus the task of drawing a representative sample is not neces
sarily difficult. The degree of care appropriate to a particular experi
ment depends upon what is already known about the variabihty of
the phenomenon in question (relative to the theoretically important
contrasts to be tested) and what particular purposes the investigator
has in mind. A medical researcher looking for reliable diagnostic
signs of some disorder may need much more precise information
about variation m contemporary human facial or body proportions
than does a paleoanthropologist who is considering how to classify
a particular fossU.
Sampling 81
Experience to date suggests that language researchers who are

working on widely spoken languages can select sufficiently repre
sentative samples in a wide variety of settings where a general mix
of members of the target community are present. In the United States,
these might include many schools and colleges, businesses, public
offices, and similar sites so long as there is no reason to suppose that
the particular site or institution might in some way attract (or repel)
individuals according to what (syntactic) dialect they use. Because
syntactic variation of the sort that generally interests syntactidans is
so far from consciousness for most individuals, it is hard to imagine
how any public institution could draw together an vmrepresentative
sample of speakers.
7.2 Linguist Informants
Perhaps the only class of sites that ought to be avoided in selecting

samples for experiments on sentence judgments are those that have a
high concentration of self-conscious language specialists (linguists,
writers, editors, language teachers, and so on). It is entirely possible that
among these groups, investigators can find the most skilled individual
informants—those who can render quick, corifident, and consistent
judgments on their native language. Although these individuals may
be extremely useful in exploratory research where the aim is to cover
many phenomena quickly, or where the needed judgments require a
sophisticated ability to focus on certain issues and ignore others, the
representativeness of an individual informant can be established orUy
through rigorous comparisons to larger samples.
7.3 Sample Size
The evidence reviewed in Chapter 1 demonstrates that judg

ments of acceptability, when collected via appropriate methods, are
highly stable. Nevertheless, the evidence of error variance in judg
ments reviewed in Chapter 2 shows that several judgment phenom
ena in English are clearly not so stable that the responses of a single
informant can reliably represent an entire community. Thus the
practical question about sampling is this: How many informants is
enough?
An important limitation on the possible answers to this ques

tion arises from constraints of experimental design. For example, the
need to counterbalance and the structure of the "thaf'-trace para
digm given in (6) above combine to require that the minimimi
experiment on this phenomenon will use four informants. It is much
better, however, to use a design that provides two different orderings
of each of the materials sets, for a total of eight informants in the
minimum "that"-trace experiment. Beyond this number, the number
of infonnants required is determined by the stability of the phenome
non itself. Some differences in acceptability are large, others small.
The ntimber of informants needed will be greatest for experiments
targeted on the smallest differences. Thus, for a phenomenon like
"that"-trace, there are three possible reasonable answers to the ques
tion of how many informants are needed for an experiment on
acceptability: (a) four, (b) eight, or (c) some higher multiple of four.
The data from a further "that"-trace experiment can be used to
clarify this issue. This experiment employed four different ratio scaling
procedures, each on a different group of irrformants, and was designed
as a test of differences among these procedures. There were 71 ii\form
ants, all of whom were tested on six sentences in each of the four relevant
sentence categories. Overall, the experiment produced essentially the
same results as the experiment described in Figvire 4 and no differences
among the procedures emerged. Nevertheless, the results show that
high overall stability is consistent with considerable diversity among
individual infonnants. Among these 71 informants, about one in ten
showed a reversal of the standard "that"-trace pattern; for these infor
mants. Subject Extraction with "that" is the most preened sentence type.
About one in five informants showed either this reversal or some other
pattern clearly different than the norm. For 16 of these 71 iiiformants,
some condition other than Subject Extraction over "that" was the least
preferred sentence type.
The impact of this level of variability was examined via a simple
resampling procedure^ in which a series of 20 random samples of size
four were drawn from the data from the 71 informants. Subject
Extraction with "that" was the least preferred sentence type in 19 out
of 20 of these samples, as in the overall result. However, the critical
interaction between "that" and Extraction Site was significant (p <
.05) in only 5 of the 20 samples. Of these 20 four-informant samples,
most would have failed to detect the "that"-trace phenomenon by
standard statistical criteria.
Sampling 83
The resuhs of a similar resampling test with samples of size

eight were different. Among these 20 samples, all showed Subject
Extraction with "that" to be the least preferred of the four sentence
types, and the critical interaction between "that" and Extraction Site
was significant in 17 of the 20 samples. That is, all but three of these
"experiments" would have successfully detected the "that"-trace
phenomenon. In short, phenomena whose effects are comparable m
magrutude and complexity to "that"-trace can be reasonably inves
tigated with samples as small as eight with the procedures used here.
Further tests were performed by combining data from two
subjacency studies (see Section 1.3.1) with data from a third experi
ment that used the same type of materials. Between the three experi
ments, there were 173 informants. An analysis of variance showed
that there was no rehable difference in overall pattern among the
three data sets. As before, 20 samples of four were randomly drawn
from this set. The Control condition was given the highest rating
among the four conditions in all of these 20 samples. The overall
ANOVA was significant in 18 of 20 of these samples. However, the
definiteness effect (i.e., the difference between the Indefirute, Defi
nite, and Specified Subject cases) was significant in only 2 of 20
samples. It appears that the large difference between the Control
condition and the three others can be reliably detected with samples
as small as four, but that the defiruteness effect requires a larger
sample.
As before, samples of eight produced effects much more hke
the full experiments reported earlier. When 20 samples of size eight
were drawn from the full pool of 173 informants, all 20 samples
showed an overall main effect, and the three-way contrast among the
Indefmite, Definite, and Specified Subject cases was significant in 13
of 20 of these samples. Thus most of these "experiments" would have
detected the defmiteness effect, although somewhat larger samples
would do better.
More formal methods for determining needed sample sizes are
available (Cohen, 1969). These methods are based on the concept of
"power" (the hkehhood that a test will detect a real difference of a given
size) and they corroborate the general picture described above.^
The practical conclusion these findings support is quite clear.
For acceptability phenomena of the magnitudes relevant to many
contemporary issues in syntactic theory, the minimvun reasonable
experiment will use eight or more informants. Some phenomena can
be detected with smaller samples, and some phenomena will require

larger informant groups, but few phenomena of current interest can
be adequately described with fewer than eight informants.
7.4 Comparing Groups
The discussion of sampling thus far has assumed that the goal
of research is to characterize general properties of groups of infor
mants. There are also linguistic issues that call for drawing samples
from two or more groups with a view to comparing the groups. If
the differences among the groups to be compared are relatively stark,
the differences may be detected with samples of sizes similar to those
discussed above. However, some differences between groups may
be stable but quite small. Where small differences are involved,
group comparisons may require samples many times larger than
those used in the studies discussed above. The size of differences can
be estimated through pilot testing and this information can be used
in estimating needed sample size.
Notes
1. A resampling procedure is one where numbers of small samples are ran

domly drawn from a larger sample to test various properties of the larger sample.
2. The methods Cohen (1969) provides for paired f-tests (pp. 60-64) can be
adapted to all of the contrasts discussed here. These methods suggest that the
"thaf'-trace phenomenon can be detected (with an 80% chance of success) with a
sample of ten. Using the same criterion, the largest effects described above (e.g., the
difference between local and remote antecedents where there is no coordinate struc
ture) can be detected with samples as small as four to six. The smallest effects (e.g.,
the pairwise contrast between definite pict«re-NP cases and specified subject cases)
require samples larger than 40. As noted earlier, whatever sample sizes the calcula
tions of power provide, practical answers will generally be rounded up to the next
higher multiple of (about) four because of constraints deriving from experimental
design issues.
8
Settings for
Experiments
In experiments on judgments (as in many psychological experiments

on humans), there is a paradoxical need to both standardize and
customize the procedure. On one hand, the experimenter's goal is to
get each informant to do the same task and to rely on the same suite
of cognitive resources to execute that task. On the other hand, each
informant is an individual who comes to the experiment with a
unique worldview, self-concept, personality, and temperament as
well as a certain range of intellectual competencies. The informant
also comes with ideas, attitudes, and opinions about language and
his or her degree of competency in language, the status of the infor
mant's own dialect or ideolect, the nature of experiments, the goals
and interests of experimenters, and a host of other issues. This stew
of factors can yield a wide range of results. While some informants
are bored, contemptuous, or even (on rare occasions) uncooperative,
others are painfully anxious to please the experimenter or intensely
85
fearful that the procedure will expose some embarrassing deficiency

(that will perhaps be secretly recorded). Thus the paradox is that, to get
informants to behave uiiiformly for the purposes of an experiment, it
may be appropriate to treat each one very much as an individual, to find
a way to engage each individual's interest, to cabn each individual's
fears and armeties, and to overcome each individual's particular mis
conceptions of the task he or she is being asked to do.
Every experiment has to negotiate a balance between the stan
dardization of materials and procedure required to meet scientific
goals and the individual customization needed to elicit the relevant
sort of uniformity in human performance.
The survey procedures with which this book is primarily con
cerned obviously emphasize standardization over customization.
Although ir\formants can raise questions or concerns that are not
answered by the questiormaire with the experimenter, the question
naire must nevertheless do most of the work of eliciting the right sort
of performance from the informant. Survey procedures compensate
for whatever added variability there is in performance by making it
feasible to collect data from many more informants than it is feasible
to test in more individualized procedures.
In situations where there is an opportunity to brief informants
individually and allow them to work in relatively private settings,
this level of customization is worthwhile.
8.1 Survey Experiments
Survey experiments are relatively easy to conduct in a variety

of institutional settings. In campus settings, a pool of informants can
be recruited in a variety of ways. Where the experiment is appropri
ate in some way to the subject matter of a course, an instructor can
sometimes justify taking class time for experiment participation,
especially if the experimenter will agree to debrief the group on the
experiment after they participate. Students can be asked to volimteer
in class settings and in a variety of other ways (i.e., bulletin board
notices, ads in a campus paper, and so on).
Wherever possible, informants should be comperxsated for their
effort. This may come m the form of money or course credit, where
Settings for Experiments 87
appropriate. Sometimes students are required to participate in one

or more experiments in coimection with a particular course.
Most public institutions that have substantial numbers of peo
ple present on a regular basis provide some mechanisms by which
iriformants can be recruited.
When a group of infonnants is assembled, it is often wise for
the experimenter to try to elicit the informants' support and coopera
tion via some brief introductory comments. Informants should be
assured that the work they are contributing to is worthwhile and that
their contribution is valuable. It is also important to ensure infor
mants that their responses will not be used in any way outside of the
experiment. Indeed, unless the nature of the experiment makes it
necessary to collect identifying information from each informant
(e.g., where each will participate twice and data from the two ses
sions must be matched up), it is wisest to ask informants not to write
their name or other identifying ir\formation on the response forms
and to reassure them that their responses are entirely anonymous.
The main hazard in briefing informants prior to their participa
tion is that it is possible to inadvertently bias informants in some way
relevant to the experiment. Experimenters need to exercise caution
to avoid saying anything that might indicate to informants which
particular kinds of sentences the experiment is focused on.
It is a minor precaution, but one worth taking, to scramble the
sequence of questioiuiaire types in a controlled way before distrib
uting them to a group of informants.
It is generally wise to distribute questioimaires in such a way
that one of each of the available types is distributed before a second
copy of any of the types is distributed, and to ensure that all types
have been used twice before any is used a third time, and so on. This
helps ensure that informants in various parts of a room, or that come
in during different parts of the day, are equally represented among
each of the informant groups.
Procedures should be designed to be done in a reasonable
amount of time. Tired, overworked, or impatient informants gener
ally will not produce as clean data as more fit ir\formants.
After informants have participated in an experiment, it is often
desirable to provide them with a debriefing report or some other
brief discussion of the experiment and its background. This is par
ticularly important in collegiate settings, although it is relevant else
where as well.
8.2 Laboratory Experiments
Generally speaking, when laboratory versions of judgment

experiments are undertaken, the same underlying concerns arise.
The comments above and the instructions to informants given in
Appendbc Ε should suffice as a guide to the organization of judgment
experiments in laboratory settmgs. The main advantage of labora
tory settings is that there is a greater opporturuty to tailor the process
to the needs of the individual informant. The iitformant has more
opportvmities to ask questions that may improve the quality of his or
her performance. In laboratory settings, it is also generally possible
to create a unique order of presentation of materials for every infor
mant. If there is another component of the task that does not involve
giving judgments, the experimenter has an opportvmity to provide
feedback to the mformant in the course of the session. This can be
helpful in ehciting good performance in a variety of timed psycho-
linguistic tasks. Needless to say, in any variety of judgment tasks,
experimenters should assiduously avoid any hint of evaluating,
critiqumg, or judging the sentence judgments informants make.
Another advantage of laboratory experiments is that there is
usually an opportimity to debrief informants after the experiment
session proper. This is sometimes arichsource of guidance for improv
ing the procedure and better imderstanding the informant's perspective
on the task. A debriefing session should also offer the informant some
mformation on the goals of the experiment. However, when ntunbers
of ii\formants are recruited from a single socially intercormected group
(e.g., a college class), informants should be asked not to share or discuss
information they gain in the debriefing session with other prospective
informants who have not yet participated.
8.3 Field Settings
One of the handier features of survey experiments is that they seem

to eUcit useful results even in rather unfavorable settings. In at least one
case, informants have been recruited to participate in a judgment experi
ment as they came through the checkout line at a university cafeteria (wifli
qviite satisfactory results). Investigators who need to recruit informants in
some sinular setting should let preliminary results guide them as to
whether informants perform normally in a nonlab setting.
9
The Organization
and Construction
of Questionnaires
This chapter will be concerned with the organization of question

naires in experiments that use written presentations of sentence
materials. The most important component of a questionnaire is a
clear and consistent set of instructions and, usually, some prepara
tory exercises. The unreachable, but approachable, goal is to ensure
that every informant has the same clear understanding of what is
wanted. Recall that any differences across informants as to how the
informant construes or approaches the task will contribute to error
variance in the experiment.
Although questionnaires can be used in various kinds of experi
ments, the discussion here will assume that the target experiment is one
that will involve a relatively large number of informants, and may be
run at several sites by several different individuals. This is a some
what extreme case that puts maximum pressure on the questionnaire
89
itself to brief informants accurately and consistently. The general

principles at work here apply also to simpler, less demanding cir
cumstances and to the other settings discussed in Qiapter 8. (At
several points, the discussion will reference an example question
naire shown in Appendbc E.)
The objective of this chapter is to give a general outline of the
logic of questionnaire construction. The discussion through Section
9.5 will leave it to the reader's imagination to determine how the
various principles and effects specified here might be achieved. The
last section of the chapter gives a kind of schematic procedure for
implementing the logic described earlier. This procedure might be
executed by hand or by way of various kinds of computer software—
from word processors and spreadsheets to database programs. Ap
pendix C provides still more concrete directions on how to construct
questionnaires in a specific spreadsheet program, Microsoft Excel.
9.1 General Instructions and Demographic Data
In addition to instructions on the judgment task, informants

usually need some guidance on various general aspects of the experi
ment. Infonnants should know who is responsible for the experiment
and whom they can contact for further ir\formation. There may be
multiple forms, documents, or other materials that informants will
need to be introduced to, or special instructions on returning the
answer sheet, and so on.
For many syntactic investigations, there will be no need to
collect demographic data. For studies of variation where such data
are relevant, it is perhaps wisest to collect the data after informants
imdertake the judgment task.
9.2 Response Training and Practice
In experiments that use category scales, it generaUy will be suffi

cient simply to describe the scale and how it is to be used. Most people,
especially those who are college age or younger, are quite fanuliar with
such procedures. However, experiments that use any variety of ratio
scaling procedures will need to provide careful instructions on the
scaling procedure itself. Unless the experiment uses instructions that
The Construction of Questionnaires 91
Table 9 Sample iiistructions for informar\ts giving sentence judgment criteria.

We need some information from you about your evaluation of some sentences
we've listed Ijelow.
We would like you to imagine that your job is to teach English to speakers of other
languages.
For each sentence listed below, we would like you to do the following. Please read the
sentence, then ask yourself if the sentence seems English-sounding or not. Suppose
one of your students were to use this sentence. If we ignore pronimciation, would the
student sound like a native speaker? Or would the sentence seem strange or unnatural
to a native speaker no matter how it was pronounced? Your task is to tell vis how
English-sounding each sentence is, using a scale.
have been used and tested previously, it will be prudent both to test
the instructioixs directly on a few ir\formants and to include some
check on the effectiveness of the instructions within the experiment
itself. One good way to help the informant master the task and also
check on the effectiveness of the instructions, informant by infor
mant, is to include practice cases such as those included on the fourth
page of the S c i m p l e questiormaire ("Practice" section) in Appendix E.
9.3 Judgment Criteria
Questionnaires must specify for informants what criteria they

are to apply in judging sentences. As noted in Oiapter 3, there is httle
direct evidence that informants can in fact exercise any very subtle
control over the criteria they use in judging sentences. Nevertheless,
there needs to be some consistent guidance that leads informants'
attention to the perceived acceptability of the sentences.
The instructions shown in Table 9 (taken from the questionnaire
in Appendbc E) have proven effective in this role, but many alterna
tives are possible.
9.4 Practice and Benchmaric Materials
Even with quite straightforward procedures (such as using a

category scale), informants usually need a Uttle practice before they
begin judging the main sentence list. Failing to provide practice or

warm-up materials introduces urmecessary noise (i.e., error vari
ance) into results from the first few sentences on the main sentence
list.
It is sometimes helpful to include in a questionnaire one or more
blocks of "benchmark" sentences that are identical across all versions
of the materials, that appear in the same linear position within the
materials to be judged (e.g., at the beginning), and that appear in
exactly the same order (within the list of benchmark sentences).
These sentences provide a baseline by which some useful compari
sons can be made across informants, and help to detect certain kinds
of error. For example, occasional informants will invert a response
scale, using the high end as the low end. If an informant's pattern on
the benchmark sentences is strongly but negatively correlated with
the average pattern on the benchmark sentences, a scale inversion
may well have occurred. A block of benchmark sentences appearing
just before the beginning of the main sentence list (i.e., the target
sentences and fillers that are the focus of the experiment) can double
as practice materials, although in this case it might be wise to ignore
results of the first half dozen or so items in the analysis to allow for
instability while the informant is becoming familiar with the task.
The first 10 sentences in the questiormaire in Appendix Ε (Items
11-20) are benchmark sentences.
9.5 The Main Sentence List
As noted in Chapter 3, the fundamental unit of accoimt in

constructing materials for sentence experiments is the token set. The
number of token sets required for a particular experiment will gen
erally be a function of two values: the number of sentences in each
token set, and the niraiber of sentences of each sentence type to be
presented to each informant. Thus an experiment that has four
sentences in each token set and that will present six instances of each
sentence type to each informant will need 24 token sets.
The experimenter will need to settle on some ratio of filler
sentences to experimental sentences. In general, there should be at
least twice as many filler sentences as there are experimental; three
to four times more fillers than experimentals is desirable where this
does not lead to an unreasonably long questionnaire. If the ratio of
Table 10 Rules for counterbalancing the sentence lists p r e s e n t e d t o different

groups of informants.
• Each informant will judge no more than one sentence from each token set.
• Each informant will judge all experimental conditions and will see equal numbers
of sentences from each condition.
• Every sentence in every token set will be judged by some informant.
fiUers to experimentals goes to two or below, there is an increasing

danger that informants will become particularly aware of the sen
tence types the experiment features, with unpredictable consequences
for the informants' patterns of response.
9.5.7 Counterbalancing
The goal of coimterbalancing is to help efface the effects of factors

irrelevant to a particular experiment and to highhght those effects the
experiment is designed to reveal. As noted earUer, informant responses
to sentences can be very sensitive to a wide variety of factors other than
the logical and structural features of sentences that interest syntactic
researchers. Nevertheless, structural features must always be imple
mented in specific sentences with specific lexical content. Thus any
single global judgment about a particular sentence is necessarily sus
pect; there is no usable way to determine how various disparate factors
contributed to its formation and thus no easy way to detemune how
that single judgment is relevant to the many theoretical questions to
which the target sentence might be relevant. Coimterbalancing aims to
distribute both the idiosyncratic and the systematic structural effects
that arise in a single sentence across the whole experiment in such a way
that the systematic effects can be reUably discriminated from the back
ground blur of idiosyncratic effects.
Implementation of counterbalancing consists in honoring three
rules in the construction of a set of sentences that mformants are to
judge. To satisfy the rules, it is necessary to divide the available
informants into "materials groups" (of roughly equal size), there
being as many of these groups as there are different sentence types
in each token set. AU the members of each "materials group" wUl
judge the same subset of the experimental materials. The counterbal
ancing rules are enumerated in Table 10.
The consequences of the rules listed in Table 10 for a complete

list of sentences are illustrated in Figure 13. In each of the scripts,
there are equal nimibers of each sentence type.
9.5.2 Blocking
Note that counterbalancing has nothing to say about the order

in which sentences appear. This leaves open the possibility (in fact,
a likelihood) that a truly random procedure for ordering the sen
tences might concentrate all or most of the sentences of a given
condition in one relatively small part of the list. This would be
undesirable because the informant's state of mind may well change
in relevant ways as she proceeds through the questiormaire. Fatigue,
boredom, and response strategies the informant may develop over
the course of the experiment can have differing effects on sentences
judged at various points in the entire procedure. The role of blocking
is to try to manage these unwanted effects by ensuring that they act
on all experimental conditions equally.
In the standard case, there will be as many blocks in a sentence
list as there are sentences of any given experimental condition in the
list. We get the same result by dividing the total number of token sets
by the number of conditions within each token set, because the total
number of token sets should be an integral multiple of the number
of conditions. In other words, if there are four sentence types repre
sented in each token set, and if each informant is to judge, say, six
instances of each of those types, the list of materials the informant
will judge will be organized into six blocks.
To implement blocking, a controlled randomization procedure
should assign each experimental sentence to a block so that only one
sentence of each type appears in each block. Each block should also
contain the number of fiUer sentences determined by the filler ratio;
that is, we multiply the number of experimental sentences in the
block by the ratio of fillers to experimental sentences to determine
how many fillers should appear in each block. Figure 13 shows two
example schematic scripts that have been organized by blocks that
respect the principles stated above.
Matwtak
Sxperimenlal MalwMt LIrt FWerLbl
Tol»iu8«L1J_S«il»noe_A Tol»n_S*UJ_S»ntone«_B RII*r_8«il*nM_1

T<it(«iu8«L2J_8«il»no»_A Toton_S«tJL/_S«il«ne«_B FIII«r_8«nl*nM_2
Toton_8«L3J_S«iHno»_A Tokw_S*L3.y_8«itonM_B FUIw_8«<lMin_3
Tot»iu8*L4J_Senl«not_A To(an_S«L4J_S«it«ne«_B Flll*r_8<nlinn_4
Flll«_6«)»WK!e_S
FHItf.SanlMiM.e
Sotpti 8αΙρΙ2
ToMnMI/SwilKmA T o k a i M I / G w i l K leaB
Toian S M 2/ Sanlino* A w B
TolanSMS/SwimaB T o k n M 3 / 8 « l k f «aA
T(ilwiS«t4/8MHnnB T ( g k K l M 4 / 8 M M f K»A
Rtac S a n a n o M
Ordered Scripts
Ortatad Script l a Ofdawd Script 2a
Blockl Tol<en_S*t^_Sanlenca> Blockl (nUar_Sanlanoa_1

Fllar_S*ntinoaj5 Rllar_SanlMica_3
Fllar_Santinoa_1 Tal(en.S*L4^_8antencaJ^
TolOK)_SaL3J_Sent«ioa_B raiar.SanlMioa.S
ram_9anHnoa_il Takan_SaUJ_8*nlano*_B
Block2 Flll*r_8entanea_e Bk)ck2 Teken.SetJL/_Sentanoe_B
Tol<en_SeL4J_S*ntai>oa_B FlUar.Sanlanoa.e
Fgiar_8anlanoe_2 Token.Set_3_/_SeiTtenoe^
Tokan_8at_1^_8anlanoa_A Flllar_Sanlanoa_2
Fmar_Senl>noa_4 Flllar_Santanc*_4
Figure 13. Schematic example of counterbalancing.

N O T E : Tfiis illustration sliows how lists of token sets and fillers can be distributed across two
scripts and t w o ordered scripts in a w a y that i s consistent with the counterbalancing rules stated
in Table 10. At the first stage, the Experimental Materials List is divided so that e v e r y sentence
from every token set gets passed on to s o m e script, and so that all the types of sentences (just two
in this case) are equally represented in each script. When the resulting lists o f experimental
materials (what I sometimes refer to as preliminary scripts) are combined with the filler list (which
is the s a m e in both cases), the result is a script, a particular subset of all the sentences used in the
experiment. T o put the materials in each script into an order, it is necessary first to organize them
into b l o c k s with one experimental sentence o f each type in e a c h b l o c k and whatever n u m b e r of
fillers needed to fill out the block (there being the s a m e n u m b e r of fillers in all blocks). Both the
assignment of the sentences in a script to blocks and the final ordering of materials within each
block are randomly determined. M a n y different ordered scripts can b e derived from any one script
(hence the lowercase letters at the e n d s o f the ordered script n a m e s ) . W h e r e this is feasible (as in
computerized laboratory experiments), each informant should see a imique ordered script.
9.5.3 Randomizing
Randomizing sentence lists is done within blocks. Generally, a

truly random procedure within the block is appropriate. Occasion
ally, it may be appropriate to control the randomization procedure
to ensure that experimental items don't appear within the first two
or three items of the first block if there are to be particularly few
warm-up or practice items.
9.5.4 H/lultiple Scripts and Versions of Scripts
There will be as many different scripts in an experiment as there

are conditions within each token set; the number of sentence types
per token set determines how many different materials groups will
be needed to satisfy the counterbalancing rules. Each script should
be constructed along the lines described above. The process that
creates ordered scripts (both the assigrunent of experimentals to
blocks and the randomization of order within blocks) is done inde
pendently for each script. Furthermore, because order (as noted
earlier) can influence the results for individual sentences, it is highly
desirable to construct two or more independent orderings of each
script. Although this exercise may seem somewhat excessive (or
obsessive), without this degree of caution, it is corrunon to find
unwanted and uninterpretable statistical interactions among the
experimental factors and those factors that represent the assignment
of informants to scripts or the assignment of token sets to conditions.
9.6 Constructing Questionnaires
This section will outline a process for constructing question

naires. It will assim\e that the process is executed in spreadsheet and
word processing software but will not be concerned with details of
implementation. Those investigators who are sufficiently familiar
with the relevant software tools will be able to carry out the process
from the description provided here. Those who would like guidance
on the use of the relevant functions in Excel will need to consult
Appendix C.
T a b l e 11 Experimental sentences projected from sentence c o m p o n e n t s .
Sentence Components Token Set 1 Token Set 2
Component A Who did the counselor Who would Nona like

(main clause) want
Component Β for for
(complementizer)
Component C to hug Chantal? to pay Sarah?
(subordinate clause.
subject trace)
Component D Chantal to hug? Sarah to pay?
(subordinate clause.
object trace)
Sentences Projected From Components
No "for," Subject Who did the counselor Who would Nona like to
Extraction case want to hug Chantal? pay Sarah?
No "for," Object Who did the counselor Who would Nona like
Extraction case want Chantal to hug? Sarah to pay?
With "for," Subject Who did the counselor Who would Nona like for to
Extraction case want for to hug Chantal? pay Sarah?
with "for," Object Who did the counselor Who would Nona like for
Extraction case want for Chantal to hug? Sarah to pay?
NOTE: Here the components of two token sets are enumerated in the upper panel of the table. In
the lower panel, spreadsheet functions have been used to assemble the components into finished
sentences of the intended types.
9.6.7 Constructing Experimental Materials Lists (Lists of Token Sets)
It is easy enough to prepare lists of token sets as simple lists or

tables of sentences in a word processor. Taking this approach, however,
creates ample opportunihes for errors to arise. A better approach is to
construct token sets as ordered sets of sentence components and then
to assemble those components into finished sentences via some auto
matic process (e.g., the concatenation functions exploited in Appendix
C, although there are many other possibilities). This makes it much
easier to ensure that the structure of each token set is consistent with the
overall plan. This procedure is illustrated in Table 11.
9.6.2 Constructing Filler Lists
No special procedures are required for generating filler sen

tences, although it may be necessary to identify and label fillers
carefully, especially if the filler hst is structured according to the level

of acceptability of each sentence.
9.6.3 Integrating Materials and Fillers by Blocks
The heart of each questionnahe is the list of sentences infor

mants will judge. There are three parts to the process of constructing
this list. First, the order of the final list of token sets to be used in an
experiment should be randomized to minimize the chance of inad
vertent similarities among items clustering together in particular
questiormaires. Identification labels should also be assigned to the
items and preserved in all further processing up to the fh\al question
naire.
Second, a counterbalanced set of preliminary scripts must be
derived from the hst of token sets that make up the experimental
materials. A preliminary script is just the list of experimental sen
tences that will ultimately go into one questiormaire. The various
sentences that make up each token set are distributed across the
complete set of preliminary scripts in a way that respects the coun
terbalancing rules. Note that in Table 12, each of the preliminary
scripts includes three sentences of each of the four types that make
up each token set. The number of preliminary scripts constructed for
an experiment should equal the number of different sentences that
make up each token set (or some integral multiple of this number).
Third, a four-step process is used to combine each preliminary
script with the hst of filler sentences. This process is illustrated in
Table 13.
The process described in Table 13 is less burdensome than it
might first appear. The spreadsheet implementation described in
Appendbc C requires less than 2 minutes to construct each new
ordered script. In Table 13, the sentences are represented only by
their identifiers. Each panel of the table includes the same materials;
only the order changes from panel to panel. In Step 1, a preliminary
script is copied into the middle column of a matrix that includes
preexisting block identifiers in the first column. Random numbers
are then generated in the third column. In Step 2, each subhst of items
of the same type (e.g., all the TSnc items, or all the F{iller) items) is
sorted according to the value of the random number assigned to each
item. Because the column of Block IDs is not included in the sort, the
Table 12 Preliminary scripts.
Preliminary Preliminary Preliminary Preliminary
Script 1 Script 2 Script 3 Script 4
NT/SE NT/OE WT/SE WT/OE

Who are the students assuming Who are the students assuming Who are the students assuming What are the students assuming
helped Vicky with the taxes? Vicky helped with the taxes? that helped Vicky with the taxes? that Vicky helped with the taxes?
Who did Sally mean had seen Who did Sally mean her father Who did Sally mean that had seen Who did Sally mean that her father
her father in the backyard? had seen in the backyard? her father in the backyard? had seen in the backyard?
Who did Nona estimate Who did Nona estimate Who did Nona estimate that Who did Nona estimate that
outscored Sarah? Sarah outscored? outscored Sarah? Sarah outscored?
NT/OE WT/SE WT/OE NT/SE

Who do they figure the Who do they figure that will Who do they figure that the Who do they figure will impress
professor will impress? impress the professor? professor will impress? the professor?
Who had the people heard they Who had the people heard that Who had the people heard that Who had the people heard would
would be fighting for? would be fighting for them? they would be fighting for? be fighting for them?
Who could Jeff remember Who could Jeff remember that Who could Jeff remember that Who could Jeff remember had
Tiffany had written to? had written to Tiffany? Tiffany had written to? written to Tiffany?
WT/SE WT/OE NT/SE NT/OE

Who had he dreamed that Who had he dreamed that I Who had he dreamed would Who had he dreamed I would
would marry me? would marry? marry me? marry?
Who did they claim that beat Who did they claim that the Who did they claim beat up Who did they claim the burglar
up the burglar? burglar beat up? the burglar? beat up?
Who might Emily declare that Who might Emily declare that Who might Emily declare had Who might Emily declare Henry
had avoided Henry? Henry had avoided? avoided Henry? had avoided?
(continued)
§ Table 12 Continued
Preliminary Preliminary Preliminary Preliminary
Script 1 Script 2 Script 3 Script 4
WT/OE NT/SE NT/OE WT/SE

Who did the broker state that Who did the broker state would Who did the broker state the Who did the broker state that
the accountant would fire? fire the accountant? accountant would fire? would fire the accountant?
Who were the children told that Who were the children told Who were the children told they Who were the children told that
they would write to? would write to them? would write to? would write to them?
Who did the counselor recall Who did the counselor recall Who did the counselor recall Who did the counselor recall that
that Chantal had hugged? had hugged Chantal? Chantal had hugged? had hugged Chantal?
NOTE: Each preliminary script constitutes a counterbalanced set of experimental sentences that would be appropriate for use in one questionnaire. Every version of
every token set is used in some preliminary script, and within each preliminary script, every sentence type is equally represented. All the sentences on each row are
members of the same token set. The labels above each block of three rows indicate the type of the sentences in that block for that column, such as NT/SE (No "thaf'/Subject
Extraction) or WT/OE (With "that"/Object Extraction).
Table 13 A p r o c e s s for i n t e g r a t i n g a n d o r d e r i n g e x p e r i m e n t a l m a t e r i a l s
a n d fillers b y b l o c k s .
Step 1 Step 2 Step 3 Step 4
Source: A Randomization Sorting by Mock Randomization

prallmlnary script within types segregates within blocks
plus the tiller list randomly assigns blocks determines final
•nd random sentences to order
numbers blocks
BIk Random BIk Random BIk Random BIk Random

ID Item 10 # ID Hem ID « ID Item 10 « ID Item I D «
1 TSIA .568 1 TS2A .158 1 TS2A .158 1 F1S .026
2 TS2A .158 2 TSIA .568 1 TS4B .772 1 Tsec .090
3 TS3A .683 3 TS3A .683 1 Tsec .090 1 TS2A . 1 5 8
1 TS4B .772 1 TS4B .772 1 TS120 .538 1 F7 .256
2 TS5B .820 2 TS5B .820 1 F15 .026 1 F18 .314
3 TS6B .856 3 TS6B .856 1 F7 .256 1 F6 .481
1 TS7C .392 1 TS8C .090 1 F18 .314 1 TS12D . 5 3 8
2 Tsec .090 2 TS7C .392 1 F6 .481 1 F16 .623
3 TS9C .517 3 TS9C .517 1 F16 .623 1 TS4B .772
1 TS10D .713 1 TBI 2D .538 1 F17 .788 1 F17 .788
2 TS11D .677 2 TS11D .677 2 TSIA .568 2 F5 .198
3 TS12D .538 3 TS10D .713 2 TS5B .820 2 F2 .295
1 F1 .626 1 F15 .026 2 TS7C .392 2 F14 .335
2 F2 .295 2 F5 .198 2 TS11D .677 2 TS7C .382
3 F3 .637 3 F12 .247 2 F5 .198 2 F11 .482
1 F4 .440 1 F7 .256 2 F2 .295 2 TSIA .568
2 F5 .198 2 F2 .295 2 F14 .335 2 F1 .626
3 F6 .481 3 F13 .305 2 F11 .482 2 TS11D .677
1 F7 .256 1 Fie .314 2 F1 .626 2 TS5B .820
2 F8 .954 2 F14 .335 2 F9 .911 2 FO .911
3 F9 .911 3 F4 .440 3 TS3A .683 3 F12 .247
1 F10 .562 1 FS .481 3 TS6B .856 3 F13 .305
2 F11 .482 2 F11 .482 3 TS9C .517 3 F4 .440
3 F12 .247 3 F10 .562 3 TS10D .713 3 TS9C .517
1 F13 .305 1 F16 .623 3 F12 .247 3 FIG .562
2 F14 .335 2 F1 .626 3 F13 .305 3 F3 .637
3 F15 .026 3 F3 .637 3 F4 .440 3 TS3A .683
1 F16 .623 1 F17 .788 3 F10 .562 3 TS10D .713
2 F17 .788 2 F9 .911 3 F3 .637 3 TS6B .856
3 F1B .314 3 F8 .954 3 F8 .954 3 F8 .954
N O T E : "BIk I D " is a block identifier that initially assigns each item to a block. "Item I D " identifies
each item by type and version; " T S I A " is Token Set 1 in version A and " F 7 " is Filler Sentence 7.
See text for further discussion. Note that in the third block constructed in Step 4, there are three
experimental sentences appearing consecutively. Experimenters should control the randomization
process to prevent such occurrences.
effect is sentences randomly assigned to blocks within each subtype

of sentence. Step 3 sorts the entire list by Block ID, which groups
together all the items that will appear in each block. Randomizing
the sequence of items within each block in Step 4 distributes the
experimental and filler items more or less uniformly through the
block. I'll call the lists that emerge from Step 4 just "scripts." Note
that by repeating Steps 2-4 for the same preliminary script, we can
generate several differently ordered versions of the same script. It is
prudent to use at least two different orderings of each script; using
only a single ordering of a script can lead to unwanted ordering
effects that may obscure the main experimental findings. This pro
cess must be applied to each of the preliminary scripts.
It is important to preserve the files generated in the course of
constructing scripts because the Item IDs in these files will later be
used to decode the data files emerging from the scarming or data
entry process.
9.6.4 Assembling Questionnaires
The final step in the process of assembling a quesHonnaire

combines a script with a master questiormaire file. The master ques
tiormaire includes all the instructions for ii\formants and other in
variant material that will appear in every questiormaire. Appendix
Ε includes a complete questiormaire. The master questionnaire can
be constructed so that a script can be copied into a predetermined
location in a copy of the file with sentences being numbered auto
matically. The number of questiormaires generated should be equal
to at least twice the number of types per token set.
10
Coding and
Decoding the Data
The responses informants produce usually require several stages of

processing before the results of an experiment can be summarized or
analyzed statistically. These processes include scanning machine-
readable forms, cleaning up the data that emerge from the scanning
process, or (with nonscannable response materials) keying in data.
As in the previous chapter, I will provide an outline of the process
experimenters need to apply here, but discussion of the specifics
of implementing these processes in software is reserved for the
appendixes.
The processes described at the beginning of this chapter are
specific to experiments that use scannable response forms in survey-
type experiments. However, these processes converge with those
arising in other types of experiments at the point represented by
Figure 18 below. From that point on, the processes described here
will apply to any judgment experiment of broadly similar design.
103
104 E X P E R I M E N T A L SYNTAX
GENERAL PURPOSE - NCS* - ANSWER SHEET

Ο wiTirucTK)
* I c D I r β Ν I J
It®®®®®®®®®® •·C 0(* ύΜ I J
*·cDI IαΜ I J
ΦΦΦΦΦΦΦΦΦΦ t©0®®©©®®®® (ί®©®©®®®Οφ® It ©0®©®®®®®®
» ®®®φ®φ®®®®
-.ΦΦΦΦΦΦΦΦΦΦΦΦ® - • C β I ( β Η t IIJ®®®®®®®φ®®
ΦΘΦφφΦφφφφφφφΟΦ A Β C β I r β Μ I I
©o®®©®®®®® ti ®0®®®®®©®©
II ® 0 © ® © ® ® ® ® ®
φθοθ8οφφφ8§θθ^Ε ®®®®®®®o®® A • c ο I r α Η I i* J ®©®®®®®®®©
-Θφφφφφφφ©®®©©® « · C ο 1 f 0 Η III J®0®®®®®®®® Μ ®®©®®®®0®®
-ΘΘΟΦΦΦΟΦΘΦΘΦΦΘ A · c ft I * α Ν I J
t®O®®®®©®®0 II ®0®®®®®®®® It ®Θ®®®®φ®®®
- ' S i s i s i ^ i i i f r ^
- k C 0 I f ®®®0 Η I t lJ® ® ® ® ® ® ® ® ® ® Μ φ®Φ®®®®®®Φ
ΦΦΦΦΦΦφΦΦΘΦΐ i ® ® ® ® ® ® ®a®Η® ®
A t C O i r O N I J
. . ΦΦΦΦΦΦΦΦΦΦΦΟ -, : D f I
II®©®®®®®®®® Μ ® ® ® ® ® ® ® ® © ®
ΦΦΦφΑφφΦφφφφφ©® : ο I I
A B c a i r a N i i
ΦΦΦΦΦΦΦΦΦΦΦΦΦΦΦ^ Jl®©®®®®®®®©
ΦΟφφφφφφφφφφφφφΙ
A · c Β I r α ·. .
__®® n®©®©®®®©®® II ®Q®®®®®®®®
_ ])φ, Η ®®®®®®®®©® A a C tl I F β >. . .
®Θ®ΘΦ©®®φ®®0©®©(
ΘΘΘΦ© φφφφφφφφφφφφφφφΙ A •CοI f0Η I J ti ®®®®®®®®®®
11®®®®®®®®®® tl ®®®©®®©®®®
ABCDIFQHIJ
A·CοI f βΝI J
Μ ®®©©®®®®®® * « ® © ® ® ® ® ® ® ® ® II ®©Φ®®®®®φ®
A · C 0 ( F 0 Η I JA · C ο I f 0 Μ I t
•I ©G®®®©®®®®
li®®®©®®®©®®
ii®O©®®O®®®0
4;®©®®©©©©®®
A •cο tr0Ν I i
)I®©®©©©®©©® I•cο I
A I e 0 I f β Ν t «®o®©®©®®®®
J Μ ® © © ® © ® ® © ® ®
Μ ®©®®®®®®®® A • C D ( f α Η • I• ©J® ® ® ® ® © ο © ©
11 I I I I I I 11 I III ®0©®®®®©®®
IIIII I iiiiitiiiiiiiri
Figure 14. A staitdard NCS 10-alternative response form.

SOURCE: NCS General Purpose Answer Sheet, © National Computer Systems, Inc., Minneapolis,
M N . Reprinted courtesy of NCS.
NOTE: The left-hand panel provides for various kinds of identifying iiiformation and demographic
data. All of these fields can b e adapted for special purposes. The right-hand panel includes 60
response items.
10.1 Scanning
It is often wise to review response forms before they are sent in

for scanning. Occasionally, informants will make stray marks on
their forms or mark their forms so carelessly that accurate scanning
is unlikely. It can sometimes save later effort to darken light marks
and erase stray marks at this stage.
Machine-readable response forms (see Figures 14 and 15) must be
scarmed by specialized equipment to extract numerical data. This
service is available on most campuses and is also commercially
available in many localities. University testing services that process
machine-scored answer sheets for course tests can usually adapt their
services for use in experimental work using machine-scored forms.
There are standard ten-alternative and five-altemative versions of
Coding and Decoding the Data 105
A B C D E F G H I J
1 ΘΘΘΘΘ®®©®®
A B C D E F G H I J
2 ® © ® ® ® ® ® ® ® ®
A B C D E F G H I J
3 ® ® ® ® ® ® ® ® ® ®
A B C D E F G H I J
4 ® ® ® ® ® ® ® ® ® ®
A B C D E F G H I J
5®®®®®®®®®®
A B C D E F G H I J
6 ® ® ® ® ® ® ® ® ® ®
F i g u r e 1 5 . A n e x p a n d e d v i e w o f six r e s p o n s e i t e m s f r o m t h e f o r m s h o w n in
Figure 14.
S O U R C E : F o m © National Computer Systems, Inc. Reprinted courtesy of NCS.
machine-scorable forms. Most campus testing services will have

available software routines for analyzing test scores. Investigators
will need to instruct the scaruiing service not to apply such analyses
to data generated in experiments but to simply construct a results file
that includes all the informant data. Usually result files can be
returned to the investigator on a floppy disk or via e-mail.
10.2 Patching, Parsing, Decoding, and Sorting Scanned Data
Ideally, all the data from each informant will appear on one line
(within one "record") in the resulting data file, and corresponding
data items (e.g., the informant's response to the 25th sentence) will
Una 1 INSTRUCTOR A
Htm 2 F 055742162 1 7 2 S S 2 3 6 7 4 S 3 S 5 9 8 8 4 9 9 5 6 4 4 9 9 6 6 8 6 64 9 6 8 8 6
Una 3 C 07007M7S 0 2 9 6 1 0 3 9 0 0 0 0 9996 00 98 04 9 0 9 9 4 5 7 8 1 3 S 7 9 9 1
Ul» 4 r 08B4874SS 190620796591929380909891290883287299
Una % Η 0317750957088(2 0792952632472627392830573824676163715
Una ( 12 1 32645411 O 0 0 0 7 7 7 7 7 7 T 7 4 7 4 3 J S 5 5 3 2 2 5 6 22 6 6 0 7 7 7 7 7 7
U » 7 Β RE
0944415800000260600041721113438562488
Un* I ε 103 6 0 5 9 67 051600790651397308472576449979660(007
Um 9 λ 093D7 4 0 7 ] 6 8 9 1 5 7 0 4 0 90 04 9 04 51 S3 9 9 4 0 9 4 4 0 9 0 9 4 4 0 4 4 4 0 4 9 4 4 0
Lin* 10 ο » · RLE Ok NM RIE
1 0 0 2 7 3 0 5 6 5 8 7751 00111100 01011100 OOUIIOIUOO100110101
BO ATO D Y X
Lin* 11 052S64000 216 9 0 3 1 ) 8 9 35 84 5663 93 54 93 9 9 7 8 9 1 9 9 9 7 9 3 4
Lin* 12 ε 061702687 0 9 3 9 4 4 7 9 3 8 9 4 9 9 9 9 4 4 9 9 9 7 9 6 9 9 999967 9 9 9 9 7
Lin* 13 c 0 7 1 4 0 0 8 9 0 1 9 0 2 6 3 0 1 9 8 9 9 9 9 9 8 3 8 0 2 4 4 7 92 0 9 5
1297274 70
D
Figure 16. Sample output from scanning.
fall in the same place (left to right) on each line, but this often is not
the case. Scanning any more than a handful of forms seems almost
always to result in at least a few anomalies (see Figure 16). For a
variety of reasons, scanning systems often report blank data fields
for imused fields and use long series of blanks as spacers in data files.
This can make the process of interpreting a file of scanned data more
challenging than it should be. Had the results represented by Figure
16 been ideal, the first 12 data lines (Lines 2-13) would all have been
patterned exactly like Line 2. Note that over Lines 4 through 11, there
are instances of missing data and spurious or surplus data (where
informants appear to have filled in items they were asked to leave
blank). Often scanned data look even less orderly on first appearance
because they are organized into very long lines that break across two
or three display lines in a word processor or editor.
Urifortimately, the only corrective for these sorts of anomalies is
usually to edit the data file by hand. Many anomalies are obvious (e.g.,
excess data) but some will require reference to the original forms. It is
generally easier to make the corrections with a word processor or text
editor in the data file before it is processed further. It is also helpful in
this process to use a monospace font (e.g.. Courier).
In general, the goal at this stage is to have every line of the file
conform to the same format, to ensure that the nth character on every
line represents the same thing as the nth character on every other line
(e.g., the third digit of the informant's ID number or the informant's
36th judgment response), to fill in any missing data that can be
recovered via inspection of the original forms, and, above all, not to
introduce any new errors via this editing process. The data file that
results from this editing process should look something like the one
shown in Figure 17. Once the file is in this form, it is ready to be
moved into a spreadsheet for further analysis.'
IKSTRVCTOR A
Β 1345411 0000777777774743355532254622660777777
F 0887455 1906207965939293809098912908832872993
Β 3058775 OOlUlOOOlOlUQOOOnilOniQOlOOllOlOl
Ε 9999999 0944415800000260600041721113438562488
6 0796476 0^6103900009996019804909945781357991
& 1005967 0516007906513973084725764499796608007
Ε 1281584 10011001101 oouinioooiooiioioouiioo
Η 0908862 0792952632472627392830573824676163715
Γ 0542162 1725523674535598849956449966866496886
IHSTRTfCTOR Β
Τ 0988587 052304492160327961971171870274347S8S2
Ε 0564000 1903338935845663935493997893999793499
F 0683862 0983108801604097409940946707610690790
& 0776478 0296103900009996009804909945781357991
Ρ 0973127 1011111011111100110110110010101011001
& 0602687 0939447938949999449997969999996799997
Β 1227470 0714008901902630198999998380244792095
Η 0742670 053722775654675848777764747S777577777
F 1107405 0719803934807199709931988908992697591
F i g u r e 1 7 . A c l e a n e d d a t a set.
When a data file of this sort is copied into a spreadsheet, all of

the characters in each line will generally be placed in a single cell (up
to the usual 256-character-per-cell limit). Thus the first spreadsheet
operation in an analysis is usually to parse the lines in the data so
that each "field" from the original appears in a separate cell of the
spreadsheet. That is, the series of digits that represent the informant's
ID (if any) should all be in the same cell of the spreadsheet, and each
single character that represents an informant response must be in a
separate cell. Some notes on the parsing process appear in Appendix
C. Figure 18 shows how the data should look after the parsing
process is complete. Note that each line now begins with a 1-charac
ter field that indicates whether the informant was in the Instructor A
group or the Instructor Β group.
Although the data are now orderly and consistent, they are not
ready for analysis because the informants represented in this table
(most of which is not visible) were using eight different question
naires, as indicated by the second character in each row. The next step
toward getting the data in analyzable form is to sort the informants
I SainulH Uolo I
t / ι ^ Λ ι J ^ ^ ^ / ι ^ / ι t / ι u ι v ι u ι v ι u t ^ A ^ A U i
§1 θ
F IHt
0
I
Μ
Ε (7te 0 I 3
A 09(0 9 9
Η S570 4 i
F S571 t ·
7I)D7 5 7
A G 1170 9 9
A 0 2972
A ε S25(
A Η loei
. 70Si
Β 95S0
C eSM .
A 2210 I
Η 75S4 0
F 1960
C 797<
<S 5112
Μ 2444
A I3S0
ε sisi
D KSI
ε 9770
S S474
D 1772
S 73tl
9 ifSI
f 9072 a t
A SJ74 I 2
C 0((2
F i g u r e 1 8 . P a r s e d d a t a in E x c e l s p r e a d s h e e t .
N O T E : T h e data collected b y Instructor A are n o w distinguished b y a character code in the first
column of the table. T h e T Y P E column iiuiicates which of eight different questioruiaires each
informant used and the I D column assigns a unique I D n u m b e r to each informant. T h e fourth
column, N S , is demograptiic data; 0 indicates that the informant is a ruitive speaker of English.
The next 12 items on each row r q i r e s e n i the first 12 sentences the informant judged. These
sentences were identical in content and order across all versions of this set o f questioiuuiires. The
columns containing data for sentences R2 through R l l have b e e n hidden. T h e remaining 24 items
on each line (S1-S24) represent both experimental and filler sentences in the linear order in which
they appeared for each informant.
so that all those who used the same ordered script appear together
in the table. Once the informants are grouped according to which
ordered script they used, codes can be added in adjacent rows to
show the type and identity of each item the informants in that block
responded to (see Figure 19).
The codes entered above the data for each group of informants
can now be used to control sort operations along the rows. By sorting
according to sentence number (or token set number), all of the rows
are aligned such that all of the responses for any given sentence (or
token set) appear in the same column of the table (see Figure 20). At
this stage, the data are ready for preliminary statistical analysis. At
this point, the data table should be preserved and further analysis
done on a copy.
• • • • • • ^ • • ^ ^ ^ • • • ^ • B Sample Dots
AIBICI Ρ Ι ί Ι F l a i Η I8ITIUIVIWI« I ζ |AA|>^|AqAI)|AE|Af|AapW|AI|AJ|At;|Al|AH*M|Afl|
U) U l <•/> — —
Cfllt9-> ρ ρ c 1 Ρ Ρ t b Ρ t Ρ g b p c Ρ Ρ Ρ Ρ a b Ρ > ρ
—> IS 43 7 8 5S 33 4 12 14 9 23 3 2 44 5 54 I 24 IS 10 11 55 1 34
A A ii3sa 0 5 3 S )
9 9
J t 1 4 ?
9
}
0 59 ?
9
i
9
4 40 9i 1 J9 i e
9 2
e
9 0
t t
9 10
a
7 9
(
Β A !9S0 0 7 0 7 0 9 I 1 i
Β A 0 7 1 ; 3 S β 5 3 1 5 e Ε 4 3 g > 6 7 7 7 4 7 1 3 8
C«ilu-> c b ρ Ρ 4 ρ c ρ Ρ d » Ρ ρ b Ρc b Ρ d t Ρ Ρ
15ρ 6· 2 9 43 23 S 5« u 54 IE 7 10 14 24 4 3912 9 55 1 34 44
A Β 32ί4 0 0 4 7 4 i 3 5 ! 5 5 i } 5 I t 1 ? ? 7 J ? J
A Β mo 0 9 0 0 9 9 9 9 0 9 D 9 9 5 9 Si 29 9 0 t09 5 5 9 0 9 9
Β Β S474 0 5 1 0 9 7 7 2 4 3 2 S 3 I 4 2 9 7 2 9 7 4 1 0 4 9 t
0 0 0 1 1 1 1 0 1 1 1 0 t I 0 1 0
d p d c p p t p t p p p e t p b i l
.. 24 93 I 3 15 7 2 44 II 4 IS 58 5 55 10 14 54 43 9 I 23 t 12 34
A c 70CS 0 9 3—7 2 i } t a 4 7 6 » I 1 5 a I 4 I a 1 2 8 i t a
Β C 7374 0 4 3 S 2 4 1 ί g 5 3 4 4 2 9 S I I 6 I 5 I 0 9 4
||414 I H >l^ D«t4 / CeHel \ D . « . v i l l i C«<»t / \E ΠΐΚΒΠΙΙΙ f"ϋβΙΗίΟϋβΙΗϊΙΙΗ Θ
Figure 19. D a t a w i t h i n f o r m a n t s g r o u p e d s u c h t h a t all t h o s e w h o used a

particular questionnaire a p p e a r together, and with coding information a b o u t
the questionnaire a d d e d .
N O T E : Sentence types α through d are the experimental sentence types, and type ρ sentences are
the fillers. T h e numerical codes (the second row of code data) are ID numbers for individual
sentences. T h u s the first item in the series S1-S24 for informants w h o used Questionnaire A w a s
Filler Sentence 16, the second w a s Filler 43, and the third w a s the sentence representing condition
c from Token Set 7.
'^-'AlBiei 6 lEI f lai Β l8ITIUIVIWI»IViriUlXBIAeiA0IAEIAflA(SIAHIAIIAJ^î;iAl|AUIAHro^
Ul Ut υΐ U 1 1 A U > U ) ( A U 1 U I U ) U > U > < / ) U >

.Ui0 0— —
— —
... W A—i n—o .—. . —. « —* ( —e o — i v w A
A A 8374 a
A A 03(0 0
7 4 • i
5 3 9
7 !
7 8
6 7 e 9 10 1
ρ
9
ρ ρ .
9 9
4.
4
ρ ρ
41 42
18
0 0
ρ ρ
51 52 5.
4 4
4 0
9 9 ! } i t 7 g 9 i 8 0 7 10
8 A 1350 α 7 α 7 3 8 6 7
9 7 9 2 9 Ρ Ρ Ρ Ρ Ρ Ρ Ρ Ρ
5 6 12 13 21 2 J 1 9 41 42 51 52 5
t5l Β β 3284 0 4 : ; 1 •7 J 7 5 J ) 4 5 7 5 7 5
A β 5474 5 1 9 3 7 1 4 6 0 2 7 7 9 3 4 4 10 4
A Β 9(50 9 0 9 9 9 S 9 9 9 9 9 9 9 9 5 Ο S Ο Ο
t§1 0 1 0 10 1 1 1 0 0 0 α Ο Ο 1 1 1 1 1
e θ sssa
H i t Ρ Ρ Ρ Ρ Ρ Ρ Ρ Ρ Ρ Ρ
κ 1 2 3 6 7 8 4 b 12 13 31 41 5
,ΤΓΙ θ C 7574 . 4 -t ) e 4 i
LUJ
i<\i\>\H\/
A C 7066 Ο
4 3
9 3
0»Uw»hCQ4<i \ S « r l * 4 D 4 l « - - f nUncaa
Ί 11Ί 'JV' 1 2 0 9 9 9 4 0
F i g u r e 20. Sorted data.

N O T E : Each block of data has been sorted along the rows so that the result for any given sentence
appears in the s a m e column o f the table for each informant. T h u s the first data item in each row
representing a response to an experimental sentence is that informant's response on s o m e sentence
from Token Set 1. W h i c h sentence from Token Set 1 w a s used differs from block to b l o c k d o w n
the table, as the codes above the token set numbers indicate.
10.3 Keying and Verifying Unscannable Responses
When infoimants respond via hne drawing, magnihide estimation,

or any other procedure that results in unscarmable hand-written re
sponses, some quaUty contiol issues arise as to how the data are translated
into an electronic form. In the case of line drawing, the lines (i.e., the
individual responses) must be individually measured. This process is
suffidentiy error-prone that it is usually wise to have two different
individuals independently measure and record the length of each hne
(without either measiuer seeing the other's results at this stage). It wul
generally be most convenient to record hne lengths in millimeters.
Hand-written line length data and magnitude estimation re
sponses from informants must next be translated to electronic form. The
handiest way to do this with contemporary technology is usually to key
the data into a spreadsheet or a database program. Occasionally, it may
be more convenient to enter the data into a word processing program
(or text editor) and later copy this file into a spreadsheet (although the
formatting of the data in the word processing program should antici
pate how the data will be translated into the spreadsheet).
In any event, it is very difficult to key in a large number of values
without error. Standard procedure in many commercial operations
cor^-onting comparable tasks is to have aU the data entered twice by
(ideally) different coders. Errors are rare enough that it is extremely
imlikely that both passes through the data set will produce the same
errors, whether the keyboarding is done by two individuals or one.
The resultmg data sets can then be compared in a spreadsheet
to identify discrepancies. The discrepant cases can then be examined
individually to determine what the correct data entry is. This process,
although somewhat labor-mtensive, seems to be both more economi
cal and more reliable than keyboarding the data once and then
checking it carefully for entry errors.
Note
1. Where resources and personnel are available, scanning services can con
struct simple software routines that will provide output formatted more like the file
in Figure 17. However, the scaiming process (together with errors informants make)
will almost always introduce errors and anomalies that must be corrected by hand
before the data are processed further.
11
Summarizing
the Data
Two kinds of data summaries are needed for most experiments on

judgments. A summary "by informants" averages each informant's
responses within each of the several sentence type categories and
usually also summarizes in some way each informant's responses to
the filler and/or benchmark materials. The table that results from
this kind of summary will have a unique row for each informant.
Along each row there will be, at minimum, one or more items of
identifying information for that informant and a series of means, each
of which is the average of all of that informant's responses in one of
the experimental conditions. If an experimenter is to do an analysis
"by materials," a further summary is also needed in which each token
set is treated as a unit. Here, the experimenter represents each token
set as though it were a single entity that was tested under several
different conditions (each condition corresponding to one of the
different sentences in the token set). The table that results from this
111
summary has essentially the same structure as that for a by-informants

analysis (identifying information followed by means), except that
each row represents one token set rather than one informant.
The raw data that go into these two different kinds of summaries
are exactly the same. The by-informants summary enters each individ
ual response score into a mean that estimates the acceptabihty of a given
sentence typefor a given informant, but effaces informahon about which
particular sentence the response came from. The by-materials summary
enters each response into a mean that estimates the acceptabihty of a
particular sentence (of a given type), but effaces information about the
individual informants that gave those responses. Each type of summary
is illustrated and discussed in more detail below.'
11.1 By-Informants Summaries
hi the simplest cases, the results of a judgment experiment can

be reduced to a table in which each informant is represented by an
ID number and the averages of the informant's responses in the
experimental conditions. For example, in Figure 21, each value that
appears in the column labeled a within the Summary Table is the
mean of the three responses that appear m the columns AB:AD^ at
the right edge of the figure. Each value in the other columns in the
Summary Table is likewise the mean of a group of individual re
sponses that appears somewhere in the spreadsheet to the right of
the region shown (the columns labeled a.. .d represent four experi
mental conditions and the columns labeled pi ., .p5 represent five
categories of filler sentence). In this example, the minimal results
table would reduce to the ID colunm and the four columns headed a
.. .d. Often, however, this most basic information will be augmented
by demographic information about the informant (the columns la
beled INST, TYPE, and NS) and summaries of the informant's re
sponses on the filler or benchmark sentences (as in pi . .. p5).^
11.1.1 Percentile Summaries
When informants use a category scale, it is sometimes inappropri

ate to represent each informant by the mean of the informant's re
sponses on that scale. As noted earher, different points on the category
Summarizing We Data 113
^ • • • ^ • H H Sample Data • • • ^ • • • • H H 131

C J U I d 6 l E l F i a i B I S I T I U i y j W I X I V I Z jAAjABIACiror
r
ζ 3 ζ 2 u Summtry Table U> U) U)
•D
ω m in — K> w
ei b c d pi P2 P3 p4 p5 Λ t t
1 2 3
θ 0 6374 0 7 4 7.33 5.33 5.33 6.00 6.50 6.00 2.00 2.00 4.33
A 0 0360 0 5 3 8.00 8.67 5.00 7.33 7.67 6.50 3.00 0.00 1.33
i9 7? 7β
Β 0 1350 0 7 0 9.00 9.00 8.00 9.00 9.00 4.50 0.00 3.50 0.33 9 9 9
A 0 2280 1 6 6 4.00 1.67 0.67 4.33 9.00 9.00 2.00 0.00 0.67 9 1 2
a s a
10 11 12
A 1 3264 0 4 4.00 2.33 5.00 4.67 4.67 4.00 6.50 6.00 6.00
i
Q
A 1 5474 0 5 1 7.67 6.33 2.33 2.67 7.33 6.00 3.00 2.50 2.00
} 5?
Β 1 9650 0 9 0 9.00 9.00 4.67 9.00 9.00 4.50 5.00 2.50 0.00 79 7
9 9
9
IP A 1 5656 0 0 1
0.00 0.33 too 0.33 0.33 0.00 1.00 1.00 1.00 0
t
7
0
a
8
0
a
9
A 2 7374 0 4 3 7.00 8.67 3.67 7.00 7.67 6.50 4.00 2.00 2.33 4 i i
Β 2 7066 0 9 3 8.67 8.00 6.33 7.33 9.00 5.50 3.00 0.50 1.33 8 9 9
A 2 6964 I 8 5 6.67 8.67 9.00 9.00 7.33 4.50 2.50 0.50 2.33 7 4 9
A 2 0662 1 4 2 1.67 1.00 3.00 2.33 0.67 1.00 1.50 1.00 0.67 4 1 0
\>\¥i\/ Sorted Date - Infertiwnta \ M B SuminaHCTd Data
Figure 21. Summarized data.

N O T E : T h e new columns that m a k e up the S u m m a r y Table (Colunms S:AA; compare with Figure
20 in Chapter 10) include summaries of informant responses in various sentence type categories
(see text for details).
scale can have different meanings for different informants. A way to

ameUorate these difficulties is to represent the results in terms of the
percentage of each informant's responses that fell at or above a given
scale value for each sentence type. In other words, we might represent
the first informant in Figure 21 as having used a scale value of seven or
higher 100% of the time in Category a, or, alternatively, as having used
a scale value of eight or higher for 33% of the a cases. If this approach is
used, the resulting table takes the same form as in Figure 21 except that
percentage values appear in the Summary Table.
Percentile sununaries are generally appropriate only where a
category scale has been used and may be less than ideal even in these
cases. Experience demonstiates that for any given sentence, the percent-
age of responses that will fall at either extreme of a category scale is
sensitive to the overall quahty of the hst within which the target sentence
is presented. If the Ust contains many sentences of low acceptabihty, the
proportion of responses at the high end of the scale will increase for
individual target sentences. However, the relative acceptability of vari-
ous sentences within a Ust of target sentences seems to remain approxi-
mately constant whether the overall quality of the list of fillers within
which the target sentences are embedded is high or low (see Cowart,
1994). One handicap associated with using percentile summaries is
that many linguists are tempted to ascribe more significance to
percentile results than is warranted. Many will expect sentences they
regard as acceptable to get a strong majority of the highest possible
responses, regardless of the content of the filler list.
Ratio scale procedures are not compatible with either simple
averages of raw scores or the percentile method described above.
11.1.2 Standardizing Informants
Another way to minimize differences in the way informants use

a scale is to convert scores expressed in terms of the original scale to
standard scores, often termed z-scores (see Appendix A), whose
values are readily interpreted relative to the overall distribution of
an informant's results. Positive values lie above the informant's
mean and negative values below. Whether positive or negative,
values near zero lie close to the informant's mean while those greater
than one lie one or more standard deviatior\s away from the infor
mant's mean, that is, they are (assimiing that the informant's scores
are approximately normally distributed) more extreme than most of
the informant's scores. One particular value of z-scores in work on
judgments is that they can help to focus discussion on relative
acceptability rather than "absolute" acceptability defined in terms of
the response scale the informant used.
Standard scores can be used with data obtained via either
category scale or ratio scale procedures.
11.2 By-Materials Summaries
A by-materials summary of the data set is constructed along the

vertical axis of tables like those used above. The ordering of sentences
(or token sets) along each row is the same for all rows (e.g., rows 12:29
in Figure 22). The Summary Table at the top of Figure 22 shows the
percentage of responses for each variant of each sentence that fell at
or above eight on the ten-point scale used here. That is, for sentence
a of Token Set 1,100% of the responses were at eight or above. For
Summarizing the Data 115
• • H M Sample Data • H H H H M H H l ^ H H B B B I i
nei r> leincrnr9 ITIUIVIWIXIVIZIAAIABIAeiAPIAEIAflACI

Summary T a b ) *
5ί „ ζ 1, 5 s s a 2 2
—> > 100» 3«» 75» 63» 50» 50» 1 3 » 2 S » 13» 63» 75» 75» 55» 77» 66»
b 63» 50» 50» 3 6 » 2 5 » 13» 50» 5 0 » 36» 63» 63» 63»
c 2S9< 25» 50» 63» 75» 36» 63» 63» 66» 75» 86» 88»
d t6!« 29)» 7l!li 66» 71*57» 5 7 » 5 7 »6 6 » 29» 43» S7»
• t « b b b c c C d 0 d Ρ |> Ρ
1 2 3 4 5 6 7 8 9 II 12 11 12 13
Β
A
0
0
•374
0360
0
0
7 4
5 3
i
10
«
6
i
9
Τ
6
i
7
4
6
4
4
7
7
ί
7
V
{
6
i
6 9
1 ί
6
t
10
I
10
β 0 I3S0 0 7 Q 10 10 10 10 10 10 ( 10 9 10 10 10 10 10 10
A 0 22«0 t 6 6 10 2 3 2 2 4 1 2 2 2 10 4 10 10 10
A 1 32M 0 0 4
b
1
b
2
b
3
t
c
4
c
5
c
6
\
d
7
)
d
6
d
9
a
10
a
11
a
12
ρ
11
(
ρ 13ρ
12
3 i 43 t i 4 1 4 5
A
Β
1
1
5474
9650
0
0
5 1
9 0
10
10
4
10
0
10 1
2
6
5
10
7
10
t1
10
3
10
6s
10
8
10
10
10
7
10
6
10
10
10
A 1 5656 0 0 1 1 2 1 2 2 2 2 1 1 1 1 1 2 1 1
A 2 7374 0 4 3
c
1
c
2
c
3
d
4
d
5
d
6
?
7
a
8
a
9
b
10
t>
11
b
12
Ρ
11
ρ
12 ρ
13
i 48 59 i 10 10 10 i 10 10
Β
A
2
2
7066
6964
0
\
9
t
3
S
7
10
7
s
10 10 ί i
9
10
10
10
6
10 6
10
5
10
10
9 9
9 10
9
10 e 10
10
10
10 10
5
A 2 0662 1 4 2 2 4 6 3 5 2 5 2 1 2 2 2 I 3
|W|V|>|W|/ Sld Sumiiwd Oala \ K S«iiiii«< H«t«r1«l» D a t a / ~ GaoW Summm Malarlala Dat|
Figure 22. By-materials data summarized in percentile terms.

N O T E : T h e values in the S u m m a r y Table show the percentage o f values that fall at or above a
criterion value (eight in this case) for each version o f each sentence.
sentence b of Token Set 2,50% of the responses were at eight or above,

and so on. A similar summary for some filler sentences is shown in
columns AE:AG. Here, however, the summaries cover the entire
experiment because every informant judged every filler sentence.
A percentile summary of by-materials data will be appropriate
only where a category scale response procedure was used. Ratio scale
data must be simamarized across informants via other methods that
are discussed in the next section.
It is generally inconvenient to do statistical analysis on a table that
takes the form of the Smnmary Table in Figure 22. Once this table is
constructed, it will usually be helpful to convert it to the form shown in
Figure 23, where all the data for each token set appear on one row.
11.2.1 Geometric Means
Because ratio scale procedures allow individual informants to

vary widely in the range of values they choose, summarizing across
Sample Data
ΑΙ Β ICTP"
2 Summary Table
Τ
Vera! on
Ύ Token Set b c d
Τ Si 63% 25% 86%
Τ 82 m%
38% 50% 25% 29%
ττ S3
S4
75%
63%
50% 50%
38% 63%
71%
86%
10 35 50% 25% 75% 71%
S6 50% 13% 38% 57%
ΤΓ
12 S7 13% 50% 63% 57%
Τ3 38 25% 50% 63% 57%
Τ* 39 13% 38% 88% 86%
•Ϊ5 310 63% 63% 75% 29%
16 311 75% 63% 88% 43%
ΎΤ 312 75% 63% 88% 57%
78
Τ9
N M I ^ | W | \ X S u m T b l Data/ GeoM||
Figure 23. By-materials d a t a s u m m a r i z e d in percentile t e r m s a n d r e f o r m a t t e d
for statistical a n a l y s i s .
NOTE: The values in the table are the same as those shown in Figure 22.
informants for a given sentence or token set cannot be done via the
methods described above. The standard method in this case is a
geometric mean.*
11.3 Summaries for Filler and Benchmark Data
Summaries of filler and benchmark data are generally straight-

forward. If the list of fillers used is structured in some way (e.g., the
fillers represent five different degrees of acceptabihty), the filler
sentences need to be segregated and summarized m ways similar to
those used for experimental sentences. The Summary Table in Figure
2 2 includes items sunmiarizmg results on different fiUers (piI, p l 2
Summarizing tfie Data 117
and so on). Benchmark materials generally do not need to be sum

marized, although they will play a role in statistical analyses. Their
uses mainly have to do with estimating the quality of performance
from individual informants.
Notes
1. The terminology for the two kinds of analyses mentioned here varies some
what in the literature. Analyses by informants are often called "by-subjects" analyses.
By-materials analyses are often called "by-sentences" analyses or "by-items" analyses.
2. Spreadsheet columns, rows, cells, and ranges of cells are referred to by way
of the marginal lat)els seen in Figure 21 and elsewhere. Thus "AB:AD" is a reference
to the range that includes all of colimuis AB, AC, and AD. The uppermost left cell in
every sheet, for example, is referred to as "cell Al." A range of cells is referenced by
w a y of the upper left and lower right comers of the range. Thus "S6:V9" is a reference
to the summaries of experimental data for the first four informants in Figure 21.
3. Each of the iUustrations in this chapter will show only that portion of the
relevant spreadsheet necessary to reveal the structure of that particular summary
under discussion at that point. The complete spreadsheets are collected in an Excel
workbook that is available at the Web site for this book, which can be foimd at
HTTP:/ / W W W . U S M . M A I N E . E D U / - L I N
4. See Section C.4 of Appendix C for instructions on calculating geometric
means.
12
Statistical
Issues
If an experiment tests four different sentence types, it is not uncom

mon for there to be no two types with exactly the same average
acceptability value. Some of these differences, however, may be
small ones that do not reflect consistent differences between the two
sentence types. Upon repeated testing, we might find no difference
at all, or a small difference in the opposite direction. To interpret a
numerical difference between the means for two sentence types, an
investigator must decide whether or not to draw the inference that
there is a consistent difference in acceptability between the two types
for the tested individual or group. Although no procedure can guar
antee the truth of the conclusions the investigator settles on, statisti
cal tests do allow the investigator to impose controls on the likeli
hood of error. Statistical procedures cannot prevent investigators
from drawing inferences that lead to errors, nor can they determine
whether or not any particular inference leads to error, but, when
119
carefully applied, they can give the investigator reasonable assur

ance that no more than a certain predetermined percentage of a set
of statistically tested inferences do in fact lead to error. Statistics, as
it is used here, is mostly about the rational control of error.'
This chapter has two purposes relative to the statistical issues
that arise in studies of acceptability. For those investigators who are
prepared to do appropriate statistical tests themselves, it will pro
vide some guidance on the handling of certain statistical issues that
arise in the context of research on judgments. For those who lack
statistical training or experience, the chapter wiU offer some general
guidance on where to fmd appropriate assistance or training.
12.1 Alternative Tests for Category Scale Data
In Chapter 6, we noted that category scale procedures do not

always produce interval level data. In particular, there may be ex
perimental chcumstances where, for example, the difference in ac
ceptability marked by scale scores of one and two on a five-point
scale will be larger {or smaller) than the acceptabihty difference
indicated by scores of four and five. Where this condition arises, the
experiment has produced, strictly speaking, ordinal level data, not
the interval level data that are assumed by most varieties of analysis
of variance and related statistical tests. If analysis of variance and
similar procedures are used with such data, the reliability of the tests
themselves is uncertain.
Where a particular contrast is numerically large compared with
variability around the relevant means, any statistical problems de
riving from a failure to achieve interval level measurement are not
likely to be consequential. There also may be circumstances where
for practical reasons an investigator will choose to use a category
scale procedure even when some small differences are expected.
In circumstances such as these, it is still reasonable and appro
priate to apply standard ANOVA tests to the resulting data. How
ever, a second kind of analysis ought to be used as well to ensure that
any possible violation of ANOVA assumptions does not affect the
outcome of the test.
This further test is based on a different stunmary of the data.
Instead of constructing means, the initial summary calculates the
percentage of responses that fell at or above some criterion in each
statistical Issues 121
informant's responses for each sentence type category. Thus, if the

criterion were set at four and an informant judged six No "that" / Sub
ject Extraction sentences on a scale of one to five, and rated two of
those sentences at four or higher, then that informant's responses in
that category would be summarized as 40%. This same kind of
summary is then constructed for all other by-informants and by-ma
terials analyses, and the same statistical tests as would be used
otherwise are applied from this point on. This sort of data is guaran
teed to meet the criteria for the interval level of measurement and
thus does not threaten the integrity of ANOVA results. Any signifi
cant main effects or differences that emerge from summaries both by
means and by percentages (of the sort suggested here) can be relied
upon (with the usual caveats about meeting the assumptions of the
tests). Any results that are significant in the means-based analysis,
but not in the percentages-based analysis, are suspect. In such a case,
it would be appropriate to do a further experiment, perhaps on a
smaller scale, using a ratio scale technique.
The main disadvantage of percentage summaries is that they
waste information. These summaries have no way to capture infor
mation about how an informant's responses are distributed above
and below the chosen criterion. Thus these summaries will generally
be less sensitive. For any given sample size, there will be small
differences that can be detected with tests based on a means sum
mary, but not with a percentages summary. Nevertheless, this ap
proach is useful as a backup to standard statistical procedures where
an investigator chooses to use a category scale technique.
12.2 Testing Generaiizations Across Informants and Token Sets
As illustrated in Chapter 11, usually the results of any experi

ment designed along the lines described in this book can be summa
rized in two ways. One way summarizes the results by informants,
providing a mean judged acceptability value for each token type for
each informant. Another way summarizes the results by token set,
providing mean judged acceptability values for each sentence in each
token set (Clark, 1973).^ For the most part, the same statistical tests are
then applied to the token set data as are applied to the informant data.
Clark's paper provides relatively simple methods for integrating
tests on informant data and token set data (in his terms, by-subjects
data and by-items data) such that an investigator can report a single
Mest or ANOVA result covering both analyses. This practice, how
ever, has met with far less than universal acceptance. When investi
gators do both kinds of tests, they quite commonly report the infor
mants and sentences tests separately. Tests done on summaries by
informant are usually presented with a subscript "1" appended (e.g.,
Fi) and summaries by token set have a subscript "2."
The rationale for doing both kinds of test is that, just as the
iiiformants actually tested in an experiment are (usually) seen as
representatives of the enthre population from which they are selected,
the token sets are likewise seen as representatives of all the relevantly
similar token sets one might construct in the same language (or
perhaps any language). Just as statistical tests on data for informants
test the reliabihty of patterns seen in those results, so tests on data
for token sets test the reUabihty of patterns seen in the summaries of
the token set data. Most often the two kinds of tests produce virtually
identical results, but especially where weak or marginal effects are
involved, it is worthwhile to ask whether an effect is reliable across
both informants and token sets. Many experiments mcorporate one
or more factors that cannot be tested in both analyses.
The analogy between informants and token sets is hardly per
fect, and the analogy between the notion of a population of people
and a population of token sets is shakier still. Thus it is not obvious
that statistical tests that are meant to test generalizations across
populations are being used equally appropriately in these two cases.
Chomsky's (1986) suggestion that languages exist only as patterns of
behavior projected from the internal states of speakers can or\ly raise
further doubts; in Chomsky's view, it appears that there is no popu
lation from which to sample. Nevertheless, Clark's recommendation
that both kinds of test be done is often soimd. There is no doubt that
experimental results on language are sometimes misinterpreted as
applying generally to all similar linguistic materials when all that has
been shown is that the result is reliable only when exactly the same
materials are used.
12.3 Variation Accounted For
Statistical tests of significance can be given greater weight than

they deserve. There are limitations associated with these measures
that need to be kept in view. Fortunately, there are also some addi
tional measures that can be derived from i-tests and analysis of
variance that help to ameliorate some of these limitations.
There is a sense in which significant results from f-tests or
ANOVAs can be highly reliable, but uninformative. It is entirely
possible, and often happens, that an ANOVA will detect a reliable
effect due to some manipulation in an experiment where that ma
nipulation has oruy a very tiny impact on the performance of sub
jects. Manipulations that control only a very small share of the
variance in an experiment, even where they are reliable, are relatively
urm\formative. Most of the variance is being affected by something,
but it isn't the factor we are manipulating.
It is also worth bearing in mind that the standard statistical
sense of "significant" has no necessary coimection to questions about
theoretical importance. Differences may be simultaneously signifi
cant (from a statistical point of view) and boring (from a theoretical
point of view). Whether a statistically significant difference matters
in some larger sense can only be determined by examiiung its rele
vance to alternative theories that bear on the situation in which the
difference arises.
Especially with a particularly stable phenomenon such as sen
tence judgments seem to be, there is a somewhat arbitrary lower limit
on the size of the effects that can be found significant.^ By increasing
sample size, progressively more and more delicate reliable differ
ences can be detected. For example, in our work on "that"-trace, it
is possible to detect a tiny but reliable difference in acceptability
between subject extraction and object extraction when no "that" is
present, provided that (by combining results from two large experi
ments) we attain a sample size of more than 1,100. Although it seems
obvious that this finding is of no theoretical interest, it raises the
question of how experimenters are to estimate which statistically
significant effects are also significant in a more general theoretical
sense.
A statistic that responds to this need is (see Appendix A for
information on calculating η^). This value indicates the share of all
the variation in a data set that is accounted for by a particular factor
in an experimental design. Thus, in an analysis of a "that"-trace
experiment, we can calculate the percentage of all the variation in the
result that is due to the "that" effect, to the interaction, and so on.
Data from analyses of two of the "that"-trace experiments mentioned
Table 14 F and statistics for two "that"-trace experiments.

Expentnent 1 Experiment 2
(N = 32) (N = 332)
F F
"That" Effect 63.5 43% 690 34%

Extractioii Site 29.6 8% 282 11%
Interaction 27.1 9% 361 13%
earlier appear in Table 14. Here the second experiment, with roughly
ten times more informants, is obviously much more sensitive. The F
values are about ten times larger than those of the first experiment.
This suggests that if there were any smaller, weaker effects in the
same design, the second experiment would be more likely to detect
significant differences for those effects. Note, however, that the val
ues of are roughly similar for the three factors despite the 1:10
difference in F values.
In fact, there were some weak effects associated with a "groups"
factor in this experiment. This was a between-informants factor in
which individuals were categorized according to which of four ver
sions of the questiormaire they responded to. Although there were
no significant effects associated with this factor in the first experi
ment, there were two reliable interactions with this factor in the
second experiment. However, showed that each of these interac
tions accounted for only about one half of 1% of the variation in the
experiment. Thus, although both were quite reliable, neither is of any
theoretical interest because they accoimt for such a very small share
of the overall variation, compared with the effects listed in Table 14.
Another application of to the same two experiments appears
in Figure 24. Here has been apphed to comparisons between four
particular pairs of means within the larger design. Each of these
individual tests is logically equivalent to a f-test covering only two
means. In these cases, the values indicate how much of the variation
in the informant means underlying the overall means is accounted
for by the contrast between the conditions, as opposed to error
variance within the conditions. Note, as before, that despite the fact
that Experiment 2 employed far more informants, the values are
generally similar, although the effect of having or not having "that"
in the object extraction cases was larger in the first experiment.
<1%
Subject Extraction / Object Extraction /
"That" Absent "That" Absent
<1%
72% 68% 38% 17%
58%
Subject Extraction / Object Extraction /
"That" Present "That" Present
62%
F i g u r e 2 4 . C o m p a r i s o n o f v a r i a t i o n a c c o u n t e d for in four p a i r w i s e c o m p a r i s o n s
in t w o " t h a f ' - t r a c e e x p e r i m e n t s .
N O T E : Ttie four innermost percentage figures show values for four pairwise compariscms from
Experiment 1 in Table 14. T h e italicized outermost values are the corresponding values for
Experiment 2.
Perhaps the best description of the role of is that it provides

an index of the informativeness of sigruficant effects in ANOVAs,
t-tests, and related procedures. If the manipulations applied in an
experiment produce reliable effects, but those reliable effects don't
control much of the variance in an experiment, we do not learn much
from knowing that the effect is reliable. If, on the other hand, an effect
is reliable and accounts for a large share of the variance in the
experiment (even if the difference in means associated with it is
small), then we know a good deal more. We know that there is a
manipulation that has an impact; we have found a factor that changes
something and we have gained greater insight into the phenomenon
at hand.
These suggestions also show how can resolve the problem
mentioned earlier where, by combining two large "that"-trace ex
periments, we were able to detect a reliable difference between
subject extraction and object extraction with "that" absent. Although
reliable when tested by way of a huge experiment, the pairwise
comparison between these means accoimts for only about one third
of 1% of the variation, even within just the pairwise test. Thus
provides clear grounds for dismissing this "significant" effect; it is of
no theoretical interest, however reliable it may be, because we learn
nothing from it.
12.4 Getting Training or Assistance in Statistics
One way investigators who lack training or experience in sta

tistical methods may apply the methods described in this book is to
seek assistance from a statistician or statistical consulting service.
Such services are available on many college and university cam
puses, often without charge. There is an enormous variety of differ
ent statistical techniques available and not all statisticians are equaUy
famihar with all of these methods. Someone famihar with factorial
repeated measures analysis of variance designs is most hkely to be
helpful. Statisticians who work frequently with experimental psy
chologists are likely to be famihar with all or most of the tests and
procedures mentioned in this book.
Investigators who want to undertake a study of statistics for
themselves should plan on working through a body of material
roughly equivalent to two courses. Analysis of variance is considered
a somewhat advanced technique, and introductory courses (and
texts) rarely give it more than a relatively brief introductory treat
ment. There are coimtless introductory statistics texts available, some
of which have been designed for self-study. To lay a foundation for
the methods used here, a statistical novice should look for a text that
covers basic descriptive statistics and graphing, basic inferential
statistics (z-tests, f-tests), elementary analysis of variance, and corre
lation. A variety of advanced texts in analysis of variance are also
available.
A nim\ber of statistical software packages are available for per
sonal computers. Access to one of these is essential. The volvune of data
in even modest-sized acceptabihty experiments quickly overcomes the
practical linuts of calculators. Not all statistical packages that incorpo
rate some analysis of variance routines support the factorial within-sub
jects (or repeated measures) designs apphed here.
Notes
1. This sentence includes a near quote of a scholar to whom I owe a consider

able debt but whose identity I have lost track o f
2. Clark (1973) pointed out that many prior experimental studies using linguis
dc materials had failed to establish the generality of their linguistic claims. It was at
that time common for investigators to (implicitly) assume that all relevantly similar
linguistic materials would produce the same results as those actually used in a given
study. Yet an experiment that compared some aspect of performance with nouns and
verbs necessarily would employ only some rather small sampling of actual nouns and
verbs in the relevant language. Clark pointed out that results obtained in this way
could not support generalizations stated in terms of linguistic categories unless
appropriate statistical tests were run on the linguisHc materials themselves, as well as
on the human subjects of the experiment. In the absence of such tests, scrupulous
investigators would confine their claims of generality to people (e.g., "Subjects iden
tify < list of specific words > more rapidly than < list of specific words >"} and avoid
statistically untested linguistic generalizations (e.g., "Subjects identify high frequency
nouns more rapidly than low frequency nouns"). This of course would be unsatisfactory
for most studies bearing on linguisHc materials, so Clark went on to suggest specific
statistical methods that could be used in cases where an investigator wishes to
generalize simultaneously to a human population and to a "population" of linguistic
materials. Clark's point has been widely accepted, although the particular statistical
solutions he advocated have enjoyed less success.
3. In statistics, effect size is not to be confused with the size of the numerical
difference between two means. For a f-test, for example, effect size is a ratio between
the difference between the means and the variability around those means. Large
numerical differences can be associated with high variability and thus small effect size
and numerically small differences between two means can be associated with little
variation and correspondingly large effect size.
Appendix A:
A Reader's Guide
to Statistics
The purpose of this appendix is to provide some brief backgroimd

material on some statistical concepts and tools that are used in this book.
This will not replace the full treatment of these notions in an appropriate
text, but it should give readers without prior backgroimd a better basis
on which to interpret the statistical findings reported here.
It is useful to distinguish between descriptive and inferential
statistical procedures. Where descriptive statistics are useful for stating
or describing some property of a sample (e.g., the location or dispersion
of a set of measurements on a group of individuals), inferential statistics
aid investigators in determining which inferences about the larger
population from which the sample was drawn are and are not war
ranted by a given pattern of results. This appendix will first review some
basic descriptive tools that are used in the book and then turn to brief
discussions of some standard inferential statistics.
A.1 Descriptive Toois
A. 1.1 Measures of Central Tendency
The most common measure of central tendency for a set of judg

ments is the mean or arithmetic average. The experiments discussed in
129
this book invariably involve collecting responses in two or more

different conditions and comparing the means of the responses in
those conditions. One generally desirable property of the mean is that
it is sensitive to all the measurements m the sample. This property is
sometimes undesirable if there are one or more outhers (extreme
measurements that lie far from most other measurements), or if the
set of measurements is skewed (scores are heavily clustered toward
the high or low end of the used range). In these cases, a useful
alternative index of central tendency is the median, which is the score
value that cuts the distribution of scores m half. If a distribution
consisted of values 3,4,6,8,9, and 26, the median would be 7 because
half the scores fall on either side of this value. The geometric mean is
useful in circumstances where a set of measurements comes from
individuals who vary considerably m average response. Thus, in
averaging responses from three individuals who have mean re
sponse levels of 3,4.5, and 34, the geometric mean, hke the median,
will reduce sensitivity to the distorting effects of extreme values.
A.1.2 Standard Deviation
One of the principal concerns in any statistical characterization

of a set of scores is the degree of dispersion or variation that occurs
within those scores. The most common statistic for describing this
property is the standard deviation. In a normally distributed set of
scores (i.e., one that has the shape of the familiar "bell curve"), a httle
more than two thirds of all the scores will he withki one standard
deviation unit above or below the mean. Thus a sample for which the
standard deviation value is large is one where the scores are more
widely distributed than one where the standard deviation value is
small.
A.1.3 Standard Scores/z-Scores
Individual informants sometimes use a single response mea

sure in different ways. The average of cdl of an informant's values will
differ from informant to ii\formant and the responses of some in
formants will be more widely distributed across the available range
than will those of another. Standard scores (also called "z-scores")
Appendix A 131
are a common tool for correcting for these sorts of differences. To

convert an original raw score value to a standard score, the score is
expressed as a distance above or below the mean, measured in
standard deviation imits. Thus a standard score value of 1.0 repre
sents a score that is one standard deviation unit above the relevant
mean. A standard score of -2.5 represents a score that is two and one
half standard deviation imits below the mean. When a set of scores
is standardized (i.e., converted to z-scores), the mean of the set of
standardized scores will always be 0 and the standard deviation of
the set of standardized scores will be 1. Thus, when the data from
each individual informant in a sample are standardized, all infor
mants now have the same mean and standard deviation. Note,
however, that this procedure loses no information about the distri
bution of the informant's scores; the pattern of the informant's scores
is completely preserved by this transformation.
A. 1.4 Standard Error of ttie Mean
One of the essential insights exploited in inferential statistics is the

observation that the variability of a set of scores in a single sample can
be used to estimate the scattering of sample means that would occur if
many similar samples were drawn from the same population. This
insight imderlies a measure called the standard error of the mean. This
value plays a somewhat similar role for a (usually hypothetical) collec
tion of sample means that the standard deviation plays for a set of scores.
If repeated samples of the same size were drawn from a population
having the same mean and standard deviation as does the actual
sample, thai a littie more than two thirds of all the new samples would
have means within a range spanning one standard error unit above and
below the given mean. Thus a mean associated with a small standard
error value is one that lies close to the true mean of the measured
population while one associated with a large standard error veilue is one
that may lie further from the population's true mean.
A.2 Inferential Statistics
In the experiments described in this book, the goal is to estimate

some property or parameter of a population on the strength of the
sample from that population that was actually tested in each experi
ment. Thus our goal is typically to detennine whether responses to
Sentence Type A indicate greater or lesser acceptabihty in the relevant
population than do responses to Sentence Type B. Given that experiments
will almost always jaeld at least some small nvunerical difference be
tween any two conditions, the basic problem of inferential statistics is
to provide investigators with some systematic guidance as to when it is
or is not appropriate to draw the inference that there is also a consistent
difference between the measured conditions in the target population.
There are two kinds of error an investigator may make in
decisions of this kind. Type I errors are those where the investigator
credits too small a difference. That is, the investigator detects a
difference between two conditions in an experiment and draws the
inference that there is a rehable difference between the two condi
tions in the target population when in fact there is no such difference.
Type Π errors are those where the investigator fails to credit a
difference detected in an experiment and falsely concludes that there
is no difference in the target population. Unfortimately, the likeli
hood of each of these two kinds of error is often linked. Insurance
against making one kind can sometimes only be purchased by in
creasing the chance of making the other. Increasing sample size can
reduce the probabihty of both kinds of error.
The test statistics used in this book address this dilemma by
trying to estimate the likelihood of the actually obtained data, given
some baseline hypothesis about conditions in the tested population.
Thus, if an investigator is mterested in the relative acceptabihty of
Sentence Type A and Sentence Type B, and if there is actually no
difference between these two in the target population, how likely is it that
the experiment could have produced the differences in average
measvu«d acceptability that actually appeared in the experiment?
This estimate is derived from the variabihty of the obtained sample.
That is, the variability within and between individuals within the
actual sample provides an estimate of the variabihty within the
population as a whole. From this it is possible to calculate how hkely
it is that differences in means of the obtained size could have arisen
if in fact there were no differences in these means in the population.
When, by procedures of this kind, we are able to show that the
obtained differences are very unhkely were there no real differences
in the population, we generally say that the obtained difference is
(statistically) sigruficant, although perhaps it is more helpful to say
Appendix A 133
that we have evidence that the difference between the tested means
is reliable.
A.2.1 t-Tests
One of the more elementary applications of this logic arises in a

family of ii\ferential statistics caU t-tests, which compare only two means
and are generally reported in a form like the following: "Sentence Type
A is more acceptable than Sentence Type B, t{26) = 2.82, ρ < .01." Here,
"26" indicates the degrees of freedom of the test (related to sample size).
The notation "p < .01" provides the most critical piece of information.
This value indicates how likely it is that the actual obtained result (the
observed difference between the two means) would have arisen if in fact
there were no difference between these conditions in the target popula
tion. Obviously, small ρ values underwrite confidence that the obtained
difference is reliable and large values (commonly considered to be those
above .05, although important complexities are being set aside here)
suggest that the obtained differences between the means could have
relatively easily arisen by chance.
A.2.2 Analysis of Variance
Analysis of variance is a more varied and versatile family of

procedures, although the underlying logic is the same as that of the
f-test. The difference is that where t-tests apply only to pairs of
means, analysis of variance (or, commorUy, ANOVA) procedures
allow an investigator to make many comparisons within a single data
set simultaneously. Thus all of the "that"-trace experiments dis
cussed in this book can be seen as two simultaneous and overlapped
pairwise experiments. One experiment compares sentences with and
without "that" and one compares those exhibiting subject extraction
with those exhibiting object extraction. By integrating these two into
a single experiment, it becomes possible to overcome a difficulty that
would arise if two or more different t-tests were done on the same
data set. If for any one t-test we set a threshold value (a in the statistics
literature) of .05, allowing ourselves to credit an obtained difference
that falls at or below this threshold and rejecting any result that falls
above it, we have ensured that the likelihood of error in drawing an
inference from our result is .05 or less. However, if we conduct 20

experiments, each comparing some pair of means, and in each case
we find that the obtamed ρ value is at or just below .05, then the odds
are that somewhere m the set of 20 comparisons there is an error.
Thus the overall chance of making an erroneous inference is not 1:20,
as it is for each individual experiment, but approximately 1:1. That
is, we are highly likely to have drawn at least one erroneous inference
in the set of 20. ANOVA techniques make it possible for an investi
gator to apply overall control to a set of comparisons so that the
threshold value the investigator selects at the outset is the one that
actually obtains in the analysis of results.
Another very valuable feature of ANOVA techniques is that they
can compare not just two but many related means at once. Thus one
experiment described in Qiapter 1 includes a definiteness factor that
has four different "levels" (roughly equivalent to a f-test that simulta
neously compares four means, asking whether any one or more of them
is reliably different than any other). A further very valuable class of
ANOVA designs allow investigators to examine interactior« among
two or more effects. That is, as happens in the "that"-trace experiments
discussed in this book, one effect (subject extraction versus object ex
traction) has a very different impact on acceptability depending upon
which "level" of ttie other factor ("that" present versus absent) is in
volved. When no "that" is present, the subject versus object extraction
effect is neghgible, and when "that" is present, it is large. ANOVA
procedures make it possible to assess the reliabihty of this sort of
interaction effect alongside simpler comparisons.
ANOVA results are reported in a format similar to that for f-tests.
In general, an identification of the effect or interaction is followed by a
summary statement of the statistical result, as in the following: "The
interaction between the Extraction Site and 'That' factors was rehable,
F(l, 22) = 8.03, ρ < .01." The notation "(1,22)" specifies the degrees of
freedom for the test, there being two values because the F statistic is a
ratio. The degrees of freedom of the test is again related to sample size,
but that relation is much less straightforward in analyses of variance.
As witti the f-test, the critical piece of information is the ρ value, which
here indicates that the observed pattern of data is very imlikely on the
assumption that there is no difference in the seunpled population. F-tests
that fail to indicate rehable results are generaUy reported with the
notation "NS" in place of the ρ value, or sometimes with no prob
abihty statement at all if the F-value is small (e.g., F < 1).
Appendix A 135
A.2.3 Assumptions Underlying Inferential Statistics
Inferential statistics are invariably defined in terms of a mathe

matical model of the situation in which the obtained data arise. These
models make assvmiptions about conditions in the experiment that
do not necessarily accurately reflect the true state of affairs. The most
sophisticated consiuners of statistical results often will want to ask
whether the data set under consideration in a particular case in fact
met all of the assiunptiorw of the tests that were applied to it. In the
context of this book, perhaps the most important assiimption of this
kind is the assumption, common to most ANOVA techniques, that
the data achieve what is called the interval level of measurement.
This notion is examined in more detail in Chapter 6. The most
common varieties of ANOVA test easo assiune (among other things)
that the data are more or less normally distributed and that variances
are approximately equal across various conditions.
A.2.4 Correlation Coefficients
The most common correlation coefficient, Pearson's r, provides

an index of how closely two continuous variables track each other.
Pearson's r ranges from -1 to 1, with larger positive values indicating
a strong direct relation between the two variables and negative
values indicating that the values are inversely related (an increase in
one predicts a decrease in the other). If, for example, we collected
many samples of vowel sounds, then directly measured fundamental
frequency for each of those soimds and by other means measured the
average spacing between harmorucs in those soimds, we expect these
two values to be very closely related; in general, we should be able
to predict one measurement from the other. If our expectation is
upheld, Pearson's r will show a large positive value. Large nvmierical
values (regardless of sign) indicate a strong relation between the
tested variables.
The reliability of the relation between two variables can be
tested. In effect, this test asks how likely it is that the obtained data
would have arisen if in fact there were no relation between the two
variables in the sampled population. Where this test is applied to an
r value, usually degrees of freedom, "df" will be reported (reflecting
the nimiber of pairs of values tested) along with a ρ value, as in the
Table 15 Calculation of η^.
Source SS DF MS F
"That" 2.0187 1 2.0187 48.7 .26

Error 1.0365 25 0.0415
Extraction Site 1.5899 1 1.5899 37.7 .21

Error 1.0548 25 0.0422
Interaction 1.1691 1 1.1691 35.5 .15

Error 0.8235 25 0.0329
Total sum of squares 7.6925
following: "The fundamental frequency and harmonic spacing mea

surements were closely related, r = .98, df = 42, ρ < .001."
A.2.S Percentage of Variability Accounted For
Very often, correlation statistics are reported in terms of

rather than r. The advantage of this statistic is that it has a much more
straightforward and intuitive inteφretation than does r itself. The
statistic can be interpreted as a direct measure of the proportion of
the variability in one measure that is accounted for by variability on
the other. In other words, once the variability of one of two correlated
measures is known, a certain proportion of the variability on the
other becomes predictable and gives a measure of that proportion.
A related statistic can also be computed for factors and interac
tions in ANOVAs. Termed η ^ this value measures the proportion of
the total variability in an experiment that is attributable to the indi
cated factor or interaction. Thus, if an experiment included a factor
of "Complexity" (having, say, two levels), and that factor is associ
ated with an value of .42, we conclude that 42% of the total
variability in the experiment is due to the manipulation of degrees
of complexity.
The calculation of is quite straightforward. As illustrated in
Table 15, the sum of squares (SS) for each factor or interaction in the
ANOVA result table is divided by the total sum of squares for the
data set as a whole.
Appendix Β:
Statistical
Supplement to
Chapter 1
This appendix reports on the statistical tests that were apphed to the
various experiments discussed in Chapter 1. The appendix is organ
ized vmder subheads relating to the sections of Chapter 1. Readers
unfamihar with the statistical terms used in this appendix will find
some notes on the puφose and interpretation of these measures in
Appendix A.
Clark (1973) argues that many experiments on linguistic materials
should be tested for their generahty across both subjects or informants
and hnguistic materials (see Chapter 12 for further discussion). Clark's
recommendation is generally followed here. Statistical tests done on
summaries by informant have a subscript " 1 " appended (e.g., Fj) and
summaries by token set have a subscript "2" appended. Significance
levels are indicated by asterisks:"»" for ρ < .05, "*»" for ρ < .01, and "*»»"
for ρ < .001. Where appropriate, the results reported below wiU be
expressed in terms of standard score (z-score) units.
B.1 Subjacency
The overall main effect in ANOVA testing differences among the

four conditions was significant, Fi(3, 261) = 194*»*, F2(3, 69) = 90.9*»*.
137
The main effect of Definiteness accoimted for 69% of overaU variance

in the experiment.
Each of the pairwise differences in Figure 2 from left to right is
significant; f i(l, 87) = 132**», Fjil, 23) = 50.1*** for the control/mdefi
nite contrast, f i ( l , 87) = 62.7*»·, fjil, 23) = 31.7*** for the mdefi
nite/defiiute contrast, and f i(l, 87) = 11.5**, Fjil, 23) = 8.15** for the
defirute/specified subject contrast.
In the second subjacency experraient (described in Figure 3),
the contrast between the Indefbiite/"of" cases and the Control was
significant, Fi(l, 40) = 46.6»**, but the contrast with the Indefinite/
"to" cases was also robust, Fi(l, 40) = 16.9***. There was no significant
difference between the two cases ending in "to." The overall pattern
of results was also significant, Fi(3,120) = 17.1»»».'
There are some reasons to doubt that category scale procedures
such as those used for the experiments reported in this section
reliably produce interval level data (see Chapter 6). There is at least
one way, however, to use category scale results that ensures that the
interval level of measurement is achieved. Instead of calculating the
mean within each category of an informant's responses, the investi
gator summarizes the data by calculating the percentage of responses
in each category that are at or above some criterial level. Thus we can
determine what percentage of trials within a given category pro
duced the highest possible response and use this value in further
analysis.^ This is, however, a very wasteful procedure. It loses all the
information encoded in the different specific score values that in
formants used.
Applying this approach to the experimental results described
above yields an outcome similar to the earher analysis. The main
effect is significant, Fi(3, 261) = 128»»», F2(3, 69) = 56.0»»». Ukewise,
each of the pairwise differences from left to right is significant; Fi(l,
87) = 135»·», F2(l, 23) = 34.7»»» for the control/indefmite conbrast,
Fi(l, 87) = 5.97», F2(l, 23) = 9.31»» for the mdefmite/defhute conbrast,
and Fi(l, 87) = 16.2***, F2(l, 23) = 4.67» for the defmite/specified
subject contrast.
B.2 "That"-Trace
Eight different pseudorandom orderings of the materials were

used in this experiment.
Appendix Β 139
For the results summarized in Figure 4, there is no difference in

acceptability between Subject and Object Extraction in the absence of
"that," but there is such a difference (favoring Object Extraction) in
the presence of "that." This difference was reflected in the interaction
between Extraction Site and the Presence/Absence of "that," Fi(l, 31)
= 30.3***, F2(l, 23) = 22.4··*, and in overall main effects of Extraction
Site, Fi(l, 31) = 31.7***, F2(l, 23) = 10.8**, and Presence/Absence of
"that," Fi(l, 31) = 96.0***, F2(l, 23) = 72.6***. Although the pairwise
comparison of the two cases with "that" is significant, Fi(l, 31) =
53.4***, F2(l, 23) = 26.1***, there is obviously no difference between
the two cases without "that," Fi^ 2 < 1· Note also that there is a
"that"-effect that co-occurs with the "that"-trace effect. Object Extrac
tion in cases with "that" is substantially less acceptable than is
extraction without "that," Fi(l, 31) = 20.0***, F2(l, 23) = 8.28**.
B.3 Coordination and Binding Ttieory
The findings simimarized in Figure 5 confirm the influence of

the Binding Theory in all three pairs of cases. The main effect of
Antecedent Location (Local versus Remote) is significant overall,
Fi(l, 41) = 125···, F2(l, 23) = 131***, and in each of ti\e three contexts.
No Coordination, Fi(l, 42) = 207···, F2(l, 23) = 151·*·, Simple Coor
dination, Fi(l, 42) = 54.8···, F2(l, 23) = 72.6···, and "Both" Coordina
tion, Fi(l, 42) = 19.9···, F2(l, 23) = 9.69··.
There was a reliable interaction between the Antecedent Loca
tion and Context factors, Fi(2,84) = 28.9**·, F2(2,46) = 22.5···, and a
main effect for Context, Fi(2,84) = 23.6···, F2(2,46) = 11.6***. Looking
only at the Remote Antecedent cases, there was a significant increase
in acceptability from the No Coordination condition to the Simple
Coordination condition, Fi(l,42) = 23.7***, F2(l, 23) = 13.8**, and from
the Simple Coordination condition to the "Both" Coordination con
dition, Fi(l, 42) = 19.6··*, F2(l, 23) = 14.6***.
B.4 Stability of Responses to Individual Sentences
Correlation tests within the two subsets of filler sentences con

firmed the impression of Figure 7; there was a strong correlation be
tween First and Second Session ratings for both the High Acceptability
items, F(l, 62) = 140***, ? = .70, and for the Low Acceptability items,
F(l,30) = 76.3»»»,r^ = .72.
Notes
1. The original data set for this experiment is no longer available; no analysis
by materials was done.
2. Other criteria can also be used, such as the percentage of responses that fall
in the two highest categories or the percentage in the lowest category.
Appendix C:
Excel as a
Syntactician's
Workbench
This appendix provides information on how Microsoft Excel (version

5), a spreadsheet program, can be used to construct and analyze
experiments on sentence acceptabihty. Although Excel's primary
applications are quantitative, it includes a rich set of text fimctions
that support the manipulations needed in constructing sentence
materials. A spreadsheet is, of course, a natural tool to use for the
extensive data manipulation process that intervenes between data
collection and statistical analysis. 1 have adopted Excel as a frame of
reference for the discussion here because it is available in nearly
identical form for both Macintosh and Windows systems, is widely
used on both those platforms, and is supported by extensive third
party docimientation. Because Excel (and its companion, Microsoft
Word) can work with a wide variety of fonts and is compatible with
multilingual technologies on both the Macmtosh and the Windows
platforms, the approaches described here are also appUcable to work
on languages using a variety of orthographies. Note, however, that
the solutions to the practical problems presented here are far from
unique. Many other approaches are possible within the framework
assumed here and a variety of other software tools provide similar
functions or alternative fimctions that could achieve similar results.
The particular processes and approaches described here are meant
141
to strike a balance between efficiency and intelligibility. Experienced

investigators may well want to alter or adapt these procedures to
better suit their individual needs.
Excel also incorporates (as do all widely used spreadsheet
programs) a macro language (Visual Basic) that allows users to
automate repetitive tasks. In some cases, this facility can be applied
via a "show me" method where the user simply executes the process
in question while Excel "watches" (i.e., records the process as a series
of program steps that can later be applied to other cases). This
appendbc will be concerned only with "manual" methods because
these are more readily adapted to a variety of goals and research
designs, but where an investigator uses certain methods frequently
on files of fairly constant structure, there are ways to automate these
methods that can provide significant time savings.
The appendix does not include basic training on Excel. A wide
variety of hitorial and reference materials are currently available,
including works that are aimed at users at every level of expertise (or
lack thereof). Most large bookstores (on campus and off) will carry a
substantial assortment of these materials. This appendix will assimie
knowledge of terms and processes that are covered in standard
tutorial materials on Excel.
The processes I will describe below have many steps. It will be
wise to save files often and to preserve intermediate steps in the
process so that later errors can be corrected without begirming from
scratch again.
The examples and illustrations provided below will often as
sume a 2 X 2 experimental design such as is applied in the "that"-trace
experiments discussed in the book. The process of adapting these
procedures to other designs is generally straightforward, although
novices may want to stay relatively close to the examples in their
early efforts.
C.I Accessing Software Tools and Sample Files
The examples used in this appendbc will be smallish and some

what toylike because more realistic ones would be far too large to be
usefully reproduced in print. Full-scale examples and other materials
relevant to this book are available, however, on the World Wide Web.
At the site maintained by the Linguistics Laboratory at The University
Appendix C 143
of Southern Mame, readers will find all of the example spreadsheets

used in the book. These include all the spreadsheet functions used to
construct the examples and some templates that can be used in
developing new experhnents. The site also includes detailed reports
on some of the experiments mentioned in the book, including listmgs
of materials and results. The web address for the site is
http://www.usm.maine.edu/~lin.
C.2 Building Questionnaires (Chapter 9)
Excel provides a concatenation function "&" that can be used

to assemble text strings appearing in different cells into a smgle
string. This can be used to assemble the various members of a token
set from a list of the common components of those sentences. A
formula of the form
=D4&" "&D6
concatenates the contents of cell D4 with a single blank character

(denoted by " "), and concatenates that with the contents of cell D6.
This function can be used to construct a token set from a series
of components, as illustrated in Figure 25. The formulas appearing
m B10:B13 reflect the contents of the cells in D10:D13.
Once a single set of components (as in D4:D7 in Figure 25) and
a set of formulas have been constructed, further token sets can be
derived from the original example. The text components for addi
tional token sets are entered in one or more columns to the right of
Colurrm D, with parallel components appearing on the same row as
in the first token set.
When the cells D10:D13 are selected (as in Figure 26), a "fill
handle" appears in the lower right comer of the selected range.
By pointing at the fill handle and dragging it directly to the right
by one or more columns, all the formulas in D10:D13 are copied
rightward over the range covered by the drag operation. Simultane
ously (because the original formulas are imphcitly defined in relative
terms), each new formula is adjusted to point to cells in the same
colimin as the formula itself. Thus the original formulas can be
extended to construct many further token sets to the right of the
original.
RppendlK C
Σ Ι I E
CMAOtffSif Token S«tCcntpomnls

D4 -· WhodW the counselor Yiart
05 - » for
06 -. lo hug Chantal?
07 -· Chantal Id hug?
Fotmuis TakeiiSet
Who dWlhecounseiorwart:
- D 4 & ' ~&0e
10 to hug Chantal?
-D4&- -&D7 Who did the counselor r « r t

II Chantal to hug?
- D 4 & " -aJDSA" - & D 6 Who did the counselor v a i t

12 tor (D hug Chanbal?
Who did the counselor wart
- D 4 & - -&D5&- -&D7
13 for Chantal to hug?
14
15 0.779329683
16
WT4 •TMK»roiectiag t > k e B »et«
Figure 25. F o r m u l a s for a s s e m b l i n g t h e s e n t e n c e s o f a t o k e n set f r o m a set

o f fixed c o m p o n e n t s .
Note that there is a numeric value in cell D15. This is generated

by the function =RAND(), which should be copied to the right along
with the formulas that construct the token sets. This will place a
random nimiber in the corresponding cell below each token set. Once
the complete table of token sets is constructed, the full table, includ-
ing the rows of sentence components, the rows of constructed token
sets, c m d the row of random numbers, should be selected and a
rowwise sort operation should be done using the row of random
numbers as the sort key.' This randomizes the sequence of token sets
from left to right to eliminate any trends or patterns that might have
affected particular subsequences of token sets within the larger set.
Appendix C 145
Who did the counselor want

to hug Chantal?
Chantal to hug?
for to hug Chantal?
for Chantal to hug?
F i g u r e 2 6 . Selected r a n g e w i t h fill h a n d l e in t h e l o w e r right c o r n e r .
After this sort operation, the random numbers can be replaced by

permanent Item ID numbers or codes.
The complete table of token sets and their ID numbers should
next be copied mto a new blank worksheet (preferably in the same
workbook). This copy operation should also accomplish two further
effects. First, the table of token sets should be copied as "values" (not
formulas) and the rows and columns should be transposed (see the
Edit/Paste Special command).^ This should provide a table in the
new worksheet that has as many columns as there are different
sentences within each token set, plus a coltuim for the ID numbers
and as many rows as there are token sets. The sentences should
appear in the cells as plain text. A schematic representation of this
list appears in Figure 27.
1^ *iΒ I r
RppendlK C
SI
c
NTSE(I)
I
NTOEd) WTSE(t) Vm)E(1>
I
32 MTSE(2> ΚΠ)Ε(2) WTSE(2> WT0E(2>
53 MTSE<3) MroE<3> WTSE(S> WT0E(3>
54 KTSE(4) MT0E<4) νΠΏΕ(4)
55 KTSE(5) KT0E(5) WTSE(5)
56 NTSE(6> ΝΤ0Ε(6} WTSE(6)
37 IIT3E(7) HrOE<7> WT3E<7) WT0E(7>
3Θ HTSE(8) ΝΤΟΕ(β) WTSE(8) WTOE(e)
10 S9 KTSE(9> MTOE(9> WTSE(9) WT0E(9)
SID NTSEdO) NTOE(tO) WTSE(IO) WTOE(IO)
11 Sti MTSEdt) MTOEdl) WTSEdl) WTOE(tl)
12
13 312 KTSEdZ) MT0Ed2) WTSE(12) WT0E(12)
Μ
MM
Figure 27. Schematic of a list of "thaf'-trace token sets.

N O T E : T h e worksheet shows 12 token sets ( S l . . . 812), each of which includes sentences o f four
different types, with each sentence being represented here b y s y m b o l s for its type aiKl token set
number, such as NTSE(3) or W T O E ( 7 ) . Note that all o f the sentences of any o n e type a p p e a r i n the
s a m e colunm.
The goal of the next operation is to replicate the structure of

Table 12, a set of four preliminary scripts. In effect, what's needed is
to "rotate" the table seen in Figure 27 to place equal nimibers of
sentences of the four types in each of the four columns of sentences.
This is achieved by shifting blocks of sentences from the right
side of the table to the left via Excel's drag and drop feature. The
operation is applied first to the three WTOE sentences for Token Sets
4-6 that appear in the range F5:F7 at the right margin of the table in
Figure 27. The block of cells is selected (see Figure 28), the SHIFT key
is depressed, the cursor is pointed at the boundary around the
selected block of cells, and the mouse button depressed. Keeping the
mouse button depressed, the cursor is then moved to the left edge of
the range C5:C7. When the cursor is pointing at somewhere close to
the middle of the left edge of this range, a large gray I-beam cursor
will appear at the boundary of Columns Β and C and indicate the
point where the block of cells will move to. When the mouse button
is released, the selected block of cells is moved to the indicated
location and all the cells lying between its old and new positions are
shifted to the right.
Appendix C 147
30 niutvo; n\ot\si— TTiutvo;

34 rfr3E<4) NT0E(4) WTSE(4) WT0E(4) 1
35 NT3E(5) NT0E(5) WT3E(5)
36 NTSE(6) NT0E(6) WT3E(6) InulHH
WTnFi'7\
F i g u r e 2 8 . Selecting a b l o c k o f cells.
RppendiH C
I τ I
31 NTSE(I) NTOEd) V/TSE(1) WT0E<1)
S2 NTSE<2) NT0E<2) WSE(2) νΠΌΕ<2)
33 NTSE(3> NT0E{3) WTSE($> Wn)E(3)
S4
S5
36
KTSE(4)
li#(Bl
NT3E(6)
MT0E(4)
IITOMu)
NT0E(6)
WTSE(4>
N'if8E(B) jϋtiH
/τϋtlέgί(sίj
57 MTSE(7) NT0E(7)
58 Ki^-(n) ΙΙΤΠΓ(8)
to 59 MTSE(9) .ΝΤρ.Ε(9).
tl 5 1 0 NTSE(io) i^£{!tdi;
12 511 M ^ d i ) )im(î\]
II. 3 1 2 NTSE(12) *tet{;f?)
14
M R • I M l / unrottt^rottHn
F i g u r e 2 9 . " R o t a t i n g " a list o f t o k e n sets.
The same operation is next apphed to the range E8:F10 that

contams the WTSE and WTOE blocks of sentences for Token Sets 7-9,
and then to range D11:F13 containing the NTOE, WTSE, and WTOE
blocks for Token Sets 10-12. The complete process is simimarized in
Figure 29.
Once this operation is complete, the structure of Table 12 should
be achieved, as in Figure 30.
The next series of operations apphes the process described m
Table 13 in Chapter 9 to each of the four preliminary scripts in Figure
30. The aim of this effort is to integrate each preliminary script with
the Ust of filler sentences and to construct an appropriate semiran-
dom ordering of the full list of experimental and filler materials. This
flppendiK C p i
ΊΣΕ
Prell ml ntrg Scripts
1 2 3 4
31 NTSE(I) NTOE(I) WTSE(1) WTOE(l)
52 NTSE{2) NT0E(2) WTSE(2) WT0E(2>
53 MTSE(3) NT0E(3) WTSE(3) WT0E(3)
34 WT0E(4) MTSE(4) ΜΓ0Ε(4) WTSE(4)

35 WT0E(5) NTSECS) ΜΓ0Ε(5) WT3E(5)
36 WT0E(6) NTSE(6) NT0E(6) WTSE(6)
37 WTSE(7) WT0E(7) NTSE(7) NT0E<7)

S8 WTSE(8> WTOE(e> NTSE(8) ΜΤΟΕ<β)
10 39 WTSE<9) WT0E(9) NTSE<9) KT0E(9)
11 StO NTOEdO) WTSE(IO) WTOE(IO) NTSE(IO)

12 311 NTOE(II) W T S E d l ) WTOE( 1 1 ) HTSEdl)
13 312 NTOEdZ) WT3E(12) WT0Ed2) ΜΤ3Ε(12)
14
rotated Τ
Figure 3 0 . Same structure as Table 12 (a counterbalanced set of four preliminary
scripts).
operation must be applied to each preliminary script to produce a

list that we will term simply a script. Each script should be used in
two or more different orderings. Table 16 includes a summary of the
key concepts cormected with scripts.
Although this process may appear tedious at first glance, it
takes an experienced user less than 2 minutes to apply all of the
following procedures to generate one ordering of one script. Because
only six to eight ordered scripts are needed for most experiments, the
total time investment in this process per experiment is not large (once
the investigator is familiar with the tools and the process itself).
Each pass through the four-step process described in Table 13
produces one urvique ordering for one script.
The Table 13 process starts with the construction of a template
like that schematized in Figure 31. The template consists of a repeat-
ing series of "Block IDs," a column of blank cells for Item IDs, a
coluirm of blank cells into which experimental and filler sentence lists
can be copied, and a column of random numbers. There should be as
many distinct Block IDs used in the left-hand colunm as there are to
be blocks in the final ordered scripts. However many Block IDs there
Appendix C 149
T a b l e 1 6 K e y c o n c e p t s related t o s c r i p t s a n d their s t r u c t u r e a n d o r d e r i n g .
Script Each script consists of a unique subset of all the experimental

sentences in the full list of token sets. Each script has exactly one
sentence from each token set, includes representatives of all the
available sentence types, and has equal numbers of each of the
several types. There are as many scripts as there are different
sentences in one token set.
Ordered Script Each ordered script is a unique sequencing of all the materials in
the script; each script should appear in two or more orderings.
Each ordering of a script is broken into a number of subsections
termed blocks, the number of blocks being the integer result of
dividing the total number of token sets by the number of
sentence types within each token set.
Block Each block includes one representative of each of the sentence types
and a number of filler sentences (the number being the integer
result of dividing total number of filler sentences by the number
of blocks). Within each block, items appear in semirandom order.
Semirandom The order of appearance of the sentences within one block is
subject to a controlled randomization process. The process of
constructing the sequence within the block is random, but only
those sequences are used where no more than two experimental
sentences (i.e., those drawn from the token sets) appear
consecutively.
are, there should be a repeatmg series of them down the column. The
number of repetitions of this series is equal to the total ntomber of
experimental and filler sentences that wiU appear in each block. The
RAND( ) function can be used to generate the column of random
numbers. Once an appropriate template is constructed for a particu
lar experiment, the template should be saved so that copies of it can
be generated for each new script and ordering needed for the full
experiment. Excel provides a template file type in the File/Save As
dialogue. When a worksheet is saved in this format, each time the
template file is opened. Excel creates a new workbook with a unique
name.
To construct one ordering of one script, the list of experimental
sentences from one preliminary script (see Figure 30) is copied into
the upper part of the sentence column in the template and the list of
fillers (usually constant for all scripts and orderings in one experi
ment) is copied into the lower part of the colunm. At the same time,
the Item IDs should be copied into the Item ID column. Notice that
the Item IDs in the schematic example in Figure 32 indicate that the
sentence is an experimental item (the initial "S"), show the ID number
HppendJH C
Μ I c I
Β
Block Item
Sentence
Random
Ο
ID ID Number
1 0760010
2 0.158528
3 0.706782
1 0.V72257
2 0.239562
3 0.551109
1 0.206695
2 0.648457
10 3 0.952483
II 1 0.111553
12 2 0.546889
IS 3 0.303434
14 1 0.515593
IS 2 0.561069
16 3 0.133546
1 0.738032
17
18 2 0.139082
0,764607
19 3
1 0.468580
20 0.231173
21 2
3 0.854345
22
1 0.416409
23 0.609614
24 2
3 0.427696
25
26
template
F i g u r e 3 1 . A t e m p l a t e for e x e c u t i n g t h e p r o c e s s d e s c r i b e d in T a b l e 1 2 .
of the token set, and indicate which particular condition this sentence
represents ("NS" for No "that," Subject Extraction, "WO" for With
"that," Object Extraction, and so on). Here the IDs for filler sentences
indicate only that the sentence is a filler and give its number. Some-
times further distinctions among filler types will also need to be
encoded.
Once the preliminary script and filler list are entered in a
template, only three sort operations remain to complete this ordering
of the current script. Each of these sort operations can be completed
very quickly by selecting the relevant range (starting from the col-
unm that holds the data that will control the sort operation) and
clicking on Excel's Sort Ascending button:
Appendix C 151
HppendiH C
A I Β I
Block Item Random
Sentence
ID ID Number
1 S 1 _ N S W h o did N o n t e o . . . 0.055481
2 3 2 _ N 3 Who are thestu... 0.259332
3 S3_J6 W h o did the det... 0.179296
1 S 4 J W 0 W h o does the nu... 0.019341
2 S 5 _ W 0 W h o does she t h . . . 0.813805
3 S 6 _ W 0 W h o had the peo... 0.157172
1 S 7 _ W S W h o might E m i l y . . 0.732924
2 S8_WS Whodldthevcl... 0.801952
10 3 S 9 _ W S W h o did the edi... 0.591400
11 t S 1 0 - N O W h o does Lou ex... 0.114735
12 2 S 1 1 _ M ) W h o did t h e c o u . . . 0.346352
13 S 1 2 _ N 0 W h o does your m... 0.890398
14 FI E x t e r n a l Jacks... 0.078323
15 F2 Matters Improv... 0.916413
16 F3 W h o the presen... 0.999352
17 F4 Ironlcallg, re... 0.2490S8
18 F5 Solution Peter... 0.456042
19 F6 The also ex pan... 0.844186
20 F7 W h o must a l s o . . . 0.301013
21 FB The absolute t... 0.496661
22 F9 W h o does Darle... 0.416967
23 F10 NAFTAvasabo... 0.962412
24 Fll W h o worshipped.. 0.653221
25 F12 W h o f o r a l l et... 0.211123
26
p r o j e c t ! no token sets
Figure 3 2 . Template with sentences.
The three sort operations have the effects of (a) randomly

assigning sentences to blocks, (b) segregating the blocks, and (c)
randomly ordering materials within each block. For the sample data
in Figure 32, these operations take the following form:
• To assign sentences to blocks, select E2:C13 (start the select

operation from cell E2 and do not include Column Β m the range to
be sorted) and chck Sort Ascending. Repeat this operation for the
filler sentences by selecting and sorting the range E14:C25.
• To segregate the blocks, select B2:D25 (starting from B2) and

click Sort Ascending.
• To randomize the order of items within each block, select the

block (starting from Column D) and sort. For the first block in Figure
32, select E2:C9 and sort. For the second and third blocks, sort the
ranges E10:C17 and E18:C25. Each of these sort operations should be
checked to ensure that it does not create sequences of more than two
consecutive experimental items. When such sequences arise, the
easiest way to remove them is to re-sort the block. After each sort
operation, the sorted range stays selected and all the random nimi
bers change^ thus, to generate a new random ordering of the range,
it is only necessary to click the Sort Ascending button again.
For a template of fixed structure that may be reused in further

experiments, it is easy to automate much of the process described
above by way of Excel's Visual Basic macro language. Note that,
where appropriate, filler sentences can be incorporated in the tem
plate because they usually do not vary from script to script within
one experiment. Where macro functions and filler sentences are built
into a template, the process of generating a new ordering of a script
can be executed in much less than a minute.
Once an ordering of a script is complete, the list of sentences
and their Item IDs should be copied to a master worksheet that will
ultimately contain all the orderings of all the scripts used in the
experiment. The Item IDs will later be extracted to guide the decod
ing of data files. Each ordered script should be assigned an identifier
as it is copied into the master list and the association between the
identifier and the ordered script should be carefully maintained in
all further operations.
A complete set of questioimaires can now be constructed by
copying each ordered script into a copy of a master questionnaire.
The master questionnaire should include all the invariant material
that is to be presented to every informant at the begirming or end of
the procedure. When the sentence list for an ordered script is copied
into a questionnaire, the items can be automatically numbered by a
variety of devices. If the sentence list for an ordered script is copied
into a Microsoft Word file as unformatted text, each sentence will be
automatically treated as a "paragraph" in Word. The list of para
graphs (sentences) in the Word document can be selected and as
signed a numbering scheme en masse automatically via the In
sert/Bullets and Numbering dialogue.
Appendix C 153
Because the various questionnahres for an experiment wiU be

very similar in appearance, it is worthwhile to add identifying infor
mation to each questionnaire file to help ensure the integrity of the
final printed questionnaires. Each questionnaire file should be given
a unique file name that reflects or incorporates the identifier for the
ordered script it contains. Each page of a questionnaire can include
this file name and a time/date stamp in a header or footer that is
automatically updated each time the document is printed.* This
ensures that each time a modified questiormaire file is printed, all the
pages of that printing will carry unique identifying mformation.
C.3 Coding/Decoding Data (Chapter 10)
The goal of the next series of steps is to convert a results file

produced by a scarming system into an orderly data file patterned
along the lines of Figure 23 in Chapter 10.
The tools needed to clean up and ahgn raw data files (see Figure
16 and Figure 17) are mostly commonplace word processing func
tions. The task is facilitated if the files are displayed in a monospace
font (e.g.. Courier) so that the nth character on each line falls at the
same horizontal position on each line. It is also often helpful to set
the word processor or editor's display so that it displays an overt
marker for space characters and nonprinting characters such as hard
return or line feed (see Word's Show/Hide [¥j button). Novices
should be aware of the distinction between the proprietary formats
that most word processors use to store formatted word processing
files and the so-called text or ASCII files. Text files are generally easier
to copy or load into a spreadsheet than are files coded in one of the
proprietary word processing formats, although most spreadsheets
have conversion functions that accommodate a variety of input
formats.
On most systems, cleaned data files in text format can be moved
into a spreadsheet by way of a Copy and Paste operation originating
with the editor or word processing program. Text files can also be
saved and opened directly by the spreadsheet program. In this case.
Excel will laimch its parsing function automatically as part of the
process of loading the file.
In general, the Copy-Paste route is more convenient. The copied
data will appear in a single column of cells in the spreadsheet, with
all the data (up to 256 characters) from each line in one cell. This
column of cells should be formatted so that the contents are dis
played in Courier or another monospace font (in Excel, use For
mat/Cells/Font). Select the single column range of cells where the
copied data appear and invoke Excel's parsing fimction (Data/Text
to Colvunns). This will launch a "wizard" (a series of dialogue boxes)
that will guide the user through the process of specifying how the
input material is to be parsed. The first dialogue box asks the user to
specify whether the different fields (data items) in each row are to be
distinguished by their linear position (the Fixed Width option) or by
marker characters such as commas or tabs appearing between the
fields (the Delimited option). In Figure 17, the first two fields (col
umns of data) on most lines could be distinguished either way
because spaces follow the single character Type field and the seven-
character ID number. However, there are no marker characters be
tween the 37 single-character response items that appear to the right
of each line (which is typical of many scanning systems). Each
character represents a single informant response. The Fbced Width
option must be used to parse data formatted like this.
The second dialogue box in the Text to Coltmms Wizard allows
the user to explicitly specify where the lines are to be parsed. In the
third dialogue box, the user indicates what format is to be applied to
each data item recovered by way of the parsing operation, and where
the array of cells that results from the parsing operation is to be
placed.
The parsed data should be labeled so that each column in the
data table has a label on the first row (e.g.. Row 2 in Figure 18). Some
of these labels can usually be generated automatically. For example,
if "SI" is typed into a single cell, the cell can be selected and its fill
handle dragged right across the cell's row or down along the cell's
column. In either case, this will yield a series of labels in successive
cells of the form "S2," "S3," and so on. This works when the string in
the initial cell begins with an alphanumeric sequence of any length
(up to the limit of cell capacity) and ends with an integer value having
not more than 10 digits.
Once the input data (from whatever source) has been parsed,
ordered, and labeled, the worksheet containing the data should be
saved and backed up. Only copies of these data should be used in
further analyses. The parsed and labeled data table should be pre
served as is.
Appendix C 155
The next step m analyzing the data is to isolate and reformat

the Item ID information that was saved with the ordered scripts so
that it can be used to identify individual responses in informant data.
Rows 3 through 10 in Figure 33 (the section labeled "Codes") contain
reformatted Item IDs. The first of each pair of rows identifies the type
of each item ("F" for filler sentences, "NO" for No "that," Object
Extraction, and so on) in an ordered script. The second row of each
pair of lines shows the item number (within the relevant type). Thus
"WO" and "10" in Column Ε is a reference to Token Set 10 m the With
"that," Object Extraction condition. "F" and "12" in Column F is a
reference to filler sentence 12. The left-to-right order of the codes in
each of the four pairs of Item ID lines corresponds to the sequence of
sentences in the relevant ordered script. When large sets of labels are
being constructed, the same parsing functions used earlier to put
scanned data into columns can be applied here. Excel's text functions
(e.g., LEFT or MID) are often also useful in constructing labels.
The most efficient way to integrate the Item ID information
with the response data is to sort the full table in Figure 33 by Script
ID and Line Type (see Note 1). This sort must be specified via the
Data/Sort dialogue. If Script ID in Column Β is specified as the Sort
By column, and the Line Type column. Column C, is specified as the
Then By column. Excel wiU group all of the hnes relevant to ordered
script 1 together and then sort among these lines according to the
values in the Line Type column. Because all of the response data lines
have been assigned a Line Type value of 3, and the two kinds of Item
ID lines have values of 0 and 1, the Item ID lines will be sorted ahead
of the data lines to which they are relevant, giving a table such as that
in Figure 34.
To order items by number (i.e.. Item ID) along each row (as m
Figure 20), it is necessary to do rowwise sort operations that are
governed by the Item Number values, for example, m Row 5 of
Figure 20. In a table structured similarly to Figure 34, the needed
sorting can be done as follows, with the four rows related to Ordered
Script 1 (Rows 4 through 7) used to iUustrate the process. First, the
entire relevant range for a given ordered script (E4:AB7 in this case)
is selected and the Data/Sort function is invoked. CUcking the Op
tions button brings up the Sort Options dialogue, where the user
should check the "Sort Left to Right" button and click OK. The Sort
operation is then specified by selecting Row 4 as the Sort By field and
BB—iHJimimt flppendiH C — — B ^ — ί
ξ1
A! Β I C m
i Ε I F I G IΗ I I IJ I Κ I I I A8 1 AC �ι
Script Line
ID Type
1 0 WO F NS F F F NS NO F
1 1 10 12 1 1 9 4 2 5 5
2 0 F WS WO F F F NO NS NS
2 1 4 4 θ 5 11 3 3 11 10
3 G F NO WD F WS F F WS NS
3 1 6 12 4 S 3 8 3 1 7
4 0 F NO NS F NO F NO F WS
4 1 3 8 6 2 9 6 7 11 10
li
12
IT SJD LT Si S2 S3 S4 S5 S6 S7 S8 S24
14 1 3 4 4 8 8 7 6 2 8 3
15 1 3 1 1 8 0 7 0 1 5 6
16 2 3 3 1 0 1 7 5 8 4 7 D
17 2 3 4 5 8 7 5 3 2 6 7 A
18 3 3 8 8 7 8 8 7 2 2 4 Τ
19 3 3 6 9 7 7 0 3 5 5 3 A
220Γ 4 3 0 2 4 3 6 9 8 8 2
3
4 5 4 5 0 1 4 6 g 1
22 FTFKlteBlP»/ Interleaved ItemlDa
JL
—
Figure 33. Item IDs formatted for use in decoding data.
N O T E : Tlie "Script I D " colunm associates each line with an ordered script in tlus experiment.
Thus the t w o lines in the data range (B14:AB21)with Script IDs of 2 represent data from ir\f ormants
w h o used ordered script 2, and the two lines in the C o d e s range {B3:AB10) with Script lEte of 2
are the Item I D s that identify the type and n u m b e r of each o f the responses from infonnants.
Columns Μ through A A have been hidden.
specifying a Descending sort on this field (this wiU put the fillers,
marked "F," after the experimentals). The No Header Row button
should also be checked. The Sort operation is then executed. This
produces an order in which all items of the same type (i.e., all the
"WO" items, all the "¥" items, and so on) appear together left to right.
In this case (because of the descending sort), all the various categories
of experimental sentences will appear first in alphabetical order by
type ID, then the filler sentences. At this point, the range representing
Ordered Script 1 should be seen as divided into two parts, a section
with data from experimental sentences and a section with data on
Appendix C 157
RppendiH C
Al Β I c Id Ε I F I G i H l I I J l Κ I L I a b T
Script Line
ID Typ« SI 32 S3 S4 S 5 36 37 3 8 324
0 \ΑΌ F NS F F F NS NO F
1 10 12 1 1 9 4 2 5 5
3 4 4 8 8 7 6 2 8 3
3 1 1 8 0 7 0 1 5 6
0 F w s WO F F F NO NS NS
1 4 4 9 5 11 3 3 11 10
to 2 3 3 1 0 1 7 5 8 4 7
tt 2 3 4 5 8 7 5 3 2 6 7
12 3 0 F NO WO F WS F F WS NS
13 3 1 6 12 4 5 3 8 3 1 7
t4 3 3 8 θ 7 8 8 7 2 2 4
ts 3 3 6 9 7 7 0 3 5 5 3
t6 4 0 F NO NS F NO F NO F WS
t? 4 1 3 8 6 2 9 6 7 11 10
te 4 3 0 2 4 3 6 9 8 8 2
t9 4 3 5 4 5 G 1 4 6 9 1
20
iBterleaved i t e t l D « / Sheetj|^|w
Figure 34. Item IDs interleaved with data rows.

N O T E : C o l u m n s Μ through A A have again been hidden to save space. Note that n o columns or
rows should b e hidden in ranges selected for use in a sort operation.
filler sentences. Two further sort operations, one for fiUers and one
for experimentals, will achieve the intended result (as m Figure 20).
In each case, the same range is selected as in the previous sort, except
that it is limited to only the filler or experimental sentences. The sort
function is invoked and directed to do the sort using only the Item
Number row. Row 5 in Figure 34. When this operation has been
applied to fillers and experimental sentences, the items will appear
left to right in numerical order within each type. The same sequence
of sorting operations is then applied to the blocks of data for other
ordered scripts.
IDI I HPPNPH-C.HCl • ^ • • ^ • • ^ • • • H
I 1 c I 0 I Ε I r I G I h I I I J l i ^ l
3 Summaries
4 for
5 Individual
Informants S u m m a r i M for Individual
7 token sets
8
9
4. 4. 4- 4,
To Tia TZa 73e T4a Γ5β Γ6β
11 > yf Tib T2b 73b T4b Γ56 TSb
12
jT Hem T y p e Iderlilier a a a b b b
14 Item Number 1 2 3 4 5 β
15
16 R e s p o n s e D a t a for I n f o r m a n t 1 ffa 11b Β 7 7 β 7 3
jT R e s p o n s e D a t a for I n f o r m a n t 2 12B 12b 9 7 β 5 β 7
18
19
20 Item T y p e Identifier b b b a a
^
21 Item Number 1 2 3 4 5 6
•22 R e s p o n s e D a t a for I n f o r m a n t 3 13a 13b 2 1 4 θ 3 7
R e s p o n s e D a t a for I n f o r m a n t 4 14a 14b 0 5 9 9 9 θ
H
2 4
1 • |M|/Flg50 /Ftg3l / Ftg S 2 / F t g 5 3 / F i g 3 4 \ F i t 3 5 / "
Figure 35. A small data table.
C.4 Summarizing Data (Chapter 11)
The procedures described in this section add new columns and

rows to tables structured along the lines of Figure 20. For conven-
ience, the essential structure of Figure 20 will be represented here by
the small data table shown in Figure 35, which adds two columns in
which to summarize the data for individual informants and two rows
in which to summarize the data for individual token sets.
This table represents a hypothetical experiment in which four
informants judged each of sbc token sets (and where each token set
included sentences of two types, denoted a and b). This section will
identify the Excel functions needed to construct the several different
kinds of mathematical summary discussed in Chapter 11, and show
how these functions can be applied to both sunimaries by informants
and by token sets. There are convenience features in Excel that can
facilitate the operations described here, but the current discussion
will focus mainly on Excel's mathematical functions rather than the
mechanics of entering them into the worksheet.
Each of the cells marked by placeholders in the summary tables
(e.g., Ila, T2b) in Figure 35 must be replaced by an Excel formula that
Appendix C 159
provides the appropriate summary of the relevant range within the

relevant column.
Means Summaries
To generate a sxmunary using means, each placeholder in the

range C16:C23 is replaced by the mean of that informant's responses to
Type a sentences. These means are generated by a formula of the form:
=AVERAGE(E16:G16)
The first of these formulas can be constructed by typing in all the

material up to the "(" and then selecting the target range with the mouse.
Excel will enter the range reference in the formula. The finished formula
can then be copied and pasted into the cell below (containing place
holder I2a) to generate the appropriate summaries for Type a sentences
on those rows. The same formula can also be copied over placeholders
I3a, I4a (which will generate svmimaries for Type b sentences, instead of
Type a). A later operation will get these values (and others in the token
set summaries) uito the appropriate colvmin.
The process can be repeated for the colunm of cells summariz
ing responses on the Type b sentences.
Summaries by token set are constructed in essentiaUy the same
fashion. In the cell occupied by placeholder Tla, the AVERAGE
formula is entered up through the "(". Then the range E16:E17 is
selected with the mouse. The formula can be completed by typing
" ) " and RETURN, or just RETURN (Excel can supply the final")" for
most simple formulas). The formula entered for placeholder Tla can
then be rephcated by selecting its cell and dragging the fill handle
rightward to highlight the ceU containing T6a. and all the intervening
cells. This replicates the formula so that each copy svunmarizes
responses m its own column. The process can be repeated for the
second row of placeholders.
The last step of the operation is to take account of the fact that
the formula entered for Ila actually summarizes responses to TJφe b
sentences for Informants 3 and 4, and that the formula for Ήα summa
rizes Type b responses for Token Sets 4,5, and 6. Once the complete table
of summary formulas is constructed, values are rearranged by an
application of the cell drag and drop feature exploited above in
cormection with the table of preliminary scripts in Figure 30. The I3a
and Ma cells (C22:C23) are highlighted, the SHIFT key is depressed,

and the selected range is dragged to the right of Colunm D (whose
contents on these rows now moves into Column C). In their new
location, the formulas still point to the same ranges they did before
they were moved. Now all the svmimaries in Column C are for Type
a sentences and all those in Column D are for Type b items. The same
process is apphed to the right half of the two rows of svimmaries by
token set. The positions of the first and second rows are exchanged.
This same general process for building formulas and construct
ing summary tables is applied in constructing the other types of
summaries discussed below. Different summary types can be con
structed in new columns and rows parallel to the summary ranges
shown in Figure 35, or the original data table can be reproduced in a
new worksheet and the new summaries constructed there.
Percentage Summaries
Percentage summaries can be constructed by way of two variants

of Excel's counting function. The COUNT function counts the num
ber of numeric values that appear in the indicated range. COUNTIF
counts the nvunber of numbers in the indicated range that satisfy
some criterion (e.g., ">1"). These can be combined to determine the
percentage of values in a range that meet a criterion. If the formula
=COUNTIF(E16:G16,">=8")/COUNT(E16:G16)
replaces the placeholder Ua, Figure 35 Excel will return a result of .33, the
proportion of responses in the indicated range that are at or above 8.
Standard Scores
To calculate standard scores, it is generally best to first calculate

a mean and standard deviation for each informant (or token set),
although this is not necessary. At the cost of some redimdancy, the
mean and standard deviation calculations can also be built into the
formulas that calculate standard scores. I will assimie that the mean
and standard deviation are to be calculated separately because these
are often needed in other processes.
The overall mean and standard deviation for Informant 1 m
Figure 35 can be calculated via these formulas:
Appendix C 161
=AVERAGE(E16:J16)
=STDEV(E16:J16)
I will assume that these values are entered in columns to the right of
the table of responses in Figure 35 (in Colunms Κ and L). With these
values in place, the standardized means for Type a and b sentences
can be calculated with the following formula:
=STANDARDIZE(AVERAGE(E16:G16),K16,L16)
Geometric Means
Geometric means can be calculated for token sets by way of the

same procedures as those described above. For placeholder Tie, the
formula takes the following form:
=GEOMEAN(E16:E17)
C.5 Transferring Data to a Statistical Program
Worksheets containing summary tables should be carefuUy

preserved. Data can be passed to statistical software in a variety of
ways. A copy of the summary worksheet can be converted to values
(i.e., all the formulas can be replaced by the then-current results of
those formulas) and superfluous material removed (e.g., colvmins
containing individual responses). The resulting summary worksheet
can be saved as a new worksheet (some statistical software cem read
some spreadsheet file formats directly), or it can be converted to text
or other file formats (Excel offers many options under the File Type
selector in the File/Save As dialogue). It is often also possible to use
copy and paste operations to move data into statistical software.
Notes
1. A general caution about SORT operations is in order. Be very careful to

always include all the columns and rows that are relevant to a particular table in each
sort operation. Inspect the results of any sort carefully before executing any other
operations. When the range to be sorted is not properly defined, sort operations can
be quite destructive. Excel provides a powerful UNDO feature that allows the user to
back out of most sort operations that go wrong, but this is most effective if it is used
immediately after the error is made.
2. The transpose operation can be avoided if desired by applying the same logic
for constructing token sets as described earlier on a vertical rather than horizontal axis.
The column of sentence components and formulas in Figure 25 can instead be laid out
along a single row. Components for new token sets are then constructed in succeeding
rows and the formulas copied downward instead of rightward. This process will save
the transposition step later on but may not be practicable urJess the investigator has
access to a relatively large monitor.
3. The random numbers are generated by the R A N D ( ) fimction, which auto
matically yields a new random number in each cell where it is used each time a sort
is performed (provided Excel is set for automatic recalculation).
4. In Word, the name of the current file is invoked by w a y of the field code
FILENAME.
Appendix D:
Tolten Set Data
From a
"That"-Trace
Experiment
3
: v :
4
V,
9 ^ 10 12
17
1 1 I f
18 19
> v
a 20
SE O E S E OE SE O E S E O E SE O E S E O E S •E O E S E I OI IE
NT W T NT W T NT W T NT W T
Figure 36. Data on each of 20 token sets for a "thaf'-trace experiment.

N O T E : E a c h graph summarizes the data for 1 o f the 2 0 token s e t s listed in T a b l e 17. O n e a c h line
in each graph, the data points represent the N o "that"/Sut)ject Extraction, N o " t h a t ' V O b j e c t
Extraction, With " t t « t " / S u b j e c t Extraction, and With "that'VObject Extraction conditions in that
order, left to right. E a c h line represents data collected at a different site ( • W a y n e S t a l e College,
Nebraska, Ν = 192; • University o f Alabama at Birmingham, Ν = 33; A Uruversity o f Southern
Maine, Ν = 107). Because each data point summarizes responses to a single sentence, o r J y atxjut
o n e quarter o f each sample contributed to any one data point. T h u s the numt)er o f irxformants
contributing to each data point varies from around 8 to 48, for the Alabama and Nebraska samples,
respectively.
Appendix D 165
T a b l e 1 7 T o k e n sets.
Item Number Sentence
1 Who was the nurse imagining (that) would find her?

/(that) she would find?
2 Who do you suppose (that) invited Ann to the circus?
/(that) Ann invited to the circus?
3 Who did Marie hope (that) would dance with Phil?
/(that) Phil would dance with?
4 Who will Jim annoimce (that) is running against him?
/(that) he is running against?
5 Who does he expect (that) will visit the neighbors?
/(that) the neighbors will visit?
6 Who do you think (that) likes John?
/(that) John likes?
7 Who do you feel (that) should take care of Fred?
/(that) Fred should take care of?
8 Who did Jane suspect (that) had kidnapped Fred?
/(that) Fred had kidnapped?
9 Who did the operator say (that) called Maxwell on the car-phone?
/(that) Maxwell called on the car-phone?
10 Who did the actor pretend (that) had killed the soldier?
/(that) the soldier had killed?
11 Who had the lawyer heard (that) would be fighting for the defendant?
/(that) the defendant would be fighting for?
12 Who are they assuming (that) helped Chris with the taxes?
/(that) Chris helped with the taxes?
13 Who did the horse owners bet (that) would t>eat Secretariat in the
contest? /(that) Secretariat would beat in the contest?
14 W h o will Dan insist (that) is in love with him?
/(that) he is in love with?
15 Who did the article proclaim (that) was attacking the criminal?
/(that) the criminal was attacking?
16 Who did the kids wish (that) could go swimming with them?
/(that) they could go swimming with?
17 Who does the teacher believe (that) hurt the boy?
/(that) the boy hurt?
18 Who did they presume (that) had rescued Bill?
/that) Bill had rescued?
19 Who did she claim (that) beat up the boy?
/(that) the boy beat up?
20 Who had you dreamed (that) would marry John?
/(that) John would marry?
N O T E : The table lists the 20 token sets used in the "that"-trace experiment described in Figure 36.
Appendix Ε:
Sample
Questionnaire for
Scannable Line
Drawing
Sentence Judgment Experiment

Tha azperimant you are partki|»tiiic in i« Wa vary much appreciaU your willingnaaa to
part ofa pilot pfqfoct on timtiwwi> aaaiat ua in thia work.
acceptability fiuuM by the (U.8.) National
Sdoneo FoamUtion. If you would liko n o n Nona of the information collactad hare will be
iafennatien about th« project, w< will ba aaaodatad with your nama in any way.
happy to tdl you n u n about the (odb of th«
worit and tfaa naulti waSra aehiavad ao far. DO NOT WRITE YOUR NAME ON THIS
You may eat in touch with ua at tha addnaa QUESTIONNAIRE OR THE ANSWER SHEET.
ahown balow.
Lancuaga Sdencaa Laboratory
Wayne Cowart Univaraity of Southern Maine
Prinapal Invoatigator 96 Falmouth S t
Portland, ME M103
Dana McDuuel
Co-Prindpal Inveatigator E-mail: cowarteusm.maine.edu
SulgectID
PreHminariee
You ahould have received a ( 2 pencil and a read Ifae inatructioas carefully and do all
green (General Piupoae - NCS) Anawer Sheet talks apedfied. A U , OF THE STEPS OF THIS
with Om queationnwie. Fleaie let the PROCEDURE ARE IMPORTANT. SKIP
experimantar know if you are misaing either of NOTHINOul
theeeitema.
If i t any point you a n nndear about anything
nie following pagee in thia quaationnaire will you a n asked to do, pleaae feel free to ask the
teach yon how to record your reapeoiea in this experimentar for clarification.
experiment ft ia eictremely important thai you
Home Town
All other naponaaa you make in this procedun will be recorded on the green Answer Sheet However, your
response to this item is to be recorded h e n on the qoaationnain.
Question: When did you begin grade sdiool (first grade)? (Please write your responses to this
question in the blanks provided below)
cmr/TOWN
STATE OR COUNTRY.
Appendix Ε 169
o t h e r D e m o g r a j u c Information
Fleue take out your green Answer Sheet now and uae it to raconj the next several responaea. Pleaae use
your #2 pencil for thia and make heavy, clear marka, aa ahown below. Make no stray roarka on the Anawer
Sheet
A B C D E F G H I J A B C D E F G H I J
Make marka like thia 0 0 0 * 0 0 0 0 0 0 âu. O O O ^ O O O O O O
Pleaae check to aee that the red number atamped in the upper left hand comer of your Anawer Sheet agreea
with the Subject 10 atamped in the box on the tint page of thia queationnaira.
In the block marked "SEX* (juat left of the heavy green line in the middle of your Anawer Sheet) mark the
appropriate category.
There ia another labeled "GRADE or EDUC* just below the box marked * 8 Ε Τ . Pleaae indicate hare how
many yeara of formal aehooliog you have completed. In the V.3., completing high achool connta aa 12, and
completing a bacfaeloi'a degree connta aa 16.
In the block marked "BIRTH DATE' (lower left), fill in the appropriate d r d e for the month of your birth,
then write in the day and year of your Urth and fill ia the appropriate cirdea below for the day and year.
Each of the oolumna A H in the bh>ck labeled 'IDENTinCATION NUMBER' win be naed (br a different
queation. Ilie queationa are liatMl below. Pleaae write in your reaponae in the empty box at the top of each
column, then fill in the appropriate bubble below. Remember Thia information ia not aaaeciated with your
name in thia atudy.
Column A: An y w a native apeaker of American Colainn F: Are any members of your immediate
Engliah? (0)yee (l)no family (biological parenU and aiblinga only)
C<daa>a B: Are you a native speaker of British left handed, ao far aa you know?
Engliah? (0)yee (l)no (0) confident aome are left-handed
Cotuinn C: Are yon a native apeaker of another (l)-not-aare, aome could be left-handed
variety of Engliah (Jamaican, Indian, (2) confident all are right-handed
AaatraUan, etc)? (0)yea (Uno Column G; What ia your academic major.
C o l n n u D: Among the parenta/caregivera of the (0) Engliah (1) a foreign language
family you grew up with, wiiat waa the bigfaeat (2) Unguistica (8) Other
level of education completed by any of thoae Cotumn B : Did you and your family move from
individuals? one city or town to anodier while you were in
(0) high achool or leat grade school (Years 1 4 ) ? (0)yea ( l ) n o
(1) college
(2) maaUra degree If you aaawered Yea to Qusation H, pleaae answer
(3) doctorate or-law or medical degree ()ueation 1 aa well
(4)don'tkBow Column I: If yo ed,how kwgdidyou
Cotnaan E : Do you ever choose to write with your continue adkool in the city/town where you
left band? (Ignore caaet where you are forced began first grade?
by drcunutancee to uae your left hand, as (0) leas than on year
when there ia temporary iqjuiy, a heavy (l)-«ne-to-two-yeara
package in your right hand, e t c ) (2)-three-to-four-yean
(Oinever (1) sometimes (2) always (3)-more-than-four-years
T h e BBepcnee IVooednre
The βΐφβηοιαιΐ you a n (bout ta jMiticipcte in lUM

an n a g a i a l n o p o o M pneadun. Tkii paga givaa , A B C O C f C H I J
you aome critkal bad«ronnd you will need to uae
thia procedun correctly. _ *e C D £ F C Η I J
Suppoaa you wen aaked to judge the relative a r w of β Η i 0 y y U 4 0
theae two circlea.
^ A B C D E F G H I J
A · 0 0 0 0 0 0 0 0 0
AecOEFGHtJ
• 0 0 0 0 0 0 0
Circle Β ia about three timea the a n a of A.
One way you could report yo
What matten in thia experiment ia that you OKT
of the niativa aiiaa ia to d n w two tinea, making
THE RATIOS RIGHT! You can make long Unas
the aecond three timea longer than the firat.
or ihort ones, just so the differencea in line length
agree with your impreaaions.
A
Β One men eompUealioa: We want you to t U a k of A
thru J as a una, but WS DONT WANT YOU TO
If we had a completely free choice, this ia how we ACTUALLY DRAW U N E S . Instead, juatahow us
would h a w you raport your imprasaiona to ua in wiien esdi of your lines wouM step. For each Hem,
thisexpeiiment UnfortnnataV, wa would then fin in tha dot that ia chisaat to when your line
have to spend months measuiing tans of thousands would end far that item. For example, tiie
of lines. diflerence between Cirdes A and Β could be
What we will do instead te this. Each time we ask indicated any of the following throe ways.
you to judge aomething, then will be an item like
thia on your Anawer Sheet A B C D E F G H I J
A B C D E F G H I J
A oioooooooo
A B C D E F G H I J
ooooooooool Β 0 0 0 0 0 * 0 0 0 0
But, WB WANT του TO I S D W O F I H E BOW A B C D E F G H I J

O F D O T 8 , A T H B U J , A e A L I N B . So, if yon wen
using this dotted Bos to deacfibe Circles A and B,
ooaooooooo
A B C D E F G H I J
you should think of it like this.
oooooooofto
A B C D E F G H I J A B C D E F G H I J
• « 0 0 0 0 0 0 0 0
A Β C D Ε F ε Η I J
•ooooooooo
A B C D E F G H I J
0 0 0 0
ooeooooooo
nwduTereneeisthatnowtheHneshavetostopon Now for some practice
one of the dots.
Looking at it this way, than a n actually
three ways that you eouU report that Cinle Β kioks
ôutthroetimaabiggereunCSicteA Inadditioa
to the ooe shown above, than a n also dwse two
other equally good ways of nporting the same
thing.
Appendix Ε 171
Pnetloe
Aa a warm-up for the main experiment, we'd like you to uae the procedure we've juat introduced to deacrilw
the cirdea bekm. Hie aeoraa you aaaign ahauld refiect the reUtiva site ofeadi drde COHPARBD TO T H E
FIRSTOMB. UaathaVK r S h M t for your naponaea. Find die place for Itam «1 en Side 1 cf the
Anawer Sheet Now HMK at Cinle t l below and dadda what acore (ramamber - think of it aa a line length)
yonwantteuaatorepreaentthaaiiearCbdatl. Fill ia the appropriate fadMe in Item « 1 on dia Anawer
Sheet For Circle t2,aelect a acore (line length) that ahowa how mud) larger t2ia,eompared t o i l .
ConUnae on and enter acores for each of tha remaining eight drcles uaing Items S through 10 on the Anawer
Sheet 171
171
171
171 171
O f P O B X A N n DDDOTBAGIBTLIIAPARTIONFLRFLM aanteoca ao that the aeon ahowa how Bach better or
TNNTFLYAALUNANIDLIMII woiw that aentance ia eomparad to the firat aantence.
«iiepneedfa«pi«BaiidoaavMadllieIVaotiaeitaH π » better a aenttnca aeema, UM b i ^ a r the acore you
•bovai ahouldose. Onea yoa'va givaa a naiMose for a
We need aome infbnnatien from YUA aboot your aeatenee, pleaae don't go back and teoonaider.
evaluation of soma lantancea we're Hated bdow. We lococniaa that the scale ia crada; it win not
We wouU like YOB te bnagine that your JOB ia to nbUa diOimoeae between aentancea.
teadi Enîah to êakera OF ethar langiiajgae. JaatdXhabeatyouemi.
For oMh aantance listed below, wa woold Uka yea to Yen dont need to worry about grammar mlaa yon
do the Mowing. Flaaaa n a d the aenlanee, than aak may have atudiadfaiadiooL TaU ua what aeema moat
yooiaalf IF the aentance aeema BngUah'WIADING or nalnial to yon. whafter or not that ia •piopar'lqr ralea
not Suppoaaaaeofyouratndcntawnetonaethia yoo may hava tearaad. Wa a n only intareated in
aentance. U wa ignora ptennadation, would the ymur aenae of what would be appropriate in ordmary
atudentaonndlika a native apeaker? Orwooldthe relaxed converaatian.
I atraage or onnatnnl to a native Alao. pleaae dont atiuggla over individual
r it waa pi«nonnoad7 Tour taak awteneaa. Make tha beat quick Jodgment you can for
ia (0 tan ua how Ei«liah-eounding eadi aentance ia eadi aaotenca and ( 0 on to HM naxt
using a aeale. Tlwre are 92 aentaneea on the qnaationnaire,
Let the l i n t aentence he your refnence. Aaaigna numbered <KTNN*lt to «102. Pleaae be aura to do them
riatetayou.Keepla all.
Pleaae record your responaea c o the Anawer Sheet
BTAIUNG with Item «11.
TfaeSeDtenom
11) This is s psiafiil movis to wstch. 28) Who did the article pioclaim that was
BtUcking the criminal?
12) Isrsol's $13,000 u n u a l par tha capita incoma
paru Wsatam Europa. 29) Who had you draamed that John would marry?
13) We had doubtad that peopla would diive up and SO) Who do you think John likes?
wait at tha border on week niefats.
31) Who, when Lisa contracted cancer found
14) It's also quiU possible to build your own herself terrible in pain, found he waa not of all
shelves bookcase, especially if have got a s u n his of theory?
knowledgeable ftiend nei^bor who's willing the
32) Extamal jacks plug of but a n next at the back
help.
cover plate.
15) Studying a the subject, we realize that Rob
33) During the second, the Black Pottery if
depended on people they being able think like.
Longshan phase, agricoltun became mon.
16) Giro Sport Design Inc., based near SanU Cruz,
34) tba alao expand aignificantly northwestwsrd
California, are making bicycle helmeU bicycle
into, besieging Vienna.
pumps in Mexico and hopes to sell them then soon.
36) Who was the nurse imagining would find her?
17) 58 percent of the businesies supported sn
import tax to make Japaneae producta lass 36) Who must also a n able to trust him, and to
competitive. believea he will then for you?
18) His brother believe they have a n on die brink 37) H i s a b i l i l y t o d i a n h e r v i e w i s a s m a n
of s bnaklhrougfa to the really big time. triumph, but one few people can claim.
19) A most extnotdinaiypsrty took place on the 38) Ironically, reform paved the way for a more
third floor second floor. radical political transformation.
20) Perhaps it took an adventunr, enigmatic and 39) At the beginning of the year the month,
reckleas, without a plan, heedless of risk, a con Lambert wanted only to make money, but at tha
man, to do «4iat he dkL end he wanted only to save his business.
21) l ^ e n has did no plot, no characters to identuy 40) Who does he expect will visit the neigfabon?
at, no hope.
41) Who for all at their resiliency, however, the
22) Women haa been assigned to non-combst seem stuck in a of rat in Stanleyville?
ships and aerved temporary duty aboard combat
42) Who had the lawyer heard that would be
ships in the past
fitting for the defendant?
23) Who the presented a n with harsh, sardonic?
43) Who did the actor pretend the aoldier had
24) Who claimed that the roaulting conflict killed?
between Serbia quidily escalated into World War
44) What hett going?
I?
45) A n then people students like Suzanne in that
26) His h i ^ r the cumber, the faster tha modem
class?
and the leaa it coats to use online servieas that
charge by the hour. 46) Many have long shared a "misconception that
everyone in Mexico is poor,' Grualva said.
26) 'It's just a big problem as see it," Freeman
said. 47) Destruction a n his only.
27) Who have questuned many of the basic 48) Who does the teacher believe that the boy hurt?
principles of Lang's theories?
49) What it take to do the and work is a computer, a
modem, and a the phone line.
Appendix Ε 173
97) He ia a teacher and a writer, a pipe-amoking

bachelor who lives in hia txwk-lined Oxford home
921 Tha "relatiooahip' that devalopa batwaan
with his brother.
Brian and Carol i> ao pathetic that it can barely be
watched. 98) The h c t that Murray is be able to combine the
qualitiea of comedy, romance and even
93) On Monday will ordering 60 women tailora
and officer* to duty aboard the aircraft carrier USS melodrama makes very rare.
Dwight D. Eisenhower. 99) Apologized and fired freelance.
9*) Who did the horse owner* bet that would beat 100) The in Mth and 16th oenturiea the, a heretical
SeereUriat in the eonteat? Chriatian sect, waa numeroua Boania in.
95) Who worahipped prayed to their ancestors, a 101) Who did she claim that the boy beat up?
Chinese custom Ut»t atill has persists?
102) Matters improve with the arrival of Claire,
96) Who did Marie hope would dance with Phil? whose name ia on the lease.
Pleaae now tear the FIBST P A I S off of thia

quwitkinnalre and i t t n m the flnt page and your
Answer Sheet loeedier to die aperimeater. The
reat of the questionnaire can l>e disposed of.
Thank You very much for your aaaiatanoe in thia

References
BaiidJ.,&Noma,E(1978).FiOTdiimenifl&qfsca/mêndpsyc^
Bard, E. G., Robertson, D., & Sorace, A. (1996). Magnitude estimation of linguistic
acceptability. Language, 72(1), 32-68.
Bemdt, R. S., Salasoo, Α., Mitchum, C. C , & Blumstein, S. (1988). The role of intonation
cues in aphasic patients' performance of the grammaticality judgment task. Brain
6· Language, 34(1), 65-97.
Bever, T. G. (1974). The ascent of the specious; or. There's a lot we don't know about
mirrors. In D. Cohen (Ed.), Explaining linguistic phenomena (pp. 173-200). Wash
ington, DC: Hemisphere.
Bever, T., & Carroll, J. M. (1981). On some continuous properties in language. In T.
Myers, J. Laver, & J. Anderson (Eds.), The cognitive representation of speech (pp.
225-233). Amsterdam: North-Holland.
Bley-Vroman, R. W., Felix, S. W., & loup, G. L. (1988). The accessibility of universal
grammar in adult language learning. Second Language Research, 4(1), 1-32.
Bloomfield, L. (1930). Linguistics as a science. Studies in Philology. 27.533-557.
Blumstein, S. E., Milberg, W. P., Dworetzky, B., Rosen, Α., & Gershberg, F. (1991).
Syntactic priming effects in aphasia: An investigation of local syntactic depend
encies. Brain end Language, 40,393-421.
Bock, J. K. (1987). Coordinating words and syntax in speech plans. In A. Ellis (Ed.),
Progress in the psychology of language (pp. 337-390). London: Lawrence Eribaum.
Bock, J. K. (1990). Structure in language: Creating form in talk. American Psychologist,
45,1221-1236.
Boring, E. G. (1953). A history of introspection. Psychological Bulletin. 50(3), 169-189.
Bransford, J. D., & Franks, J. J. (1971). The abstraction of linguistic ideas. Cognitive
Psychology, 3,33\-350.
Garden, G. (1970). A note on conflicting idiolects. Linguistic Inquiry, 1(3), 281-290.
Garden, G. (1973). Dialect variation in abstract syntax. In R. W. Shuy (Ed.), Some new
directions in linguistics (pp. 1-34). Washington, DC: Georgetown University Press.
Carroll, J. M., Bever, T. G., & Pollack, C. R. (1981). The non-uniqueness of linguistic
intuitions. Language, 57,368-383.
Chomsky, N. (1957). Syntactic structures. The Hague, the Netherlands; Mouton.
Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge: ΜΓΓ Press.
Chomsky, N. (1973). Conditions on transformations. In S. Anderson & P. Kiparsky
(Eds.), A festschrift for Morris Halle (pp. 232-286). New York: Holt, Rinehart &
Winston.
175
Chomsky, N. (1977). On wh-movement. In P. Culicover, T. Wasow, & A. Akmajian

(Eds.), Formal syntax (pp. 71-132). New York: Academic Press.
Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris.
Chomsky, N. (1986). Knotvledge of language: Its nature, origin and use. New York:
Praeger.
Chomsky, N., & Lasnik, H. (1977). Filters and control. Linguistic Inquiry, 8,425-504.
Clark, H. H. (1973). The language as fixed effect fallacy: A critique of language
statistics in psychological research. Journal of Verbal Learning and Verbal Behavior,
12,335-359.
Cohen, J. (1%9). Statistical power analysis for the behavioral sciences. New York: Aca
demic Press.
Cowart, W. (1989a). Ilhcit acceptability in 'picture' NPs. Papers from the regional
meetings, Chicago Linguistic Society, 25,27-40.
Cowart, W. (1989b). Notes on the biology of syntactic processing. Journal of Psycholin
guistic Research, 18,89-103.
Cowart, W. (1994). Anchoring and grammar effects in judgments of sentence accept
ability. Perceptual and Motor Skills, 79,1171-1182.
Ellis, R. (1991). Grammaticahty judgments and second language acquisition. Studies
in Second Language Acquisition, 13,161-182.
Ferreira, F., & Henderson, J. M. (1991). Recovery from misanalyses of garden-path
sentences. Journal of Memory & Language, 30(6), 725-745.
Fowler, A. E. (1988). Grammaticahty judgments and reading skill in Grade 2. Annals
of Dyslexia, 38,73-94.
Frazier, L., & Chfton, C. J. (1996). Construal. Cambridge: MIT Press.
Gerken, L., & Bever, T. G. (1986). Linguistic intuiHons are the result of interactions
between perceptual processes and hnguistic universals. Cognitive Science, 10,
457-476.
Geschekier, G. A. (1976). Psychophysics: Method and theory. Hillsdale, NJ: Lav^rcnce Erlbaum.
Heringer, J. T. (1970). Research on quantifier-negative idiolects. In M. CampbeU, J.
Lindhohm, A. Davison, W. Fisher, L. Furbee, J. Lovins, E. Maxwell, J. Reighard,
& S. Straight (Eds.), Papersfrom the sixth regional meeting, Chicago Linguistic Society
(pp. 287-295). Chicago: Chicago Linguistic Society.
Hill, A. A. (1961). Grammaticahty. Word, 17,61-73.
Katz, J. J., & Bever, T. G. (1976). The faU and rise of empiricism. In D. T. Langendoen,
J. J. Katz, & T. G. Bever (Eds.), An integrated theory of linguistic ability (pp. 11-64).
New York: Crowell.
Kerlinger, F. N. (1973). Foundations of behavioral research (2nd ed.). New York: Holt,
Rinehart & Winston.
Kirk, R. E. (Ed.). (1972). Statistical issues: A reader for the behavioral sciences. Monterey,
CA: Brooks/Cole.
Labov, W. (1972). Some principles of linguistic methodology. Language in Society, 1,
97-120.
Labov, W. (1975). Empirical foundations of linguistic theory. In R. P. Austerlitz (Ed.),
The scope of American linguistics: Papers of the first Golden Anniversary Symposium
of the Linguistic Society of America, held at the University of Massachusetts, Amherst,
on July 24 and 25,1974 (pp. 77-133). Lisse: Peter de Ridder.
Langendoen, D. T., & Bever, T. G. (1973). Can a not imhappy person be called a not
sad one? In S. R. Anderson & P. Kiparsky (Eds.), A festschrift for Morris Halle. New
York: Holt, Rinehart & Winston.
Linebarger, M. C , Schwartz, M. F., & Saffran, E. M. (1983, December). Syntactic
processing in agrammatism: A reply to Zurif and Grodzinsky. Cognition, 15(1-3),
215-225.
References 177
Lodge, Μ. (1981). Magnitude scaling, quantitative measurement of opinions. Beverly Hills,

CA: Sage.
Marks, L. E. (1974). Sensory processes: The new psychophysics. New York: Academic Press.
Nagata, H. (1987a). Long-term effect of repetition on judgments of grammaticality.
Perceptual & Motor Skills, 65(1), 295-299.
Nagata, H. (1987b). Change in the modulus of judgmental scale: An inadequate
explanation for the repetition effect in judgments of grammaticality. Perceptual
& Motor SkUls, 65(3), 907-910.
Nagata, H. (1988). The relativity of linguistic intuition: The effect of repetition on
grammaticality judgments. Journal of Psycholinguistic Research, 17(1), 1-17.
Nagata, H. (1989a). Effect of repetition on grammaticality judgments under objective
and subjective self-awareness conditions. Journal of Psycholinguistic Research,
18{3), 255-269.
Nagata, H. (1989b). Repetition effect in judgments of grammaticahty of sentences:
Examination with ungrammatical sentences. Perceptual & Motor Skills, 68{1),
275-282.
Nagata, H. (1989c). Judgments of sentence grammaticality and field-dependence of
subjects. Perceptual & Motor SkUls, 69(3, Ft. 1), 739-747.
Nagata, H. (1989d). Judgments of sentence grammaticality with differentiation and
enrichment sh'ategies. Perceptual & Motor SkUls, 69(2), 463-469.
Nagata, H. (1990). Speaker's sensitivity to rule violations in sentences. Psychologia, 33,
179-184.
Nagata, H. (1991). On-line judgments of grammaticality of sentences involving rule
violations. Psychologia: An International Journal of Psychology in the Orient, 34(3),
171-176.
Nagata, H. (1992). Anchoring effects in judging grammaticality of sentences. Percep
tual & Motor Skills, 75(1), 159-164.
Newmeyer, F. J. (1983). Grammatical theory. Chicago: University of Chicago Press.
Popper, K. (1959). The logic of scientific discovery. New York: H a φ e r . (Original work
published 1935)
Ross, J. R. (1970). On declarative sentences. In R. A. Jacobs & P. S. Rosenbaum (Eds.),
Readings in English transformational grammar (pp. 222-272). Waltham, MA: Ginn.
Ross, J. R. (1979). Where's EngUsh? In C. J. FiUmore, D. Kemper, & W. S. Wang (Eds.),
Individual differences in language ability and language behavior (pp. 127-163). New
York: Academic Press.
Schachter, J., & Yip, V. (1990). Grammaticality judgments: Why does anyone object to
subject extraction? Studies in Second Language Acquisition, 12,379-392.
Schiitze, C. (19%). The empirical base of linguistics: Grammaticality judgments and linguis
tic methodology. Chicago: University of Chicago Press.
Snow, C , & Meijer, G. (1977). On the secondary nature of syntactic intuitions. In S.
Greenbaum (Ed.), Acceptability in language (pp. 163-177). The Hague, the Nether
lands: Mouton.
Stevens, S. S. (1975). Psychophysics. New York: John Wiley.
Stevens, S. S., & Galanter, E. H. (1957). Ratio scales and category scales for a dozen
perceptual continua. Journal of Experimental Psychology, 54(6), 377-411.
Stokes, W. (1974). All of the work on quantifier-negation isn't convincing. In M. W.
La Galy, R. A. Fox, & A. Bruck (Eds.), Papers from the tenth regional meeting, Chicago
Linguistic Society (pp. 692-700). Chicago: Chicago Linguistic Society.
Wulfeck, B., & Bates, E. (1991). Differential sensitivity to errors of agreement and word
order in Broca's aphasia. Journal of Cognitive Neuroscience, 3(3), 258-272.
Zribi-Hertz, A. (1989). Anaphor binding and narrative point of view: English reflexive
pronouns in sentences and discourse. Language, 65,695-727.
Author Index
Baird,J.,74,77(nl)
Galanter, E. H., 7 1 , 7 3
Bard, E. G., 6 9 , 7 3 , 7 4
Gerken, L , 6,29(n6)
Bates, E., 5
Gershberg, F., 5
Bemdt, R. S.,5
Gescheider, G. Α., 71, 7 4 , 7 7 ( n l ) , 77(n2)
Bever,T.G.,5,6,8,29(r\6)
BIey-Vroman,R.W.,56
Bloomfield, L., 1
Henderson, J. M., 5
Bltmistein, S., 5
Heringer, J. T., 4
Bock, J. K . , 5 1
Hill, A. Α., 4 , 3 2
Boring, E. G., 3
Bransford, J. D., 50
loup, G. L., 56
Garden, G., 4 , 5 6
Carroll, J. M., 5 , 6 , 9 , 2 9 ( n 6 ) Katz,J.J.,29(n6)
Chomsky, N., x, 1 , 7 , 1 5 , 1 8 , 2 0 , 3 0 ( n l l ) , Keriinger,F.N.,43
32,122
Kirk, R. E., 77(n2)
Clark, H. H., 1 2 1 , 1 2 2 , 1 2 6 ( n 2 ) , 137
CUfton, C J . , 21
Cohen,J.,83,84(n2)
U b o v , W., 4 , 5
Cowart, W., 6 , 9 , 3 0 ( n l 3 ) , 52
Langendoen, D. T., 8
U s n i k , H., 18
Linebarger, M. C , 5
IJworetzky, B., 5
Lodge, M., 74
ElUs, R., 5
Marks, L. E., 77(nl)
Meijer, G., 4 , 5
Milberg, W. P., 5
Felix, S. W., 56
Mitchum, C. C , 5
Ferreira, F., 5
Fowler, A. E., 5
Franks, J. J . , 50
Nagata,H.,5,6,9,29(n3)
Frazier, L., 21
Newmeyer, F. J . , 5,29(n2), 56
Noma, E., 7 4 , 7 7 ( n l )
179
Pollack, C . R . , 5
Schwartz, M. R , 5
Popper, K., xii

Snow, C , 4 , 5
Sorace, Α., 69
Stevens, S. S., 7 1 , 7 3 , 7 4 , 77(nl)
Robertson, D., 69
Stokes, W., 4
Rosen, Α., 5
Ross, J. R . , 4 , 3 0 ( n l 8 ) , 3 2
Wulfeck, B., 5
Saffran, E. M . , 5
Salasoo, Α., 5
Yip, v., 5
Schachter, J., 5
Zribi-Hertz,A.,30(nl8)
Schiitze, C , 2 , 9 , 2 9 ( n l ) , 29(n5), 29(n6), 67
Subject Index
Acceptability:
By-subjects summaries. See
absolute, 9
By-informants summaries
continuous data and, 1 8 , 4 4 , 7 0
experiment design for, 39-53
relative, 22-26
Categorical variables, 44
scaling of, 68-70

Category scales, 70-72
threshold of, 72
data summaries for, 112-114,115
See also Judgments; Stability, of

respoiwe training for, 90
judgments
statistical tests and, 120-121
Analysis of variance (ANOVA), 83,

Central tendency, measures of, 129-130
133-135,136
See e/so Means
category scale methods and, 120-121

Coding/decoding, of data, 103-110
for informant/token-set data, 122

in Excel, 153-157
sigiuficance and, 1 2 3 , 1 2 5
Cognitive resources, and judgments,
See also Variance

7-10,59
ANOVA. See Analysis of variance

Compensation, of iiiformants, 86-87
Antecedents, and judgment stability,

Competence, compared to
19-22
performance, 7
Arithmetic average. See Means

Computer software:
ASCII fUes, 153

for coding/decoding data, 105-109,
Averages. See Means

153-157
for constructing questiormaires, 90,
96,143-153
Benchmark sentences, 9 2 , 1 1 7
for presentation modes, 64
Between groups variance, 4 3 , 4 4 - 4 5

for scaiming line drawings, 74-75
Binding Theory, and judgment

for summarizing data, 158-161
stability, 19-22,139
spreadsheet, 107-110,112-116
Blocking, 94,98-102
spreadsheet. Excel, 90, %, 109,
in Excel, 146-152
117(n3), 141-161
By-informants summaries. 111, 112-114,

statistical, 126,161
121-122
word processor, 1 0 6 , 1 1 0 , 1 5 2 - 1 5 3
By-items summaries. See By-materials

Concatenation fimction, in Excel, 143
summaries
Confidentiality, 87
By-materials summaries, 111-112,

Context, control of, 51-52
114-116,121-122
Continuity, of acceptability data, 18,44,70
181
Control groups, 15-18,30(nl2), 39

for constructing questioimaires,
Coordinate NPs. See Coordinadon

143-153
experiments
for summarizing data, 158-161
Coordination experiments:
Web sites for, 117(n3), 142-143
factorial designs and, 49-50

Experiment design, 10-11
instructions for, 56-57

blocking and, 94,98-102
judgment stability and, 19-22,139

central problem in, 39-40
Correlation coefficients, 24,135-136

context control in, 51-52
Counterbalancing:
counterbalancing and, 1 3 , 9 3 - 9 4 , 9 5 ,
procedures for, 13, 98

98
rules of, 9 3 - 9 4 , 9 5
factorial, 4 4 , 4 7 - 5 1 , 5 2
Counting function, in Excel, 160

goal of, 47
Customization, of procedures, 85-86

groups in, 3 9 - 4 0 , 4 5 - 4 6 , 8 4
informant instructions in, 5 6 - 5 9 , 7 1 ,
89,90,91,168-171
Data:
meaningful results achieved by, 40
coding/decoding of, 103-110,153-157

presentation modes for, 63-65
summaries of, 111-117,120-122,

response methods in, 67-77
158-161
scaling issues in, 68-77
types of, 18,68-70

sentence materials in, 46-51
Debriefing, of informants, 87
settings and, 75-76,85-88
Decoding. See Coding/decoding of data

variance and, 40-53
Defirute NP experiments:
verbs in, 50-51
judgment stability and, 15-18

within subject, 53
sample size and, 83

See also Error variance;
Demographic data, 9 0 , 1 1 2 , 1 6 8 , 1 6 9
Questionnaires; Sampling
Descriptive statistics, 129-131

Experts, as informants, 81
Design of experiments. See Experiment

compared to naive informants,
design
35-36,59-60
Difference scores, 41. See also Variance

Extralinguistic influences, 4 7 , 5 6 , 5 9 . See
Discrete data, 18
also Judgments, factors affecting
Extraneous systematic variance, 45-51
Electronic format. See Computer
software; Scannable line

Factorial designs, 44, 4 7 - 5 1 , 5 2
drawings; Scannable response

Field settings, 8 8
forms
Filler sentences, 1 3 , 2 4 , 5 1 - 5 2
Error bars, 13-14

data summaries for, 116-117
Errors:
questiormaire construction and,
frequency of, 40
92-93,97-102
in scanning response forms, 106

ratio of, 92-93
in statistical inference, 119-120,132

Forced choice responses, 67
Error variance, 6,31-37

F-tests, 1 2 4 , 1 3 4
extraneous variance and, 4 5 , 4 6
modeling of, 40
response practice and, 92

Generative linguistics, and judgment
within groups variance as, 4 0 , 4 3

methods, 1,4
See also Variance

Geometric means, 1 1 5 - 1 1 6 , 1 3 0 , 1 6 1
Eta-squared (η^) statistic, 123-125,136

Grammars:
Excel software, 9 0 , 9 6 , 1 0 9 , 1 4 1 - 1 4 2
school, 5 6 , 5 7
for coding/decoding data, 153-157

shared, ix-x
Grammaticality, 5, 7, 8
Subject Index 183
response methods and, 67-68 factors affecting, 7-10,4fr47,55-59,85-86

Grammatical theory, and judgments, 2, grammatical theory and, 2 , 7 - 1 2
7-12 importance of, 1-2
Groups, in experiment design, 39-40, objective methods and, ix-x
45-46,84 theory of, in experiments, 10-11
variety of, 55-56,59
See also Error variance; ReUabiUty;
ID information, 98-102,112 Stability, of judgments
in Excel, 1 4 5 , 1 4 8 , 1 4 9 - 1 5 0 , 1 5 2 , 1 5 5
I-language, χ
Indefinite NPs, and judgment stability, Laboratory settings, 7 5 - 7 6 , 8 8
15-18 Levels, of data, 68-70
Inferential statistics, 119-120,131-136 Line drawings, 74-75,110,167-173
Informants: Linguists:
compensation of, 86-87 as informants, 59-60,81
confidentiality of, 87 training of, 59-60,126
data records for, 105-109 Seefl/soExperts
data summarized by. 111, 112-114, Location, statistical, 41. See also Means
121-122
debriefing of, 87
differences among, 46 Machine readable response forms,
factors affecting, 7-10,46-47,55-59, 103-109
85-86 Macro langviage, in Excel, 1 4 2 , 1 5 2
instructions to, 5 6 - 5 9 , 7 1 , 8 9 , 9 0 , 9 1 , Magnihide estimation, 7 3 - 7 4 , 1 1 0
168-171 Master questionnaire, 1 0 2 , 1 5 2
naive vs. expert, 35-36,59-60 Materials, sentence. See By-materials
recruitment of, 86-87,88 summaries; Sentences; Token sets
representativeness of, 79-81 Means, 41,129-130
selection of, 45-46 comparisons Ijetween, 1 1 9 , 1 2 4
single presentation coi«traint on, 50 for data summaries, 115-116,120,
Informativeness, and statistical 159-160,161
significance, 123-125 inExceL 159-160,161
Instability. See Stability, of judgments Median, 130
Instructions, to informants, 56-59,71 Microsoft software. See Excel software;
for scaimable hne drawings, 168-171 Word software
in questionnaire construction, 89,90,9 1 Multiple coders, 110
Intersubjective methods, ix-x. See also Multiple instances, of data collection,
Judgments 32-37,50. See a/so Rehabihty
h\terval data, 6 9 , 7 2 , 1 2 0 - 1 2 1 Multiple scripts, 96,98-102
Intervals, on scales, 70-71
Interval scales, 6 9 , 7 0
Introspection. See Judgments Naive informanb, compared to experts,
Intuitive instructions, 56-57 35-36,59-60. See also Informants
Items, simmiaries by. See By-materials Nominal data, 68
summaries Nominal scales. See Category scales
Object extraction:
factorial designs and, 4 8
Judgment criteria, 91. See also judgment stabihty and, 18-19
Judgments, factors affecting
Judgments:
cogniHve resources and, 7 - 1 0 , 5 9 Objective methods, ix-x
doubts about, 2-7 "Of," sentences ending in, 18
Order, in response scales, 68-69

Random number function, in Excel,
Order, sentence, 6 , 1 3 , 1 5 - 2 3
144,149
blocking and, 94
Random variation, 6 , 4 0 . See also Error
extraneous variance and, 4 6 , 5 1 - 5 2

variance
scripts and, 9 6 , 1 4 8
Ratio data, 69
Ordinal data, 6 8 , 7 2 , 1 2 0
Ratio scales, 6 9 , 9 0
Ordinal scales, 6 8 - 6 9 , 7 0
data summaries for, 1 1 4 , 1 1 5
methods for, 72-76
Recorded presentation mode, 64(table)
Parsing function, in Excel, 1 5 3 , 1 5 4

Records, for data, 105-109. See also Excel
Partitioning, of variance, 41-44

software
Pearson's r, 135-136
Recruitment, of informants, 8 6 - 8 7 , 8 8
Percentage of variability, 123-125,136

Reflexives, and judgment stability, 19-22
Percentage summaries, 120-121,160

Relative acceptability, within token
Percentile summaries, 112-114,115

sets, 22-26
Performance, compared to competence, 7

Reliability, 2 , 9
Pic/«re-NPs, 15-18
as statistical significance, 13,132-133
Pilot testing, 84
error variance and, 32
Populations, and statistical inference,

informant instructiorw and, 56
39,79,131-136
informativeness and, 125
Power, and sample size, 83

nahire of, 28(nl)
Practice, for responses, 9 1 , 9 2 , 1 7 1

sampling and, 79
Prepositions, at end of sentences, 17-18

separate sessions and, 2 2 - 2 6 , 3 2 , 3 5
Prescriptive instructions, 56-57

See also Stability, of judgments
Presentation, modes of, 63-65

Repeatability, and error variance, 32
Procedures:
See e/so Reliability
coimterbalancing and, 1 3 , 9 8
Representativeness:
experiment settings and, 85-88

of informants, 79-81
Proportions:
of sentences, 47
in magnitude estimation, 73-74

Resampling, 8 2 - 8 3 , 8 4 ( n l )
of total variability, 123-125,136

Research design. See Experiment design
Psychology, judgment data in, 3-4,5-6

Respondents. See Informants
Psychophysics, 6
Response forms, scaniuible, 103-109
factorial design for, 49-50

Response methods, 67-77
scaling for, 72-73

Response h'aining, 90-92,170-171
ρ values, 133-136
r statistic, 135-136
statistic, 136
Questionnaires:
assembling of, 102

Sample size, 80,81-84
consh^ction of, 89-102,143-153

for comparing groups, 84
distribution of, 87
grammaticality and, 68
Excel and, 143-153

miiumum reasonable, 83-84
for scaniuble line drawings, 167-17 3

significance and, 1 2 3 , 1 2 4
master, 1 0 2 , 1 5 2
Sampling, 3 9 , 4 0
most important component of, 89

re-, 82-83,84(nl)
number of, 102

representativeness and, 79-81
presentation mode of, 64(table)

variability and, 8 0 , 8 2
Scale inversion, 92
Scaling:
Randomization, in token sets, 9 4 , 9 6 ,

category methods of, 7 0 - 7 2 , 9 0 ,
147-149,152
112-114,115,120-121
Subject Index 185
data levels produced by, 70

Stability, of judgments:
ratio methods of, 6 9 , 7 2 - 7 6 , 9 0 , 1 1 4 ,

as a theory-relative construct, 11-12
115
Binding Theory and, 19-22,139
response training and, 90

experimental findings on, 12-27,
statistical tests and, 120-121

137-140
types of, 68-70

experiment design and, 10-11
Scannable line dravirings, 74-75,167-173

grammatical theory and, 7-12
Scannable response forms, 103-109. See

in linguistics literature, 4 - 5 , 6
also Unscannable responses

in psychology literature, 3 - 4 , 5 - 6
School grammar, and instructions, 56,

in psychophysics literature, 6
57
number of responses and, 6-7
Scripts, 9 6 , 9 8 - 1 0 2
on coordination, 19-22,139
in Excel, 146-150,155-157
on subjacency, 15-18,137-138
Selection, of informants, 45-46

on "that"-trace, 18-19,23-26,138-139
Semirandom orderings, 147,149(table)

over separate sessiotw, 22-26
Sentences:
theory vs. observation of, 11-12
benchmark, 9 2 , 1 1 7
within sentence types, 22-26
design of, in experiments, 46-51

Standard deviation, 130
filler, 1 3 , 2 4 , 5 1 - 5 2
Standard error of the mean, 131
filler, and data summaries, 116-117

Standardization, of procedures, 85-86
filler, and questionnaire

Standard scores, 1 4 , 1 1 4 , 1 3 0 - 1 3 1
conshiiction, 92-93,97-102
in Excel, 160-161
individual, viîthin token sets, 22-26,

Statistical significance, 13,132-133
139-140
theoretical importance and, 122-125
order of, and blocking, 94

Statistical software, 126,161. See also
order of, and extraneous variance,

Computer software
46, 51-52
Statistical tests, 13,132-136
order of, and judgment stability, 6,

error control and, 119
13,15-23
for category scale data, 120-121
order of, and scripts, 9 6 , 1 4 8

integrated, for informant and token
representativeness of, 47
set data, 121-122
types of, and experiment design,

on variation accounted for, 122-125,
47-51
136
types of, and judgment stability, 6,

Statistics:
15-22
as control of error, 120
verbs for, 50-51

descriptive, 129-131
See also Token sets

inferential, 119-120,131-136
Settings, of experiments, 75-76, 85-88

training in, 126
Shared grammars, ix-x

See also Experiment design
Significance, statistical, 13,132-133

Subjacency:
theoretical importance and, 122-125

judgment stability and, 15-18,
Single presentation constraint, 50

137-138
Sorting, 107-108
sample size and, 83
in Excel, 150-152,155-157
Subject extraction:
Specified subjects, and judgment

error variance and, 33
stability, 15-18
factorial designs and, 48
Spoken presentation mode, 64(table)

judgment stability and, 1 8 - 1 9 , 2 6
Spreadsheets:
sample size and, 82-83
Excel, 9 0 , 9 6 , 1 0 9 , 1 1 7 ( n 3 ) , 141-161
Subjectivity, ix-x
for coding/decoding data, 107-110

Subjects. See Informants
for summarizing data, 112-116,

Summaries, of data, 111-117,120-122
117(n3)
in Excel, 158-161
Survey experiments, procedures for, significance and, 1 2 3 , 1 2 5
86-87
Types I and Π errors, 132
Syntactidans, trairung of, 59-60
Synthesized speech, for presentation
mode, 64(table)
Unacceptability. See Acceptability
Systematic variance, extraneous, 45-51

Unscarmable responses, 110
Technology. See Computer softvâre

Variability;
Text editors, 1 0 6 , 1 1 0
percent accounted for, 123-125,136
Text files, 153
sampling and, 8 0 , 8 2
"Thaf'-trace experiments:
See also Stability, of judgments
error variance and, 32-35
Variables, 44. See also Data
Variance:
judgment stability and, 18-19,23-26,

between groups, 4 3 , 4 4 - 4 5
138-139
controlling of, 40-53
sample size and, 82-83

definition of, 41
token set data from, 163-165

experiment design and, 40-53
Threshold, of acceptability, 72
extraneous systematic, 45-51
Time, devoted to procedures, 87

Timed Usks, 88
partitioning of, 41-44
"To," sentences ending in, 18

total, 4 1 , 4 3
Token sets:
within group, 4 0 , 4 3 , 4 5
blocking in, 94,98-102

See also Analysis of variance; Error
counterbalancing procedures for, 13,

variance
98
Veits, in experiment design, 50-51
counterbalancing rules for, 9 3 - 9 4 , 9 5

Visual Basic macro language, 1 4 2 , 1 5 2
data from, example of, 163-165
individual sentences in, 22-26,
139-140
Warm-up. See Response training
in Excel, 143-152
Within group variance, 40, 4 3 , 4 5
number of, 92
Within subject experiment design, 53
questionnaire construction and,

Wizard function, in Excel, 154
92-96,97-102
Word processors:
randomization in, 9 4 , 9 6 , 1 4 7 - 1 4 9 , 1 5 2
for coding/decoding data, 1 0 6 , 1 1 0 ,
relative acceptability within, 22-26

153
single presentation constraint on, 50

for questionnaire construction,
summaries by. See By-materials

152-153
summaries
Word software, 152-153
Total variance, 4 1 , 4 3
Worid Wide Web site, 117(n3), 142-143
Training:
Written presentation mode, 63,
for responses, 90-92,170-171

64(table). See also Questiormaires
for statistics, 126
of linguists, 59-60,126
Treatment, for control groups, 39

Zero point, 69
f-tests, 84(n2), 1 2 2 , 1 3 3
z-scores, 1 4 , 1 1 4 , 1 3 0 - 1 3 1
About the Author
Wayne Cowart is Associate Professor of Linguistics at the University

of Southern Maine, where he is also Director of the Core Curriculum.
He holds a Ph.D. m linguistics from the City University of New York
(1983). His research interests include various aspects of the psycho
logy and biology of language and philosophy of mind. His work has
appeared in Language, Memory, and Cognition, the Journal of Psycho-
linguistic Research, Cognition, Perceptual, and Motor Skills, and various
anthologies. Funding for his work has been provided by the National
Science Foimdation and the National Institutes of Health.
187
CPSIAinformationcanbeobtainedatwww.ICGtesting.com I I Ml I II I I I
Printed in the USA III I I I I I Nil I I II
241716LV00004B/13/A 9 780761 900436

Experimental Syntax

Uploaded by

Copyright:

Available Formats

Experimental Syntax

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Experimental Syntax

Uploaded by

Copyright:

Available Formats

:perinienta

Applying Objective Methock

For information address:

SAGE Publications, Inc.

2455 Teller Road

Thousand Oaks. California 91320

SAGE Publications Ltd

London ECIY ISP

SAGE Publications India Pvt Ltd

B-42, Panchsheel Enclave

New Delhi 110 017

Printed in the United States of America

Library of Congress Cataloging-in-Publication Data

Includes bibliographical references and index.

ISBN 0-7619-0042-X (alk. paper).—ISBN 0-7619-0043-8 (pbk.:

Methodology. 3. Acceptability (Linguistics) I. Title.

Acquiring Editor: Alex Schwartz

1. Introduction: Are Judgments Stable? 1

1.4. Stability of Responses to Individual Sentences 22

1.5. Outline of the Book 27

2. Error Variance in Sentence Judgments 31

3. Designing Experiments on Acceptability 39

3.4. Controlling Context 51

4. The Sentence Judgment Task 55

5. Presenting Sentence Materials to Informants 63

6. Response Methods and Scaling Issues 67

8. Settings for Experiments 85

9. The Organization and Construction of

9.2. Response Training and Practice 90

10. Coding and Decoding the Data 103

Scanned Data 105

11. Summarizing the Data 111

12. StatisHcal Issues 119

12.2. Testing Generalizations Across Informants and

Token Sets 121

Appendbc A: A Reader's Guide to Statistics 129

Appendix B: Statistical Supplement to Chapter 1 137

B.l. Subjacency 137

Appendbc C: Excel as a Syntactician's Workbench 141

C.l. Accessing Software Tools and Sample Files 142

C.2. Building Questionnaires (Chapter 9) 143

C.3. Coding/Decoding Data (Chapter 10) 153

C.4. Summarizing Data (Chapter 11) 158

Appendbc D: Token Set Data From a "That"-Trace

Appendix E: Sample Questionnaire for Scannable

Author Index 179

Subject Index 181

About the Author 187

This book provides an introduction to objective methods by which

and execute meaningful experiments on judgments of acceptability.

grammar that they hold, more or less, in common. In terms of

tions from many sources. My greatest debt is to my colleague Dana

Web Site for This Book

Some Excel spreadsheets and other tutorial

So far as research practice is concerned, one of the most striking

that could never have been acquired without reliance on judgment

Before launching into a practical explication of the method,

1.1 Doubts About the Stability of Judgments

Linguists were not the first to use introspective judgments in a

A number of investigators have found evidence that individual

Sentences such as (lb), (Ic), and (Id) appear to involve extrac

The final set of preliminary observations concerns an interac

The reliable differences among judgments of sentence types de

undergraduates at Wayne State College (NE), the University of Ala

1. In the context of psychometrics especially, psychologists are often con

8. Even if linguists have wildly misconstrued the origin of judgments (sup