Probation

NBER WORKING PAPER SERIES
ABILITY, GENDER, AND PERFORMANCE STANDARDS:

EVIDENCE FROM ACADEMIC PROBATION
Jason M. Lindo
Nicholas J. Sanders
Philip Oreopoulos
Working Paper 14261

http://www.nber.org/papers/w14261
NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
August 2008
We thank Chris Ayres, Scott Carrell, Doug Miller, Marianne Page, Ann Huff Stevens, and seminar
participants at the UC Davis Applied Microeconomics Lunch Series, the 2008 WEAI Conference,
and the University of British Columbia for their helpful comments and suggestions at various stages
of this paper. The views expressed herein are those of the author(s) and do not necessarily reflect the
views of the National Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been peer-
reviewed or been subject to the review by the NBER Board of Directors that accompanies official
NBER publications.
© 2008 by Jason M. Lindo, Nicholas J. Sanders, and Philip Oreopoulos. All rights reserved. Short
sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided
that full credit, including © notice, is given to the source.
Ability, Gender, and Performance Standards: Evidence from Academic Probation
Jason M. Lindo, Nicholas J. Sanders, and Philip Oreopoulos
NBER Working Paper No. 14261
August 2008
JEL No. D80,I20
ABSTRACT
We use a regression discontinuity design to examine students' responses to the negative incentive brought
on by being placed on academic probation. Consistent with a model of introducing performance standards
in which agents respond differently based on ability, we find that being placed on probation at the
end of the first year discourages some students from returning to school while improving the performance
of those who return. Contrary to the predictions of the model when ability is known, we find that heterogeneous
discouragement effects result in high ability students having a greater overall dropout rate near the
cutoff than lower ability students. The result can be explained by extending the model to allow for
the performance standard to also affect self confidence (ability expectations). We also consider effects
by gender and find that being placed on probation more than doubles the probability that men drop
out but has no such discouragement effect for women.
Jason M. Lindo Philip Oreopoulos

Department of Economics Department of Economics
University of California, Davis University of British Columbia
One Shields Avenue Canada
Davis, CA 95616 and NBER
jlindo@ucdavis.edu philip.oreopoulos@ubc.ca
Nicholas J. Sanders
Department of Economics
University of California, Davis
One Shields Avenue
Davis, CA 95616
njsanders@ucdavis.edu
1 Introduction
Academic probation is a nearly universal tool used by universities to ensure currently enrolled
students achieve minimum academic standards. The general structure of such programs is
simple—if a student’s grade point average (GPA) is below a certain threshold, the student
is placed on academic probation which serves as a wake-up call and can lead to escalat-
ing penalties. Our paper uses longitudinal data from three campuses at a large Canadian
university to estimate the causal impact of being placed on academic probation by exploit-
ing the discontinuous nature of the policy in a regression discontinuity (RD) design. More
generally, our paper serves as an analysis of how individuals respond to a threat of punish-
ment, as failure to improve one’s grades after being placed on probation leads to suspension.
Due to concerns about welfare, experiments providing such a large negative incentive are
not often viable. Further, examining students’ responses to academic probation provides a
rare opportunity to examine the impacts of a negative incentive in an important real world
setting.
Despite the prevalence of academic probation, the existing literature on its impact on
students is limited to a single paper which compares the mean retention rate of engineering
students on probation to the mean retention rate of those in good academic standing (Scalise,
Besterfield-Sacre, Shuman, and Wolfe 2000). While there is a fairly extensive literature on
the causal e↵ects of other explicit university policies, these studies have focused primarily
on positive incentives and services.1 We o↵er the first study examining the causal e↵ects of
a negative incentive program.
Once a student has been placed on academic probation, she must earn a GPA above the
Examples include papers analyzing the e↵ects of advising and merit scholarships (Angrist, Lang, and
1
Oreopoulos 2007; Leuven, Oosterbeek, and van der Klaauw 2006), tuition costs (Hakkinen and Uusitalo
2003; Heineck, Kifmann, and Lorenz 2006; Garibaldi, Giavazzi, Ichino, and Rettore 2007), financial aid
(DesJardins, Ahlburg, and McCall 2002), and appointment to the Dean’s List (Seaver and Quarton 1976;
revisited by Cook and Campbell 1979).
2
campus-set standard in the next term or she will be suspended from the university for one
year. As such, placing students on academic probation is equivalent to setting a minimum
standard for their future performance. In this sense, our results have implications for the
wide variety of circumstances in which setting a standard might be used as a means of af-
fecting performance. In addition to school administration, applications include management,
parenting, and health and safety regulation.
Our results also provide an opportunity to empirically examine theoretical models of
imposing performance standards. Bénabou and Tirole (2000) outline one such model that
predicts that setting a performance standard involves an inherent tradeo↵ between motivat-
ing some agents to improve their performance and discouraging other agents from making
any attempt at all. The model further predicts that a performance standard will be most dis-
couraging for agents of relatively low ability who do not believe they can meet the standard.
We explore the extent of this tradeo↵ by analyzing the e↵ect of being placed on academic
probation on students’ decision to drop out and on the subsequent performance for those
who remain, using students’ high school grades as a measure of ability.
This paper also contributes to the growing literature on gender di↵erences in response to
educational incentives. Previous studies have found that women are more responsive to pos-
itive incentives than men: women respond to advising and scholarship programs while men
do not (Angrist, Lang, and Oreopoulos 2007); tuition reductions impact college completion
rates for women more than men (Dynarski 2005); and the e↵ects of high school achievement
awards appear limited to women (Angrist and Lavy 2002). However, because all of these
papers focus on policies providing positive incentives, little is known about gender di↵erences
in response to negative incentives.
Our RD research design is motivated by the idea that students with a GPA just above
the academic probation threshold after the first year provide a good counterfactual for those
who have a GPA just below. As long as characteristics related to student outcomes are
3
continuous through the threshold, we can measure the e↵ect of being placed on probation in
a RD design framework.
Our results indicate that the e↵ects of academic probation are remarkably heterogeneous.
Consistent with Bénabou and Tirole’s model, being placed on academic probation after the
first year discourages some students from returning to school and motivates those who remain
to improve their subsequent performance. Contrary to the predictions of the model, we find
that academic probation results in students of higher relative ability dropping out at a greater
rate than students of lower relative ability. We discuss an extension of their model that can
explain our results by allowing academic probation to have heterogeneous impacts on self-
confidence across ability. We also find that the impact on students’ decisions to drop out of
school di↵ers by gender and native language. Being placed on academic probation doubles
the probability that men drop out but has no such e↵ect on women. Similarly, it discourages
English speakers from returning but has no such e↵ect non-native English speakers.
Finally, we consider the e↵ects of being placed on academic probation on graduation rates.
Because being placed on probation simultaneously motivates some students to improve their
grades while causing others to drop out, there is no clear prediction of a positive or a negative
e↵ect on the probability of graduation. While we find insignificant e↵ects overall, we find a
strong negative e↵ect for students of higher ability.
The rest of the paper is organized as follows. Section 2 describes the university and its
academic probation program in more detail. Section 3 reviews Bénabou and Tirole’s (2000)
model of performance standards and considers an extension. Sections 4 and 5 describe the
data and our empirical strategy. Section 6 presents our main results. Section 7 discusses our
main results. Section 8 concludes.
4
2 Institutional Background
Our data comes from a large Canadian university made up of three individual campuses—
one central campus and two smaller satellite campuses. The central campus (Campus 1) has
an acceptance rate of about 55% while the two satellite campuses (Campus 2 and Campus
3) have acceptance rates of approximately 77%. The central campus resembles a large
US state college while the satellite campuses have more part-time and commuter students.
Campus 1 and Campus 2 share identical rules regarding academic probation—students with
a cumulative GPA below 1.5 grade points are placed on academic probation. Campus 3
has a GPA cuto↵ at 1.6 grade points. For the purposes of our analysis, students from all
three campuses have been combined into a single sample.2 We account for the di↵erence in
cuto↵ points in our RD analysis by using students’ distance from their campus’ cuto↵ as the
running variable rather than absolute GPA.
When a student is placed on academic probation, a letter is sent notifying them of
their current academic standing. The letter specifies why the student has been placed on
probation, how to regain good academic standing, and the consequences of failing to improve.
The letter also encourages students to continue at the university and improve their academic
performance, and lists various services provided by the university aimed at helping them to
do so. A copy of the letter is included in the appendix.
Because many freshman classes span the entire year, students’ academic standing is not
evaluated for the first time until the end of their first year.3 At Campus 1 and Campus
2, students’ academic standings are evaluated again at the end of every subsequent full
2
In an earlier version of the paper, we used this across-campus variation to examine the e↵ects of aca-
demic probation in a di↵erence-in-di↵erence framework. These results were similar to the RD estimates but
imprecise due to a smaller sample size.
3
Students also must attempt a minimum number of credits before they are evaluated. We omit all students
who have not yet been evaluated by the end of their first full year. It is possible this is endogenous, though
unlikely. If a student drops a course before the deadline, the course grade will not be counted in determining
academic probation (or GPA in general). If this incentive to drop is continuous through the cuto↵, however,
this should not pose a problem.
5
scholastic year and summer term. At Campus 3, students’ academic standings are evaluated
again at the end of each subsequent term.
Students on academic probation face the threat of suspension after subsequent sessions
if their grades do not improve. At all campuses, students on probation can avoid suspension
and return to good academic standing by bringing their cumulative GPA up to the cuto↵.
They can also avoid suspension after the next evaluation by achieving a per-session GPA
above a particular minimum specified by their campus.4 Students who avoid suspension but
fail to improve their cumulative GPA above the cuto↵ remain on academic probation and
continue to face the threat of suspension until they regain good academic standing.
Students failing to sufficiently improve their grades are suspended for one full academic
year. If suspended students choose to return to the university and again fail to sufficiently
improve their grades, they can be suspended for three years. A third failure to meet the
GPA requirement can lead to permanent suspension from all campuses.
3 Theoretical Background
In this section, we first review Bénabou and Tirole’s (2000) model of agents’ responses to a
performance standard. The model is framed as a game between a principal and an agent,
where the principal has the ability to set standards for the agent. While Bénabou and Tirole
consider the model from the perspective of both the agent and the principal, we focus only
on the agent. We then relate this model to academic probation and consider an extension.
Consider an agent facing a choice between three possible paths: option 1, option 2, or
neither. If the agent chooses neither option, both her costs and benefits are 0. If the agent
attempts option i, she incurs cost ci and, if successful, gains the benefit Vi . If the agent
attempts option 1 or 2 and fails, she still incurs a cost but receives no benefit. Option 1 is
4
At Campuses 1 and 2, the GPA cuto↵ is 1.7. At Campus 3, it is 1.6.
6
an easy option with a low potential benefit while option 2 is a difficult option with a high
potential benefit. Costs and benefits can be summarized as
0 < c1 < c2 0 < V1 < V2 (1)
Ability is expressed as the probability of successfully completing either option, where

higher ability translates into a higher probability of success. The probability of success for
either option is ✓. Assuming that the agent is perfectly informed and knows ✓, she solves
max{0, ✓V1 c1 , ✓V2 c2 }. (2)
Let ✓ be the level of ability for which the agent is indi↵erent between attempting neither
option and attempting option 1 and ✓ be the level of ability for which the agent is indi↵erent
between pursuing option 1 and pursuing option 2. With the following assumption,
c1 c2 c1
✓⌘ <✓⌘ < 1, (3)
V1 V2 V1
which ensures that both options are optimal for at least some ✓, it can be shown that
the lowest ability individuals (✓ < ✓) choose neither option, the highest ability individuals
(✓ < ✓) choose the difficult option, and the remaining individuals (✓ < ✓ < ✓) choose the
5
easier option.
If the principal removes option 1 as a possible course of action, perhaps by forbidding it
or imposing an additional cost on it such that it is always inferior to other options, then the
5
Interestingly, the same results arise from a model in which the agent is sure to be successful in whichever
task she chooses but the cost of each task is inversely related to the ability measure ✓. That is, the agent would
solve max{0, V1 c✓1 , V2 c✓2 }. Expressed in this way, the model can feature decreasing marginal returns to
ability. Analogously, the same results can arise from a model in which the probabilitypof success is p increasing
in ability and the cost is decreasing in ability. For example, if the agent solves max{0, ✓V1 p c1
✓
, ✓V2 p c2
✓
}.
7
agent will choose option 2 if and only if
c2
✓ ⌘ ✓? (4)
V2
and pursue neither option otherwise. Within the range of agents who would choose option 1
if it remained a possibility, the model predicts those with the higher ability will work harder
(engaging in option 2) while those with the lower ability will give up (pursuing neither
option).6
This model naturally lends itself to analyzing how students might respond to being placed
on academic probation. Consider the choice faced by two students (agents) whose first year
GPAs were near the academic probation threshold, one just above and one just below. Since
the student just above the threshold remains in “good academic standing,” their options are
unrestricted. These options can be placed into three categories: return to school with the
intent of achieving some low GPA (option 1), return with the intent of achieving some high
GPA (option 2), or dropping out of school (neither option 1 nor 2).
As a result of being placed on probation, the student just below the cuto↵ faces a di↵erent
set of choices. We can think of academic probation as the administration forbidding, or
placing an extremely large negative incentive on, pursuing option 1. In the framework of
the model, if the student below the cuto↵ chooses to pursue option 1, she will be suspended
from the university.
The testable implications of Bénabou and Tirole’s theoretical framework are the following:
1. Low ability students should be more likely to drop out, whether option 1 is forbidden
or not.
2. Forbidding option 1 will increase the overall probability of students dropping out and
6
Specifically, those with ✓ in [✓? , ✓] will now choose option 2 while those with ✓ in [✓, ✓? ] will now choose
not to pursue either option.
8
increase the performance of those who return.
3. Forbidding option 1 will cause relatively low ability students to drop out and relatively
high ability students to return and work harder.
4 Data
The data used in the analysis is from a administrative data set of college students from the
large Canadian university described in Section 2. Observations are at the student level and
cover a ten-year period from 1995 to 2005 with each scholastic year broken into Fall, Winter,
and Summer terms. The data includes student term registration status, GPA, gender, age,
first language, and a measure of high school performance. We use the set of background
variables for falsification tests and as controls in some regression specifications.
We restrict the sample to students we can potentially observe for two years. Since the
data spans through the end of the 2005 school year, we restrict to students who entered in
the 2004 school year or earlier. This leaves nine cohorts of students. We omit students with
missing data for any variables of interest.7 We also restrict the sample to students entering
the university between the ages of seventeen and twenty-one (99% of the remaining sample).
Additionally, we keep only students who have had their academic standing evaluated at
the end of their first year (98% of the remaining sample).8 Finally, we limit the sample to
students those within a bandwidth of 1.4 grade points of their academic probation cuto↵.
This e↵ectively drops students who failed all of their first year classes in addition to students
who cleared the cuto↵ by a wide margin.9 The resulting sample includes 32,697 students.
7
The largest impact this has on the sample is due to omitting students with missing high school grade
information which reduces the sample by 16%. The primary reason high school grades are missing for
students is that they attended high school outside the province.
8
In almost all cases, a student is not evaluated due to an insufficient number of attempted credits.
9
As a robustness check, some regressions have further bandwidth restrictions. This is discussed in the
following section.
9
The first column of Table 1 shows the descriptive statistics for the sample. The students
average entry age is 18.7 years. Approximately 38% are male, 73% have English as their
first language, and 88% were born in North America. 53% of the students attend Campus 1
(the central campus).10 Of the remaining students, 20% attend Campus 2 and 27% attend
Campus 3. For the most part the means at the limits of the cuto↵ are most relevant and
these are discussed in subsequent sections. Before proceeding, we should note that a sizable
fraction of students (23%) are placed on academic probation after their first year. While it is
unfortunate that so many students fail to meet their campus’ academic requirements, this is
a positive aspect for our research design because it improves our ability to obtain estimates
on both sides of the probationary cuto↵.
5 Empirical Strategy
We begin by estimating the impacts of being placed on academic probation after the first
year. At the end of the first year the probation status for student i at campus c is a
deterministic function of their GPA, which can be expressed as
P ROBicyear 1 = 1(GP Ayear

ic
1
< GP ACU Tc ) (5)
where GP Ayear
ic
1
is student i’s cumulative GPA at the end of the first year and GP ACU Tc
is the probationary cuto↵ at the student’s campus. Because the discontinuity in probation
status is “sharp,” as we will show in Section 6.1, as long as other student characteristics
related to the outcomes are continuous through the threshold, the treatment e↵ect for stu-
dents near the threshold can be obtained by comparing the outcomes of students just below
the threshold to those just above.
10
If a student’s first language is not English, it is usually an Asian language. Despite being a Canadian
institution, less than one percent of students have French as their first language.
10
The following equation can be used to estimate the impact of being placed on academic
probation after the first year on student outcomes:
Yic = m(GP Ayear

ic
1
) + 1(GP Ayear
ic
1
< GP ACU Tc ) + uic , (6)
where Yic is an outcome for student i at campus c, m(GP Ayear

ic
1
) is a continuous function
of students’ standardized first year GPAs (the distance from their campus’ probationary
cuto↵s), 1(GP Ayear
ic
1
< GP ACU Tc ) is an indicator equal to one if the student’s GPA is
below the probationary cuto↵, and uic is a random error term. The coefficient of interest is
, the estimated impact of being placed on academic probation after the first year. Since
GPA data are in hundredths of a grade point, we face a running variable with a discrete
support. Following Lee and Card (2008), we estimate the previous equation by assuming a
low-order polynomial functional form for m() and cluster our standard errors on students’
normalized first year GPAs. Specifically, m() is a second order polynomial that is flexible on
each side of the cuto↵.11 As a robustness check we also estimate the e↵ects controlling for
observables.
As noted by Imbens and Lemieux (2008), the main concern with such an approach is that
the results may be sensitive to outcome values for observations far away from the cuto↵. For
this reason, our main estimates use only data within a bandwidth of 1.4 from the cuto↵,
dropping individuals who fail all (or nearly all) of their first year classes and those who
11
Our regression equation is given by:
year 1 year 1
Yic = ↵ + le 1(GP Aic < GP ACU Tc ) + (GP AN ORMic )
year 1
+ 1 (GP AN ORMic ) ⇥ 1(GP Ayear
ic
1
< GP ACU Tc )
year 1 2
+ 2 (GP AN ORMic ) ⇥ 1(GP Ayear
ic
1
< GP ACU Tc )
+ uic
where le is the estimated impact of being placed on academic probation after the first year. Our figures
suggest this functional form is a good fit. In prior versions of this paper, we found similar results using
higher order polynomials.
11
cleared the cuto↵ by more than 1.4 grade points. We also show estimates using bandwidths
of 1.0 and 0.6 to verify that our results are not sensitive to choice of bandwidth. While we
present results using rectangular kernel weights, we have also verified that the results are
robust to performing local linear regression with triangular kernel weights.
We use the strategy described thus far to estimate the impacts of being placed on aca-
demic probation after the first year. However, we are also interested in analyzing the impacts
of ever being placed on academic probation. The approach described above does not produce
such an estimate because “control” students just above the cuto↵ after their first year may
be placed on probation after later evaluations. As a result, we must use a slightly di↵erent
approach.
To estimate the impact of ever being placed on academic probation, we rely on the fact
that students’ first year GPAs create a discontinuity in whether or not a student is ever
placed on academic probation (as shown in Section 6.1). Since this discontinuity is fuzzy
(the change in the probability of treatment is less than one), the impact on student out-
comes can be obtained by adjusting estimated discontinuities in outcomes by the estimated
discontinuity in treatment status. This is equivalent to a two-stage least squares estimate
in which the variation in treatment status is driven by the discrete change occurring at the
threshold (Hahn, Todd, and Van der Klaauw 2001). We obtain estimates and their stan-
dard errors based on two-stage least squares regression in which the excluded instrument is
1(GP Ayear
ic
1
< GP ACU Tc ).12
6 Results
In this section, we begin by testing the validity of the regression discontinuity design. We then
examine the e↵ects of being placed on academic probation on students’ decisions to drop out,
12
These estimates take the same approach to bandwidth choice, kernel weights, and standard errors as the
other estimates.
12
subsequent GPAs, probabilities of suspension, and graduation rates and the heterogeneity
of these e↵ects across student ability, gender, and native language.
6.1 Tests of the Validity of the RD Approach
Non-random sorting across the cuto↵ is a main concern with any RD design in which the
threshold could be known to those who might be a↵ected by the policy. In our case, this
would appear if students just below the cuto↵ were actively influencing their GPAs to avoid
probation (e.g. convincing teachers to give them a higher grade to raise their GPA above the
cuto↵ point) or expending just enough e↵ort to get grades above the cuto↵. By focusing on
academic probation status at the end of first year, we decrease the likelihood of this concern.
First year students are less familiar with campus policies and, thus, less likely to know what
grades would be required to avoid academic probation. In addition, though Campus 3 has
semester-length first year courses, at Campuses 1 and 2 most first year courses span the
entire year and the majority of the grade is based on evaluation at the end of the term,
making it difficult for students to correctly “hit” a performance point just above the cuto↵,
especially given that their overall GPA is calculated over several courses.
If sorting were a problem, we would expect to see a discontinuity in the distribution of
grades at the cuto↵, as a disproportionate number of students would fall just above the cuto↵
relative to the number of students just below the cuto↵.13 Figure 1 shows the distribution
of students’ first year grades relative to their campus cuto↵s, with cell sizes of 0.05 grade
points. Using each of these cells as an observation, the figure also shows the predicted cell
sizes based on local linear regressions using triangular kernel weights and a bandwidth of
0.6. The estimated discontinuity at the threshold is not statistically significant, indicating
13
At the same time, this might not be the case if the distribution of grades is fixed. For example, there
might be a situation in which some students manipulate their grades to get themselves above the cuto↵ but,
as a result of their e↵orts, other marginal students are pushed below the cuto↵.
13
that the distribution of students is continuous through the threshold.14 15
Our research design requires both observable and unobservable characteristics related to
student outcomes to be continuous through the threshold. Table 2 and Figure 2 explore the
extent to which a wide range of observable characteristics are continuous through the cuto↵.
Significant discontinuities would indicate that students with particular characteristics are
more or less able to manipulate their grades so as to avoid being placed on probation. As
a whole, these estimates support the validity of our research design. We find no significant
discontinuities in students’ high school grades, credits attempted in the first year at the
university, age at entry, gender, birthplace, or campus attended. The only variable for which
there is some evidence of a discontinuity is whether or not English is a student’s first language
and the estimate is only statistically significant in some specifications.16
The graphs in Figure 2 are also useful because the slopes of the curves inform us about
the relationship between students’ characteristics and their performances in their first year
of college. Most relevant to the results that follow, the strong positive relationship between
high school grades and college grades indicates that high school grades are a good predictor
of a student’s success in college. Within our sample, a one-percentile increase in the high
school grade measure is associated with a 0.1 increase in first year GPA. This motivates
our use of the label “high ability” in reference to students with high school grades above
the sample median and “low ability” for those with grades below the sample median.17 In
Section 7, we will return to this issue and further consider whether high school grades are a
good measure of ability conditional on students’ first year grades at the university.
14
We have also tested for discontinuities in the distribution using cell sizes of 0.01, 0.02, and 0.04, grade
points and with varying bandwidths. These results were never statistically significant.
15
This test is similar to that proposed by McCrary (2008).
16
As in Figure 1, we have also verified that there are not significant discontinuities in the distributions of
students for whom English is their first language and students for whom it is not. Similarly, we do not find
significant discontinuities in the distributions of students for any of the other subgroups we consider in our
analysis.
17
Note that the sample of low ability students near the academic probation threshold will be greater than
the sample of high ability students near the threshold since the median is based on all entering students.
14
6.2 First Year GPAs and Academic Probation
Figure 3 and the first panel of Table 3 show the estimated discontinuity in probation status at
the end of the first year. Because the discontinuity is sharp, discontinuities in other student
outcomes can be interpreted as the causal e↵ect of being placed on probation at the end of
the first year. Such estimates are equivalent to the following thought experiment: what if
students’ academic standings were not evaluated after the first year and, instead, were first
evaluated after the completion of their next term?
In looking at the e↵ects on long-term measures of achievement, we are also interested in
the e↵ect of ever being placed on academic probation. This question is distinct from the
former because students just above the cuto↵ after their first year are placed on probation
if their cumulative GPA falls below the cuto↵ in the future. As shown in Figure 4 and the
second panel of Table 3, 25 to 30% of those just above the cuto↵ at the end of their first
year are placed on probation at some point in their academic careers. This implies that a
substantial fuzzy discontinuity exists and we can estimate the impact of ever being placed on
probation on long-run measures of student achievement using standard techniques associated
with fuzzy regression discontinuity designs (as discussed in Section 5).
6.3 The Immediate Response to Academic Probation
The immediate question that all students face at the end of their first year is whether or
not to continue at the university.18 Students who have been placed on academic probation
have been informed that they will be suspended if their GPA does not meet the campus-set
standard in their next term. However, at this point in time, continued enrollment is not
impeded in any other way.
Table 4 and Figure 5 show the estimated impact on students’ decisions to permanently
18
Manski (1989) and Altonji (1993) develop extensive models that explore the sequential nature of the
schooling decision under uncertainty.
15
leave the university after their first year. The estimates, which are statistically significant at
the one percent level in all specifications, indicate that being placed on academic probation
at the end of the first year increases the probability that a student leaves the university by
2.2 to 2.7 percentage points, or by more than 50% of the control mean.19
Table 5 and Figure 6 explore the extent to which di↵erent subgroups of students respond
to being placed on probation after the first year in di↵erent ways.20 These results suggest
that the average e↵ects shown in Table 4 and Figure 5 mask substantial heterogeneity.
The estimated impact on the “low ability” group (those with high school grades below
the median of students entering the university), significant at the ten percent level in all
specifications, suggests that being placed on probation increases the probability of dropping
out by 1.6 percentage points, or approximately a third. We find a much greater e↵ect
on students of higher ability (high school grades above the median of students entering
the university), for whom being placed on academic probation doubles the probability of
dropping out. In fact, the impact of academic probation is so great for high ability students
near the cuto↵ that they are more likely to drop out than their lower ability counterparts.
This is not the case for students just above the cuto↵.
We also find heterogeneous e↵ects across gender and native language. Our estimates
indicate that being placed on academic probation more than doubles the probability that
men drop out after the first year but has no e↵ect on women’s decision to continue at the
university.21 It is unlikely that this finding reflects a general unresponsive nature of females
at the university we study—a prior study at the same university found that women were
19
It should be noted that the control mean is not the overall drop-out rate but, rather, the drop-out rate
for students just above the cuto↵.
20
The table is limited to estimates using a bandwidth of 1.4—in general, the estimates are not sensitive
to the choice of bandwidth.
21
The estimated impacts on subgroups defined by the interaction between ability and gender indicate that
the prior results by ability are mostly driven by males. The estimated impacts on both high and low ability
females are insignificant. Being placed on probation approximately doubles the probability that low ability
men drop out and triples the probability that high ability men drop out.
16
more responsive to experimentally o↵ered peer advising and organized study group services
(Angrist, Lang, and Oreopoulos 2007). In results not shown, we find this gender di↵erence
does not appear to be driven by di↵erences in the types of courses that men and women
attempt in their first year.22 The results that follow indicate that it is not that women do
not respond to being placed on probation but that they respond di↵erently. Lastly, we find
that probation increases the probability of dropping out for native English speakers but that
there is no impact on non-native English speakers.
6.4 The Impact on Subsequent GPA
Students on academic probation who choose to remain at the university have a substantial
incentive to improve their grades to the campus-set standard—failing to do so results in being
suspended from the university for a year. The first panel of Table 6 and Figure 7 show the
impact of being placed on academic probation on students’ subsequent GPAs. Specifically,
the outcome variable is a student’s GPA in the next session in which they are evaluated.23
The estimates suggest that being placed on academic probation causes students to improve
their GPAs by approximately 0.2 grade points, a 66% greater improvement than students
who just surpass the probationary cuto↵.
It is important to note that these estimates might be biased by the e↵ect of academic
probation on the composition of students who continue to enroll in the university. For
22
Specifically, we have separately considered the impact for students who took first year courses that
disproportionately enrolled men and for students who took first year courses that disproportionately enrolled
women. Regardless of the gender composition of the peers in their first year courses, we find no evidence that
being placed on academic probation impacts the decision to drop out for women and we do find evidence
that it impacts men. We have also analyzed the impact for students who took at least half of their first year
credits in science courses versus those who did not; similar gender di↵erences also appear in these results.
23
Note that the “next session” will be a summer session if students are enrolled in summer classes and a
second year session if they are not. If we were to use only second year GPAs, there would be missing data
for students placed on academic probation after their first year who do poorly in summer school and are
suspended for their second year. While summer school grades tend to be higher than grades during the rest
of the year, in results not shown but available upon request, we find no statistically significant impact of
academic probation on the probability that a student takes summer courses.
17
example, if academic probation results in the attrition of relatively low ability students,
then we would expect the estimated impact on GPAs to be positive even if being placed on
probation has no e↵ect on individual behavior beyond the choice to drop out. However, in
the previous section we showed that it is actually the higher ability students who are most
discouraged from returning to the university which suggests that our estimates represent a
lower bound of the true e↵ect on GPAs. As an additional check, the second and third panels
of 6 show that the impact on GPAs remains positive and significant under very conservative
assumptions about what GPAs would have been observed for students who did not return
for another session after being placed on probation.24
Table 7 and Figure 8 show the estimated impacts on students’ subsequent GPAs sepa-
rately for low ability students, high ability students, men, women, native English speakers,
and non-native English speakers. All of the estimates suggest that being placed on academic
probation after the first year improves students’ subsequent GPAs.25 Since we found no e↵ect
on the continued enrollment of women and non-native English speakers, the estimated im-
pact of academic probation on subsequent GPAs is unlikely to be driven by composition bias
for these groups. Thus, at least for these groups of students, we can be relatively confident
that the impact on students’ GPAs is truly positive.
6.5 The Impacts on Suspension and Graduation
The previous sections analyzed the impact of being placed on academic probation after the
first year on students’ outcomes immediately following their first evaluation. When those
outcomes are measured, students are either currently on probation or in good standing. In
24
Using the same RD approach, we have also looked at the difficulty of the classes that returning students
take using average course grades as a proxy. There is no evidence that being placed on probation causes
students to take courses in which grades tend to be higher.
25
We cannot rule out that the e↵ects are the same across groups. Even if the estimates were substantially
di↵erent for a particular group, we would not be able to ascertain the extent to which the di↵erence was due
to composition bias.
18
this section, in which we consider long-run e↵ects, we will continue to estimate the impact of
being placed on academic probation after the first year. Since these outcomes are measured
at a later date, some students who were in good standing after their first year have been
placed on academic probation in subsequent evaluations. We exploit the fuzzy discontinuity
in whether or not students are ever placed on academic probation to estimate the impact of
ever being placed on probation.
Table 8 and Figure 9 show the estimated impacts on ever being suspended. Not sur-
prisingly, being placed on probation substantially increases the probability of ever being
suspended. One important thing to note is that the control means are not zero—some of the
students who just avoided academic probation after their first year are placed on probation
after later terms, fail to improve their grades, and are subsequently suspended.
Table 9 and Figure 10 show the estimated impacts on ever being suspended by subgroups.
These results indicate that academic probation increases the probability of suspension for all
of the subgroups considered but that the magnitude of the e↵ect varies across groups. We
find that the e↵ect on suspension is greater for high ability students than low ability students.
Previous results suggested that probation increases the probability that high ability students
drop out but has no e↵ect on the continued enrollment of the low ability students. As such,
this result cannot be attributed to relatively fewer high ability students who are at risk of
suspension after returning to school. This finding is consistent, however, with our previous
estimates suggesting that probation has a greater positive impact on the grades of low ability
students.
We also find that the estimated e↵ect of being placed on probation is larger for women
than men, but this may be a result of the di↵erences in academic probation’s e↵ect on drop
out rates documented in Section 6.3. Specifically, since probation increases the probability
that men leave the university before registering for any future terms, there are relatively
fewer men at risk of being suspended in future terms.
19
Lastly, we find that academic probation has a greater impact on the probability of sus-
pension for students whose first language is not English. Like the results by gender, this
finding might be partially explained by the di↵erences in academic probation’s impact on
students’ decisions to return to the university that were documented in Section 6.3.
While it is relatively simple to analyze the impact of academic probation on the proba-
bility of graduation, it is difficult to unpack the various mechanisms that are set into motion
by the treatment. The estimates in Section 6.3 demonstrated that a sizable share of students
around the threshold drop out from the university as a result of being placed on probation.
If this were the only e↵ect, then we would expect academic probation to either reduce the
probability of graduating or to have no e↵ect if the only students who drop out are those who
would not have graduated regardless. However, if academic probation motivates the remain-
ing students to improve their grades, as our estimates suggest in Section 6.4, then academic
probation might increase the probability of graduating. Finally, academic probation might
reduce the probability of graduation because it increases the probability of suspension, which
in turn likely increases the probability that a student drops out before finishing their degree.
Figure 11 and Table 10 show the estimated impacts of academic probation on whether
or not a student has graduated within four, five, or six years of their initial enrollment. For
these estimates, our sample is restricted to students in cohorts who are observed for at least
four, five, or six years, respectively, whether they graduate or not. While the point estimates
are consistently negative, we cannot reject that being placed on academic probation has no
e↵ect on the probability of graduation on average. In addition, relatively large standard
error estimates imply that we cannot rule out sizable e↵ects in either direction.
Figure 12 and Table 11 show the estimated impacts on graduation rates for a variety of
subgroups. In most cases, we cannot rule out that the impacts are the same across groups.
However, the estimated impact on the probability of graduating within five or six years is
especially large for students with high school grades above the median. For these students,
20
ever being placed on academic probation reduces the probability of graduation within five
years by 14 percentage points and reduces the probability of graduation within six years by
20 percentage points.26 While these magnitudes are very large, they are less surprising given
that we found these students are especially discouraged from continuing their enrollment
immediately after being placed on probation and that we found only weak evidence that
being placed on probation has a positive impact on the GPAs of those who return.
7 Discussion
As a whole, the most striking feature of our results is the heterogeneity in students’ responses
to academic probation. Consistent with Bénabou and Tirole’s (2000) theoretical model of
performance standards, we do find that some students are discouraged from putting forth
any e↵ort to meet the standard (choosing to drop out) while those who return are motivated
to improve their GPAs. Bénabou and Tirole’s model also predicts that the individuals with
the lower relative ability should be the most likely to give up. This is fairly intuitive since
these individuals are least likely to be able to meet the standard. However, we find that
the performance standard imposed by academic probation leads to higher ability students
having the larger drop out rate.
This unexpected result raises the question: are high school grades a good measure of
ability for students near the cuto↵? Given that all of the students entered the same university
and that students near the cuto↵ earned approximately the same first year GPA, some
concern about this labeling may be warranted. For example, the students with relatively low
high school grades may have been admitted to the university based on other characteristics
that signal high ability to admissions officers that we cannot observe. These signals might
26
Like the estimated impact on the probability of dropping out at the end of the first year, these estimates
are driven primarily by males. For high ability males, ever being placed on academic probation at the end
of the first year reduces the probability of graduating within six years by more than a third.
21
include the difficulty of a student’s high school, the number of advanced placement classes
taken, or involvement in extracurricular activities.
To answer this question, we can compare the subsequent performance of students in
our “high ability” group and “low ability” who barely avoided being placed on academic
probation at the end of their first year.27 The control means from Tables 5, 7, 9, and 11
indicate that, even though they earned the same first year GPA as the students we label
“low ability,” the students we label “high ability” are less likely to leave the university at
the end of their first year, earn higher GPAs in their next term, are less likely to ever be
suspended, and are more likely to graduate within four, five, and six years. As a whole,
these results suggest that our measure of ability e↵ectively predicts academic performance
for students at the university, even conditional on first year GPA.
Given that we are confident about our measure of ability, we suspect that Bénabou
and Tirole’s model does not predict the results that we observe across ability because the
assumption of perfect information does not hold in this setting.28 That is, students may
not know their true ability (or their probability of meeting the performance standard). We
can reconcile the model with our findings if we extend it to allow agents to act according
to self-perceived ability, where self-perceived ability can be a↵ected by being subjected to a
performance standard.
In a situation in which students do not have perfect information and act according to
self-perceived ability, being placed on academic probation can be thought of as a negative
signal sent by the administration to the student. Thus, being placed on academic probation
not only reduces a student’s second year choice set but also negatively a↵ects her perceived
ability.29
27
We focus on control students to avoid confounding di↵erences in performance due to ability with di↵er-
ences due to heterogeneous treatment e↵ects.
28
Falk, Hu↵man, and Sunde (2006) find that individuals are substantially uncertain about their abilities
in an experimental setting.
29
This feature is similar to models of performance rewards o↵ered by principals and their impacts on the
22
In this framework, an imperfectly informed student acts according to her self-perceived
˜ which is a function of her true ability, ✓, and whether or not she has been placed
ability, ✓,
on academic probation. Here, an unrestricted student solves
˜ AP )V1
max{0, ✓(✓, ˜ AP )V2
c1 , ✓(✓, c2 }. (7)
Analogous to the previous version of the model, students with the lowest self-perceived
abilities will choose to drop out, students with the highest self-perceived abilities will pursue
option 2, and the remainder will pursue option 1 if it is allowed. However, it is not necessarily
the case that those with the highest self-perceived ability have the highest true ability. While
we would generally expect a positive relationship between self-confidence and true ability,
subjecting agents to a performance standard might alter this relationship.
Consider two students: Hiabil and Loabil. Hiabil earned straight A’s and was a top
student throughout high school. Due to an uncharacteristically bad first year at the univer-
sity, Hiabil performs worse than she expected and is placed on academic probation. Loabil,
having consistently earned a B+ average throughout high school, believes she is a mediocre
student. Lobail performs below her expected level as well and is also placed on academic
probation.30
As a negative signal sent by the administration, we would expect being placed on aca-
demic probation to reduce the self-confidence of both students. However, the impact of the
signal may vary due to their prior experiences. Since Hiabil is accustomed to success, being
placed on academic probation may come as a great shock and reduce her self-confidence by a
self-confidence of agents discussed in Bénabou and Tirole (2003). In a related study, Eriksson, Poulsen,
and Villeval (2008) find that agents rarely quit the provided task regardless of di↵erences in performance
feedback. However, the decision to quit in their experiment is quite di↵erent from the decision faced by the
students in our analysis. Whereas dropping out of school has important long-term implications, the primary
incentive for individuals to quit the experimental task was the chance to use the remaining time available to
read a newspaper (or materials they brought with them).
30
Due to the random element involved in first year GPAs, students may have a wide range of abilities
despite their similar first year performance.
23
large amount. However, it might have a smaller impact on Loabil’s self-confidence since, due
to her prior levels of scholastic success, she expected to experience difficulty. If the di↵erence
in the impacts is great enough, Hiabil could have a lower post-signal self-confidence than
Loabil even though she has higher ability. In this manner, a performance standard that
a↵ects students’ perceived abilities heterogeneously could change the relative ordering of ✓˜
such that ✓˜h < ✓˜l even though ✓h > ✓l . As a result, high ability individuals could be more
likely to give up than individuals of lower ability.31 32
In this extended model, the decision to return to school can serve as a measure of revealed
˜ allowing us to further consider the impacts of being subjected to the
self-confidence (✓),
performance standard of academic probation. Our data has clear analogues to each of the
three main variables considered: agents who are subjected to the performance standard
(academic probation) and similar agents who are not, a measure of agents’ abilities (✓, high
˜ the decision to return to school). A
school grade percentile ranking), and self-confidence (✓,
more vivid picture of the relationship between these variables can be obtained by adding
the measure of ability as well its interaction with being placed on academic probation to our
previous regression equations.
Using these estimates, Figure 13 demonstrates the relationship between self-confidence
and ability for students on academic probation and those in good standing. Note that the
curves are intended to demonstrate the relationship between ability and self-confidence within
31
Similarly, our findings can also be reconciled using a two-skill model. Consider students with two types
of skill important for high school and college, Shs and Sc , respectively, which are positively correlated.
Students understand that the skills for high school and college are not one in the same. Thus, once students
are informed whether or not they are placed on probation, they evaluate their Sc based on their Shs in
addition to the new information. The signal provided by academic probation, of course, is negative. If
the magnitude of the “shock” is increasing in Shs , the model would be consistent with the top high school
students being most likely to drop out even though, on average, they have a greater aptitude for college.
32
It is also possible that the di↵erences we observe could be driven by di↵erential returns to education by
ability, as documented by Altonji and Dunn (1996) and Ashenfelter and Rouse (1998). For this to explain
the behavior that we observe, the return to education would have to be higher for high ability students in
good standing than low ability students in good standing, while the opposite would have to be true for those
on academic probation.
24
each group, but not across groups (since we would expect a lower probability of returning
for those on probation even if their self-confidence was unchanged). Within the framework
of the model, students on a given curve are directly comparable since they face the same set
of choices. The slopes of the curves indicate the following:
1. For students who are not subjected to the performance standard, there is little rela-
tionship between ability and self-confidence.
2. Being subjected to the performance standard causes agents of higher ability to have
lower self-confidence than those of lower ability.
The lack of a relationship between revealed self-confidence and ability for students in
good academic standing is consistent with the psychology literature on estimated intelli-
gence, which tends to find only a weak positive relationship.33 The negative relationship
between revealed self-confidence and ability for those on academic probation diverges from
previous findings. However, as described above, it is consistent with a model in which high
ability agents respond more negatively to being subjected to a performance standard. This
result may also di↵er from previous studies because we focus on a measure of revealed self-
confidence rather than a traditional survey in which participants su↵er little or no penalty
from misrepresenting their true beliefs about their ability. More research along these lines is
needed.
The two enumerated findings both raise the question: which students more accurately as-
sess their own ability—those of high or low ability? We cannot answer this question directly
since we do not know the dropout rates that we would observe if students had perfect infor-
mation. In an experimental study of students at the University of British Columbia, Heine,
Takata, and Lehman (2000) found that, even when confronted with information implying
that their performance was below average, students still thought they could outperform most
33
See Furnham (2001) for a review of the literature.
25
of their classmates. Following this finding, if we assume that all students are overconfident,
then these findings would imply that low ability students are especially overconfident.
Though these models have no testable predictions regarding gender, the heterogeneity
in changes to perceived self-confidence discussed above may be relevant for the gender dif-
ferences we find as well. Under the interpretation that the decision to continue at the
university is determined by self-perceived ability, our findings suggest that men have higher
self-perceived ability than women in the absence of probation and lower self-perceived ability
after being placed on probation. As with the di↵erences across abilities, this suggests that
men may have a stronger negative reaction to being placed on probation because their self
confidence is greater to begin with.
Our findings might also be related to gender di↵erences in overconfidence. Studies have
repeatedly shown that men are more likely to overestimate their scholastic abilities while
women are more accurate in their assessments.34 If being placed on academic probation
causes students to more accurately assess their own abilities, which is plausible given the
high stakes that are involved, then the gender di↵erence would be expected since men would
be required to make a greater adjustment.
8 Conclusion
We show that being placed on academic probation significantly discourages students from
returning to school after their first year of university study. Surprisingly, it is the highest
ability individuals who are discouraged most. The e↵ect on students with high school grades
above the median of entering students is so large that, as a group, they are less likely to
return than students of lower ability. We discuss how this finding can be reconciled with the
34
See Lundeberg, Fox, and Punćochaŕ (1994), Pajeres (1996), Beyer (1999), Furnham (2001), and Cheva-
lier, Gibbons, Thorpe, and Hoskins (2007). Male overconfidence has also been documented in non-scholastic
environments such as finance (Barber and Odean 2001).
26
theory of performance standards.
Consistent with a sizable literature on students’ responses to positive incentives, we
show that there are also gender di↵erences in response to a negative incentive. We find
that academic probation doubles the probability that men drop out but has no such e↵ect
on women. We also find evidence that being placed on probation improves the grades for
individuals of both genders.
Ideally, a performance standard might just “weed out” those who have no chance at suc-
cess and serve as a motivation for the others. In our setting, it would be a tool that improves
graduation rates. While we do find positive e↵ects on the subsequent GPAs for students who
continue to enroll, we find no evidence that being placed on academic probation improves
graduation rates. In fact, we find a very large negative impact on high ability students near
the cuto↵—ever being placed on probation reduces their probability of graduation by 20
percentage points. Like the e↵ects on the probability of dropping out, this e↵ect is so large
that these students are less likely to graduate than students of lower ability.
Our results have important consequences for the wide variety of circumstances in which
standards might be applied as a means of improving performance. Our findings on gender
suggest that, as with positive incentives, negative incentives are more likely to yield the
desired result with women than with men. Considering the current concerns regarding
gender di↵erences in academic success in both early and advanced education, this raises an
important question—can incentive schemes be designed that “work” for men?
We also confirm that a performance standard can entail a tradeo↵ between causing some
to work harder and causing others to give up. While this outcome might be anticipated by
many potential standard setters, our results suggest an additional consideration that is more
unexpected. Specifically, since high ability individuals are disproportionately likely to react
by giving up, a standard-setter can be left with individuals of lower ability, o↵setting the
improved performance of those who remain.
27
References
Altonji, Joseph G., “The Demand for and Return to Education When Education Out-
comes are Uncertain,” Journal of Labor Economics, 1993, 11 (1), 48–83.
and Thomas A. Dunn, “The E↵ects of Family Characteristics on the Return to Edu-
cation,” The Review of Economics and Statistics, 1996, 78 (4), 692–704.
Angrist, Joshua D. and Victor Lavy, “The E↵ect of High School Matriculation Awards:
Evidence from Randomized Trials,” 2002. NBER Working Paper No. 9389.
Angrist, Joshua, Daniel Lang, and Philip Oreopoulos, “Incentives and Services for
College Achievement: Evidence from a Randomized Trial,” 2007. NBER Working Paper
No. 12790.
Ashenfelter, Orley and Cecilia Rouse, “Income, Schooling, and Ability: Evidence from
a New Sample of Identical Twins,” The Quarterly Journal of Economics, 1998, 113 (1),
253–284.
Barber, Brad M. and Terrance Odean, “Boys Will be Boys: Gender, Overconfidence,
and Common Stock Investment,” Quarterly Journal of Economics, 2001, 116 (1), 261–292.
Bénabou, Roland and Jean Tirole, “Self-Confidence and Social Interactions,” 2002.
NBER Working Paper No. 7585.
and , “Intrinsic and Extrinsic Motivation,” The Review of Economic Studies, 2003, 70
(3), 489–520.
Chevalier, Arnaud, Steve Gibbons, Andy Thorpe, and Sherria Hoskins, “Students
Academic Self Perception,” 2007. UCD Geary Institute Discussion Paper Series.
Cook, Thomas D. and Donald T. Campbell, Quasi-experimentation: Design & Anal-

ysis Issues for Field Settings, Houghton Mi✏in Company, 1979.
DesJardins, Stephen L., Dennis A. Ahlburg, and Brian P. McCall, “Simulating the
Longitudinal E↵ects of Changes in Financial Aid on Student Departure from College,”
The Journal of Human Resources, 2002, 37 (3), 653–679.
Dynarski, Susan, “Building the Stock of College-Educated Labor,” 2005. NBER Working
Paper No. 11604.
Eriksson, Tor, Anders Poulsen, and Marie-Claire Villeval, “Feedback and Incentives:
Experimental Evidence,” 2008. IZA Discussion Paper No. 3440.
Furnham, A., “Self-estimates of intelligence: culture and gender di↵erence in self and
other estimates of both general (g) and multiple intelligences,” Personality and Individual
Di↵erences, 2001, 31 (8), 1381–1405.
28
Garibaldi, Pietro, Francesco Giavazzi, Andrea Ichino, and Enrico Rettore, “Col-
lege Cost and Time to Complete a Degree: Evidence from Tuition Discontinuities,” 2007.
NBER Working Paper No. 12863.
Hahn, Jinyong, Petra Todd, and Wilbert Van der Klaauw, “Identification and
Estimation of Treatment E↵ects with a Regression-Discontinuity Design,” Econometrica,
2001, 69 (1), 201–209.
Häkkinen, Iida and Roope Uusitalo, “The e↵ect of a student aid reform on graduation:
a duration analysis,” 2003. Mimeo.
Heine, Steven J., Toshitake Takata, and Darrin R. Lehman, “Beyond Self-
Presentation: Evidence for Self-Criticism Among Japanese,” Personality and Social Psy-
chology Bulletin, 2000, 26 (1), 71–78.
Heineck, Martin, Mathias Kifmann, and Norman Lorenz, “A duration analysis of

the e↵ects of tuition fees for long term students in Germany,” 2006. Mimeo.
Imbens, Guido W. and Thomas Lemieux, “Regression discontinuity designs: A guide

to practice,” Journal of Econometrics, 2008, 142 (2), 615–635.
Lee, David S. and David Card, “Regression discontinuity inference with specification
error,” Journal of Econometrics, 2008, 127 (2), 655–674.
Leuven, E., H. Oosterbeek, and V. van der Klaauw, “The E↵ect of Financial Rewards
on Students’ Achievement: Evidence from a Randomized Experiment,” 2003. CEPR
Discussion Paper no. 3921.
Lundeberg, Mary A., Paul W. Fox, and Judith Punćochaŕ, “Highly Confident,
but Wrong: Gender Di↵erences and Similarities in Confidence Judgments.,” Journal of
Educational Psychology, 1994, 86 (1), 114–121.
Manski, Charles F., “Schooling as experimentation: a reappraisal of the postsecondary

dropout phenomenon,” Economics of Education Review, 1989, 8 (4), 305–312.
McCrary, Justin, “Manipulation of the running variable in the regression discontinuity

design: A density test,” Journal of Econometrics, 2008, 142 (2), 698–714.
Pajares, Frank, “Self-Efficacy Beliefs and Mathematical Problem-Solving of Gifted Stu-

dents,” Contemporary Educational Psychology, 1996, 21 (4), 325–344.
Scalise, Alejandro, Mary Besterfield-Sacre, Larry Shuman, and Harvey Wolfe,

“First Term Probation: Models for Identifying High Risk Students,” Frontiers in Educa-
tion Conference, 2000, 1, F1F/11–F1F/16.
Seaver, W. Burleigh and Richard J. Quarton, “Regression Discontinuity Analysis of

Dean’s List E↵ects,” Journal of Educational Psychology, 1976, 68 (4), 459–465.
29
Figure 1
1500 1000 Distribution of student grades relative to their cuto↵
frequency count
500 0
−1.5 −1 −.5 0 .5 1 1.5

1st Year GPA minus probation cutoff
Notes: Each small hollow circle indicates the number of students with a distance
from their cuto↵ within 0.025 points (including the lower but not the upper
endpoint). Using each of these cells as an observation, the curve is predicted
from local linear regressions with a bandwidth of 0.6 using triangular kernel
weights. The estimated control mean is 608 and the estimated discontinuity is
58 (p-value = 0.488).
30
Figure 2
Estimated discontinuities in observable characteristics
High School Grade Percentile Rank
(a) (b) (c)
20
Credits Attempted in 1st Year
5
60
Age at Entry
40
19
4
20
18
0
3
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
1st Year GPA minus probation cutoff 1st Year GPA minus probation cutoff 1st Year GPA minus probation cutoff
(d) (e) (f)

.6
.9
1
English is 1st Language

Born in North America
Male
.4
.8
.7
.2
.6
.5
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
(g) (h) (i)

1
1
.8
.8
.8
At Campus 1
At Campus 2
At Campus 3
.6
.6
.6
.4
.4
.4
.2
.2
.2
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Notes: The curve is predicted based on second order polynomial that is flexible
on each side of the cuto↵. Each hollow circle is the mean of the outcome in an
interval of 0.05 around the point (including the lower but not the upper endpoint).
31
Figure 3
1 Probation status at the end of the 1st year
Probation Status
.5
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Notes: Same as Figure 2.
32
Figure 4
1 Ever on academic probation
Ever on Academic Probation
.5
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
33
Figure 5
.2 Student voluntarily drops out of school at the end of the first year
Left university voluntarily
.1 0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
34
Figure 6
Stratified results for voluntarily leaving school at the end of the first year
High school performance ranking

Lower 50% Upper 50%
.3
.3

.2
.2
.1
.1
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
1st Year GPA minus probation cutoff 1st Year GPA minus probation cutoff
Gender
Men Women
.3
.3

.2
.2
.1
.1
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Native language
English Other
.3
.3

.2
.2
.1
.1
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
35
Figure 7
1 GPA in next enrolled term
Subsequent GPA Minus Cutoff
0 −1
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
36
Figure 8
Stratified results for GPA in next enrolled term

Lower 50% Upper 50%
1
1

0
0
−1
−1
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Gender
Men Women
1
1

0
0
−1
−1
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Native language
English Other
1
1

0
0
−1
−1
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
37
Figure 9
.6 Ever suspended from the university
Ever Suspended
.2
0 .4
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
38
Figure 10
Stratified results for ever being suspended

Lower 50% Upper 50%
.6
.6
Ever Suspended
Ever Suspended
.4
.4
.2
.2
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Gender
Men Women
.6
.6
Ever Suspended
Ever Suspended
.4
.4
.2
.2
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Native language
English Other
.6
.6
Ever Suspended
Ever Suspended
.4
.4
.2
.2
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
39
Figure 11
1 Graduation rates
Has Graduated
.5
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Within 4 years Within 5 years Within 6 years
40
Figure 12
Stratified results for graduation rates

Lower 50% Upper 50%
1
1
Has Graduated
Has Graduated
.5
.5
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Within 4 years Within 5 years Within 6 years Within 4 years Within 5 years Within 6 years
Gender
Men Women
1
1
Has Graduated
Has Graduated
.5
.5
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
Native language
English Other
1
1
Has Graduated
Has Graduated
.5
.5
0
0
−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50 −1.50 −1.00 −0.50 0.00 0.50 1.00 1.50
41
Figure 13
Predicted self-confidence (probability of returning after the first year)
1
Predicted self−confidence for students near the cutoff
.98
(Probability of returning)
.96
.94
.92
.9
0 .2 .4 .6 .8 1
Theta (high school grade percentile / 100)
On Probation Not on Probation
Notes: The regression equation is

year 1 year 1
Returnic =m(GP AN ORMic )+ 1 HSGRADEic + 2 1(GP AN ORMic < 0)
year 1
+ 3 HSGRADEic ⇥ 1(GP AN ORMic < 0) + uic ,
where m() is a second order polynomial flexible on each side of the academic probation cuto↵. The lines
year 1
plotted in the figure are the regression estimates evaluated at the cuto↵ (GP AN ORMic = 0). The
resulting equation is:
P r[Return] =0.95766 +0.00005 ⇥ HSGRADE 0.01624 ⇥ AP 0.00030 ⇥ HSGRADE ⇥ AP

(0.00401) (0.00005) (0.00955) (0.00015)
The curves are intended to demonstrate the relationship between ability and self-confidence within each
group. Because the groups face di↵erent choice sets, the predicted probability of returning is not a valid
measure of self-confidence across groups.
42
Table 1
Summary Statistics
Mean Standard Deviation
Characteristics
High School Grade Percentile 40.12 25.36
Credits Attempted in First Year 4.49 0.52
Age at Entry 18.71 0.73
Male 0.38 0.48
English is First Language 0.73 0.45
Born in North America 0.88 0.33
At Campus 1 0.53 0.50
At Campus 2 0.20 0.40
At Campus 3 0.27 0.45
Outcomes
Distance from Cuto↵ in 1st Year 0.48 0.65
On Probation After 1st Year 0.23 0.42
Ever on Academic Probation 0.28 0.45
Left University after 1st Evaluation 0.05 0.22
Distance from Cuto↵ Next Evaluation 0.71 0.84
Ever Suspended 0.12 0.32
Graduated by Year 4 0.37 0.48
Note: The sample consists of the 32,697 students within 1.4 grade
points of the cuto↵ in their first year.
43
Table 2
Estimated Discontinuities in Observable Characteristics
Bandwidth 1.40 1.00 0.60

Observations with non-zero weight 32,697 23,861 13990
(1) (2) (3)
Dependent variable: High school grade (percentile ranking)

1st Year GPA < probationary cuto↵ 1.290 1.045 1.920
(1.213) (1.375) (1.651)
Constant (control mean) 30.639*** 30.857*** 30.788***
(0.738) (0.845) (1.060)
Dependent variable: Credits attempted in 1st year

(0.079) (0.088) (0.110)
(0.051) (0.057) (0.071)
Dependent variable: Age at entry

(0.035) (0.040) (0.051)
(0.020) (0.025) (0.035)
Dependent variable: Male

1st Year GPA < probationary cuto↵ 0.005 0.013 -0.015
(0.023) (0.027) (0.033)
(0.011) (0.013) (0.013)
Dependent variable: Born in North America

(0.012) (0.014) (0.018)
(0.008) (0.009) (0.011)
Dependent variable: English is 1st Language

1st Year GPA < probationary cuto↵ -0.035 -0.055** -0.049
(0.022) (0.026) (0.034)
(0.014) (0.017) (0.021)
Dependent variable: Attending Campus 1

(0.035) (0.038) (0.043)
(0.023) (0.025) (0.026)
Dependent variable:Attending Campus 2

1st Year GPA < probationary cuto↵ -0.011 -0.001 -0.016
(0.026) (0.029) (0.036)
(0.017) (0.019) (0.022)
Dependent variable: Attending Campus 3

1st Year GPA < probationary cuto↵ -0.002 -0.018 0.008
(0.046) (0.052) (0.062)
(0.031) (0.035) (0.037)
Notes: Same as Table 3

* significant at 10%; ** significant at 5%; *** significant at 1%
44
Table 3
Estimated Discontinuity in Probation Status
Bandwidth 1.40 1.40 1.00 1.00 0.60 0.60

Observations with non-zero weight 32,697 32,697 23,861 23,460 13,990 13,990
Control variables no yes no yes no yes
(1) (2) (3) (4) (5) (6)
Dependent variable: On academic probation after first evaluation

1st Year GPA < probationary cuto↵ 0.996*** 0.996*** 0.996*** 0.996*** 0.997*** 0.996***
(0.002) (0.002) (0.002) (0.002) (0.003) (0.003)
Constant (control mean) 0.001 - 0.001 - 0.000 -
(0.000) - (0.001) - (0.001) -
Dependent variable: Ever on academic probation

(0.012) (0.012) (0.012) (0.012) (0.015) (0.014)
Constant (control mean) 0.339*** - 0.368*** - 0.392*** -
(0.012) - (0.012) - (0.015) -
Notes: Estimated standard errors, clustered on GPA, are displayed in parentheses. Estimates are based
on regressions including a second order polynomial in adjusted first year GPA flexible on each side of
the cuto↵ using rectangular kernel weights and bandwidths noted in the table header. Constants are
included in all specifications, though they but are not shown for models with covariates as they do not
represent the control mean in such cases.
45
Table 4
Estimated E↵ect on the Decision to Leave after the First Evaluation
Bandwidth 1.40 1.40 1.00 1.00 0.60 0.60

(1) (2) (3) (4) (5) (6)
(0.008) (0.008) (0.008) (0.009) (0.008) (0.009)
(0.004) - (0.004) - (0.004) -

46
Table 5
Estimated E↵ect on the Decision to Leave after the First Evaluation for Subgroups
Subgroup HS Grades HS Grades Men Women Native Non-native

< Median >= Median English English
(1) (2) (3) (4) (5) (6)
1st Year GPA < probationary cuto↵ 0.017* 0.048*** 0.045*** 0.014 0.033*** 0.010
(0.009) (0.017) (0.014) (0.010) (0.011) (0.011)
Constant (control mean) 0.045*** 0.029*** 0.035*** 0.044*** 0.047*** 0.024***
(0.005) (0.006) (0.006) (0.005) (0.005) (0.006)
Notes: Estimated standard errors, clustered on GPA, are displayed in parentheses. Estimates are based on re-
gressions including a second order polynomial in adjusted first year GPA flexible on each side of the cuto↵ using
rectangular kernel weights and a bandwidth of 1.4.
47
Table 6
Estimated Discontinuities in Subsequent GPA
Bandwidth 1.40 1.40 1.00 1.00 0.60 0.60

(1) (2) (3) (4) (5) (6)
No adjustment for students not returning

(0.027) (0.026) (0.032) (0.031) (0.039) (0.038)
(0.016) - (0.021) - (0.025) -
Assuming students who left after being placed on AP would have performed the same in year 2 as in year 1
(0.024) (0.022) (0.029) (0.027) (0.035) (0.033)
(0.016) - (0.021) - (0.025) -
Assuming students on AP who left would have GPA change in the 10th percentile of similar students
1st Year GPA < probationary cuto↵ 0.062** 0.058** 0.082*** 0.074*** 0.066* 0.058*
(0.025) (0.023) (0.029) (0.027) (0.035) (0.032)
(0.016) - (0.021) - (0.025) -
on regressions including a second order polynomial in adjusted first year GPA flexible on each side of the
cuto↵ using rectangular kernel weights and bandwidths noted in the table header. Constants are included
in all specifications, though they but are not shown for models with covariates as they do not represent
the control mean in such cases. In the third panel, students that left after being placed on probation
have their second year GPA imputed as their first year GPA minus 0.52—only 10% of students within
0.6 of the cuto↵ who return after being placed on probation do worse.
48
Table 7
Estimated Discontinuities in Subsequent GPA for Subgroups

(1) (2) (3) (4) (5) (6)
1st Year GPA < probationary cuto↵ 0.205*** 0.127* 0.134*** 0.222*** 0.177*** 0.216***
(0.030) (0.076) (0.046) (0.035) (0.036) (0.052)
(0.019) (0.035) (0.025) (0.023) (0.032) (0.037)

49
Table 8
Estimated E↵ects on Ever Being Suspended
Bandwidth 1.40 1.40 1.00 1.00 0.60 0.60

(1) (2) (3) (4) (5) (6)
E↵ect of probation after the first year

(0.013) (0.014) (0.015) (0.016) (0.020) (0.022)
(0.006) - (0.006) - (0.008) -
E↵ect of ever being placed on probation

Ever on probation 0.148*** 0.148*** 0.144*** 0.139*** 0.139*** 0.141***
(0.019) (0.020) (0.024) (0.026) (0.033) (0.036)
(0.010) - (0.012) - (0.017) -
on regressions including a second order polynomial in adjusted first year GPA flexible on each side of the
cuto↵ using rectangular kernel weights and bandwidths noted in the table header. Constants are included
in all specifications, though they but are not shown for models with covariates as they do not represent
the control mean in such cases. The e↵ect of ever being placed on probation is based on a 2SLS estimate
as described in Section 5
50
Table 9
Estimated E↵ects on Ever Being Suspended for Subgroups

(1) (2) (3) (4) (5) (6)

(0.012) (0.026) (0.019) (0.018) (0.019) (0.020)
(0.012) (0.009) (0.011) (0.007) (0.006) (0.010)
E↵ect of ever being placed on probation

Ever on probation 0.138*** 0.184*** 0.131*** 0.159*** 0.118*** 0.220***
(0.025) (0.037) (0.030) (0.025) (0.028) (0.030)
Constant (control mean) 0.079*** 0.038** 0.104*** 0.047*** 0.079*** 0.040**
(0.012) (0.016) (0.019) (0.011) (0.012) (0.017)
rectangular kernel weights and a bandwidth of 1.4. The e↵ect of ever being placed on probation is based on a 2SLS
estimate as described in Section 5.
51
Table 10
Estimated E↵ects on Graduation
Bandwidth 1.40 1.40 1.00 1.00 0.60 0.60

(1) (2) (3) (4) (5) (6)
Dependent variable: Graduated after 4 years

1st Year GPA < probationary cuto↵ -0.016 -0.022 -0.017 -0.025 -0.052** -0.060***
(0.017) (0.015) (0.020) (0.018) (0.025) (0.022)
(0.011) - (0.012) - (0.014) -
E↵ect of ever being on probation
Ever on probation -0.026 -0.035 -0.028 -0.042 -0.089** -0.102***
(0.027) (0.024) (0.032) (0.029) (0.043) (0.038)
(0.019) - (0.022) - (0.027) -

1st Year GPA < probationary cuto↵ -0.016 -0.021 -0.020 -0.029 -0.032 -0.040
(0.025) (0.022) (0.029) (0.026) (0.039) (0.036)
(0.015) - (0.018) - (0.025) -
Ever on probation -0.026 -0.033 -0.033 -0.048 -0.055 -0.068
(0.039) (0.034) (0.048) (0.043) (0.066) (0.060)
(0.026) - (0.033) - (0.046) -

1st Year GPA < probationary cuto↵ -0.016 -0.023 -0.017 -0.025 -0.031 -0.043
(0.022) (0.021) (0.027) (0.025) (0.037) (0.035)
(0.015) - (0.019) - (0.028) -
Ever on probation -0.026 -0.036 -0.028 -0.042 -0.054 -0.072
(0.035) (0.033) (0.044) (0.042) (0.063) (0.059)
(0.025) - (0.034) - (0.050) -
rectangular kernel weights and bandwidths noted in the table header. Constants are included in all specifications,
though they but are not shown for models with covariates as they do not represent the control mean in such cases.
The e↵ect of ever being placed on probation is based on a 2SLS estimate as described in Section 5
52
Table 11
Estimated E↵ects of Ever Being Placed on Probation on Graduation for Subgroups

(1) (2) (3) (4) (5) (6)
Dependent variable: Has graduated after 4 years

Ever on probation -0.029 0.010 -0.056 -0.006 -0.047 0.021
(0.030) (0.062) (0.044) (0.040) (0.030) (0.050)
(0.034) (0.046) (0.033) (0.023) (0.021) (0.034)
Ever on probation 0.010 -0.135** 0.018 -0.051 -0.024 -0.050
(0.045) (0.068) (0.064) (0.045) (0.043) (0.055)
(0.034) (0.041) (0.047) (0.027) (0.027) (0.038)
Ever on probation 0.024 -0.198** -0.051 -0.014 -0.020 -0.067
(0.041) (0.079) (0.066) (0.044) (0.045) (0.052)
(0.033) (0.042) (0.050) (0.028) (0.027) (0.039)
rectangular kernel weights and a bandwidth of 1.4. As many students with high school grades above the median
are in the upper end of the overall grade distribution, the bandwidth restriction results in a substantially smaller
sample size compared to students with high school grades below the median. The e↵ect of ever being placed on
probation is based on a 2SLS estimate as described in Section 5.
53
A Appendix
A.1 Probation letter sent to students at Campus 2
Dear <first name> :
Your academic record indicates that you are experiencing challenges with your studies at
xxxxxxxxxxx. As a result, you have been placed ”On Probation” at the end of the xxxxxxx
session. “On Probation” is an academic status applied to a student if he or she:
1. Is having difficulty achieving a term average of at least 1.7 GPA or a yearly average
of 1.5 CGPA.
2. Is having difficulty meeting performance expectations and/or deadlines as outlined
by the course instructor.
3. Is having difficulty achieving the minimum grades required for graduation.
A student who at the end of any session during which they are on probation has a cumulative GPA
of less that 1.5 and a sessional of less that 1.7 shall be suspended therefore it is imperative that
you seek assistance to improve your academic standing to avoid further sanction.
Rest assured that you can improve this status and that xxxxxxxxxxxxx o↵ers assistance at many
junctions. First, you can access help by making an appointment with an academic advisor in the
Office of the Registrar to develop strategies to improve your academic record. Book an appointment
at xxx-xxx-xxxx or online at www.xxxxxxxxxx. Second, contact xxxxxxxxx for assistance with
study habits, note taking, e↵ective research, time management, study groups and peer mentors.
Finally, the xxxxxxxxxx o↵ers skills and interest testing which can help you focus on your strengths.
We know that you are capable to academic success, based on your academic record at admission.
A good academic record is essential for entry to Limited Enrolment programs, graduate school, and
professional schools. Let us review your goals and help you develop a plan to achieve them.
You have the opportunity and available support to be successful. Please utilize our services to
insure your future success.
54
For further information on academic status, please refer to xxxxxxxx of the Academic Calendar
or here: http://www.xxxxxxxxxxxxxx
55

Probation

Uploaded by

Copyright:

Available Formats

Probation

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

Probation

Uploaded by

Copyright:

Available Formats

NBER WORKING PAPER SERIES

ABILITY, GENDER, AND PERFORMANCE STANDARDS:

Working Paper 14261

NATIONAL BUREAU OF ECONOMIC RESEARCH

Jason M. Lindo Philip Oreopoulos

0 < c1 < c2 0 < V1 < V2 (1)

Ability is expressed as the probability of successfully completing either option, where

max{0, ✓V1 c1 , ✓V2 c2 }. (2)

P ROBicyear 1 = 1(GP Ayear

Yic = m(GP Ayear

where Yic is an outcome for student i at campus c, m(GP Ayear

6.1 Tests of the Validity of the RD Approach

6.3 The Immediate Response to Academic Probation

6.4 The Impact on Subsequent GPA

6.5 The Impacts on Suspension and Graduation

Cook, Thomas D. and Donald T. Campbell, Quasi-experimentation: Design & Anal-

Heineck, Martin, Mathias Kifmann, and Norman Lorenz, “A duration analysis of

Imbens, Guido W. and Thomas Lemieux, “Regression discontinuity designs: A guide

Manski, Charles F., “Schooling as experimentation: a reappraisal of the postsecondary

McCrary, Justin, “Manipulation of the running variable in the regression discontinuity

Pajares, Frank, “Self-Efficacy Beliefs and Mathematical Problem-Solving of Gifted Stu-

Scalise, Alejandro, Mary Besterfield-Sacre, Larry Shuman, and Harvey Wolfe,

Seaver, W. Burleigh and Richard J. Quarton, “Regression Discontinuity Analysis of

−1.5 −1 −.5 0 .5 1 1.5

(a) (b) (c)

(d) (e) (f)

English is 1st Language

(g) (h) (i)

−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50

1st Year GPA minus probation cutoff

Notes: Same as Figure 2.

−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50

1st Year GPA minus probation cutoff

Notes: Same as Figure 2.

−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50

1st Year GPA minus probation cutoff

Notes: Same as Figure 2.

High school performance ranking

Left university voluntarily

Left university voluntarily

Left university voluntarily

Notes: Same as Figure 2.

−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50

1st Year GPA minus probation cutoff

Notes: Same as Figure 2.

High school performance ranking

Subsequent GPA Minus Cutoff

Subsequent GPA Minus Cutoff

Subsequent GPA Minus Cutoff

Notes: Same as Figure 2.

−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50

1st Year GPA minus probation cutoff

Notes: Same as Figure 2.

High school performance ranking

Notes: Same as Figure 2.

−1.50 −1.00 −0.50 0.00 0.50 1.00 1.50

1st Year GPA minus probation cutoff

Within 4 years Within 5 years Within 6 years

Notes: Same as Figure 2.

High school performance ranking