Peer Effects and Pupil Attainment: Evidence from Secondary School Transition

Stephen Gibbons*, Shqiponja Telhaj**

July 2005

* Department of Geography and Environment and Centre for Economic Performance, London School of
Economics. * *Centre for Economic Performance and Centre for Economics of Education, London School of Economics Address for Correspondence: Centre for Economics Performance, London School of Economics, Houghton Street, London, WC2A 2AE, United Kingdom. Telephone: +44 (0) 20 7955 6245. Email: s.gibbons@lse.ac.uk; Acknowledgements: We thank the Department for Education and Skills for funding and providing the data for this research

-1-

Abstract It is a common belief that children will thrive if educated amongst better class and schoolmates. It is a belief that guides many parents in their choice of school, and has important implications for policy on school choice and organisation. Many studies have tried to measure this peer-group effect, but the enterprise is plagued by empirical difficulties. In this study, we use the population of state Secondary school pupils in England to tease out how pupil attainments at age 14 respond to differences in the prior, age-11 attainments of their current school grade peer-group. Data on home addresses and school attendance allow us to compare outcomes of children who live in the same street, or who attended the same Primary school up to age 11, but then move on to different Secondary schools with different peergroup quality. These peer-group effects seem to exist, but that they are small in magnitude a 1 s.d. increase in peer-group prior attainments allows a pupil to improve their own score by barely 0.08 of a standard deviation. We tackle various gnarly empirical problems arising in regression models of pupil attainments that incorporate individual and group prior attainments as explanatory variables. Estimates from such models are seriously biased by transient components in prior pupil attainments, correlation between current and prior peer-group characteristics and by ability sorting into Secondary schooling. We address these issues by using teachers predictions as instruments for prior attainments, defining a pupils current peer-group in terms of those school mates with whom he or she has had no contact in the past, and by predicting current peer-group attainments with the productivity of their origin Primary schools, measured by the gain in attainments of different cohorts between ages 7 and 11.

-2-

1. Introduction Schools seem often to be judged on the kind of children they enrol, rather than on the quality of their teaching or the other facilities they offer. This observation has led many to argue that the background and abilities of a pupils school-mates must have an important influence on his or her own achievements at school. Motivated by this, a rich international literature has evolved to try to model and measure the consequences of social interactions between pupils so called peer-group effects spanning the economics, education, sociological and psychological fields. The issue is a critical one in respect of current educational policy which favours expansion of school choice, because choice based on peer-group quality can, in theory at least, lead to a high degree of sorting across schools along lines of prior ability [Epple and Romano (2000)]. This will exacerbate educational inequalities if peer-group quality has real impacts on personal achievement. An understanding of peer effects is also important because they can mean that educational interventions that appear beneficial to the individual pupil may be even more effective when rolled out to the population [Glaeser, Sacerdote et al. (2003)]. Our paper extends the evidence base by providing estimates of the influence of innovations to a pupils peer-group at the time when they transfer from Primary to Secondary schooling. 2. Background Peer-group effects are a distinct class of influences arising from social interactions a broad term which encompasses any type of individual behaviour that involves interdependency with the behaviour or characteristics of others. Economists have long shown an interest [Becker (1974)], but there has been a rapid growth in the field since the 1990s with contributions in theory and empirical work. Theoretical research seems motivated by a desire

-3-

to widen the scope of economic thought to encompass aspects of behavioural modelling more commonly attributed to sociology and psychology. Empirical work constrained by the data is generally concerned with finding evidence for the existence of such effects, rather than the precise pathways by which they occur. The term peer-groups usually indicates social interactions of children or young adults with people of similar age, rather than broader neighbourhood effects or interactions with superiors, family or teachers. We continue to use the term in this way. The range of outcomes that have interested researchers is diverse, including smoking [Alexander and et al. (2001); Ellickson, Bird et al. (2003)], joke-telling [Angelone, Hirschman et al. (2005)], sexual behaviour [Selvan, Ross et al. (2001); Prinstein, Meade et al. (2003)], purchase of a retirement plan [Duflo and Saez (2000)] and more commonly education. On reflection, it seems very likely that many decisions are linked to similar decisions by a friend or other associate (in same cases fairly explicitly, like the decision to have sex, be in a gang or play tennis), and many consumption decisions rely on other consumers participating (e.g. video phones). However, the more interesting possibility is that group behaviour or attributes can modify individual actions in relation to important social and economic decisions that will affect their life chances especially achievement in education. Although the literature on peer effects in education dates back to 1960s with the publication of the famous Coleman Report (1966), the importance of peer-group effects is still disputed. Some very bold claims have been made about the potency of peers in child development [Rich Harris (1999)], yet the results of numerous studies are very mixed, finding strong, weak or non-existent effects across a wide range of outcomes. This reflects the difficulty in defining the peer-group, isolating causal peer-group effects from other influences, lack of appropriate data, and different identification methodologies adopted by

-4-

researchers. Indeed, as Manski (1993) and Moffit (2001) argue, the empirical analysis of social interactions is plagued by conceptual and data problems. The first key issue is that measures of peer-group characteristics may be good proxies for unobserved individual, family background or institutional factors that can affect student attainment, making peer effects look important when they may not be. Secondly, group membership is very likely to be endogenous to the outcome under study since people choose their school and their friends, leaving group and individual characteristics highly correlated. Thirdly, peer interaction is simultaneous in that a student affects and is affected by his or her peers (The reflection problem of Manski (1993)) although if peer effects are structurally unimportant, this source of bias vanishes. Lastly, there are conceptual and data-related problems in defining the peer-group is it the whole school, the childs year group or class, or some narrower delineation requiring information on personal friendship networks (with even more serious problems of endogenous group membership)? The earliest studies on peer effects in educational attainment [Hanushek (1971); Summers and Wolfe (1977); Henderson, Mieszkowski et al. (1978)] had mixed findings, but took relatively few steps towards overcome problems of peer-group endogeneity. Many more recent studies use instrumental variables approaches to try to overcome this, though it is hard to find plausible instruments. For instance, Dills (2005) predicts peer-group changes from introduction of magnet schools that selects high quality students, yet the average ability of remaining pupils will be decreasing in the proportion of high quality pupils that leave the school. Similarly Fertig (2003) instruments the coefficient of variation of peers with variables measuring whether a school selects pupils upon entry and whether the schools are in the private-sector. Goux and Maurin (2005) find good source of exogenous variation in peergroup attributes the average age and show that this matters for pupil achievement, but the cause could be average attainments or the average group age itself.

-5-

Other approaches are on offer. Hanushek et al. (2003) try to eliminate the problem of simultaneity by employing specifications based on lagged peer achievement, and tackle group selection problems using a fixed effect strategy. Similarly, McEwan (2003) includes school fixed effects strategy. Hoxby (2000) relies on the exogenous variation across cohorts in peer composition at the school grade-level in Texas elementary schools. Some other studies have exploited the random assignment of peer to individual students to find a solution to the problem of endogenous sorting of students. For example, Sacerdote (2001) and Zimmerman (2003) use the random assignment of roommates in colleges to find a positive association between roommates academic attainment and students own achievement. Cullen, Jacob et al. (2003) exploit the randomised lotteries that determine high school admission in the Chicago Public Schools, finding no systematic pattern of positive achievement and high quality peer-group effects. Sanbonmatsu, Kling et al. (2004) utilise a randomised housing mobility experiment in Boston, Baltimore, Chicago, Los Angeles and New York to isolate the impact of residential neighbourhood characteristics on student educational outcomes. They find that being given the option to move to a richer neighbourhood did not improve pupils academic performance. Even empowered with these more sophisticated estimation methods and richer data than earlier studies, researchers are still divided on the importance of peer effects. Some [e.g. Angrist and Lang (2004) ; Arcidiacono and Nicholson (2005)] find no significant relationship between peers and own achievement whilst others [e.g. Hoxby (2000); Zimmer and Toma (2000); Sacerdote (2001); Winston and Zimmerman (2003)] report positive effects. We should emphasise that generally even those studies that find positive peer effects find that they are small. Below we provide the range of peer effects estimates in some recent studies.

-6-

Table 1: A summary of some recent peer effect estimates

Studies Hoxby (2000) Context Texas schools, US Dartmouth College US Outcome Test scores, Peer-group or treatment Classmates test scores Roommates Grade Point Average Classmates School grade Methodology Cohort gender and race composition Random assignment to rooms School fixed effects School-bygrade fixed effects Random assignment IV using neighbours age Random assignment to rooms Assignment by lottery Approx order of magnitude 1.s.d. 0.4 s.d.

Sacerdote (2001)

College Grade Point Average,

1.s.d.

0.07 s.d.

McEwan (2003) Hanushek (2003)

Chile US

Test Scores Test Scores

1.s.d. 1.s.d.

0.27 s.d. 0.02.s.d.

Sanbonmatsu et al. (2004) Goux and Maurin (2005) Zimmerman (2003) Cullen, Jacob and Levitt (2003)

Moving to Opportunity experiment France Williams College, US Chicago public schools

School Test Scores Held back a grade in school College Grade Point Average Test scores, and others

Opportunity to move to new neighbourhood Neighbourhoods Roommates prior SAT scores Attendance at oversubscribed schools

Near zero and insignificant 1.s.d. 1 s.d. 0.1.s.d. 0.05 s.d.

Near zero and insignificant

Our approach in this paper combines some of these methods. Whilst we have no explicit randomisation in our identification strategy, we believe we can isolate sources of variation in the distribution of peer-groups across Secondary schools that are exogenous to a pupils own choices and abilities, and which we argue can be used to identify peer-group influences. The source of this variation is differences in the quality of Primary schools that supply pupils to Secondary schools in England, and the fact that there is a considerable degree of compulsory assignment in the allocation of Secondary places. The level of detail in our data allows us to compare outcomes for pupils who go on to attend different Secondary schools, but who live in the same street or attended the same Primary school; and we use information on Primary school effectiveness to predict components of each pupils new peer-group quality that are uncorrelated with their own abilities. In summary, we identify Secondary school peer-group effects from the fact that Secondary school pupils come from Primary schools of different
-7-

quality, and there is random variation in the composition of this group of feeder schools induced by education authority admissions policies. An advantage of our approach is that we can use a large representative sample of pupils drawn from 99% of the standard state schools in England. 3. Methods Empirical estimation of peer-group effects and the influence of social interactions in general is notoriously difficult, because peer-groups form endogenously in ways that are usually related to the outcome in question. This is also true in the context that we study: the effect of the average prior attainments of a pupils peer-group at the time they enter Secondary school (age 11-12), on his or her subsequent progress over three years. Average peer-group prior attainments in a pupils Secondary school will be linked to a pupils own unobservable characteristics because similar pupils choose the same school, because pupils at the same school will have family background attributes (like income) in common, and because many will have shared early experiences in the same Primary school. However, we have a lot of data to bring to bear on the problem, and can, we argue, find variation in peergroup attainments at age-11 that is uncorrelated with own ability at 14, except through Secondary school peer-group effects. This variation arises because of cohort-cohort variation in the Secondary school intake, in terms of the quality of Primary schools from which they originate. Our methods are based on estimating pupil-level models of attainments that control for prior attainments and include peer-group quality as explanatory variables. Such models are mis-specified if these prior test scores are intended to measure accumulated human capital or individual ability effects [Todd and Wolpin (2003)]. The problem is exacerbated when the model also includes the prior attainments of the contemporaneous peer-group. We devise

-8-

some strategies that, we argue, address these empirical issues. In order to explain these clearly, we set out our empirical model in some detail below. A child accumulates human capital in Primary school (prior to age 11). His or her level of human capital accumulation hi at the end of this phase depends on initial human capital (perhaps innate ability or motivation) ai , the human capital of Primary school peers api , Primary school quality p , individual characteristics xi and a neighbourhood or unobserved family background effect r . The notation
p i

here indicates that the peer-group is defined

relative to child i within the Primary school group p . Each child takes a common test at age 11, yi , which yields a noisy measure of human capital.

hi = ai + api + xi + p + r yi = hi + i

(1)

After completing the Primary school phase, pupils move to a new school and begin their Secondary education. Here they meet new friends and classmates, though usually some of this new peer-group will be familiar from their Primary school or neighbourhood. Human capital accumulation builds partly on prior human capital, driven on by the association with new peers, pupil characteristics (e.g. age, income, ethnicity), Secondary school teaching quality s , Primary school effects p that capture to persistent human capital growth trends amongst peers originating from the same Primary school, and neighbourhood or common family background effects r . Attainment is assessed at age 14 to provide a Secondary-phase test score zi , which has unobserved individual-specific components i . These might include new shocks to individual ability or motivation, plus random errors.
s zi = hi + h i + xi + p + r + s + i

(2)

The goal of our research is to estimate the parameters of this model, in particular the
s influence of peer-group human capital hi at the start of the Secondary phase. However,

-9-

individual human capital hi is unobservable and must be proxied by prior test scores.
s Similarly, Secondary school peer-group hi requires some statistic that is representative of the

group traditionally the mean test scores of other group members. The model is now:
s s zi = yi + y i + xi + p + r + s + i i i

(3)

Looking at (3) reveals that many of the unobservable components are correlated with the explanatory variables. Individual and group test scores are correlated with: i , pi through the measurement error process; p and r because pupils from the same Primary school or postcode share common influences on human capital growth in both phases (for example if a teacher in Primary school instilled a passion for learning that influenced attainments at the end of Primary school, and future growth rates of attainments, or if there are common neighbourhood or family background effects); s because of selection by prior attainments on Secondary school quality (if for example, high-human capital children choose schools with the best teaching record). We assume that i is uncorrelated with yi because any persistent components of ability are already embodied in prior human capital. However, we want to
s s allow the possibility that i is correlated with group prior ability ai (and hence h i ) to

capture the idea that Secondary school enrolment may involve sorting along lines of unobserved ability. It is straightforward to eliminate institutional factors like p , r and s by differencing the variables from group means (the standard within-group estimator). But this creates problems

- 10 -

of its own when the regressors are correlated with the group means of unobservables in the model. For example, if we difference out Primary school effects the model now looks like1
s sp zi z p = yi y p + y i y i + xi x p

+ r rp

) ( ) ( ) ) + ( s s ) + ( ) ( ) (
p p p i i

(4)
s i sp i

sp in which, in addition to the problems already discussed, ip and i are correlated with

s y i because the Primary school group and Secondary school group have some members in

common. One partial solution is to consider in the reference group only those members of the Secondary school group who are not also members of the Primary school group (or other group which is used when carrying out the within-groups transformation2). But this still leaves the noise components negatively correlated with individual and group Primary test scores, and the possibility of individual shocks correlated with group ability. These factors mean that the within-group estimator of (4) is biased: downward by measurement error, upward by ability sorting3.

This problem is analogous to that usually encountered in small T panel data models with individual

fixed effects, when the specification includes a lagged dependent variable. In this case the lagged dependent variable is correlated with the time-averaged component of the individual average error term.
2

This obviously wont work in the case of s, but the Secondary school peer-group is sufficiently large

(the mean is around 200) that we can ignore the problem.

3

We note also that an additional identification problem arises when using group fixed effects with peer-

group influences. This is caused by the severe collinearity between individual variables and peer-group averages of these variables when the data is transformed to deviations from group means. Consider the extreme case when your peer-group k comprises all other members of a group J of size n and we difference the data zij and the peer-group characteristics z ij from the group mean. The variables after the within-group transformation are
k

- 11 -

However, we have instruments available. Firstly, note that the prior expectation of test score attainments in (1), conditional on observable pupil characteristics, is a plausible instrument for pupil human capital since it is uncorrelated with test-score measurement error. Our data provide teacher assessments of expected attainment in the age-11 tests, expressed in terms of the National Curriculum Key Stage Levels (see Section 4 below; we propose to use these as instruments for individual pupil achievement.
s Secondly, aggregating (1) by Secondary peer-group suggests that effectiveness pi of

the Primary schools attended by peers is an appropriate instrument for Secondary peer-group human capital, purged of any innate group ability effects. This assumes that p is determined by resource allocation, teaching quality, leadership and other institutional factors in Primary
s schools. We propose to measure pi using the school-mean gain in attainments between ages

7 and 11 using different cohorts from those used to estimate the main equation (2)4. The identifying assumption of this strategy is that this estimate of the age 7-11 value-added of peers Primary schools is only correlated with a pupils own attainments at age 14 because of

zi = zi

1 nJ

z
iJ

ij

, z ik = zi

1 zj n 1 j i , jJ
J

If all groups J are the same size, then zi = n 1 z i and the peer-group effect cannot be identified
J k

separately from the effect of the individuals own characteristic. The issue becomes less severe as the number of kids in each street varies, and as the street contains a mixture of kids from different schools, and as your classmates are drawn from different streets. But it is another good reason to consider only the peer-influence of schoolmates who originate from different residential neighbourhoods, who came from different Primary schools, but end up at the same Secondary school from the observation pupil. This observation is related to the discussion in Lee (2004) in the context of identifying spatial models.
4

In principle, Primary school value-added would also provide an instrument for individual attainments

(purged of individual ability effects), but in practice this has insufficient predictive power at the individual level.

- 12 -

its relationship with peers prior attainments. Moreover, the relationship between peer-group prior attainments and the current age-11 cohort value-added is attributable only to the fact that they attended the same Primary schools (conditional on pupils own prior attainments, characteristics, and Primary school, Secondary school and neighbourhood fixed effects included in (2)). In other words, we assume that pupils do not choose Secondary schools based on the future value-added of younger cohorts in the Primary schools attended by pupils who will form their Secondary school peer-group. This seems credible to us, since prospective pupils can have no information on which pupils are to be admitted from which schools in a given year, and any predictable components (e.g. if there are some well known feeder schools) will be accounted for in our estimation by Secondary school fixed effects. In summary, we propose to estimate model (2) using teacher expectations of attainments at age 11, and the age-7-age 11 value added of peers origin Primary schools, as instruments for the accumulated human capital of pupils and their Secondary school peer-group. We further extend the empirical specification beyond that in (2) to allow for contextual peers group effects from pupil demographic characteristics (i.e. we include xi ). Again, Secondary

school peer-group is defined in terms of those pupils who do not live in the same postcode as the individual, and who did not attend the same Primary school. Lastly, our specification allows for Primary school, residential (postcode) and Secondary school fixed effects, which means we are basing our estimation on differences in outcomes for pupils who live in the same street but attend different Secondary schools, pupils who attended the same Primary school but attend different Secondary schools, and pupils who attend the same Secondary school in different years.

- 13 -

4. Data and institutional context

In England, state compulsory-age education is organised into five National Curriculum Key Stages and spread over two phases. Primary schooling starts at age 4/5 and continues to age 10/11, spanning Foundation Stage and Key Stage 1 (5-7), and Key Stage 2 (8-11). Compulsory Secondary schooling runs from age 11/12 to age 15/16, spanning Key Stages 3 (12-14) and Key Stage 4. Pupil progress is assessed by standard tests at the end of each Key Stage. Funding of schools is organised largely through central government grant distributed to Local Education Authorities (LEAs), and these LEAs handle most of the school admissions and other administrative procedures. In a few LEAs the Primary/Secondary distinction is somewhat blurred by the prevalence of Middle schools, which typically bridge part of Key Stage 2-Key Stage 3, though the exact age range varies. For the purposes of this paper we describe the schools at which a pupil takes the Key Stage 3 tests (age-14) as Secondary, and the school at which they take the Key Stage 2 tests (age-11) as Primary. The picture is made more complex by the institutional differences between schools at all phases. Most schools (65%) are designated Community schools which means, essentially, that they are non-selective in admission and are administered by the LEA. Some other schools have religious affiliations and are allowed to select on the basis of religious commitment, and a smaller number are run by other types of charitable institution but still come under the state-school umbrella. A number of state schools in some LEAs are allowed to select pupils by academic ability (e.g. traditional Grammar schools). In addition, there is a small, but not inconsequential, number of private sector (Independent) schools. We will focus on Community schools as an initial step to reduce selection issues induced by parental choice: Although pupils and their parents can express preferences over which Community school they would like to attend, many end up at schools that were not their first choice (or even any of their choices) because the most popular schools are over-subscribed.

- 14 -

Information on pupil Key Stage tests results is collected by the Governments Department of Education and Skills (DfES), who use the data to publish school performance league tables. The DfES have also, since 2002, conducted a Pupil Level Annual Census with information on pupil demographics for the current school population (in attendance on Census day), which can be linked to pupil test records held in a National Pupil Database. The Census is based on a day in January in 2002 and 2003 and records pupil characteristics, home postcodes and school identifiers. From this database, we extract the two cohorts aged 14 in 2002 and 2003 for our main sample, plus the two cohorts age 11 in 2002 and 2003 for calculation of the Primary school value-added used in our instrumentation strategy. The age-14 cohort census data are spliced to pupils age-14 Key Stage 3 results, and the Key Stage 2 results they recorded at age 11. The age-11 cohort census data is linked to pupils Key Stage 2 test scores and the Key Stage 1 test scores these pupils achieved at age 7. Additional school-level information such as admissions policy and school type is derived from the DfES Edubase system which holds details on all educational establishments. As we have said, our sample is restricted to Community schools only, avoiding distinctive school types that may be preferred by distinct groups in the population, or schools where there is explicit selection of pupils by ability. Also note that since we include postcode fixed effects and need meaningful variation in the Secondary school peer-group within the postcode, there must be at least two pupils attending at least two different Secondary schools. Hence, we restrict the estimation sample to postcodes where pupils attend at least as many Secondary schools as there are resident pupils. A postcode is typically 14 addresses (the median in England) corresponding to a contiguous group of houses on one side of a street in an urban area.

- 15 -

Note however, that peer-group quality measures are constructed using weighted group means from the two full population cohorts, not this postcode-restricted sample. The weighted peer-group mean of a variable x for pupil i in Secondary school cohort s is defined as the within-s sum of x across pupils from postcodes and Primary schools other than the is own, divided by the total number of pupils in s.
5. Results

5.1. Descriptive statistics The estimation sample we described above is much reduced in size from the full population in our data around 10% of the full Secondary population and about 15% of the population in Community Schools. But, it contains around 120,000 pupils and is highly representative of the population in Community schools as a whole. In fact, the 15% of pupils in the sample are drawn from 99% of the Community schools across the country. Because we focus on Community schools our sample is slightly biased towards denser urban areas with higher proportions of pupils on free school meals and in non-white ethnic groups. The geographical distribution of the 43000 postcodes in the sample is shown in Figure 1. The number of pupils per postcode ranges from 2 to 18, with a mean of 2.8. For ease of comparison and interpretation, all variables except our instruments are transformed into percentiles of the distribution in the Secondary school pupil population. The means and standard deviations of these percentiles in the estimation sample are shown in the main tables of regression coefficients, but for reference, Table 1 summarises these actual data. 5.2. Regression estimates of the educational production function We now turn to the central regression estimates of the model of English and Mathematics attainments at age 14 (Key Stage 3) based on Equation (2) and presented in

- 16 -

Table 3 and Table 4. In all specifications, we condition on prior attainments in the corresponding subject at age-11 (Key Stage 2), age in months, and include a set of dummy variables as descriptors of pupil ethnicity (7 categories), free-school-meal entitlement, gender, special educational needs (4 categories), and English as a first language. In both Tables, Column 1 presents ordinary least-squares estimates without controlling for school or geographical fixed effects, Column 2 differences the variables from residential postcode means, Primary school means and Secondary school means. In Columns 3-7 we introduce instrumental variables into this within-groups specification to correct for transient components in the prior test scores and selection by unobserved ability on unobservable components of peer-group ability. It is evident from Column 1, Row 2 that we find a moderate basic association between pupil age-14 English attainments and the English attainments of their peers as measured at age 115. There are links too with the other characteristics of the peer-group, though peergroup prior attainments dominate in both magnitude and statistical significance. In Maths, the link with peers prior attainments is less marked, and peer-group income (measured by freeschool meal entitlement) has a stronger relationship. It seems unlikely that these relationships are causal, for all the reasons outlined in Section 36. Moving to Column 2, the specification controls for residential postcode, Primary and Secondary school fixed effects and there are some notable changes: For English, there is a slight fall in the coefficient on peer-group

For simplicity we do not consider the cross-elasticities with pupil attainments in other subjects, though

we recognise that these may be relevant too. Inclusion of prior attainments in other curriculum areas in the regressions does not change the basic message of these results.
6

The reader worried about sample selection issues should note that the OLS estimates using the full

sample of 1869 Community schools are 0.159 (7.21) for English and 0.062 (4.47) for Maths, based on pupil sample sizes of 602895 and 623484.

- 17 -

attainments, and otherwise only free-school meal entitlement and the proportion with English as an additional (not their first) language seem important. For Maths, somewhat surprisingly, the within-group transformation raises the coefficient on peer-group age-11 Maths attainments and its statistical significance to bring it in-line with that for English. As anyone would expect, a pupils own prior attainments are the strongest predictor of attainment at age 14, and are responsible for over 95% of the R-squared shown at the bottom of the table. In the OLS estimates, the magnitude of this coefficient is a bit disheartening the coefficient of 0.65 for English indicates that any improvement in attainment at age-11 is halved within 5 years through human capital depreciation7. However, one of the main difficulties in interpreting these results stems from the use of these transient measures of pupil prior attainments, which will bias things downwards8. Selection on unobservable characteristics of the peer-group that are correlated with prior attainment may also be an issue. So, in Columns 3-7 we turn to our IV strategy outlined in Section 3, using Primary teacher assessments of pupil ability, and Primary school age-7 to age-11 value added as predictors of pupil and peer-group attainments. The value-added instruments are simply the schoolmean gain in attainments between age 7 and 11 (measured on a point score system used by the DfES), for pupils aged 10-11 in our sample years9.

If the intervention raises attainments at t = 0, then the effect at t=T is 0.65T. The period between tests

spanned by our data is 3 years so the annual depreciation factor is 0.87.

8

This can be thought of as similar to the standard linear measurement error problem, as the problem of

transient incomes discussed in the intergenerational mobility literature, or estimation of dynamic panel data models with a first differenced error term, and is discussed in Todd and Wolpin (2003).
9

We also tried as an instrument the conditional value added obtained as the Primary school-mean

residual from a regression of age-7-11 value-added on pupil characteristics: the results were almost identical.

- 18 -

For transparency, we present the first stage for each of the endogenous variables (own test scores in Column 4, peer-group test scores in Column 5), and the reduced form in Column 6. Overall, these results suggest that transience in the test-scores is the dominant problem IV leads to substantially higher coefficients on own prior attainments (about 20% up for English and 12% for Maths). The annual depreciation factor on human capital is now 0.92 for English and 0.95 for Maths. This means it takes 13-14 years to halve any gains in Maths made at age11. The IV coefficients on peer-group attainments are almost double their OLS counterparts: A 10 percentile move up the distribution of peer-group attainments is corresponds to a 2.4 2.8 percentile improvement in individual pupil attainments. Contextual effects from demographic characteristics now seem unimportant for English, but are still significant for Maths. However in terms of magnitude, the effect of prior attainments dominates all the other peer-group influences. The first stages for both endogenous variables are encouraging, with large coefficients and high t-statistics for the relevant instruments. It is reassuring to note too that pupils own age-11 attainments are uncorrelated with peer-group Primary school value-added in Row (13) of Column 4, and that teacher assessments of pupil abilities do not have a strong association with peer-group attainments relative to their power in predicting own-attainments. Inspection of the reduced form in Column 6 in relation to the IV in Column 2 is also informative, in particular because this shows that peer-group income (free meal entitlement) only affects age-14 attainments in so far as it has already had a bearing on prior attainments. 5.3. Testing for selection on Primary school value added One concern is that a Secondary school pupil selects his or her school partly on expectations of the quality of school from which their peers will originate. In fact, such selection would be rational if peers Primary school quality is really beneficial to his or her own education as our results indicate. If true, then this means that peer-group effects must be
- 19 -

positive, but our estimates will be upward biased. We can go some way to allaying such concerns by considering the schoolmates who arrive in Secondary school some time after the initial transfer between Primary and Secondary schooling. Such late arrivals may occur because children have moved home, arranged for a school transfer, or have been held back for some reason for example if awaiting a place a their school of choice. It is difficult to believe even if pupils choose schools based on peers prior school quality that they can anticipate the compositional changes induced by late arrivals to the peer-group; so we would expect big differences between the peer-group effect we measure using the peer-group as a whole, and that which can be estimated from late arrivals. To test this, we define late-arrivals as pupils who joined the Secondary school after the median start date, and introduce into the regressions an additional peer-group variable computed from the mean prior attainments of this late-arrivals group. Table 5 presents the key coefficients from this exercise for age-14 Maths and English scores. If selection is the main issue, our assumptions imply significant negative coefficients on the late-arrivals peer-group (perhaps too because of shorter exposure times). However, for attainments in English there is no difference between the general peer-group effect and that measured off late arrivals, whether we consider raw prior attainments (Column 1) or Primary school value-added (Column 2) as the peer-group characteristics. For Maths, the prior attainments of late arrivals seem to provide an additional kick to the peer-group effect (Column 1), but the effect vanishes when we turn to Primary school value-added as the group characteristic (Column 2). Nothing here suggests that the main effect of Secondary school peer-group is an illusion caused by selection on expected group composition. 5.4. Non-linearities and complementarities A common theme in the literature on peer-group effects is the degree to which they are non-linear. Certainly, this is an important consideration since the policy implications when the
- 20 -

marginal effect of peer-group quality is increasing in peer-group quality are very different from when there are diminishing marginal returns. In the first instance, policy that improves peer-group quality in the best groups provides greater gains than at the bottom of the distribution and segregation is efficient; in the second, integration and equity is more efficient. For example, Zimmerman (2003) and McEwan (2003) provide support for the policy mixing students of different ability. However, there are many other studies that find no evidence of nonlinearities [e.g. Hoxby (2000)] Similar considerations make complementarities between own attainments and peer-group quality interesting since these will reinforce educational inequality across individuals and drive sorting by ability across schools. We address both these concerns in a basic non-parametric fashion by replacing the linear peer and prior attainment effects in our regressions with a dummy variable set for the joint distribution of teacher assessments of pupil Key Stage Level at age-11, and peer-group age-11 attainment quartiles. Teacher assessments are grouped into Level 1-2, Level3, Level 4 and Level 5+; peer attainments are divided into quartiles. Unfortunately we are not able to instrument peer-group quality effectively, so we present the coefficients from the OLS estimates with Primary, Secondary and residential fixed effects and are forced to rely on these estimates10. The coefficients from this regression are shown in Table 6a (English) and Table 6b (Maths). The reported coefficients show the percentile of the age-14 test score relative to the baseline pupils those who were assessed at Level 1-2 at age 11, and are in Secondary schools with peers in the lowest quartile of age-11 attainment.

10

We noted from the IV regressions above, that the main source of bias is the use of transient test

measures of prior attainment [c.f. Hanushek and et al. (2003)], which we implicitly correct for here by using teacher assessment of student ability

- 21 -

The most obvious feature emerging from these numbers is that the peer-group effects are unsurprisingly dwarfed by a pupils own prior attainments; though that is not our main concern. The interesting issues are whether the peer-group effects are non-linear, conditional on prior attainments, and whether there are complementarities between peer-group and ability. Considering the first issue, inspection of Table 6 suggests some non-linearities in the sense that the gap between each peer-group quartile within any ability band is non-constant; but the patterns are not particularly striking and where they exist tend to suggest large effects at the lower end or middle of the peer-group distribution. There is more to say about complementarities, in that the least able pupils do not benefit from peer-group improvements, whereas the middle and higher ability groups do (observe the F-statistics for the joint test of the significance of the coefficients in each column, and the overall difference between the top and lowest quartile). In Maths, only those pupils expected to reach the age-11 target Level in the national curriculum (Level 4) show any benefit from peer-group attainments. These findings offer some understanding of the reasons why lower ability pupils (or their parents) might be less pro-active in there efforts to secure better peer-groups: These children have little to gain from better peer-groups. This is a main concern of those who criticise school choice on the basis that it leads to increased school segregation Walford (1996). One explanation for our findings is that these peer-group effects operate through competition for teaching resources, rather than direct social interaction between pupils. Suppose that the speed of learning of the lowest ability children is constrained by their own abilities, but that the speed of learning of more able children is governed by the rate at which teaching progresses, taking into account the average mix of abilities. This would explain the kind of patterns seen in Table 6, with no influence from peer-group attainments on the lowest achievers, but stronger effects on other pupils. As peer-group ability increases, teachers become less constrained in what they are able to teach.

- 22 -

6. Discussion and conclusions

Pupils seem to do better in their early stages of Secondary school when their new schoolmates have a good record of achievement. Our reading of this result is that there is some form of social interaction between pupils that promotes higher attainments. Manski (2000) provides a useful classification of social interactions in groups those due to pupils preference to act like their friends, those due to competition for constrained resources like teacher time, and those due to the information that group behaviour provides about the expected consequences of individual action but we are unable here to be precise which of these mechanisms prevails. Perhaps individual behaviour is mutable under group influences, and a move to a new school with high-attaining children opens up new challenges with the individual drawn into higher achievement by the expectations of the group. However, since our findings seem to reflect responses to peer-group prior attainments and not other contextual differences and because the effects are zero for the lowest ability groups, we conjecture that the most likely explanation may be more mundane: teaching can proceed faster in higher ability groups, or can start from a higher base-line when the groups prior attainments are higher. If expectations or norm (preference) related factors were important we see no reason why other group characteristics should not have an equally strong influence or why the lowest ability pupils show no response. Prior research in the educational literature has often cited low-income of peers measured by free-school meal entitlement as an important contextual influence on pupil attainments [e.g. Strand (2002)]. Our results show that the influence of peer-group free-meal entitlement on pupil attainments works through the prior attainments of the peer-group. We find no evidence that there are direct influences from peer-group free-school meal entitlement, and other group demographics have insignificant or relatively small effects. These are

- 23 -

encouraging results for policy makers because pupils prior attainments are surely more amenable to early interventions than socioeconomic and demographic characteristics. Our results are also useful in suggesting that policy that raises standards at Primary schools will have more persistent effects on treated pupils if they remain together in Secondary schools, than if they are widely dispersed. In other words, policy is more effective on aggregate than we might think from looking at the relationship between a pupils attainments at Primary school and their subsequent scores at age 14. There is a social multiplier effect [Glaeser, Sacerdote et al. (2003)] in terms of the influence of Primary school effectiveness. Roughly speaking, a technological innovation that led to a 10-percentile improvement in attainments at Primary school would lead to an 8- percentile improvement in pupil age-14 attainments directly, plus a 2.5 percentile effect attributable to better Secondary school peers. The peer-group effect more than compensates for the depreciation of human capital that seems to take place between age-11 and age-14. Having said that, the contribution of peer-group abilities to the distribution of attainments in the short run seems very small. A one standard deviation improvement in peergroup quality relates to a mere 0.08-0.09 standard deviation increase in pupil attainments at the end of our three-year periods. In the long run though, a pupil who benefits from better peer-groups throughout his or her school career may be at more of an advantage than this would at first suggest. Suppose a pupils peer-group is one standard deviation above the mean throughout his or her 12 years of schooling, and that the effects of prior attainments (own and peer-group) on attainments at the end of each 3-year period are roughly in line with our results: his or her end-of-school attainments will be some 0.24 standard deviations above the

- 24 -

mean, since the effects are cumulative over the period11. Even under this scenario, peergroups play a fairly limited role in the distribution of educational attainments: if peer-group quality was perfectly correlated over the years for each individual it would only account for 5.5% of the variance of educational attainments across individuals at the end their compulsory schooling years12. Given the magnitude of these effects it is hard to believe that the efforts to which some parents go to secure schools with a good peer-group are worthwhile, purely in terms of the improvement in educational achievement that better quality peer-groups can offer. Better peer-groups perhaps provide other benefits physical safety, emotional security, familiarity or simply exclusivity which make them desirable commodities, aside from any small educational advantages they offer.

11

This assumes a model of the form yit = 0.8 yit 1 + 0.08 yit 1 where the variables are standardised.

Over four periods the effect of a persistent 1 s.d. increase in peer-group attainments y is, from the sum of a

finite geometric series, 0.08

1 0.84 = 0.236 . More plausibly given what we can infer from our data, a 1 0.8

correlation coefficient of 0.5 between mean peer-group attainments in each period would suggest that a pupil who starts off in a peer-group that is 1 s.d. above the mean, will end up with attainments that are
2 3 0.08 0.125+0.8 0.08 0.25+0.8 0.08 0.5+0.8 0.08=0.093 s.d. above the mean.
12

0.2362

- 25 -

7. References
Alexander, Cheryl and et al. (2001). 'Peers, schools, and adolescent cigarette smoking', Journal of Adoloescent Health, vol. 29(1), pp. 22-30. Angelone, D.J, Hirschman, Richard, et al. (2005). 'The Influence of Peer Interactions on Sexually Oriented Joke Telling', Sex Roles, vol. 52(3-4), pp. 187-199. Angrist, Joshua D. and Lang, Kevin (2004). 'Does School Integration Generate Peer Effects? Evidence from Boston's Metco Program', American Economic Review, vol. 94(5), pp. 1613-34. Arcidiacono, Peter and Nicholson, Sean (2005). 'Peer Effects in Medical School', Journal of Public Economics, vol. 89(2-3), pp. 327-50. Becker, Gary S (1974). 'A Theory of Social Interactions', Journal of Political Economy, vol. 82(6), pp. 1063-93. Cullen, Julie Berry, Jacob, Brian A., et al. (2003). 'The Effect of School Choice on Student Outcomes: Evidence from Randomized Lotteries', National Bureau of Economic Research, Inc, NBER Working Papers. Dills, A. (2005). 'Does Cream Skimming Curdle the Milk? A Study of Peer Group Effects', Economics of Education Review, vol. 24(1), pp. 19-28. Duflo, Esther and Saez, Emmanuel (2000). 'Participation and Investment Decisions in a Retirement Plan: The Influence of Colleagues' Choices', National Bureau of Economic Research, Inc, NBER Working Papers, 7735. Ellickson, Phyllis P., Bird, Chloe E, et al. (2003). 'Social Context and Adolescent Health Behavior: Does School-level Smoking Prevalence Affect Students Subsequent Smoking Behavior?' Journal of Health and Social Behavior 2003,, vol. 44(4), pp. 525-535. Epple, Dennis and Romano, Richard (2000). 'Neighborhood Schools, Choice, and the Distribution of Educational Benefits', National Bureau of Economic Research, Inc, NBER Working Papers, 7850. Fertig, Michael (2003). 'Educational Production, Endogenous Peer Group Formation and Class Composition - Evidence from the PISA 2000 Study', IZA Discussion Papers, 714, Bonn. Glaeser, Edward L., Sacerdote, Bruce I., et al. (2003). 'The Social Multiplier', Journal of the European Economic Association, vol. 1(2-3), pp. 345-353. Goux, Dominique and Maurin, Eric (2005). 'Close Neighbors Matter: neighbourhood effects on early performance at school', Unpublished Working Paper. Hanushek, E. (1971). 'Teacher Characteristics and Gains in Student Achievement: Estimation Using Micro Data', American Economic Review, vol. 61(2), pp. 280-288. Hanushek, Eric A. and et al. (2003). 'Does Peer Ability Affect Student Achievement?' Journal of Applied Econometrics, vol. 18(5), pp. 527-44. Henderson, Vernon, Mieszkowski, Peter M., et al. (1978). 'Peer Group Effects and Educational Production Functions', Journal of Public Economics, vol. 10(1), pp. 97-106. Hoxby, Caroline (2000). 'Peer Effects in the Classroom: Learning from Race and Gender', NBER Working Paper, W7867. Lee, Lung-fei (2004). 'Identification and Estimation of Spatial Econometric Models with Group Interactions, Contextual Factors and Fixed Effects', Unpublished Working Paper, Ohio State University. Manski, Charles (2000). 'The Economics of Social Interactions', Journal of Economic Perspectives, vol. 14(3), pp. 115-136. Manski, Charles F. (1993). 'Identification of Endogenous Social Effects: The Reflection Problem', Review of Economic Studies, vol. 60(3), pp. 531-42. McEwan, Patrick J. (2003). 'Peer Effects on Student Achievement: Evidence from Chile', Economics of Education Review, vol. 22(2), pp. 131-41. Moffit, R (2001). 'Policy Interventions, Low-Level Equilibria, and Social Interactions', Social Dynamics, (S. N. Durlauf and H. Young eds.), Cambridge MA: MIT, pp. 45-82.

- 26 -

Prinstein, Mitchell J, Meade, Christina S, et al. (2003). 'Adolescent oral sex, peer popularity, and perceptions of best friends' sexual behavior', Journal Of Pediatric Psychology, vol. 28(4), pp. 243-249. Rich Harris, Judith (1999). The Nurture Assumption: Why Children Turn Out the Way They Do, Bloomsbury. Sacerdote, Bruce (2001). 'Peer Effects with Random Assignment: Results for Dartmouth Roommates', Quarterly Journal of Economics, vol. 116(2), pp. 681-704. Sanbonmatsu, L., Kling, J., et al. (2004). 'Neighbourhoods and Academic Achievement: Results from the Moving to Opportunity Experiment, Industrial Relation Section', Princeton University Working Paper, 492. Selvan, Ross, et al. (2001). 'Study of perceived norms, beliefs and intended sexual behaviour among higher secondary school students in India', AIDS Care, vol. 13(6), pp. 779-788. Strand, S (2002). 'Pupil Mobility, Attainment and Progress During Keys Stage 1: a cautious interpretation', British Education Research Journal, vol. 28(1), pp. 63-78. Summers, A. and Wolfe, B. (1977). 'Do Schools Make a Difference?' American Economic Review, vol. 67(4), pp. 639-652. Todd, Petra E. and Wolpin, Kenneth I. (2003). 'On the Specification and Estimation of the Production Function for Cognitive Achievement', Economic Journal, vol. 113(485), pp. F3-33. Walford, G (1996). 'School Choice in England and Wales', School Choice and the Quasi Market, Oxford: Triangle, vol. 6, pp. 49-62. Winston, Gordon C. and Zimmerman, David J. (2003). 'Peer Effects in Higher Education', National Bureau of Economic Research, Inc, NBER Working Papers, 9501. Zimmer, R. and Toma, E. (2000). 'Peer Effects in Private and Public Schools Across Countries', Journal of Policy Analysis and Management, vol. 19(1), pp. 75-96. Zimmerman, David J. (2003). 'Peer Effects in Academic Outcomes: Evidence from a Natural Experiment', Review of Economics and Statistics, vol. 85(1), pp. 9-23.

- 27 -

Figure 1: Geographical distribution of postcodes in the estimation sample

- 28 -

Table 2: Descriptive statistics for the largest estimation sample

Mean Age-14 English percentile Age 14 Maths percentile Age-11 English percentile Age-11 Maths percentile Primary school value-added points English level, teacher assessment 1 2 3 4 5 6 Maths level, teacher assessment 1 2 3 4 5 6 Peer-group proportion girls Peer-group mean age Peer-group white Peer-group English as first language Peer-group eligible for free school meals Sample size 0.002 0.039 0.279 0.506 0.173 0.002 0.500 161.4 months 0.767 0.865 0.217 117750 0.270 months 0.003 0.552 0.287 0.502 0.151 0.001 44.41 45.11 44.50 45.16 35.58 s.d. 27.83 27.39 27.97 28.21 1.50

- 29 -

Table 3: Secondary pupil progress and peer-groups, pupils aged 14 in 2001/2-2002/3, Community schools, English
(1) OLS Own age-11 test (percentile) Peer-group age-11 mean attainments Peer-group proportion girls Peer-group mean age Peer-group white Peer-group English first language Peer-group eligible for freeschool meals Teacher assessment Fstatistic F(5,1850) Peers Primary value-added Within-R2 0.680 (241.74) 0.181 (7.11) 0.001 (0.04) -0.023 (-3.24) 0.021 (2.09) -0.034 (-2.97) -0.041 (-4.99) 0.588 (2) OLS-within 0.654 (241.06) 0.148 (7.17) -0.002 (-0.33) -0.006 (-1.08) -0.000 (-0.00) -0.027 (-3.14) -0.028 (-3.72) 0.526 (3) IV-within 0.783 (212.76) 0.237 (2.55) -0.008 (-1.09) -0.020 (-1.52) 0.003 (0.40) -0.030 (-3.26) 0.003 (0.14) 0.512 (4) IV Stage 1a 0.002 (0.61) 0.009 (2.16) -0.012 (-2.46) -0.001 (-0.16) -0.025 (-6.24) 21976.21 0.037 (0.59) 0.613 (5) IV Stage 1b 0.060 (30.92) 0.128 (49.57) 0.003 (0.97) 0.044 (10.90) -0.203 (-65.32) 18.05 0.898 (21.87) 0.710 (6) Reduced-form 0.007 (1.46) 0.017 (3.32) -0.006 (-0.67) -0.022 (-2.42) -0.064 (-10.39) 8583.2 0.199 (2.39) 0.444 (7) IV-restricted 0.782 (211.98) 0.221 (5.33) * * * * * 0.529 Means 45.40 (27.74) 39.44 (9.64) 52.96 (28.55) 54.08 (28.00) 47.73 (30.36) 47.56 (30.16) 63.33 (27.30) 35.59 (1.50)

Regression at the pupil level. t-stat in brackets, clustered on Secondary school; Dependent variable is Key Stage 3 (age-14) English test score percentile. Mean 44.54, s.d. 27.81;Controls are: pupil gender, ethnic group, free-school-meal entitlement, special educational needs, age in months, year dummy; * indicates variables excluded from 1st and 2nd stage; Number of pupils 112894; Number of Secondary schools 1851; All columns except (1) allow for residential postcode, Primary school and Secondary school fixed effects.

- 30 -

Table 4: Secondary pupil progress and peer-groups, pupils aged 14 in 2001/2-2002/3, Community schools, Maths
(1) OLS Own age-11 test (percentile) Peer-group age-11 mean attainments Peer-group proportion girls Peer-group mean age Peer-group white Peer-group English first language Peer-group eligible for freeschool meals Teacher assessment Fstatistic F(5,1851) Peers Primary value-added Within-R2 0.778 (352.85) 0.065 (3.81) 0.020 (4.05) -0.010 (-1.82) -0.014 (-2.07) -0.011 (-1.46) -0.094 (-18.61) 0.751 (2) OLS-within 0.771 (378.16) 0.145 (10.46) 0.015 (5.10) -0.015 (-3.43) -0.018 (-3.85) -0.012 (-2.29) -0.028 (-6.12) 0.697 (3) IV-within 0.866 (337.57) 0.278 (4.87) 0.013 (4.25) -0.036 (-3.69) -0.017 (-3.80) -0.018 (-3.13) 0.007 (0.57) 0.689 (4) IV Stage 1a 0.007 (2.35) -0.002 (-0.41) -0.002 (-0.44) 0.000 (0.03) -0.024 (-6.70) 27027.11 0.065 (1.13) 0.652 (5) IV Stage 1b 0.060 (30.92) 0.128 (49.57) 0.003 (0.97) 0.044 (10.90) -0.203 (-65.32) 11.75 1.024 (26.32) 0.698 (6) Reduced-form 0.020 (5.68) 0.005 (1.08) -0.016 (-2.70) -0.005 (-0.82) -0.063 (-14.26) 16729.93 0.296 (4.48) 0.565 (7) IV-restricted 0.866 (335.42) 0.249 (5.33) * * * * * 0.687 Means 45.17 (28.21) 40.035 (9.05) 52.93 (28.53) 54.19 (28.01) 47.73 (30.35) 47.64 (30.18) 63.68 (27.21) 35.58 (1.50)

Regression at the pupil level. t-stat in brackets, clustered on Secondary school; Dependent variable is Key Stage 3 (age-14) Maths test score (adjusted) percentile. Mean 45.11, s.d.27.40;Controls are: pupil gender, ethnic group, free-school-meal entitlement, special educational needs, age in months, year dummy; * indicates variables excluded from 1st and 2nd stage; Number of pupils 117750; Number of Secondary schools 1852; All columns except (1) allow for residential postcode, Primary school and Secondary school fixed effects.

- 31 -

Table 5: Robustness to pupil sorting on Primary school quality

Age-14 English (1) OLS-within All peer-group age-11 mean attainments Age-11 mean attainments of late arrivals All peers Primary value-added Late-arrivals Primary value-added Pupils 0.101 (7.67) 0.020 (1.35) 110383 (2) Reduced form 0.464 (5.05) -0.092 (-0.79) 110383 Age-14 Maths (3) OLS-within 0.096 (11.77) 0.020 (2.95) 115153 (4) Reduced Form 0.505 (7.14) 0.035 (0.49) 115153

Table notes as for Tables 1 and 2, within-group estimates

- 32 -

Table 6a: Interactions between own and peer-group abilities: English

Teacher assessment of age-11 ability Level 1-2 Peer-group Quartile 1 0 Peer-group Quartile 2 Peer-group Quartile 3 Peer-group Quartile 4 Q4-Q1 Peer-group effect F(2, 1852) test 1.001 (1.79) 0.549 (0.90) -0.707 (-0.96) -0.707 3.02 (0.049) Level 3 6.509 (14.67) 6.943 (14.22) 7.748 (15.91) 8.336 (15.66) +2.277 8.73 (0.000) Level 4 26.873 (54.33) 27.835 (54.23) 28.731 (55.96) 30.126 (58.91) +3.253 30.71(0.000) Level 5+ 52.361 (90.18) 52.680 (91.95) 53.773 (94.00) 53.828 (97.58) +1.467 4.45 (0.004)

Table shows the coefficients and t-statistics on dummy variables for own attainment/peer attainment quartile interactions in the OLS regression similar to Column 2 in Table 1; Sample size 117201; Other controls as in Table 1

Table 6b: Interactions between own and peer-group abilities: Maths

Teacher assessment of age-11 ability Level 1-2 Peer-group Quartile 1 0 Peer-group Quartile 2 Peer-group Quartile 3 Peer-group Quartile 4 Q4-Q1 Peer-group effect F(2, 1853 test) -0.322 (-0.72) -0.581 (-1.15) -0.740 (-1.28) -0.740 0.29 (0.746) Level 3 9.202 (27.14) 9.108 (25.10) 9.575 (25.21) 9.750 (23.81) +0.548 2.00 (0.112) Level 4 32.972 (86.15) 33.806 (89.03) 34.967?? (91.05) 34.967 (91.05) +1.995 63.13 (0.000) Level 5+ 61.113 (130.45) 61.688 (136.30) 62.674 (145.12) 63.940 (153.70) +2.827 19.45 (0.000)

Table shows the coefficients and t-statistics on dummy variables for own attainment/peer attainment quartile interactions in the OLS regression similar to Column 2 in Table 2; Sample size 121977; Other controls as in Table 2

- 33 -