Professional Documents
Culture Documents
Author(s): Jon Rasbash, George Leckie, Rebecca Pillinger and Jennifer Jenkins
Source: Journal of the Royal Statistical Society. Series A (Statistics in Society), Vol. 173,
No. 3 (JULY 2010), pp. 657-682
Published by: Wiley for the Royal Statistical Society
Stable URL: https://www.jstor.org/stable/40666280
Accessed: 23-01-2019 11:58 UTC
REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/40666280?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
Royal Statistical Society, Wiley are collaborating with JSTOR to digitize, preserve and
extend access to Journal of the Royal Statistical Society. Series A (Statistics in Society)
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
J.R.Statist. Soc. A (2010)
173, Part 3, pp. 657-682
Summary. School effectiveness analyses have largely ignored the role of the family as an impor-
tant source of variation for children's educational progress. Sibling analyses in developmental
psychology and behavioural genetics have largely ignored sources of shared environmental
variation beyond the immediate family. We formulate a multilevel cross-classified model that
examines variation in children's progress during secondary schooling and partitions this variabil-
ity into pupil, family, primary school, secondary school, local education authority and residential
area. Our results suggest that about 50% of what has been labelled as pupil variation in school
effectiveness models is really between-family variation and that about 22% of the total variance
is due to shared environments beyond the immediate family.
1, Overview
Children are raised in complex social environments that involve multiple layers
(Bronfenbrenner, 1 977). The advent of statistical techniques that deal with nested dat
has made it possible to differentiate between elements of this complex social environ
school effectiveness research can differentiate between variance that is attributable to schools
and individuals (Aitkin and Longford, 1986; Goldstein et al., 1993). Sibling studies, within
family research, have differentiated between variance that is attributable to families and indi-
vidual children (Lynch et al, 2006; Georgiades et al, 2008; Jenkins, Simpson, Dunn, Rasbash
and O'Connor, 2005). Neighbourhood research has differentiated neighbourhood effects from
those of individuals (Xue et al, 2005). These research areas work in parallel, ignoring the
influence of aspects of the complex social environment that are not their focus. This may result
in a misattribution of influence. For instance, school effectiveness studies have largely ignored
the role of the family in children's attainment. They find substantial variance that is attribut-
able to unmeasured pupil level characteristics but, as family influences have not been explicitly
considered, much of what is referred to as 'pupil level characteristics' may instead be attribut-
able to the family. Many school effectiveness studies have included family level covariates, such
as household socio-economic status (Garner and Raudenbush, 1991) in the model. However,
the designs still have only one child per family, so the estimated 'child' level variance is an
Address for correspondence: George Leckie, Centre for Multilevel Modelling, University of Bristol, 2 Priory
Road, Bristol, BS8 1TX, UK.
E-mail: G. Leckie@bristol.ac.uk
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
658 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
amalgamation of unobserved child and family level effects. This is suggested by results from
family studies indicating that siblings show high degrees of similarity. Family studies are
confounded in a different way from school effectiveness studies. These studies use a single
catch-all term for the shared environment (the environment that is shared by siblings) and fail
to differentiate between family, school or neighbourhood influences. The confounds that are
evident in school effectiveness and family studies occur because of limitations of data as well as
statistical models. In the present paper we create a data set that included children nested within
neighbourhoods, schools and families. Through the use of cross-classified multilevel models we
differentiate between different types of environmental influence.
The goals of this paper were threefold. First, we present the application of a cross-classified,
multilevel model that allows us to differentiate between family, school (primary and second-
ary), neighbourhood and local education authority (LEA) influences. Although in our study
we do not have a genetically sensitive design, which would enable separation of genetic effects
as discussed more fully in Section 2.2, we describe the way in which the model can be extended
to provide an estimate of genetic influence. Second, we improve on previous school effective-
ness studies through our differentiation of family and pupil level effects. Third, we improve on
family research involving sibling designs, by differentiating between types of shared influence:
family, neighbourhood and school. Specifically the two research questions that we address are
as follows.
(a) How much of the variation in progress during secondary schooling is attributable to
shared environmental influences beyond the immediate family, where the wider shared
environment is operationalized as secondary school, neighbourhood, LEA and carry-
over effects from primary school?
(b) How much of what has been thought of as pupil level variation in progress during sec-
ondary schooling is really due to family level factors?
Estimation of the family component was achieved through the identification of twin pairs in
our data set. Two limitations of this use of twin data should be borne in mind. First, zygosity
information, i.e. the degree of genetic relatedness, for twin pairs, was not available and so it
was not possible to distinguish mono-zygotic (MZ) or identical twins from di-zygotic (DZ) or
non-identical twins. This means that the family effect that we present confounds the social and
genetic influence of families. Second, we cannot identify a separate family component for non-
twin children and so our results regarding the family component of variation are only directly
generalizable to other twins. Our findings are applicable to non-twin siblings in so far as twin
and non-twin siblings are equivalent. This is an issue that we return to in the discussion.
2. Background
In this section we describe the existing literature on school effectiveness studies and sibling
designs, pulling out key points that are relevant to our subsequent arguments.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 659
2009). Findings from the analysis of school effectiveness studies have shown that around
5-20% of unexplained variability in pupils' progress in achievement is attributable to unmea-
sured school level factors with the remaining 80-95% attributable to unmeasured pupil level
characteristics. This gives an intraschool correlation coefficient in the range 0.05-0.20. These
approximate ratios apply in both primary schools (see for example Goldstein et al (2007)) and
secondary schools (see for example Leckie (2009)). Including additional pupil and school level
predictor variables and their interactions may further reduce the unexplained variability in the
model, usually by no more than 10% or so. However, the intraschool correlation coefficients
typically remain in the range 0.05-0.20, indicating that the majority of the unexplained
variability in progress is attributable to pupil level processes. Both intuitively and from behavi-
oural genetic (BG) studies of cognitive traits (Plomin et ai, 2003; Ko vas et ai, 2005), which we
discuss below, we expect that, if family is included in school effectiveness models, much of the
unexplained between-pupil variation will be attributable to differences between families.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
660 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
tWhere more than one response was examined in a study, we have selected just one
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 661
(and perhaps progress). We here briefly examine how the findings of stu
compare with those of studies of achievement, but we shall return to this rese
later, where it will be used to illuminate the discussion of Section 6.
Plomin (1999) and Plomin and Petrill (1997) presented arguments in favour
itary component to cognitive ability, that is estimable by BG methods. Bouch
(1981), who are often quoted as providing strong evidence of a substantial he
to cognitive ability, summarized the findings from 144 studies exploring fam
in cognitive ability. A subset of these findings is shown in Table 2, repor
cognitive ability across pairs of twins and siblings.
Here we can see similar findings to those of Table 1 : substantial correla
0.47-0.86) between siblings. Again we note that these correlations are relative
with the intraschool correlations for educational progress, obtained from sch
studies, typically in the range 0.05-0.2.
One criticized assumption of BG models applied to data from studies w
reared together is the assumption that environmental factors affect non-tw
DZ twins equally. We shall return to this point in Section 4 where we presen
the BG model.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
662 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
include additional classifications which may additionally explain test score variability. Obvious
choices for additional classifications are LEA, residential neighbourhood and, in the case of
secondary school tests, primary school attended. However, with one exception that is described
below, we have been unable to identify either an example in the literature or a data set with the
necessary linkage, where both family and school membership, let alone additional classifica-
tions, are readily available to apply such a cross-classified model.
Ideally, data and models that allow us to separate child and family variance components are
required if we are to obtain a true estimate of the size of family effects on educational progress.
In studies with one child per family, we can obtain an estimate of the effect of family level
covariates that we have measured, but we have no measure of the full extent of the variance
arising from unobservable family level characteristics.
2.4. Duncan et al. (2001) study using the adolescent health data set
One data set that does provide the necessary linkage for estimating these models is the National
Longitudinal Study of Adolescent Health (http : / /www. epe . une . edu/pro j ects/add
health). This data set includes all students from a sample of 132 US schools. A random
sample of approximately 200 children from each school was selected for in-home interviews,
producing the core in-home sample of 12000 adolescents. This sample was further boosted
to include siblings, MZ twins and DZ twins to estimate genetic and family effects better. The
boosted sample was 20745 children.
Duncan et al (2001) analysed the data, separately estimating correlations on child height,
intelligence, which we discuss below, and delinquency between 226 MZ pairs, 166 DZ pairs,
534 non-twin sibling pairs, 1357 nominated best friend pairs, 268906 neighbourhood pairs and
291 704 'grademate' (school) pairs. Correlations from the first three pair types give indications
of the size of family shared environmental and shared genetic effects whereas correlations from
the last three give indications of the magnitude of peer, neighbourhood and school effects. One
shortcoming of their analytic strategy is that each pair set is analysed separately, so results are
not properly adjusted. For example, the familial correlations included correlations due to sib-
ling pairs attending the same school (grademate) and living in the same neighbourhood. Thus
part of the familial correlations will be attributable to some shared school and neighbourhood
influences. The non-familial correlations are adjusted for a range of family covariates; however,
this will by no means remove all the family effects from the resulting non-familial correlations.
Fitting a multilevel model where children are nested within a cross-classification of family,
school, neighbourhood and best friend pair would achieve a proper separation of effects.
Despite the methodological limitations that were described above, the findings from Duncan
et al (2001) are instructive. Results for the Peabody picture vocabulary test are summarized
as this is the closest outcome to that described in the present paper. The correlations for MZ,
DZ, full sibling, neighbourhood and school pairs are 0.78, 0.64, 0.51, 0.06 and 0.07 respectively.
The relative sizes of these correlations indicate that familial effects are considerably larger than
neighbourhood and school effects. We note also the similarity to the Bouchard and McGue
(1981) findings that were presented in Table 2, which give correlations, for intelligence quotient
IQ, of 0.86, 0.60 and 0.47 for MZ twins, DZ twins and full siblings on cognitive ability. The
difference in the correlations between MZ and DZ twins suggests some role for genetic influence.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 663
since pupils are nested within schools which are strictly nested within LEAs. Fielding et al
(2006) analysed educational achievement by using a large-scale data set of over 80000 pupils
in England who attended 534 secondary schools in 25 LEAs. In their models with no predictor
variables, they found that LEAs explain only 3% of the variation in test scores compared with
22% for secondary schools. Yang and Woodhouse (2001) conducted a multiple-cohort study of
academic progress between GCSE examinations taken at age 16 years and General Certificate of
Education Advanced level (A level) examinations taken at age 18 years (fuller details of GCSE
and A levels are given in Section 3). They found that LEAs account for less than 1% of the
variation in progress compared with 5% for secondary schools.
Neighbourhood effects are often discussed as important potential influences on educational
achievement. However, few value-added studies incorporate neighbourhoods in their analysis.
Notable exceptions were Garner and Raudenbush (1991), Fielding et al (2006) and Leckie
(2009). Garner and Raudenbush (1991) provided an early analysis of cross-classified data for
2500 pupils in Scotland nested within a cross-classification of 17 schools and 524 neighbour-
hoods. However, rather than estimating a random-effects cross-classified model, they estimated
a two-level random-effects model of pupils (level 1) nested within neighbourhoods (level 2). In
their model without co variâtes, neighbourhoods account for 20% of the total variation in scores.
However, after they adjusted for prior achievement, family background, neighbourhood social
deprivation and school dummy variable fixed effects, just 4% of the remaining variation lies
between neighbourhoods. In a reanalysis, Raudenbush (1993) estimated a full random-effects
cross-classified model that partitions the total variability between pupils, neighbourhoods and
schools. A caveat is that, with just 17 schools, the school component of variation is likely to
be imprecisely estimated. In their model without covariates, neighbourhoods (13.9%) explain
nearly twice as much variation as schools (7.4%); however, 79% of the variation remains between
pupils. Adjusting for the same predictors as Garner and Raudenbush explains most of this var-
iation and the ratio of school to neighbourhood variance increases from a half to two-thirds.
This rise is expected because the study adjusts for a neighbourhood level variable but does not
adjust for any school level variables.
More recent studies by Fielding et al (2006) and Leckie (2009) both used cross-classified
models in an analysis of different samples of English secondary school pupils. Both studies
found that neighbourhoods explain significant variation in pupils' educational achievement
and progress, with greater variation found for smaller scales of neighbourhood. However, in
contrast with Raudenbush (1993), both Fielding et al (2006) and Leckie (2009) reported that
schools are more important than neighbourhoods. Comparisons between these three studies are
somewhat limited as the studies refer to different contexts and different samples of pupils and
importantly make different adjustments for pupil background characteristics. However, what
is certain is that pupils make by far the largest contribution to variation in test scores.
Cross-classified models are also required to model any sustained or carry-over effects of
schools that are attended in an earlier phase of education on pupils' current progress. Goldstein
and Sammons (1997) considered the persistence of junior school effects on pupils' progress in
secondary schools. They found that the variance of junior school effects was greater than that
for secondary schools in both their models without covariates and in their models that adjust
for prior achievement and other pupil background characteristics. They suggested that junior
schools are so variable partly because of the importance of early schooling, but also because
they tend to be smaller. The smaller size of junior schools may lead them to capture pockets of
heterogeneity in the population to a greater extent than the larger secondary schools.
The importance of earlier school membership was also reported by Rasbash and Goldstein
(1994) and in a reanalysis by Browne et al (2001). These two studies estimated cross-classified
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
664 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
models for primary and secondary schools in Fife, Scotland. In models without covariates, they
found that the primary school variance was three times that of secondary schools. However,
it should be noted that, with just 19 secondary schools, the secondary school component of
variation is likely to be imprecisely estimated. Goldstein et al. (2007), using larger data sets than
all of the above studies, looked at infant school effects on progress during junior schooling. In
cross-classified models that adjust for prior achievement and other pupil background charac-
teristics, they found inconsistent results across the two geographic areas that they considered
with the infant school variance being relatively larger in one area, but smaller in the other.
In all the studies that are discussed here, whether LEAs, neighbourhoods, primary schools
or secondary schools are included in the analysis, pupils remain by far the largest source of
variation in test scores. In this paper we expect that, when family is included in a multilevel
analysis of secondary school effectiveness, much of what is typically regarded as between-pupil
variation will be reattributed to variation between families and the size of the family effect will
be substantially larger than that for LEAs, residential neighbourhoods, primary schools and
secondary schools.
The examination data are taken from the national pupil database, which is a census o
the English state education system provided to us by the Department for Children
Families. The national pupil database holds data on pupils' test score histories and a
ber of pupil level characteristics. From this database, we extract the cohort of pu
born between September 1st, 1990, and August 31st, 1991, who took their GCSE e
in 2007 and key stage 2 (KS 2) examinations in 2002. These examinations are taken
of secondary schooling (age 16 years) and primary schooling (age 1 1 years) respec
ful GCSE results are often a requirement for taking A levels which in turn are a
of university entrance determinant. For those who leave school at 16 years of age
are their main job market qualification. To GCSE scores, we merge data from
pupil level annual school census data sets which give pupils' background charac
ther information on the national pupil database and pupil level annual school c
and how to access them can be found at http : / /www. bris . ac . uk/cmpo/
The initial sample consists of the 593 754 pupils who took their GCSE exami
mainstream secondary schools. Thus our sample excludes, for example, schools wit
dard age ranges such as middle schools and those that cater only for pupils w
cational needs (SENs). We exclude for convenience pupils who have missing v
the predictor variables that were used in the analysis. This reduces the sample by
pupils.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children 's Educational Progress 665
twin pairs is improved further by restricting the definition of a twin pair to those pupils who addi-
tionally share the same ethnicity and the same time varying pattern of eligibility for free school
meals (FSMs) (see Section 3.3 for details of all the variables that were used in the analysis) over
the observation period. FSM eligibility is determined by family income and is therefore constant
within family between twins. Unfortunately, we could not identify any non-twin siblings that
are present in our sample (i.e. because both siblings were born in the sampled academic year).
We do identify a very small number of triplets and quadruplets (335 pupils) in the data. If we
included these without changing the structure of the model, the family component would apply
to twin, triplet or quadruplet families, and it may not be sensible to group these three family
types as different processes may be going on for each type. The alternative, which would allow
for these different processes, would be to estimate separate family components of variation for
each type, but this would greatly complicate the model. We therefore exclude these individuals
from the sample, giving a final sample size of 551 220 pupils.
We identify 5116 twin pairs (70% are the same sex) in the data corresponding to 9.37 twin
births per 1000 maternities. This is slightly less than the 1 1 .54 twinning rate that was seen in the
general population for 1990 and 1991, the calendar years when our sample of pupils were born
(Tables 3 and 4). One reason for this underestimate is our inability to identify twin pairs where
the twins live apart. A further reason could be that we cannot identify twin pairs where only
one twin attends a mainstream state school.
Table 3. Number of maternities with multiple births in England and Wales for the calendar
years 1990 and 1 991 1
|The source of data is the Office for National Statistics (http : / /www. s
uk/statbase/). The figures include maternities where live births and/or stillb
Numbers per thousand are given in parentheses. The data are taken from Table
Birth Statistics, historical series FMI, volume 13, Table 6.3 in Birth Statistics,
FMI, volumes 11-16, and Table 6.4 in Birth Statistics, historical series FMI, vo
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
666 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
Table 4. Distribution of maternities by age of mother separated for twin and non-twin
births in England and Wales for the calendar years 1990 and 1991 1
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 667
4. Statistical model
In this section we formulate a general model for pupils' test scores which contains
traditional school effectiveness model and the standard BG model. Our general
poses the total variability in pupils' test scores at secondary level into the parts that
LEAs, secondary schools, residential neighbourhoods, primary schools, families
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
668 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
Table 5. Summary statistics for the main variables of interest, reported separately
for twins and non-twinsf
age and ethnicity (pupil level) and/or school denomination or school proportion of female pupils
(school level). The terms Uj and e¿j are respectively the school level and pupil level random effects
which are assumed normally distributed, independent of one another and independent of any
predictor variables that are included in the model. Estimates of school effects are derived from
posterior predictions by empirical Bayes procedures (Goldstein, 2003).
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 669
where y¡j is the GCSE score for the /th pupil within theyth family. The model has three random
effects: Uj ; is a shared environmental effect that is common to all children in they'th family, e¿j is
an environmental child-specific random departure (non-shared environment) and g¡j is a genetic
effect for the /th child in the yth family, i.e. a child's genetic make-up contributes to that child's
response by an amount gif, the genetic effects are another type of child level residual. For two
children (/ 1 , ¿2) in the same family, the non-shared environmental covariance is cov(e/j j, e¿2j) = 0;
however, the genetic covariance for the same two individuals is cov(#/17, gi2j) = r{i'j,i2j)(J^ where
r0 1 j¿2J) *s ^e relationship coefficient between two individuals. The relationship coefficient equals
0.000, 0. 125, 0.250, 0.500 and 1 .000 for unrelated individuals, cousins, half-siblings, full siblings
or DZ twins, and MZ twins respectively. The correlations between MZ twins, full siblings or
DZ twins, half-siblings, cousins and unrelated individuals are therefore modelled as
(jI + vI
correlation between MZ twins = - =
*2 + <^ 0 + <^
(J2U+O.5(T2Q
correlation between full siblings or DZ twins = - =
al + 0.25a2Q
correlation between half-siblings = - =
al + 0A25a2Q
correlation between cousins = - =
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
670 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
4.4. Combining the school effectiveness model with the multilevel family model
The two-level school effectiveness model, equation (1), and the two-level family model adapted
for our data, equation (4), are both special cases of a more general multilevel model of achieve-
ment that partitions unexplained variations in test scores between several levels. We write this
general model of achievement, for the case of a single predictor, using the 'classification' nota-
tion of Browne et al (2001), which avoids a proliferation of subscripts when many random
classifications are present, as follows:
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children 's Educational Progress 67 1
GMGMt .0)- I
where y, is the test score for the /th pupil in the data set and jc, is their prior achievement and
there are N pupils. Equation (5) has random effects at six different classifications: pupil, fam-
ily, primary school, neighbourhood, secondary school and LEA. In models and results to be
discussed, the LEA refers to that of the secondary school that is attended by the pupil and
not necessarily that of the primary school. We number these classifications from 1 to 6. Hence,
the '(2)', '(3)', '(4)', '(5)' and '(6)' superscripts and subscripts identify any variables or random
effects that are associated with each of the classifications above the pupil level where the '(1)'
superscripts and subscripts are implicit for convenience of notation. The classification func-
tion secondary« denotes the secondary school that the /th pupil in the data set attends. The
units within the secondary school classification are indexed from 1 to 7(5) and «^secondar (/) ^s
the secondary school effect for the school that is attended by the /th pupil. The classification
functions and random effects for the other classifications are similarly defined. Random effects
are assumed normally distributed, independent across classifications and independent of any
predictor variables that are included in the model. family(/) denotes the family which child /
comes from and therefore ¿/famiiy(/) in the 'classification' notation of equation (5) is equivalent
to dj in the conventional multiple-subscript notation of equation (4).
Since classification notation does not show the multilevel structure in the data, 'classifica-
tion diagrams' are typically presented in addition to the model equation (Browne et al, 2001).
Classification diagrams have one node for each classification in the model. Two nodes con-
nected by an arrow indicate a nested relationship whereas two unconnected nodes indicate a
crossed relationship. Fig. l(a) depicts a classification diagram for the two-level hierarchy that
is assumed by equation (1), the school effectiveness model. The pupil and school classifica-
tions are represented by two separate nodes and the arrow from the pupil to the school node
indicates the nesting of pupils within schools. Fig. l(b) depicts the corresponding classifica-
tion diagram for the two-level hierarchy that is assumed by equation (4), the multilevel family
model adapted for our data. This diagram is drawn separately for twins and non-twins. Fig. l(c)
extends this to a six-classification model of pupils, families, primary schools, neighbourhoods,
secondary schools and LEAs, drawn separately for twins and non-twins. If the /th child is not a
twin (dfamiiy(i) = 0), then a five-classification model is fitted with pupils-families nested within
a three-way cross-classification of secondary schools by neighbourhoods by primary schools.
The arrow between the secondary school and LEA node indicates that secondary schools are
further nested within LEAs. However, the lack of a connecting arrow between neighbourhoods
and LEAs or between primary schools and LEAs indicates that neither neighbourhoods nor
primary schools are nested within LEAs; the relationship between LEAs, neighbourhoods and
primary schools is also crossed. The pupil-family node represents a single classification where
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
672 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
Secondary school
Pupil
(a)
Pupil-family | Fami|y
Pupil
(b)
Primary school Neighbourhood Secondary school Primary school Family Secondary school
Pupil-family Pupil
(C)
Fig. 1 . Classification diagrams for (a) the school effectiveness model, equation (1 ), (b) the multilevel family
model, drawn separately for non-twins (c/fam¡iy(/) = °) and twins (^famiiyi/) = 1 )> equation (4), and (c) the six-clas-
sification models of pupils, families, primary school, neighbourhood, secondary schools and LEAs, drawn
separately for non-twins (tffamiiy(/) = °) and twins (^famiiy(/) = 1)> equation (5)
we have only identified one pupil per family. This group is made up of pupils who have siblings
in different academic years, pupils who are only children, pupils who are unidentified twins
pairs and pupils who are siblings born in the same academic year. For all these cases we cannot
separate pupil and family effects, and so we estimate a composite pupil-family effect. If the
ith child is a twin (¿/family (i) = 1)> then a six-classification model is fitted with pupils now nested
within a three-way cross-classification of secondary schools by families by primary schools.
As before secondary schools are also further nested within LEAs, but now also families are
nested within neighbourhood and form an additional classification in the child's 'home envi-
ronment' hierarchy. The lack of connecting arrows between the family node and the primary
school, secondary school or LEA nodes indicates that two children from the same twin family
may attend different primary schools, different secondary schools and even different LEAs. For
further examples of the use of classification notation and classification diagrams for describing
complex data structures, see Browne et al (2001), Fielding and Goldstein (2006) and Rasbash
and Browne (2008).
The models that are described here include a single pupil level predictor, but further predictors
that are measured at any level or classification can be easily added to the model. These models
can also be easily extended to incorporate random slopes at one or more of the classifications.
For example, setting the coefficient of prior achievement to be random at the secondary school
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 673
classification will allow the effectiveness of schools to vary over the distribution of intake ability.
More generally, the models could be extended to include, for example, further classifications,
discrete responses and multivariate responses (Goldstein, 2003; Raudenbush and Bryk, 2002).
Where pupils change schools mid- way through an academic year, the end-of-year test score can
be seen as belonging to multiple schools and such information can also be potentially incorpo-
rated by allowing for multiple-membership structures (see Leckie (2009) and Goldstein et al
(2007) for recent examples). However, we do not pursue such extensions in this paper.
Estimation of cross-classified models by existing maximum likelihood approaches runs into
important computational limitations, especially when large numbers of units are involved
(Browne et al., 2001). As a result, in the analyses to follow, models are fitted by using Mar-
kov chain Monte Carlo based algorithms as implemented in the MLwiN package (Rasbash
et al., 2009). Starting values for the fixed parameters are estimated from the simpler two-level
model given by equation (1) by using a maximum likelihood approach, iterative generalized
least squares (Goldstein, 1986), in MLwiN. Further details of the Markov chain Monte Carlo
estimation methodology are given in Browne (2009). The data have a complex data structure
and many units at each classification. The models that we have proposed are slow to fit with
such data and so, during model exploration, we used a 10% random subsample of the non- twins
and only used the full sample for the final models. We note that the general pattern of results
was preserved when we moved from the subsample to the full sample.
5. Results
Table 6 presents resulting parameter estimates after fitting to the data a serie
labelled A, B, C and D. Table 7 presents a partition, into percentages, of the tot
iance attributable to the various classifications. In each model, the response is t
Since our focus is on pupils' progress during secondary schooling, rather th
achievement at the end of secondary schooling, we adjust for prior achieve
also adjust for being a twin by including a binary indicator variable as a main
interaction with the KS 2 score. These additional adjustments reflect that twin
than non-twins at GCSE and that they make more progress during seconda
Section 3.4).
Model A is the two-level school effectiveness model, equation (1), where pupils (level 1) are
treated as nested within secondary schools (level 2) (see Fig. l(a)). Model B is the simple two-
level family model, equation (4), where we have specified separate submodels for twins and non-
twins (see Fig. l(b)). In the twin submodel we partition unexplained variation in GCSE scores
between families (level 2) and between children within families (level 1). However, for non- twins
we observe only one child per family and so the child and family level are confounded. Thus, in
the non-twin submodel we estimate only a single overall variance. Model C is the more general
cross-classified model of achievement, equation (5), which partitions variation in unexplained
progress between six classifications: pupil, family, primary school, neighbourhood, secondary
school and LEA (see Fig. l(c)). As with model B, we have specified separate submodels for
twins and non-twins. Model D extends model C by additionally adjusting for a range of pupil
and neighbourhood level predictor variables. At the pupil level we adjust for gender, ethnicity
as a series of binary indicators of different ethnic groups, FSMs, age and SENs. At the LSOA
neighbourhood level we adjust for the IDACI-measure of social deprivation. The choice of
variables is essentially the full set available to us in our data. This set also closely matches those
employed in other studies in the school effectiveness literature which use our data (Goldstein
et ai, 2007; Leckie, 2009; Leckie and Goldstein, 2009).
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
674 J. Rasbash, G. Leckie, R. PHIinger and J. Jenkins
Table 6. Parameter estimates for models A-D|
Fixed part
Intercept -0.003 (0.00 1 ) -0.003 (0.00 1 ) 0.00 1 (0.008) -0.039 (0.007)
Twin 0. 1 79 (0.007) 0.177 (0.008) 0. 1 62 (0.007) 0. 1 54 (0.007)
KS 2 score 0.729 (0.001) 0.730 (0.001) 0.701 (0.001) 0.641 (0.001)
KS 2 score x twin 0.000 (0.007) -0.040 (0.007) -0.027 (0.006) -0.020 (0.009)
Within academic year age -0.012 (0.000)
Female 0.184(0.002)
FSMs -0.248 (0.003)
SENs -0.231 (0.003)
Ethnicity (reference, wh
Black 0.225 (0.006)
Asian 0.429 (0.005)
Chinese 0.556(0.015)
Mixed 0.045 (0.005)
Other 0.403 (0.010)
Neighbourhood deprivation (IDACI) -0. 103 (0.001)
Random part
Twins and non-twins
LEA 0.005 0.005
(0.004, 0.008) (0.003, 0.007)
Secondaryschool 0.065 0.043 0.035
(0.062, 0.069) (0.040, 0.045) (0.033, 0.037)
Primary school 0.035 0.025
(0.034 ,0.036) (0.025, 0.026)
Neighbourhood (LSOA) 0.008 0.002
(0.007, 0.008) (0.002, 0.002)
Pupil 0.402
(0.398, 0.405)
Twins
Family 0.238 0.168 0.157
(0.226, 0.251) (0.158, 0.179) (0.147, 0.167)
Pupil 0.160 0.157 0.150
(0.154, 0.166) (0.151, 0.163) (0.144, 0.156)
Non-twins
Pupil 0.468 0.383 0.357
(0.466, 0.470) (0.381, 0.384) (0.356, 0.359)
fAll models are fitted by using Markov chain Monte Carlo esti
of 450000. Standard deviations are given in parentheses for
credible intervals are given for the random-part parameter e
malized to have mean 0 and variance 1 . The age variable ranges in
youngest pupil in the academic year (born on August 31st) and
year (born on September 1st). A 1-unit change in the age varia
MLwiN screen shots of the model equation window for mo
bristol . ac . uk/ MLwiN /mo de Is . shtml.
In the fixed part of model A, the school effectiveness model, the results in Table 6 show t
twins perform significantly better than non-twins at GCSE and that, for both groups of pupil
1 -standard-deviation increase in the KS 2 score is associated with 0.729 of a standard deviation
increase in the GCSE score. This link between GCSE and prior achievement is very strong and
effectively changes the interpretation of all subsequent variables in the model from explaining
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 675
Twins
LEA 1.3 1.2
(0.8, 1.8) (0.8, 1.8)
Secondary school 14.0 10.3 9.5
(13.2, 14.7) (9.7, 10.9) (8.9, 10.0)
Primary school 8.5 6.8
(8.1,8.8) (6.5,7.1)
Neighbourhood (LSOA) 1.8 0.5
(1.7,2.0) (0.4,0.6)
Family 59.8 40.4 41.8
(58.0, 61.6) (38.6, 42.2) (40.0, 43.7)
Pupil 86.0 40.2 37.8 40.2
(85.3, 86.8) (38.4, 42.0) (36.2, 39.3) (38.5, 41.8)
Non-twins
LEA 1.1 1.1
(0.7,1.6) (0.7,1.5)
Secondary school 14.0 9.0 8.3
(13.2.14.7) (8.6,9.5) (7.9,8.8)
Primary school 7.4 6.0
(7.2, 7.7) (5.8, 6.2)
Neighbourhood (LSOA) 1.6 0.5
(1.5,1.7) (0.4,0.6)
Family
Pupil 86.0 100 80.8 84.1
(85.3.86.8) (100,100) (80.2,81.4) (83.6,84.6)
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
676 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
Model C, the general cross-classified model of achievement, which includes all six random
classifications, allows us to address the two key research questions that we posed in Section 1.
One of our research questions asked how much of the variation in progress during secondary
schooling is attributable to the wider shared environment beyond the immediate family. We
define school, area and LEA as the wider shared environment. This is broadly true as 94% of
twin pairs attend the same secondary school, 98% attend the same primary school, 100% come
from the same LSOA and 99% attend either the same secondary school or secondary schools
within the same LEA. Comparing the results for twins in model B with model C we see that the
introduction of the effects for the wider environment beyond the family substantially reduces
the shared family variance from 0.238 to 0.168 and in a model with the same fixed effects is
now 40.4% of effectively the same total residual variance rather than 59.8% and only margin-
ally reduces the between-pupil within-family variance (from 0.160 to 0.157), indicating that the
wider environmental effects can be thought of as shared between twins, and thus confirming our
definition of these wider environments as wider shared environments. The variance partition
results for model C in Table 7 show that for twins 10.3% is attributable to the secondary school,
1.8% is attributable to the LSOA neighbourhood, 1.3% is attributable to the LEA and 8.5% is
attributable to carry-over effects from primary school. Therefore 21.9% of the total variation
in twins' progress is attributable to these wider shared environments. From a BG perspective
these are substantial effects although smaller than the composite family effect for twins which
accounts for 40.4% of variation in educational progress. The remaining 37.8% of the variance
for twins is at the pupil level. This variance is actually a composite of three effects: test measure-
ment error, non-shared genetic effects and non-shared environmental effects. The responses are
based on combining scores across the pupils' public examination grades and we do not have
reliability estimates for these measures. However, allowing for any measurement error reduces
the pupil level variance and hence will increase variance partition coefficients (the percentage
of variance that is attributable to a level) for all the other levels in the model.
Another research question asks how much of what had been thought of as pupil level
variation in school effectiveness models is attributable to family effects. Model C shows that for
twins 40.4% of the total variation is at the family level and 37.8% is at the pupil level. If we omit
family from this model the family component becomes reattributed to the pupil level so from
this study, for twins, around 50% (40.4/(40.4 + 37.8)) of what has been thought of in school
effectiveness research as individual level variation is really due to family level factors.
The 86.0% of the variance due to pupil level for twins in model A is reduced by more than a
half to 37.8% in model C and these figures are directly comparable since the models have the
same set of fixed effects. Most of this reallocated variance (40.4% of the total variance) now lies
at the family level, but a sizable amount (8.5%) lies at the primary school level, and very small
amounts are due to area (LEA and LSOA neighbourhood). Model C suggests that model A
wrongly attributed a small amount of variation at the LEA, primary school, LSOA neighbour-
hood and family levels to secondary school, since the percentage due to secondary school has
also reduced, from 14.0% in model A to 10.3% in model C.
The results of model C indicate that models A and B incorrectly attribute unexplained varia-
tion in progress associated with omitted levels to the levels that are present. This suggests that
previous studies that fitted two-level school effectiveness models overestimate pupil level effects,
whereas previous studies that fitted two-level family models overestimate the effect of family
membership.
In summary, model C shows that for twins the largest components of variation in pupils'
unexplained GCSE scores after adjusting for prior achievement and being a twin are, in decreas-
ing order of magnitude: family, pupil, secondary school and primary school. Including the
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 677
family level in the model shows that the pupil component, for twins, is considerably smaller
than previously thought. The amounts of variation due to family and pupil are roughly equal
and together account for around 80% of the total unexplained variation, whereas the amounts
due to secondary and primary school are again roughly equal and account for most of the
remaining 20% of variation. Area (LEA and LSOA neighbourhood) accounts for a very small
amount of the unexplained variation: less than 2%.
Model D adjusts for a fuller set of pupil and neighbourhood level predictor variables: gender,
ethnicity as a series of binary indicators of different ethnic groups, FSM, age, SEN and IDACI.
We do this for two reasons. The first is that we want to exclude the portions of the components of
variation in progress at the LEA, secondary school, primary school and LSOA neighbourhood
levels which are due to different schools or areas having different compositions of pupils with
respect to gender, ethnicity, FSM, age and SEN. This is a check of robustness; we would like to
verify that the variation at these levels is not purely due to variations in composition of individ-
ual and family level variables: that there are genuine school and area effects on progress. The
second reason is that we are interested to see how much of the variation at the family and pupil
levels can be explained by gender, ethnicity, FSM, age and SEN and how much remains unex-
plained and thus due to other factors which are unmeasured in our data set. Similarly, we are
interested to see how much of the variation at the LSOA neighbourhood level can be explained
by IDACI. The estimated coefficients of these predictor variables are very similar to those seen in
the literature (see, for example, Goldstein et al. (2007), Leckie (2009) and Leckie and Goldstein
(2009)): girls and younger pupils make significantly greater progress than boys and older pupils.
Those eligible for FSMs and those with SENs make significantly less progress than non-FSM
and non-SEN pupils whereas all ethnic groups, particularly Chinese pupils, make considerably
more progress than white pupils. Finally, those living in more deprived neighbourhoods make
less progress.
Model D gives similar findings for our two research questions to those of model C. In terms of
our first research question, model D shows that, for twins, 9.5% of variation in progress during
secondary schooling is attributable to secondary school, 6.8% is attributable to carry-over effects
from primary school, 0.5% is attributable to the LSOA neighbourhood and 1 .2% is attributable
to the LEA; therefore a total of 18.0% of variation in learning progress is attributable to these
wider shared environments after conditioning on demographic and social factors. In terms of
our second research question, model D shows that, for twins, 41.8% of the total variation is at
the family level, so, for twins, somewhere around 50% of what has been thought of in school
effectiveness research as individual level variation is really due to family level factors.
If we compare model C with model D, we can see that, whereas the component of variation at
the LEA level remains virtually unchanged after controlling for these additional covariates, the
component of variation at each of the other levels has decreased a little. This indicates that these
additional covariates were responsible for some of the variation at these levels that we found in
model C. Proportionately the biggest decrease is in the LSOA neighbourhood level variation:
this has gone from 0.008 in model C to 0.002 in model D, so almost all the variation at the
neighbourhood level that we observed in model C was due to differences in neighbourhoods
with respect to these additional covariates. Just one of these covariates, IDACI, is measured at
the neighbourhood level. The other covariates are pupil level or family level, so any reduction
in the neighbourhood level variance due to introducing these would indicate that a correspond-
ing amount of the variance at the neighbourhood level before introducing the covariates came
about because different neighbourhoods have different make-ups of pupils with respect to these
covariates (which influence pupils' progress), rather than because the neighbourhood that a
pupil lives in has an influence on their progress.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
678 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
The same is true of the reduction in variation at the secondary and primary school levels in
model D compared with model C: we have not included any school level variables, primary or
secondary, so these portions of the variances at the school levels that we found in model C must
be due to differences in composition of schools, with respect to pupil, family and LSOA neigh-
bourhood level variables, rather than to school effects. However, there remain non-negligible
amounts of variation at both the primary and the secondary school levels in model D, which
could be due to other pupil, family and area level variables that were not included, or could be
real school effects (or could be a little of both).
The family and pupil level variation both decrease with the introduction of the additional
covariates. Some of the variation at these levels that we found in model C was thus due to gen-
der, ethnicity, FSM status, age, SEN status and IDACI. However, there remains considerable
variation at both these levels: most of the influence of family and pupil on progress remains
unexplained.
It is interesting to note that, after we control for this set of covariates mostly measured at the
family or pupil level, whereas the amounts of variation at the family and pupil levels decrease
(as we would expect), the proportion of variation at these levels increases (for both twins and
non-twins). This indicates that although the introduction of the set of covariates has reduced
the amount of variation at all levels, it has made (proportionately) more of a reduction for the
secondary, primary and LSOA neighbourhood levels than for the family and pupil levels. This
result is due to the uneven spread of pupils, in terms of their individual and family characteristics,
across schools and across neighbourhoods.
6. Discussion
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 679
The largest variance component is the composite family effect. The question arises how much
of this is attributable to shared genetic factors and how much to the shared environmental effects
of the immediate family? We can look to the literature for some indications here. BG studies,
covering a wide range of developmental traits (but most frequently looking at IQ and other
specific cognitive abilities such as verbal, spatial, processing speed and memory), consistently
find high effects for genetics and the non-shared environment and little evidence for effects of
the shared environment. For example, surveying the field, Plomin (1986) stated
'Environmental influences that affect development are not shared by family members'.
According to these findings we would expect most of the family effect to be genetic. However,
we note that our finding of a substantial variance component for the wider shared environ-
ment beyond the immediate family for learning progress contradicts these BG findings of the
lack of importance of shared environmental factors. We also note that, of the 12 studies that
were presented in Table 1 which partitioned various measures of achievement, seven found the
genetic component to be larger than the shared environmental component, five found the shared
environmental component to be larger than the genetic component and one found these com-
ponents to be almost equal. These studies thus do not give such a clear picture of large genetic
effects and non-existent shared environmental effects. The shared environmental effect in these
studies incorporates school and area effects as well as family effects, but eight of the 12 studies
estimated the shared environmental effect as considerably higher than the 20% or so that the
present analysis and the findings of school effectiveness research suggest is the maximum likely
to be due to school and area effects. Therefore, in broadbrush terms, accounting for likely upper
values for school and area variance partition coefficients (the broader shared environment), the
studies in Table 1 suggest that significant variation is still attributable to the immediate shared
family environment.
We have not analysed absolute educational achievement as, from a school effectiveness point
of view, this is not the most socially relevant measure. What matters more from a social point
of view is the progress that children make during schooling, which we have operationalized as
progress over the secondary schooling period. Our analysis of learning progress assesses only
the composite family effect. However, if it were possible to partition this we might expect similar
patterns to those for the studies for achievement in Table 1 . If this were so then our composite
family effect may have an appreciable shared family environment component.
From a school effectiveness point of view it is interesting that, for twins, the largest single
variance component is the composite family effect.
An important question is to what extent these findings can be generalized to non-twins. Given
that non-twin sibling correlations on cognitive traits are very substantial but lower than for twins
(see Table 2) we would expect to find a reduced but still substantial composite family effect for
non-twin siblings. Our expectation is that the effect of the non-family shared environment will
be similar for twin and non-twin siblings as our results show that the percentage of variation
for twins and non-twins that is attributable to these levels does not differ.
To separate out the composite family effects into shared family environment and shared
genetic effects (from the BG perspective) while simultaneously taking account of the broader
environment and to address questions of generalizability to non-twin siblings, we need a richer
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
680 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
data set. We require data with multiple children per school and multiple children per family and a
genetically informative design, i.e. a design which includes families with differential genetic relat-
edness compositions, e.g. MZ and DZ twins, full siblings, half-siblings and unrelated siblings.
Where researchers have such data sets available we would recommend that they consider the
use of cross-classified multilevel modelling as an analytical technique to progress these research
questions.
Acknowledgements
We are grateful for the very helpful and detailed comments that were provided by the three
referees and the Joint Editor. This work was funded under the UK Economic and Social Research
Council's National Centre for Research Methods programme.
References
Aitkin, M. and Longford, N. (1986) Statistical modelling issues in school effectiveness studies (with dis
J. R. Statist. Soc. A, 149, 1-42.
Bouchard, T. J. and McGue, M. (1981) Familial studies ol intelligence- a review. Science, 212, 1055-10
Bronfenbrenner. U. (1911) Toward an exoerimental ecoloev of human-develooment. Am. PsvchoL. 32. 513-531.
Brooks, A., Fulker, D. W. and Defries, J. C. (1990) Reading performance and general cognitive-ability - a mult
variate genetic-analysis of twin data. Persnlty Individ. Diff., 11, 141-146.
Browne, W. J. (2009) MCMC Estimation in MLwiN v2.13. Bristol: Centre for Multilevel Modelling.
Browne, W. J., Goldstein, H. and Rasbash, J. (2001) Multiple membership multiple classification (MMM
models. Statist. Modllng, 1, 103-124.
Cleveland, H. H., Jacobson, K. C, Lipinski, J. J. and Rowe, D. C. (2000) Genetic and shared environmental contr
butions to the relationship between the home environment and child and adolescent achievement. Intelligenc
28, 69-86.
Deary, J. J. (2006) Educational performance in twins. Br. Med. J., 333, 1080-1081.
Deary, I. J., Strand, S., Smith, P. and Fernandes, C. (2007) Intelligence and educational achievement. Intelligen
35, 13-21.
Devlin, B., Daniels, M. and Roeder, K. (1997) The heritability of IQ. Nature, 388, 468-471.
Duncan, G. J., Boisjoly, J. and Harris, K. M. (2001) Sibling, peer, neighbor, and schoolmate correlations as
indicators of the importance of context for adolescent development. Demography, 38, 437-447.
Fielding, A. and Goldstein, H. (2006) Cross-classified and multiple membership structures in multilevel models:
an introduction and review. Research Report RR791. Department for Education and Skills, London. (Available
from http: //www. dcsf.gov.uk/research/data/uploadfiles/RR791.pdf.)
Fielding, A., Thomas, H., Steele, F., Browne, W., Leyland, A., Spencer, N. and Davison, I. (2006) Using cross-
classified multilevel models to improve estimates of the determination of pupil attainment: a scoping study.
Research Report. School of Education, University of Birmingham, Birmingham. (Available from http: //
www.socscistaf f .bham.ac .uk/f ielding/Df es-Scoping.report-School-of _Ed_ version.
Augusto 6 .pdf.)
risher, R. A. (19 18) lhe correlation between relatives on the supposition of Mendelian inheritance. Trans. R.
SocEdinb., 52, 399^33.
Friend, A., DeFries, J. C, Wadsworth, S. J. and Olson, R. K. (2007) Genetic and environmental influences on
word recognition and spelling deficits as a function of age. Behav. Genet., 37, 477-486.
Garner, C. L. and Raudenbush, S. W. (1991) Neighborhood effects on educational attainment: a multilevel
analysis. Sociol Educ, 64, 251-262.
Georgiades, K., Boyle, M. H., Jenkins, J. M., Sanford, M. and Lipman, E. (2008) A multilevel analysis of whole
family functioning using the McMaster family assessment device. J. Famly Psychol, 22, 344-354.
Goldstein, H. (1986) Multilevel mixed linear model analysis using iterative generalized least squares. Biometrika,
73, 43-56.
Goldstein, H. (2003) Multilevel Statistical Models, 3rd edn. London: Arnold.
Goldstein, H., Burgess, S. and McConnell, B. (2007) Modelling the effect of pupil mobility on school differences
in educational achievement. 1 R. Statist. Soc. A, 170, 941-954.
Goldstein, H., Rasbash, J., Yang, M., Woodhouse, G., Pan, H., Nuttall, D. and Thomas, S. (1993) A multilevel
analysis of school examination results. Oxf. Rev. Educ, 19, 425-433.
Goldstein, H. and Sammons, P. (1997) The influence of secondary and junior schools on sixteen year examination
performance: a cross-classified multilevel analysis. Schi Effectiv. Schi Improvmnt, 8, 219-230.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
Children's Educational Progress 681
Guo, G. and Wang, J. M. (2002) The mixed or multilevel model for behavior genetic analysis. Behav. Genet., 32,
37^9.
Haworth, C. M. A., Kovas, Y., Dale, P. S. and Plomin, R. (2008) Science in elementary school: gener
and school environments. Intelfaence. 36. 694-701.
Haworth, C. M. A., Kovas, Y, Petrill, S. A. and Plomin, R. (2007) Developmental origins of low mathematics
performance and normal variation in twins from 7 to 9 years. Twin Res. Hum. Genet., 10, 106-1 17.
Jenkins, J. M., Dunn, J., O Connor, T. G., Rasbash, J. and Behnke, P. (2005) Change in maternal perception ot
sibling neeativitv: within- and between-familv influences. J. Famlv PsvchoL, 19, 533-541.
Jenkins, J. M., Rasbash, J. and O'Connor, T. G. (2003) The role of the shared family context in differential
parenting. Devlpmntl PsychoL, 39, 99-113.
Jenkins, J., Simpson, A., Dunn, J., Rasbash, J. and O Connor, T. G. (2005) Mutual influence of marital conflict
and children's behavior problems: shared and nonshared family risks. Chid Devlpmnt, 76, 24-39.
Kempthorne, O. (1997) Heritability: uses and abuses. Genetica, 99, 109-112.
Kovas, Y, Hayiou-Thomas, M. E., Oliver, B., Dale, P. S., Bishop, D. V. M. and Plomin, R. (2005) Genetic influ-
ences in different aspects of language development: the etiology of language skills in 4.5-year-old twins. Chid
Devlpmnt, 76, 632-651.
Leckie, G. (2009) The complexity of school and neighbourhood effects and movements of pupils on school
differences in models of educational achievement. J. R. Statist. Soc. A, 172, 537-554.
LecKie, u. ana uoiastein, n. {¿vw) i ne limitations oi using scnooi league taoies to ìniorm scnooi cnoice. j. k.
Statist. Soc. A, 172, 835-851.
Lynch, S. K., Turkheimer, E., D'Onofrio, B. M., Mendie, J., Emery, R. E., Slutske, W. S. and Martin, N
A genetically informed study of the association between harsh punishment and offspring behavioral
J. Famly PsychoL, 20, 190-198.
McGue, M., Keyes, M., Sharma, A., Elkins, I., Legrand, L., Johnson, W. and lacono, W. G. (2007)
ronments of adopted and non-adopted youth: evidence on range restriction from the Sibling Interac
Behavior Study (SIBS). Behav. Genet., 37, 449-462.
O Connor, T. G, Dunn, 1, Jenkins, J. M., Pickering, K. and Rasbash, J. (2001) Family settings and ch
adjustment: differential adjustment: within and across families. Br. J. Psychiatr., 179, 1 10-115.
van den Oord, E. (2001) Estimating effects of latent and measured genotypes in multilevel models. Stat
MedRes., 10, 393^07.
van den Oord, E. and Rowe, D. C. (1997) An examination of genotype-environment interactions for academic
achievement in an US national longitudinal survey. Intelligence, 25, 205-228.
Pam, A., Kemker, S. S., Ross, C. A. and Golden, R. (1996) The "equal environments assumption in MZ-DZ
twin comparisons: an untenable premise of psychiatric genetics? Acta Genet. Med. Gemell, 45, 349-360.
Petrill, S. A. and Thompson, L. A. (1993) The phenotypic and genetic-relationships among measures of cogni-
tive-ability, temperament, and scholastic achievement. Behav. Genet., 23, 511-518.
Petrill, S. A. and Thompson, L. A. (1994) The effect of gender upon heritability and common environmental
estimates in measures of scholastic achievement. Persnlty Individ. Diff., 16, 631-640.
Plomin, R. (1986) Development genetics and psychology. In Development, Genetics, and Psychology. Hillsdale:
Erlbaum.
Fiomin, K. (lyyy; uenetics ana generai cognitive aoiiity. nature, ¿w¿, <^z:>-^zy.
rlomin, K., JJernes, J. C, Craig, 1. W. ana JvlcUumn, r. (/UU3) Behavioural Uenetics in we rost-genomic hra.
Washington DC: American Psychological Association.
Plomin, R., Dehnes, J. C, MacClearn, Ci. b. and Mcuumn, F. (2UU8) Behavioural Uenetics, Mh edn. JNew York:
Worth.
Plomin, R. and Petrill, S. A. (1997) uenetics and intelligence: what s new/ Intelligence, 24, 55-1 1.
Fosthuma, D., De ueus, b. J., Bleicnrodt, N. and tfoomsma, u. 1. (ZUUU) iwin-smgleton ditlerences in intelligence/
Twin Res., 3, 83-87.
Provine, W. B. (2001) The Origins oj Theoretical Population Uenetics. Chicago: University ot Chicago Press.
Rasbash, J. and Browne, W (2001) Modelling non-hierarchical structures. In Multilevel Modelling of Health
Statistics (eds A. H. Leyland and H. Goldstein). Chichester: Wiley.
Rasbash, J. and Browne, W. J. (2008) Non-hierarchical multilevel models. In Handbook of Multilevel Analysis (eds
J. De Leeuw and E. Meijer). New York: Springer.
Rasbash, J. and Goldstein, H. (1994) Efficient analysis of mixed hierarchical and cross-classified random structures
using a multilevel model. J. Educ. Behav. Statist., 19, 337-350.
Rasbash, J., Jenkins, J., O'Connor, T. G, Reiss, D. A. and Tackett, J. (2010) A social relations model of
observed family negativity and positivity using a genetically-informative sample. J. Persnlty Sod PsychoL,
to be published.
Rasbash, J., O'Connor, T. and Jenkins, J. (2003) Multilevel models for family data. Centre for Multilevel Model-
ling, University of Bristol, Bristol. (Available from http://www.crran.bristol.ac.uk/research/
Family/ jrgene tics .pdf.)
Kasbasn, J., Meele, t., Browne, w. J. ana uoiastein, n. {¿vvy) a users uuiae to mlwu' v/.w. tsnstoi: centre
for Multilevel Modelling.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms
682 J. Rasbash, G. Leckie, R. Pillinger and J. Jenkins
Räuden bush, S. W. ( 1 993) A crossed random effects model for unbalanced data with applications in cross-sectional
and longitudinal research. J. Educ. Behav. Statist., 18, 321-349.
Raudenbush, S. W. and Bryk, A. S. (2002) Hierarchical Linear Models: Applications and Data Analysis Methods,
2nd edn. New York: Saee.
Rescorla, L. and Rosenthal, A. S. (2004) Growth in standardized ability and achievement test scores from 3rd to
10th erade. 1 Educ. PsvchoL. 96. 85-96.
Thompson, L. A., Detterman, D. K. and Plomin, R. (1991) Associations between cognitive-abilities and sc
achievement - genetic overlap but environmental differences. Psychol Sci., 2, 158-165.
Thompson, L. A., Detterman, D. K. and Plomin, R. (1993) Differences in heritability across groups diffe
ability, revisited. Behav. Genet., 23, 331-336.
Wainwright, M. A., Wright, M. J., Geffen, G. M., Luciano, M. and Martin, N. G. (2005) The genetic
academic achievement on the Queensland core skills test and its shared genetic variance with IQ. Behav. Gen
35, 133-145.
Williamson, G. L., Appelbaum, M. and Epanchin, A. (1991) Longitudinal analyses of academic-achieve
J. Educ. Measmnt, 28, 61-76.
Xue, Y. G., Leventhal, T, Brooks-Gunn, J. and Earls, F. J. (2005) Neighborhood residence and menta
problems of 5- to 11 -year-olds. Arch. Gen. Psychiatr., 62, 554-563.
Yang, M. and Woodhouse, G. (2001) Progress from GCSE to A and AS Level: institutional and gender diffe
and trends over time. Br. Educ. Res. J., 27, 245-267.
This content downloaded from 158.109.94.211 on Wed, 23 Jan 2019 11:58:32 UTC
All use subject to https://about.jstor.org/terms