Child Development, November/December 2005, Volume 76, Number 6, Pages 1217 – 1233

Analysis and Interpretation of Twin Studies Including Measures of the

Shared Environment
Eric Turkheimer and Brian M. D’Onofrio Hermine H. Maes and Lindon J. Eaves
University of Virginia Virginia Commonwealth University

Recent reports using a classical behavior genetic research design in which twin data are combined with a
measured characteristic of their shared family environment have made striking claims about estimating envi-
ronmental influences on behavior with genetic effects controlled. Such claims are overstated for two related
reasons. First, when a variable is measured at the family level in a way that makes it necessarily equivalent for
twins reared together, it is not possible to partition it into genetic and environmental components. Second,
although structural equation modeling and DeFries – Fulker analysis are sound tools for the analysis of many
types of twin data, they do not control for genetic or environmental confounds when estimating the effect of
measured family-level variables.

Although it has long been promised that behavior appear to apply to entire families, like socioeconomic
genetic studies will help elucidate salient environ- status (SES) or martial status, may be differentially
mental influences on human variation (Heath, Ken- experienced by children in the same family. In many
dler, Eaves, & Markell, 1985; Plomin, 1994; Reiss, instances, however, variables are only measured at
Plomin, & Hetherington, 1991), the most frequent the level of the family, and therefore can only vary
conclusion reached about the environment on the between families, not within them. Our review is
basis of genetically informed studies is that envi- motivated by both older and more recent articles that
ronmental factors shared by siblings do not influence have added such family-level variables to twin
children at all (Harris, 1998; Plomin & Daniels, 1987; studies. Recently, some of these studies have claimed
Rowe, 1994; Turkheimer, in press; see Rutter, 2000, to estimate the environmental influence of a mea-
for a balanced review of the controversy). sured family-level variable on an outcome in twin
Previous papers have reviewed the limited suc- children, controlling for genetic influences on the
cess of studies attempting to specify nonshared en- same outcome.
vironmental influences (Turkheimer & Waldron, Whereas an earlier review of the nonshared en-
2000). This article explores the potential and limita- vironment was concerned primarily with the em-
tions of twin studies for the exploration of shared pirical outcome of studies including a measured
environmental factors. By ‘‘shared,’’ we simply mean nonshared environmental variable (Turkheimer &
environmental factors that are jointly experienced by Waldron, 2000), in this paper we demonstrate that
siblings raised together, although as we have noted twin studies including a family-level variable (which
previously (Turkheimer & Waldron, 2000), there are we refer to as the measured C design, C referring to
significant definitional problems in understanding the shared environmental term in traditional twin
shared and nonshared environment. Any purported models) cannot provide estimates of environmental
environmental variable can contain both shared and effects unbiased by genetic factors. Based on our
nonshared variability. Even variables that would review, we specify what twin studies including
measures of the shared environment can and cannot
do to elucidate environmental processes, and illus-
Preparation of this article and the analyses were supported by trate our conclusions with analyses of twin data from
grants from the National Institute of Mental Health (MH67300) the Virginia 30,000 data set. We conclude with rec-
and William T. Grant Foundation. Data collection for the Virgina ommendations for future research designs and ana-
30,000 data set was supported by Grants GM-30250, AG-04954, lytic methods.
AA-06781, MH-40828, and HL-48148 from the National Institutes
of Health and a gift from RJR Nabisco. Because some of our comments will be quite
Correspondence concerning this article should be addressed to critical, we should be very explicit about the scope of
Eric Turkheimer, Department of Psychology, University of Vir- our concerns. Our comments are limited to the par-
ginia, P.O. Box 400400, Charlottesville, VA 22904-4400. Electronic
mail may be sent to ent3c@virginia.edu. A commentary on this
paper written by A. Caspi, A. Taylor, and S. Jaffe is available from r 2005 by the Society for Research in Child Development, Inc.
Dr. Caspi at a.caspi@iop.kcl.ac.uk. All rights reserved. 0009-3920/2005/7606-0007
1218 Turkheimer, D’Onofrio, Maes, and Eaves

ticular subtype of twin analysis we have already (e.g., Neale & Cardon, 1992). Double-headed arrows
described, in which a classical twin analysis is represent covariances and single-headed arrows
combined with a measured shared environmental specify regressions of one variable on another. The
variable. Moreover, we do not intend to suggest that squares, T1 and T2, represent the phenotypic mea-
either this research design or the usual statistical sures of the twins. The A latent variable represents
methods of analyzing its results are fundamentally additive genetic effects. Therefore, the parameter
flawed. Rather, the difficulties arise in the interpre- connecting the twins’ additive genetic variance
tation and generalization of the results, which in components is set at 1.0 for the MZ twins and 0.50 for
some recent cases have been insufficiently conserv- the DZ twins. The parameter a represents the influ-
ative. Another treatment of methodological issues in ence of genetic component on the phenotype. The C
this design can be found in Purcell and Koenen (in latent variable represents the shared environmental
press). More basic reviews of the advantages and component, and its path coefficient c influences both
disadvantages of genetically informative designs for twins to the same extent. E denotes the nonshared
the study of purported environmental factors can be environment. The e path estimate represents the in-
found elsewhere (D’Onofrio et al., 2003; Eaves, Last, fluence of environmental variation that is unique to
Young, & Martin, 1978; Rutter, Pickles, Murray, & each twin. Twin studies commonly report the pro-
Eaves, 2001). portion of the total phenotypic variance accounted
for by each variable. Therefore, the influence of ge-
netic factors (a2) is referred to as the heritability, the
Standard Twin Design
influence of c2 is the shared environmental influence,
First, we will very briefly review the standard twin and e2 is the nonshared environmental influence. The
model. Classical twin studies compare the similarity path model makes the critical assumption that the
of identical (monozygotic, MZ) twins and fraternal effects of genes and the shared environment are in-
(dizygotic, DZ) twins. MZ twins share all of their dependent. If they are correlated (‘‘passive geno-
genes and, on average, DZ twins share half of their type – environment correlation’’), estimates of c2 are
genes. Therefore, to the extent that genetic factors biased in studies of twins reared together (Jinks &
influence a trait, then MZ twins will be more similar Fulker, 1970).
than DZ twins. By comparing the covariation among Early twin studies were conducted to determine if
and between MZ and DZ twins, basic twin models genetic factors influenced traits or behaviors. Over
partition the variance of a measured trait, called a the last quarter century, researchers have illustrated
phenotype, into additive genetic, shared environ- that genes influence most, if not all, behaviors and
mental, and nonshared environmental components traits (Plomin, DeFries, McClearn, & McGuffin, 2000;
(Eaves, 1982). Turkheimer, 2000) and have demonstrated the ubiq-
Figure 1 is an example of the structural equation uitous importance of the nonshared environmental
model (SEM) for a basic univariate twin analysis variance component, in contrast to the relatively

Figure 1. Structural equation model for the standard twin design.

 Twin Studies with Measured Environments 1219

smaller influence of the shared environmental com- shared environmental measure. One would not want
ponent (Daniels & Plomin, 1985; Dunn & Plomin, to conclude, however, that SES was in general a
1990; Turkheimer, 2000; Turkheimer & Waldron, completely environmental variable, as it is well
2000). Apparently, most environmental influences known that most purported measures of the envi-
cause siblings to be less alike. ronment include genetic variability (Plomin &
Bergeman, 1991); twin children are just not geneti-
cally informative about their parents’ SES. The
Twin Design With Measured Shared Environments measured C model can also be parameterized with
an arrow between the measured environment and
Figure 2 is an example of a model used to study a
the latent shared environmental factor, but the fit of
measured C twin design (Neale & Cardon, 1992).
the two models is exactly the same.
The model is equivalent to the standard twin model
described above, except that it includes a specific
measure of the family environment (Cf, the f sub-
Studies Reporting SEMs of the Measured C Design
script noting that it is a family-level variable, jointly
describing both twins). Note that assigning the Kendler, Neale, Kessler, Heath, and Eaves (1992a)
measured variable the status of ‘‘shared environ- were the first to include a measured family-level
mental’’ is somewhat arbitrary. In designs such as we variable into a SEM of twins. (A similar model is
are considering here, twins are necessarily perfectly referred to as the age correction model in Neale &
concordant for the variable in question, not as a Cardon, 1992.) The analyses included childhood
matter of its actual composition but as a simple parental loss, either through separation or death, as a
consequence of the research design. (It is also pos- ‘‘specified’’ shared environment in their univariate
sible to include covariates that differ within families, twin analyses of various adult psychological disor-
such as a measure of parental treatment. These ders, including depression, generalized anxiety dis-
studies present somewhat different issues and will order, and panic disorder. When basic twin models
be discussed later.) Parental SES, for example, is (Figure 1) were fit without the Cf measure, there was
necessarily the same for members of a twin pair as- no significant influence of the latent shared envi-
sessed at the same point in time, so within the lim- ronmental variable on the adult disorders (i.e., the
ited context of a study of twin children, it is a purely confidence intervals around the estimate for the

Figure 2. Structural equation model of twin design that includes a measured family-level factor.
1220 Turkheimer, D’Onofrio, Maes, and Eaves

shared environment [c] included zero). For example, partially genetically mediated’’ (p. 1102) in the dis-
a twin model of major depression that dropped the cussion section of the manuscript. In contrast, the
shared environmental parameter did not result in a abstract stated, ‘‘a multivariate model showed that
significant loss in fit compared with the full twin adult domestic violence accounted for 2% and 5% of
model (Kendler, Neale, Kessler, Heath, & Eaves, the variation in children’s internalizing and exter-
1992b). However, when the parental loss variables nalizing problems, respectively, independent of genetic
were included as specified shared environmental effects’’ (p. 1095, italics added).
variables (Figure 2), the parameters associated with Thapar et al. (2003) used the measured C design to
the parental loss were statistically significant for explore the association between maternal smoking
major depression, generalized anxiety disorder, and during pregnancy and attention deficit hyperactivity
panic disorder. The authors concluded, ‘‘A model disorder (ADHD). The parameter associated with
that includes parental loss as a form of ‘specified’ smoking during pregnancy was statistically signifi-
family environment shows that, if it is truly an envi- cant. The discussion includes a careful description of
ronmental risk factor for adult psychopathological condi- the limitations of the analysis: ‘‘. . .even twin or
tions, it can account for between 1.5% and 5.1% of the adoption designs cannot be used to test whether
total variance in liability to these disorders. . .’’ maternal smoking during pregnancy has a truly
(p. 109, italics added). Note that this conclusion causal relationship with offspring ADHD symptoms,
draws attention to the essential assumption that the independent of genetic factors, not even where maternal
specified variable is purely environmental. The arti- ADHD is assessed. . . . Thus, we are careful in stating
cle went on to discuss how the association between that we observe an association between maternal
parental loss and the adult conditions could be me- smoking during pregnancy and offspring ADHD
diated by other environmental (family dynamics symptoms and do not conclude that this necessarily
correlated with parental loss) or genetic (predispo- implies causality’’ (p. 1988, italics added). The ab-
sition to poor marital processes in the parents and stract, however, makes a much stronger claim:
psychiatric problems in the offspring) factors. ‘‘Maternal smoking during pregnancy appears to
As time has gone by, however, researchers have show an association with offspring ADHD symp-
become less assiduous in their attention to the am- toms that is additional to the effects of genes and not
biguity of the shared environmental status of mea- attributable to [other confounds]’’ (p. 1985, italics
sured family-level variables in twin studies. Caspi, added).
Taylor, Moffitt, and Plomin (2000) utilized a similar In a study of the relation between domestic vio-
model to explore the influence of socioeconomic lence and IQ in young children, Koenen, Moffitt,
variables on emotional and behavioral problems in Caspi, Taylor, and Purcell (2003) utilized a measured
2-year-old twins. In this case, basic twin models il- C model to demonstrate a significant relation be-
lustrated that the shared environmental latent vari- tween domestic violence and children’s intelligence.
able accounted for 20% of the variance in behavior In the abstract, the authors claim, ‘‘Structural equa-
problems. The researchers then included a composite tion models showed that adult domestic violence
index of measured neighborhood factors, which ac- accounted for 4% of the variation, on average, in
counted for 5% of the variance in behavioral prob- child IQ, independent of latent genetic influences’’
lems. Although it was noted in the discussion that (p. 297, italics added). In this case, no mention was
the approach assumes that the measure of socioeco- made of the assumption that the measure of
nomic deprivation is a pure environmental variable, adult domestic violence was a purely environmental
the abstract stated, ‘‘A nationwide study of 2-year- variable.
old twins shows that children in deprived neigh- Finally, Kim-Cohen, Moffitt, Caspi, and Taylor
borhoods were at increased risk for emotional and (2004) investigated the influence of stimulating ac-
behavioral problems over and above any genetic liabil- tivities in the home on a measure of cognitive resil-
ity. The results suggest that the link between poor ience. The phenotype was a measure of cognitive
neighborhoods and children’s mental health may be resilience to SES (residuals cognitive ability after SES
a true environmental effect. . .’’ (p. 338, italics added). had been partialed). Although the authors stated, ‘‘A
Jaffee, Moffitt, Caspi, Taylor, and Arsenaeault caveat is in order. Although we assume that stimu-
(2002) used a measured C model to study the relation lating activities is an environmental variable, it is
between domestic violence and internalizing and possible that this variable is also influenced by pa-
externalizing problems in twin children. As in the rental IQ, which is partly heritable’’ (p. 662), they go
previous study, the authors briefly cited the ‘‘possi- on to conclude that their analysis ‘‘demonstrate[s]
bility that the effect of domestic violence may be that the environment does play an important role in
 Twin Studies with Measured Environments 1221

children’s cognitive resilience to SES adversity be- ronmental term is discounted by the proportion of
yond any heritable influences’’ (p. 662, italics added). the variance accounted for by the covariate. The
analysis, therefore, is a way of conducting a twin
analysis on a phenotype after removing the portion
Analysis of SEMs of the Measured C Design of the shared variance attributable to a measured
Our review of the published articles using the variable. Comparing the results of such an analysis
measured C twin SEM reveals that, over time, the to an analogous twin model without the measured
assumption that the measured variable is a purely variable (Figure 1) tells you how much of the total
environmental influence on outcome has been shared environmental effect on the phenotype is
somewhat underemphasized. Whereas the earliest accounted for by the measured covariate, which is
articles were very cautious about the interpretation the way it was interpreted in the early studies. The
of the results, more recent papers have tended to regression of the phenotype on the measured shared
mention the assumptions only briefly, if at all, and environmental variable does not control for any-
also to include stronger claims about the results’ thing, because it takes exactly the same value
significance. In the following sections, we endeavor whether or not the biometric portion of the model is
to clarify two points about SEMs of the measured included.
C design. First, we specify the quantity estimated by Conducting this analysis on the actual twin data
the parameter associated with measured shared en- will help clarify the point. As an example, we used a
vironmental variables in SEMs of the measured sample of twins from the VA 30,000 data set (see
C design. Second, we show that misspecifying the Truett et al., 1994, for more details of the sample; our
action of the measured covariate as purely environ- concerns here are methodological rather than sub-
mental when in fact it is not can have serious con- stantive) The analyses explored years of education
sequences for the rest of the analysis, leading to completed by twins and a measure of parental sep-
incorrect conclusions about the roles of genes and aration and divorce (a family-level variable because
environment in the genesis of the phenotype being the twins are necessarily concordant for the mea-
studied. sure). The sample included 576 MZ and 742 DZ twin
To answer these questions, it is necessary to derive pairs with complete data for both twins and marital
the estimated values of the path coefficients in Figure status data for their parents. We fit the full model
2 from the observed covariances among the twins’ illustrated in Figure 2, as well as two submodels, one
phenotypes and the observed covariates in the MZ with only the biometric decomposition and without
and DZ twins. Applying standard path diagram divorce (Figure 1), and one with divorce but without
tracing rules to Figure 1, we obtain the following the biometric decomposition (figure not shown).
system of equations: The results are shown in Table 1. The biometric
decomposition of educational attainment has no ef-
rMZ ¼ a2 þ c2 þ f 2 ; fect at all on the small but significant regression of
rDZ ¼ 12 a2 þ c2 þ f 2 ; years of education on parental divorce status. The
ð1Þ
rCOV ¼ f; small percentage of twin covariance for educational
status that can be attributed to parental divorce is
a2 þ c2 þ e2 þ f 2 ¼ 1:
simply subtracted from the shared environmental
Solving for the values of the paths, we obtain component. It bears repeating that even this desig-
nation of the divorce effect to the shared environ-
a2 ¼ 2ðrMZ
rDZ Þ;
ment is arbitrary, a consequence of the fact that twins
c2 ¼ 2rDZ
rMZ
r2COV ;
ð2Þ
e2 ¼ 1
rMZ ;
Table 1
f ¼ rCOV : Parameter Estimates of Twin Models for Years of Education That Include
a Measure Parental Divorce
The path for f, the regression of the twins’ phenotype
on the measured environmental variable, is equal to Model A C E f
the observed correlation between the phenotype and
the environmental variable, so the genetically in- Full model .66 .589 .46
.068
formative part of the analysis has no effect on it. In Biometric only .66 .586 .46 F
Divorce only F F F
.068
the genetically informative portion, the genetic and
nonshared environmental parameters are un- Note. The full model is presented in Figure 2. Figure 1 represents
changed from their usual forms. The shared envi- the biometric-only model. The divorce-only model is not shown.
1222 Turkheimer, D’Onofrio, Maes, and Eaves

are not genetically informative about parental char- parents (e.g., mid-parent education) is then included
acteristics that they share. Our conclusion from the in the model, and it correlates with the phenotype at
analysis is that a small portion of the shared family r 5.5, accounting for .25 of the variability in child
variance in educational attainment can be explained phenotype. Because all of the shared variability in
by the observed regression of educational attainment the phenotype has been hypothesized to be genetic,
on parental divorce, if one assumes that the effects of the correlation with the parental variable must be
parental divorce are purely environmental. But that mediated genetically. However, an actual researcher
is all we can conclude. The observed regression of would have no way of knowing this, because as we
educational attainment on divorce is just that; the have shown, the measured C design is not informa-
estimate does not control for genetic or environ- tive about the biometric construction of the covari-
mental effects on either divorce or educational at- ate. So, in practice, the investigator would fit the
tainment. SEM specified above, and the variance attributable
Most of the studies we have cited using the to the covariate would be subtracted from the c
measured C design do not report the regression of term, resulting in a2 5 .5, c2 5
.25, e2 5 .5, f 2 5 .25.
the phenotype on the measured covariate without The negative variance estimate for c2 would be an
the biometric decomposition of the phenotype; obvious violation of the twin model, leading to
therefore, it is not possible to document that the consideration of additional parameters for genetic
same phenomenon occurs in the published litera- dominance or epistasis that would be a spurious
ture. One report that does include such a regression result of falsely assuming that the pathway between
is the paper by Jaffee et al. (2002). These authors the covariate and the phenotype was entirely envi-
described a series of SEM in which they dropped ronmental.
parameters in order to find the best model for the The important point, however, is not the particu-
relation between domestic violence and offspring lar misspecifications that would result in this or any
psychopathology. The magnitude of the parameter of the many other examples that might be con-
(f ) for the regression of externalizing (.22) and in- structed. Rather, the crucial issue is a conceptual one
ternalizing (.15) on domestic violence did not change about statistical modeling: The SEMs used to analyze
as each of the latent biometric variables (A, C, and E) the measured C design entail an assumption that is,
were removed. on the one hand, untestable and, on the other hand,
We now turn to our second point about structural is capable of producing misleading results when it is
equation analyses of the measured C design. In the violated. When this is combined with our first point,
foregoing, we analyzed the effect of simultaneously that the twin analysis has no effect on the estimate of
conducting a biometric decomposition of a pheno- the regression of the phenotype on the covariate, the
type with a regression of the phenotype on a mea- wisdom of continuing to use these models is called
sured family-level covariate, and concluded that the into serious question.
regression on the covariate is exactly the same
whether or not the twin information is included. Our
Regression Analyses of Twin Data With Measured
next concern is the converse: how does including a
Environments
family-level covariate affect a biometric analysis of a
phenotype? The answer, as we will see, is that rather Another analytic technique, called DeFries –
than having no effect at all, under many circum- Fulker (DF) analysis, has also been used to analyze
stances the inclusion of a family-level covariate the measured C design (e.g., Jaffee, Caspi, Moffitt, &
will lead to incorrect conclusions in the rest of the Taylor, 2004; Jaffee, Moffitt, Caspi, & Taylor, 2003;
analysis. McGue & Lykken, 1992; Maughan, Taylor, Caspi, &
In the analysis above, we showed that the vari- Moffitt, 2004). DF analyses of twin and family data
ance explained by the covariate, f2, is subtracted from estimate biometric components without resorting to
the shared environmental component, c2, when the latent variable structural equation modeling. The
covariate is included in the model. This attribution of approach was originally used to analyze samples of
the covariate to the shared environmental term is twins in which one twin is selected as a proband
arbitrary, however, and it is simple to construct cir- (DeFries & Fulker, 1985), and has since been ex-
cumstances in which it is incorrect. Consider the panded for use in random samples of twins (Cherny,
following circumstances. A twin study of a pheno- DeFries, & Fulker, 1992; Labuda, DeFries, & Fulker,
type shows that rMZ 5 .5 and rDZ 5 .25, resulting in a 1986; Rodgers & McGue, 1994).
simple biometric decomposition of a2 5 .5, c2 5 0, Before examining the slightly more complex form
e2 5 .5. A family-level covariate describing both of DF analysis that recent studies have utilized for
 Twin Studies with Measured Environments 1223

the measured C design, it will be useful to review the So we have

basics of why the method works for the simple case. a2 þ c2
In DF analysis, one first double enters the twins as MZ : ¼ b1 þ 1:0b3 ;
a2
þ c2 þ e2
two variables in the data set and regresses one on the 1 2
ð7Þ
a þ c2
other. In the dependent variable, which we will refer DZ : 22 2 ¼ b1 þ :5b3 ;
to as the double-entered vector, the phenotype of an a þ c þ e2
arbitrary member of each pair is included first, fol- and therefore
lowed by the second member of the pair as a sepa- c2
rate observation. The independent variable, which ¼ b1 ;
a2 þ c2 þ e2
we will call the reverse-entered vector, is the same as ð8Þ
a2
the first except that the order of the twins is reversed ¼ b3 ;
within each pair, so across the two vectors each twin a2 þ c2 þ e2
is paired twice with his or her cotwin, once in each which shows that the b1 term estimates the stand-
order. Then, a regression model is estimated in ardized shared environmental component, and the
which the double-entered vector is regressed on the interaction term (b3) estimates the standardized ad-
reverse-entered vector, the zygosity of the pair ex- ditive genetic component.
pressed as the degree of genetic relationship (rg, 1.0
for MZ pairs; .5 for DZ pairs), and the interaction Published Regression Analyses of Measured C Design
between the two.
^
yij ¼ b0 þ b1 yji þ b2 rg þ b3 yji rg : ð3Þ It has long been known that the DF model can be
extended in a variety of ways. Rowe and Waldman
Although recent applications of DF analysis have (1993) suggested that researchers add a shared envi-
utilized advances in statistical software to calculate ronmental measure (F) to the regular DF equation to
more accurately the standard errors of the estimated explore how specific measures of the shared environ-
parameters in light of the double entering (Kohler ment could mediate the influence of the latent factor.
& Rodgers, 2001), the most important issues in- X
n
volving the analysis are not estimation and statistical yij ¼ b0 þ b1 yji þ b2 rg þ b3 yji rg þ b4 F Xi : ð9Þ
inference, but rather substantive interpretation of i¼1
the results. Rowe and Waldman (1993) stated correctly that when
The simplest way to derive the expectations of the the measured shared environmental variable is added
DF model is to consider the simple regression of to the basic DF model, the influence of the shared
one (double-entered) twin on the other, separately environmental parameter (b1) should decrease because
for the MZ and DZ pairs. The covariance between the ‘‘measured variable ‘takes up’ some of the ex-
the double- and reverse-entered vectors is simply the planatory variance associated with abstract shared
covariation between members of twin pairs. In the environmental influence’’ (p. 366). However, the au-
MZ pairs, the pair covariance is equal to thors did not analyze a measured C data set in their
a2 þ c 2 ð4Þ review chapter.
McGue and Lykken (1992) used a sample of twins
and the regression of one of the double-entered
to explore genetic and environmental contributions
vectors on the reverse-entered vector is equal to this
to divorce. The authors explained their use of re-
covariance over the phenotypic variance, that is,
gression analyses as follows: ‘‘In order to consider
a2 þ c2 simultaneously the influence of all family back-
: ð5Þ
a2 þ c 2 þ e2 ground data, we used a logistic function of birth
In the DZ twins, the equivalent regression is equal cohort, zygosity, co-twin’s divorce status, parents’
to divorce status, spouse’s parents’ divorce status, and
1 2 all two-way interaction terms to predict divorce risk’’
þ c2
2a
: ð6Þ (p. 37). Note that parents’ divorce status was neces-
a2 þ c 2 þ e2 sarily shared by both twins, but the twins could
In the full model encompassing both kinds of differ on the spouse’s parents’ divorce status. The
twins, including the rg term for zygosity and the regression analyses resulted in a significant param-
interaction between the reverse-entered vector and eter for the interaction between zygosity and the
rg, the two regressions are expressed as a main effect cotwin’s divorce status and the main effects of the
b1 common to both zygosities and an in interaction other predictors. The authors concluded, ‘‘spouse’s
term b3 that expresses the difference between them. family background of divorce and respondent’s
1224 Turkheimer, D’Onofrio, Maes, and Eaves

family background of divorce contributed independ- account for different levels of smoking), other po-
ently to the prediction of marital dissolution’’ (p. 370, tential confounding variables (maternal ASB, pater-
italics added). nal ASB, maternal depression, and SES
Recent use of the expanded DF model has also disadvantage), and the basic variables from the DF.
focused on family-level factors. Jaffee et al. (2003) Once again, the smoking during pregnancy and the
explored the impact of father absence and parental confounding variables were shared influences by
antisocial behavior (ASB) on young children’s ASB. definition. The full regression model illustrated that
The regression analyses added measures of father’s the parameters associated with smoking during
ASB, mother’s ASB, father’s presence, and the in- pregnancy were reduced substantially when the
teraction of father’s presence and ASB to the stan- confounding and DF variables were included but
dard DF regression analysis. Each of these familial were still sizeable. The paper concludes, ‘‘Around
variables was necessarily shared by the twin chil- half of the observed association between prenatal
dren. The full DF model resulted in significant re- smoking and young children’s conduct problems
gression coefficients associated with the interaction was attributable to correlated genetic effects. But the
between genetic relatedness and twin’s ASB, the results were also clear in showing that, even after
main effects of father’s ASB and mother’s ASB, and controlling for genetic influences, prenatal smoking
the interaction between father’s ASB and father’s continued to be significantly linked to children’s
presence. Based on the results of the regression behavior outcomes’’ (p. 841, italics added).
models, the authors concluded: ‘‘Children experi-
ence a double whammy of risk for antisocial be-
Analytical Review of DF Analysis of Measured
havior. They are at genetic risk because antisocial
C Design
behavior is highly heritable. In addition the same
parents who transmit genes also provide the chil- As was the case with SEM analysis of the design, the
dren’s rearing environment. We found that a father’s important question is how to interpret the parameter
antisocial behavior accounted for his children’s be- associated with the measured family-level ‘‘envi-
havior problems independent of any genetic risk he may ronmental’’ variable (b4). It should be noted from the
have imparted. . .’’ (p. 120, italics added). outset that adding a covariate shared by both twins
The relation between parental maltreatment and (by definition) to a model in which the phenotypes of
young children’s ASB has also been explored using the twins are already regressed on each other seems
expanded DF analysis (Jaffee et al., 2004). Although to be a paradoxical thing to do. Conceptually, the
parental maltreatment can differ among twins, these reason double-entered twin vectors are regressed on
analyses appeared to enter the variable as a single each other in DF analysis is as a way of estimating
family-level variable. The regression weights asso- the phenotypic variance the twins have in common,
ciated with physical maltreatment and the interac- with the difference between the magnitude of this
tion of genetic relatedness and the cotwin’s ASB shared variance in the MZ and DZ twins estimating
were significant. Consequently, the authors con- the heritability, and the residuals from the regression
cluded, ‘‘the effect of physical maltreatment was estimating the variance not shared by members of
significant after controlling for any genetic transmission the twin pair, or e2. So by adding the shared family-
of antisocial behavior. . .’’ (p. 50, italics added). level covariate to this model, one is attempting to
Finally, the relation between smoking during predict shared twin variability in a model that al-
pregnancy and childhood conduct problems was ready has a term specifically designed to account for
explored using DF analyses (Maughan et al., 2004). all of it. What could be left to estimate? It turns out
Although many studies have documented an asso- that the inner workings of DF analysis are somewhat
ciation between smoking during pregnancy and more complex than the results it produces, so to
offspring difficulties, a number of researchers have understand exactly what takes place we need to
questioned whether genetic factors may confound consider the details of the regressions on which DF
the intergenerational relation (e.g., D’Onofrio et al., analysis is based.
2003; Fergusson, 1999; Sexton, Fox, & Hebel, 1990; In a DF analysis of the measured C design, one
Silberg, Parr, Neale, Rutter, Angold, & Eaves, 2003; regresses the double-entered vector on the reverse-
Thapar et al., 2003). To account for genetic factors entered vector and on a family-level covariate. It is
and other processes that may mediate the inter- simplest to begin by considering a sample of MZ
generational association, Maughan et al. (2004) in- twins, and a measured covariate that has been
cluded smoking during pregnancy (as a family-level standardized to a mean of zero and a variance of one.
variable represented by a set of dummy codes to Figure 3 is a path diagram of the model. The double-
 Twin Studies with Measured Environments 1225

jth pair can be expressed in terms of the pair mean

and a deviation from the pair mean:


y1j
y2j
y:j þ : ð13Þ
Reverse Family 2
Entered Level
Vector Covariate
This twin is to be predicted from its cotwin, with
deviation ðy2j
y1j Þ=2 from the pair mean. The re-
gression coefficient, as we have seen, is equal to the
intraclass correlation rI. The residual from this re-
gression, therefore, is equal to

 

y1j
y2j y2j
yij
y:j þ
rI y:j þ
2 2
Double ð1 þ rI Þ

Entered ¼ ð1
rI Þy:j þ y1j
y2j ð14Þ
Vector
2
and the mean of residuals for twin pair j is simply
Figure 3. Structural equation model of DF analysis. equal to
ð1
rI Þy:j : ð15Þ
entered vector, reverse-entered vector, and measured Although the procedure of double-entry regres-
family covariate have a variance covariance matrix sion succeeds in estimating rI, it does not control for
equal to anything: in the residuals from the double-entered
2 3 regression, the pair means, which are the only por-
s2y
4 a2 þ c2 þ f 2 s2 5: tion of the data with which a family-level variable
ð10Þ
y can possibly covary, are correlated 1.0 with the pair
f f 1:0 means in the original data. So a measured covariate
Applying standard formulas for the values of with covariation f with phenotype in the original
bivariate regression coefficients, we find that the scores has the same relation with phenotype in the
regression of the double-entered vector on the re- residuals, except that it is rescaled by the proportion
verse-entered vector is given by of nonshared variance. This coefficient is zero when
either f or the nonshared variance is zero.
a2 þ c2
; ð11Þ The uninformativeness of the f parameter is not
s2y
f 2 redeemed by consideration of the MZ and DZ twins.
which is the proportion of the y variance not ex- The expectation of 1
rI in Equation (12) in terms of
plained by the covariate that is shared between the ACE parameters is different in the MZ and DZ twins.
twins. The regression of the double-entered vector In the MZ twins, 1
rI is equal to e2, whereas in the
on the family-level covariate is given by DZ twins it is equal to 12 a2 þ e2 . Because the expec-
! tations for the main effects of the covariate are dif-
s2y
f 2 ð1
rI Þ; ð12Þ ferent in the MZ and the DZ pairs, one could include
sy
f 2 both a main effect of the covariate and an interaction
with genetic relatedness (rg), although the interaction
which is the simple regression of the phenotype on
has not generally been included in actual analyses. If
the measured family variable, rescaled by the pro-
it were however, the main effect of the covariate
portion of the phenotypic variance not attributable
would have the expectation
to the covariate that is not shared by the twins. As !
was the case with SEM models of the measured C s2y
2 
f 2 a þ e2 ; ð16Þ
design, the DF estimate of the regression of the sy
f 2
phenotype on the covariate does not control for any-
thing, genetic or environmental, although in DF and the interaction with zygosity would have the
analysis it is linearly rescaled. expectation
!
To understand why this result occurs, consider the s2y
2
simple regression of a set of double-entered twins on
f 2 2
a ; ð17Þ
sy
f
the reverse-entered vector, without the covariate. If
the twin scores are expressed as deviations from the which still only reexpresses the original regression
population mean, the score of the ith member of the parameter in terms of a different variance.
1226 Turkheimer, D’Onofrio, Maes, and Eaves

This whole point is crucial to our argument, so we tween Analyses 2 and 3 (
.19 to
.065) is a result of
will take the time to repeat it at a conceptual level statistical control of shared genetic and/or environ-
before proceeding to an example from the education mental effects achieved by including the reverse-
and divorce data in the VA 30,000 data set. When ordered twin vector, but it is not. In fact, the coeffi-
twin children vary in some outcome, a sibling model cient of .065 is simply the product of the original
allows us to partition the variability in outcome into regression coefficient for divorce status in Analysis 2
a portion varying between families (essentially the (
.189), and the proportion of phenotypic variance
variance of the family means) and a portion varying not attributable to the covariate that is not shared
within them (essentially the variance of children between siblings. (The apparent similarity between
around their family mean). Using twin zygosity al- the coefficient of .655 in Equation (18) and .065 in
lows us to further partition the between and within Equation (20) is coincidental.) The variance of edu-
variances into ACE components. The variance of a cation in the combined sample is 8.00, of which
family-level covariate, however, which is necessarily .1892 5 .036 is accounted for by divorce status. The
equal for members of a twin pair, is by definition intraclass correlation for education equals .655.
completely shared in both MZ and DZ twins (there is Therefore, the coefficient for standardized divorce
no variation around the family mean), and therefore status in the model including the reverse-entered
cannot be further partitioned into genetic and envi- twin vector will be equal to
ronmental components. It does not make sense to ask  
8:0ð1
:655Þ
whether the deviations of twins’ education scores
:189 ¼
:065: ð21Þ
8:0
:0357
from the pair means covary with their parents’ di-
vorce status, because if one twin has a deviation of This is exactly the value estimated in Analysis 3.
1x, the other twin has a deviation of
x, whereas
their parent’s marital status is exactly the same. On
Discussion
the bottom line, a variable that has no within-family
variation cannot covary with the within-family var- Our review of the literature on the measured C de-
iation in another variable. sign and the various statistical and interpretive ap-
Working the analysis in the divorce-years of ed- proaches to it was motivated by a number of factors.
ucation data will help clarify the point. We double First, we agree with researchers in behavior genetics
entered the twins, and then conducted three analy- (e.g., Eaves, Last, Martin, & Jinks, 1977; Plomin,
ses: (1) a regression of one double-entered twin DeFries, & Loehlin, 1977; Rutter et al., 1997; Rutter &
vector on the other; (2) a regression of the twin vector Silberg, 2002; Scarr & McCartney, 1983) and devel-
on a standardized dummy variable coding for di- opmental psychology (e.g., Booth, Carver, & Grang-
vorce; and (3) a regression in which the twin vector er, 2000; Collins, Maccoby, Steinberg, Hetherington,
was regressed simultaneously on the reverse-entered & Bornstein, 2000) that studies need to explore how
vector and the dummy variable for divorce. genetic and environmental factors act and interact.
Analysis 1 results in a regression equation of Although the debate over whether genetic factors
Twin1 ¼ 5:02 þ :655 Twin2 : ð18Þ influence individual differences on behavioral char-
acteristics is over (Turkheimer, 2000), the need to
The regression coefficient of .655 is the intraclass explore the mechanisms between what are con-
correlation, or the proportion of the variance that is sidered to be environmental risk factors and indi-
shared by twin pairs. In Analysis 2, the simple re- vidual adjustment is paramount. Our own research
gression of standardized divorce status on the dou- program, outlined below, is representative of our
ble-entered twins results in desire to explore such processes.
Educ ¼ 14:7
:19Div: ð19Þ Second, we understand the fundamental motiva-
Therefore, without controlling for anything, each tion behind both the older and more recent reports
standard deviation of divorce status is associated describing twin studies with measured family-level
with a decrease of about a fifth of a year in the factors. The theoretical rationale outlined by many of
children’s education. Then, in Analysis 3, the double- the researchers (clearly expressed in Jaffee et al.,
entered twin vector is regressed on the reverse- 2004) needs to be appreciated by all social scientists,
entered vector and divorce status, with the result because genetic factors may indeed confound sta-
tistical associations between risk factors and child
Educ1 ¼ 5:11 þ :65Educ2
:065Div: ð20Þ
adjustment. Third, we are sympathetic with the
At first glance, it might seem as though the reduction desire to respond to researchers who have claimed
in the magnitude of the effect of divorce status be- that shared environmental factors are unimportant
 Twin Studies with Measured Environments 1227

(e.g., Plomin & Daniels, 1987; Harris, 1998; Rowe, when the reverse-entered twins are included in the
1994), based, in part, on the null findings for shared regression has the effect of reducing its magnitude
environmental influences in classical twin designs. from the value in the simple regression without the
Properly qualified and analyzed, standard twin reverse-entered vector. This reduction has convinced
studies in general do have the potential to demon- researchers that they are controlling for genetic or
strate the importance of both genetic and environ- environmental factors when in fact they are not.
mental factors and to show how they can confound
statistical associations between purported risk fac-
tors and child adjustment. Hierarchical Linear Models
Nevertheless, it must be recognized that twin A much greater degree of clarity about the struc-
studies with family-level environmental character- ture of this research design can be achieved by
istics are unable to explore genetic mediation of as- turning to a more modern form of regression than
sociations between the family-level variables and DF analysis: hierarchical linear modeling (HLM). For
individual adjustment, and moreover that widely a review of HLM see Raudenbush and Bryk (2002),
employed methods for analyzing the design can lead and for applications of HLM to family data see van
to inaccurate results. Many of the conclusions that den Oord (2001) and Guo and Wang (2002). There is
have been reached about the implications of such increasing recognition in the field that HLM is ide-
studies have been overstated. We are not claiming ally suited to the analysis of family data, in which
that all statistical associations that have been ex- sibling or twin children are nested within families,
plored in the measured C reports are attributable to and therefore correlated with each other. Therefore,
genetic factors; rather, we conclude that the claims we can say for the ith child in the jth family, the
made in the recent literature cannot be supported by predicted score is given by an intercept for the family
the methodology and analyses that were used. In the bj and random variation of children in the family
remainder of this paper, we apply what we have around the family intercept:
learned about the design, analysis, and interpreta-
yij ¼ bj þ s2w : ð22Þ
tion of the measured C design to make some rec-
ommendations for future progress in the area. The family intercept bj can then be modeled as a
SEM analyses of family-level variables and out- function of covariates like maternal smoking or paren-
comes in twins only provide an estimate of the un- tal education, plus a second error term describing the
controlled phenotypic relation between them, and variability of the family means around the grand mean:
cannot determine whether the effect of the measured bj ¼ b0 þ bC COV þ s2B : ð23Þ
variable is genetic or environmental. Therefore, re-
searchers should desist from describing the results of Several conceptual benefits of this approach are
such analyses as showing the influence of an envi- immediately apparent. First of all, the approach
ronment ‘‘independent of genetic factors’’ (e.g., provides a name for the type of covariate we have
Thapar et al., 2003, p. 1988), which represents an been discussing, one that is necessarily shared by
assumption of the model, not a testable hypothesis. both members of a twin pair: it is a family level, or
In general, we recommend that future analyses re- level 2, variable, which describes the data at the
frain from estimating the biometric parameters higher level of nesting rather than at the level of
(a random variance parameter can be estimated to individual participants. It is exactly like a variable
take into account that there are two children per describing the classroom in the paradigmatic HLM
family; see the discussion of hierarchical linear model in which individual students are nested
models) when shared environmental influences are within classrooms. Second, in an HLM structure it
included in a twin data set. Doing so only fosters the becomes very clear that a family-level variable can
false impression that something is being controlled only be used to predict variability in y at the family
in the estimation of the relation between the family- level, that is, the variability of the family means
level variable and the twin outcome, and as we have around the grand mean. Neither genetic nor envi-
shown inclusion of the covariate can produce seri- ronmental variability in y can be controlled.
ously misleading results in the biometric analysis.
Expanded DF analyses do not overcome the lim-
Other Extensions of DF Analysis
itations of the SEM approach. In some ways, DF
analysis has proved even more misleading, because Apparently, the recent analyses of the measured C
the rescaling of the regression parameter associated design using DF analysis have confused two things
with the covariate in terms of the nonshared variance that DF analysis can do with something it cannot.
1228 Turkheimer, D’Onofrio, Maes, and Eaves

One common extension of DF analysis, explored speaking, teasing apart genetic and environmentally
most fully by Rodgers and colleagues, can be applied mediated effects would require examining offspring
to research designs including a covariate that varies exposed to maternal smoking in utero where the
both between and within families (Rodgers, Rowe, & intrauterine environment was provided by a non-
Li, 1994; Rodgers, Rowe, & May, 1994), for example, genetically related ‘mother’ (surrogate; that is a ‘be-
a measure of maternal interaction that can be mea- fore-birth’ adoption design). That is clearly not a
sured separately for each of the children within a feasible design in humans’’ (p. 1988).
family (e.g., Caspi et al., 2004), or a measure of a We certainly agree that a maternal smoking study
family-level construct like SES or parental marital with ‘‘experimental control’’ is impossible, but there
status that is expressed in terms of the experience of are designs that can separate genetic and environ-
individual siblings within a family. In this situation, mental processes in the consequences of maternal
the within-family regression of child outcome on the behavior. The main limitation of the twin studies
covariate does control for genetic and environmental reviewed here is that, like smoking in women preg-
between-family covariation. This is clearest in the nant with twins, there is by definition no genetic or
case of identical twins discordant for a predictor nonshared environmental variation in the parental
within the family: perhaps they were differentially measure. One way to overcome this limitation is to
exposed to an environmental toxin. The within- include measures of the twins’ parents along with
family regression of an outcome on the predictor family-level environmental factors and twin charac-
describes the relation between the two independent teristics (Eaves et al., 1978), a design referred to as
of genetic and shared environmental variance, be- the twin-family design. Genetic and shared envi-
cause both of these equal zero within MZ twin pairs. ronmental variation in the family-level variable is
In an HLM model of the design, one would simply based on the biometric parameters of the offspring’s
include the covariate at both the between- and phenotype, which are assumed to be the same in the
within-family levels of the model and interpret the parental generation (e.g., Kendler et al., 1996; Meyer
resulting parameters separately (Lynch et al., sub- et al., 2000; Taylor, McGue, & Iacono, 2000). The
mitted; Mendle, Turkheimer, D’Onofiro, Lynch, & strength of this design lies in its ability to estimate
Emery, submitted). environmental effects while controlling for genetic
The other applicable extension of DF analysis, less effects on both the parents and children (Rutter et al.,
commonly used, is to use it to detect gene by envi- 1997), but it also entails some major assumptions and
ronment interaction (Rodgers, Rowe, & Li, 1994; limitations (D’Onofrio et al., 2003; Rutter et al., 2001).
Rowe, Almeida, & Jacobson, 1999; Rowe, Jacobson, & The same genetic and environmental factors are as-
van den Oord, 1999). In this analysis, one adds to the sumed to influence child characteristics and parental
basic DF model containing a two-way interaction behavior (requiring the same phenotype to be
between genetic relatedness and the reverse-entered measured in both generations), and twin-family
twin vector a three-way interaction between this models with only one parental characteristic may
term and the measured covariate. The three-way underestimate the noncausal, intergenerational ge-
interaction estimates the extent to which the two- netic pathway (Kendler et al, 1996; Meyer et al.,
way interaction (i.e., the heritability of phenotype) 2000).
varies as a function of the covariate, which is one Instead of including parents of twins as in the
way of characterizing gene by environment interac- twin-family approach, twin studies can also be ex-
tion. In HLM, such models are referred to as het- panded by adding the children of twins, referred to
eroscedasticity models, because they involve as the children of twins (CoT) design. The major
modeling the magnitude of the between- and within- advantage of the CoT Design is that it provides ge-
family error terms as a function of a covariate. See netic and environmental variation in family-level
Guo and Wang (2002) for examples of gene by en- risk factors, because unlike parents of twins, twin
vironment interaction analyses via HLM, and Purcell parents can vary in family-level variables like marital
(2002) for related analyses using SEM. status and smoking during pregnancy. Therefore,
the statistical association between a family-level
variable and offspring adjustment can be decom-
Extended Twin Designs
posed into environmental processes specifically re-
Thapar et al. (2003) lament the limitations of stand- lated to the risk factor (consistent with a causal
ard twin models for exploration of intergenerational hypothesis), environmental influences common to
relations, in particular offspring characteristics re- both twins and the offspring (consistent with the
lated to smoking during pregnancy: ‘‘Strictly selection hypothesis), and genetic factors (also
 Twin Studies with Measured Environments 1229

consistent with the selection hypothesis) that are proach that incorporates gene by environment in-
shared by the parents and the offspring (reviews in teraction and gene – environment correlation, see
D’Onofrio et al., 2003, submitted-a; Gottesman & Eaves and Erkanli (2003) and Eaves, Silberg, and
Bertelsen, 1989; Heath et al., 1985; Rutter et al., 2001). Erkanli (2003).
Our laboratory is actively engaged in CoT analyses
of genetic and environmental pathways from parental
Combining Behavior Genetic Designs
behavior to outcomes in children. One analysis of the
association between father absence and age of men- Tests of causal effects of parenting behavior on chil-
arche in the female offspring concluded that the dren are greatly strengthened when several family
phenotypic relation is not due to environmental designs are combined, with their varied strengths
processes related to the risk factor; rather familial and weaknesses (Rutter et al., 2001). For example,
confounds (either genetic or common environmental) adoption studies represent a powerful method for
mediate the relation (Mendle et al., submitted). In exploring these processes (e.g., Plomin, 1995). The
contrast, analyses using the CoT design suggest that adoption design also includes major assumptions
the association between harsh punishment and off- (Rutter et al., 2001), but combining the adoption and
spring drug, alcohol, and behavioral problems is due CoT designs could take advantage of the unique
to environmental processes specifically associated advantages of each, while limiting the assumptions
with the parenting behavior (Lynch et al., submitted). when each is analyzed separately. The CoT design
Analyses of the consequences of parental divorce can also be combined with traditional studies of
have also illustrated that the processes underlying twins as children to explore both passive and active
intergenerational relations depend on the specific rGE (Neiderhiser et al., 2004).
association being explored. Whereas the results of Studies using many levels of genetic relatedness
CoT analyses of parental divorce are consistent with a among family members also have the power to dis-
causal relation with lifetime history of psychopa- criminate genetic from environmental transmission.
thology (D’Onofrio et al., submitted-a), some life For example, a number of investigators have used
course patterns associated with parental marital dis- the family data available in the National Longitudi-
solution, such as cohabitation, appear to be due to nal Survey of Youth to study a variety of normative
genetic and environmental confounds (D’Onofrio et and pathological developmental processes (Rodgers,
al., submitted-b). The limitations of the CoT design Rowe, & Buster, 1999; Rodgers, Buster, & Rowe,
include relatively low statistical power and difficul- 2001). Another extensive combination of different
ties in accounting for effects arising in the spouses of behavior genetic designs is the ‘‘Stealth Model’’ de-
twins (reviews in D’Onofrio et al., 2003; Heath et al., veloped by Eaves and colleagues to explore the in-
1985; Rutter et al., 2001). tergenerational transmission of religious practices,
personality characteristics, and body mass index
(e.g., D’Onofrio et al., 1999; Eaves, Heath, Martin,
Measured Genotype Designs
Maes, et al., 1999; Eaves, Heath, Martin, Neale, et al.,
Some very interesting recent studies have reported 1999; Kirk et al., 1999; Lake, Eaves, Maes, Health, &
interactions between measured family-level risk Martin, 2000; Maes, Neale, & Eaves, 1997; Truett
factors and measured genotypes (i.e., DNA was et al., 1994). The design can also be used explore
collected and analyzed) using singletons (Caspi bivariate intergenerational associations (Maes, Ne-
et al., 2002, 2003) or twins (Foley et al., 2004). These ale, Martin, Heath, & Eaves, 1999). The model uses
studies investigated the main effect of the family- upwards of 30 different family relationships that
level measure, the main effect of the measured gen- vary in their degree of environmental and genetic
otype, and the interaction of the two. These studies relatedness to test for the mechanisms involved in
are able to test the hypothesis that the expression of intergenerational associations while including esti-
the measured genotype depends on the family-level mates of assumptions that hinder many behavior
covariate. As the interaction between the measured genetic designs.
genotype and the family-level covariate does not
depend on the partitioning of genetic and shared
Conclusion
environmental variability in the covariate, the criti-
cisms we have outlined in this paper do not apply to Differentiating actual social causation (i.e., direct
this design, although researchers must be cautious environmental influences) from genetic or environ-
about simply assuming that family-level variables mental confounds remains one of the fundamental
are environmental in origin. For an analytical ap- problems facing the social sciences. Conclusions
1230 Turkheimer, D’Onofrio, Maes, and Eaves

about causal mechanisms linking environmental risk depression: Moderation by a polymorphism in the
factors to childhood adjustment are crucial because 5-HTT gene. Science, 301, 386 – 389.
they inform both public policy and personal deci- Caspi, A., Taylor, A., Moffitt, T. E., & Plomin, R. (2000).
sions about persistent and vexing choices, like pa- Neighborhood deprivation affects children’s mental
rental divorce (Amato, 2000; D’Onofrio et al., health: Environmental risks identified in a genetic de-
submitted-a), father absence (Jaffee et al., 2003; sign. Psychological Science, 11, 338 – 342.
Mendle et al., submitted), parenting practices (Ba- Cherny, S. S., DeFries, J. C., & Fulker, D. W. (1992). Multiple
regression analysis of twin data: A model-fitting ap-
umrind, Larzelere, & Cowan, 2002), parental mal-
proach. Behavior Genetics, 22, 489 – 497.
treatment (Jaffee et al., 2004; Lynch et al., submitted),
Collins, W. A., Maccoby, E. E., Steinberg, L., Hetherington,
parental psychopathology (Gottesman, 1991; Jacobs
E. M., & Bornstein, M. H. (2000). Contemporary research
et al., 2003), prenatal smoking and drug use (Fe- on parenting: The case for Nature and Nurture. Ameri-
rgusson, 1999), social support (House, Landis, & can Psychologist, 55, 218 – 232.
Umberson, 1988), and poverty (Dohrenwend, 1992). Cook, T. D., & Shadish, W. R. (1994). Social experiments:
Evaluating potential causal mechanisms without Some developments over the past fifteen years. Annual
the use of randomized experiments is extremely Review of Psychology, 45, 545 – 580.
difficult (Cook & Shadish, 1994). Merely correlating Daniels, D., & Plomin, R. (1985). Differential experience of
environmental risk factors with child outcomes us- siblings in the same family. Developmental Psychology, 21,
ing traditional family studies is of limited utility 747 – 760.
because of ubiquitous environmental and genetic D’Onofrio, B. M., Eaves, L. J., Murrelle, L., Maes, H. H., &
confounding factors. Therefore, multiple research Spilka, B. (1999). Understanding biological and social
strategies and designs are required, especially ge- influences on religious affiliation, attitudes and be-
netically informed approaches (Rutter et al., 2001). haviors: A behavior-genetic perspective. Journal of Per-
Although the current review has focused on the sonality, 67, 953 – 984.
limitations of one widely used behavior genetic de- D’Onofrio, B. M., Turkheimer, E., Eaves, L. J., Corey, L. A.,
sign, we remain excited by the potential of the vari- Berg, K., Solaas, M. H., et al. (2003). The role of the
Children of Twins design in elucidating causal rela-
ous extended family designs to advance the
tions between parent characteristics and child out-
methodology of twin studies and the empirical sci-
comes. Journal of Child Psychology and Psychiatry, 44,
ence of parenting and child development. Never-
1130 – 1144.
theless, vigilance regarding the assumptions and D’Onofrio, B. M., Turkheimer, E. N., Emery, R., Slutske, W.,
limitations of research designs and statistical analy- Heath, A., & Martin, N. (in press a). A genetically in-
ses will remain the sine qua non of continued em- formed study of marital instability of offspring psy-
pirical progress. chopathology? Journal of Abnormal Psychology.
D’Onofrio, B. M., Turkheimer, E. N., Emery, R., Slutske, W.,
References Heath, A., & Martin, N. (in press b). A genetically in-
formed study of the association between parental di-
Amato, P. R. (2000). The consequences of divorce for adults
vorce and offspring life course patterns. Developmental
and children. Journal of Marriage and the Family, 62, 1269 –
Psychology.
1287. DeFries, J. C., & Fulker, D. W. (1985). Multiple regression
Baumrind, D., Larzelere, R. E., & Cowan, P. A. (2002). Or-
analysis of twin data. Behavior Genetics, 15, 467 – 473.
dinary physical punishment: Is it harmful? Psychological Dohrenwend, B. P., Levav, I., Shrout, P. E., Schwartz, S.,
Bulletin, 128, 580 – 589. Navch, G., Link, B. G., et al. (1992). Socioeconomic status
Booth, A., Carver, K., & Granger, D. A. (2000). Biosocial and psychiatric disorders: The causation-selection issue.
perspectives on the family. Journal of Marriage and the Science, 255, 946 – 952.
Family, 62, 1018 – 1034. Dunn, J., & Plomin, R. (1990). Separate lives: Why siblings are
Caspi, A., McClay, J., Moffitt, T. E., Mill, J., Martin, J., Craig, so different? New York: Basic Books.
I. W., et al. (2002). Role of genotype in the cycle of vio- Eaves, L. J. (1982). The utility of twins. In V. E. Anderson,
lence in maltreated children. Science, 297, 851 – 854. W. A. Hauser, J. K. Penry, & C. G. Sing (Eds.). Genetic
Caspi, A., Moffitt, T. E., Morgan, J., Rutter, M., Taylor, A., basis of the epilepsies (pp. 249 – 276). New York: Raven.
Arseneault, L., et al. (2004). Maternal expressed emotion Eaves, L., & Erkanli, A. (2003). Markov Chain Monte Carlo
predicts children’s antisocial behavior problems: Using Approaches to analysis of genetic and environmental
monozygotic-twin differences to identify environmental components of human developmental change and G 
effects on behavioral development. Developmental Psy- E interaction. Behavior Genetics, 33, 279 – 299.
chology, 40, 149 – 161. Eaves, L., Heath, A., Martin, N., Maes, H., Neale, M.,
Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Kendler, K., et al. (1999). Comparing the biological and
Harrinton, H., et al. (2003). Influence of life stress on cultural inheritance of personality and social attitudes in
 Twin Studies with Measured Environments 1231

the Virginia 30,000 study of twins and their relatives. Journal of the American Academy of Child and Adolescent
Twin Research, 2, 62 – 80. Psychiatry, 41, 1095 – 1103.
Eaves, L. J., Heath, A. C., Martin, N. G., Neale, M. C., Jinks, J. L., & Fulker, D. W. (1970). A comparison of the
Meyer, J. M., Silberg, J. L., et al. (1999). Biological and biometrical genetical MAVA and classical approaches to
cultural inheritance of stature and attitudes. In C. R. the analysis of human behavior. Psychological Bulletin,
Cloninger (Ed.). Personality and psychopathology (pp. 73, 311 – 349.
269 – 308). Washington, DC: American Psychiatric As- Kendler, K. S., Neale, M. C., Kessler, R. C., Heath, A. C., &
sociation. Eaves, L. J. (1992a). Childhood parental loss and adult
Eaves, L. J., Last, K., Martin, N. G., & Jinks, J. L. (1977). A psychopathology in women. Archives of General Psychi-
progressive approach to non-additivity and genotype- atry, 49, 109 – 111.
environmental covariance in the analysis of human Kendler, K. S., Neale, M. C., Kessler, R. C., Heath, A. C., &
differences. British Journal of Mathematical and Statistical Eaves, L. J. (1992b). A population-based twin study of major
Psychology, 30, 1 – 42. depression in women. The impact of varying definitions of
Eaves, L. J., Last, L. A., Young, P. A., & Martin, N. B. (1978). illness. Archives of General Psychiatry, 49, 257 – 266.
Model-fitting approaches to the analysis of human be- Kendler, K. S., Neale, M. C., Prescott, C. A., Kessler, R. C.,
havior. Heredity, 41, 249 – 320. Heath, A. C., Corey, L. A., et al. (1996). Childhood pa-
Eaves, L., Silberg, J., & Erkanli, A. (2003). Resolving mul- rental loss and alcoholism in women: A causal analysis
tiple epigenetic pathways to adolescent depression. using a twin-family design. Psychological Medicine, 26,
Journal of Child Psychology and Psychiatry, 44, 1006 – 1014. 79 – 95.
Fergusson, D. M. (1999). Prenatal smoking and antisocial Kim-Cohen, J., Moffitt, T. E., Caspi, A., & Taylor, A. (2004).
behavior. Archives of General Psychiatry, 56, 223 – 224. Genetic and environmental processes in young chil-
Foley, D. L., Eaves, L. J., Wormley, B., Silberg, J. L., Maes, H. dren’s resilience and vulnerability to socioeconomic
H., Kunh, J., et al. (2004). Childhood adversity, monoa- deprivation. Child Development, 75, 651 – 668.
mine oxidase A genotype, and risk for conduct disorder. Kirk, K. M., Maes, H. H., Neale, M. C., Heath, A. C.,
Archives of General Psychiatry, 61, 738 – 744. Martin, N. G., & Eaves, L. J. (1999). Frequency of church
Gottesman, I. I. (1991). Schizophrenia genesis: The origins of attendance in Australia and the United States: Models of
madness. New York: Freeman & Company. family resemblance. Twin Research, 2, 99 – 107.
Gottesman, I. I., & Bertelsen, A. (1989). Confirming unex- Koenen, K. C., Moffitt, T. E., Caspi, A., Taylor, A., & Purcell,
pressed genotypes for schizophrenia. Archives of General S. (2003). Domestic violence is associated with environ-
Psychiatry, 46, 867 – 872. mental suppression of IQ in young children. Develop-
Guo, G., & Wang, J. (2002). The mixed or multilevel ment and Psychopathology, 15, 297 – 311.
model for behavior genetic analyses. Behavior Genetics, Kohler, H. P., & Rodgers, J. L. (2001). DF-analyses of her-
32, 37 – 49. itability with double-entry twin data: Asymptotic
Harris, J. R. (1998). The nurture assumption: Why children standard errors and efficient estimation. Behavior Ge-
turn out the way they do. London: Bloonsbury. netics, 31, 179 – 191.
Heath, A. C., Kendler, K. S., Eaves, L. J., & Markell, D. Labuda, M. C., DeFries, J. C., & Fulker, D. W. (1986).
(1985). The resolution of cultural and biological inher- Multiple regression analysis of twin data obtained from
itance: Informativeness of different relationships. Be- selected samples. Genetic Epidemiology, 3, 425 – 433.
havior Genetics, 15, 439 – 465. Lake, R. E., Eaves, L. J., Maes, H.H, Heath, A. C., & Martin,
House, J. S., Landis, K. R., & Umberson, D. (1988). Social N. G. (2000). Further evidence against the environmental
relationship and health. Science, 241, 540 – 545. transmission of individual differences in neuroticism
Jacobs, T., Waterman, B., Heath, A., True, W., Bucholz, K. K., from a collaborative study of 45,850 twins and relatives
Haber, R., et al. (2003). Genetic and environmental effects on two continents. Behavior Genetics, 30, 223 – 233.
on offspring alcoholism: New insights using an off- Lynch, S. K., Turkheimer, E., Emery, R. E., D’Onofrio, B. M.,
spring-of-twins design. Archives of General Psychiatry, 60, Mendle, J., Slutske, W., et al. (2005). A genetically in-
1265 – 1272. formed study of the association between harsh punish-
Jaffee, S. R., Caspi, A., Moffitt, T. E., & Taylor, A. (2004). ment and offspring behavioral problems. Manuscript
Physical maltreatment victim to antisocial child: Evi- submitted for publication.
dence of an environmentally mediated process. Journal McGue, M., & Lykken, D. T. (1992). Genetic influence on
of Abnormal Psychology, 113, 44 – 55. risk of divorce. Psychological Science, 3, 368 – 373.
Jaffee, S. R., Moffitt, T. E., Caspi, A., & Taylor, A. (2003). Maes, H. M., Neale, M. C., & Eaves, L. J. (1997). Genetic
Life with (or without) father: The benefits of living with and environmental factors in relative body weight and
two biological parents depend on the father’s antisocial human adiposity. Behavior Genetics, 27, 325 – 351.
behavior. Child Development, 74, 109 – 126. Maes, H. H., Neale, M. C., Martin, N. G., Heath, A. C., &
Jaffee, S. R., Moffitt, T. E., Caspi, A., Taylor, A., & Eaves, L. J. (1999). Religious attendance and frequency
Arseneault, L. (2002). Influence of adult domestic vio- of alcohol use: Same genes or same environments: A
lence on children’s internalizing and externalizing bivariate extended twin kinship model. Twin Research, 2,
problems: An environmentally informative twin study. 169 – 179.
1232 Turkheimer, D’Onofrio, Maes, and Eaves

Maughan, B., Taylor, A., Caspi, A., & Moffitt, T. E. (2004). Rodgers, J. L., Rowe, D. C., & Buster, M. A. (1999). Nature,
Prenatal smoking and early childhood conduct prob- nurture, and first sexual intercourse in the USA: Fitting
lems. Archives of General Psychiatry, 61, 836 – 843. behavioral genetic models to NLSY kinship data. Journal
Mendle, J., Turkheimer, E., D’Onofrio, B. M., Lynch, S. K., of Biosocial Science, 31, 29 – 41.
& Emery, R. E. (2005). Stepfather presence and age at Rodgers, J. L., Rowe, D. C., & Li, C (1994). Beyond nature
menarche: A children of twins approach. Manuscript versus nurture: DF analysis of nonshared influences
submitted for publication. on problem behaviors. Developmental Psychology, 30,
Meyer, J. M., Rutter, M., Silberg, J. L., Maes, H. H., 374 – 384.
Simonoff, E., Shillady, L. L., et al. (2000). Familial Rodgers, J. L., Rowe, D. C., & May, K. (1994). DF analysis of
aggregation for conduct disorder symptomatology: The NLSY IQ/achievement data: Nonshared environmental
role of genes, marital discord, and family adaptability. influences. Intelligence, 19, 157 – 177.
Psychological Medicine, 30, 759 – 774. Rowe, D. C. (1994). The limits of family influence: Genes, ex-
Neale, M. C., & Cardon, L. R. (1992). Methodology for genetic perience, and behavior. New York: Guilford Press.
studies of twins and families. Dordrecht: Kluwer. Rowe, D. C., Almeida, D. M., & Jacobson, K. C. (1999).
Neiderhiser, J. M., Reiss, D., Pederson, N. L., Lichtenstein, School context and genetic influences on aggression in
P., Hansson, K., Cederblad, M., & Elthammer, D. (2004). adolescence. Psychological Science, 10, 277 – 280.
Genetic and environmental influences on mothering of Rowe, D. C., Jacobson, K. C., & van den Oord, E. (1999).
adolescents: A comparison of two samples. Develop- Genetic and environmental influences on vocabulary IQ:
mental Psychology, 40, 335 – 351. Parental education level as moderator. Child Develop-
Plomin, R. (1994). Genetic research and identification of ment, 70, 1151 – 1162.
environmental influences: Emanuel Miller Memorial Rowe, D. C., & Waldman, I. D. (1993). The question of
Lecture 1993. Journal of Child Psychology and Psychiatry, ‘‘How?’’ reconsidered. In R. Plomin & G. E. McClearn
35, 817 – 834. (Eds.), Nature, nurture, & psychology (pp. 355 – 374).
Plomin, R. (1995). Genetics and children’s experiences in Washington, DC: American Psychological Association.
the family. Journal of Child Psychology and Psychiatry, 37, Rutter, M. (2000). Psychosocial influences: Critiques, find-
695 – 704. ings, and research needs. Development and Psychopathol-
Plomin, R., & Bergeman, C. (1991). The nature of nurture: ogy, 12, 375 – 405.
Genetic influence on ‘‘environmental’’ measures. Be- Rutter, M., Dunn, J., Plomin, R., Simonoff, E., Pickles, A.,
havioral and Brain Sciences, 14, 373 – 427. Maughan, B., et al. (1997). Integrating nature and nur-
Plomin, R., & Daniels, D. (1987). Why are children in the ture: Implications of person – environment correlations
same family so different from each other? Behavior and and interactions for developmental psychopathology.
Brain Sciences, 10, 1 – 16. Development and Psychopathology, 9, 335 – 364.
Plomin, R., DeFries, J. C., & Loehlin, J. C. (1977). Geno- Rutter, M., Pickles, A., Murray, R., & Eaves, L. J. (2001).
type – environment interaction and correlation in the Testing hypotheses on specific environmental causal
analysis of human behavior. Psychological Bulletin, 84, effects on behavior. Psychological Bulletin, 127, 291 – 324.
309 – 322. Rutter, M., & Silberg, J. (2002). Gene – environment inter-
Plomin, R., DeFries, J. C., McClearn, G. E., & McGuffin, P. play in relation to emotional and behavioral distur-
(2000). Behavioral Genetics. New York: Worth Publishers. bance. Annual Review of Psychology, 53, 463 – 490.
Purcell, S. (2002). Variance components models for gene – Scarr, S., & McCartney, K. (1983). How people make their
environment interaction in twin analysis. Twin Research, own environments: A theory of genotype – environment
5, 554 – 571. effects. Child Development, 54, 424 – 435.
Purcell, S., & Koenen, K. (2005). Environmental mediation Sexton, M., Fox, N.L, & Hebel, J. R. (1990). Prenatal expo-
and the twin design. Behavior Genetics, 35, 491 – 498. sure to tobacco: II Effects on cognitive functioning at age
Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear three. International Journal of Epidemiology, 19, 72 – 77.
models: Applications and data analysis methods (2nd ed.). Silberg, J. L., Parr, T., Neale, M. C., Rutter, M., Anglold, A.,
Thousand Oaks, CA: Sage Publications. & Eaves, L. J. (2003). Maternal smoking during preg-
Reiss, D., Plomin, R., & Hetherington, E. M. (1991). Ge- nancy and risk to boys’ conduct disturbance: An ex-
netics and psychiatry: An unheralded window on the amination of the causal hypothesis. Biological Psychiatry,
environment. American Journal of Psychiatry, 148, 53, 130 – 135.
283 – 291. Taylor, J., McGue, M., & Iacono, W. (2000). Sex differences,
Rodgers, J. L., Buster, M. A., & Rowe, D. C. (2001). Genetic assortative mating, and cultural transmission effects on
and environmental influences on delinquency: DF adolescent delinquency: A twin family study. Journal of
analysis of NLSY kinship data. Journal of Quantitative Child Psychiatry and Psychology, 41, 433 – 440.
Criminology, 17, 145. Thapar, A., Fowler, T., Rice, F., Scourfield, J., van den Bree,
Rodgers, J. L., & McGue, M. (1994). A simple algebraic M., Thomas, H., et al. (2003). Maternal smoking during
demonstration of the validity of DeFries – Fulker analy- pregnancy and attention deficit hyperactivity disorder
sis in unselected samples with multiple kinship levels. symptoms in offspring. American Journal of Psychiatry,
Behavior Genetics, 24, 259 – 262. 160, 1985 – 1989.
 Twin Studies with Measured Environments 1233

Truett, K. R., Eaves, L. J., Walters, E. E., Heath, A. C., He- N. Press, & A. Chapman (Eds.) Wrestling with behavioral
witt, J. K., Meyer, J. M., et al. (1994). A model system for genetics: Science, ethics, and the prospect of public conver-
analysis of family resemblance in extended kinships of sation. Baltimore: Johns Hopkins University Press.
twins. Behavior Genetics, 24, 35 – 49. Turkheimer, E., & Waldron, M. (2000). Nonshared envi-
Turkheimer, E. (2000). Three laws of behavior genetics and ronment: A theoretical, methodological, and quantita-
what they mean. Current Directions in Psychological Sci- tive review. Psychological Bulletin, 126, 78 – 108.
ence, 9, 160 – 164. van den Oord, E. (2001). Estimating effects of latent and
Turkheimer, E. (in press). Mobiles: A gloomy view of the measured genotypes in multilevel models. Statistical
future of research into complex traits. In E. Parens, Methods in Medical Research, 10, 393 – 407.