Child Development, September/October 1999, Volume 70, Number 5, Pages 1151–1162

Genetic and Environmental Influences on Vocabulary IQ: Parental Education Level as Moderator

David C. Rowe, Kristen C. Jacobson, and Edwin J. C. G. Van den Oord

This article examines how parental education level moderates the genetic and environmental contributions to variation in verbal IQ. Data are from 1909 non-Hispanic Whites and African American sibling pairs from the National Longitudinal Study of Adolescent Health, which obtained nationally-based samples of identical (MZ) , fraternal (DZ) twins, full and half siblings, cousins (in the same household), and biologically unrelated siblings. In the whole sample, the variance estimate for (h2 ϭ .57, SE ϭ .08) was greater than that for shared environment (c2 ϭ.13, SE ϭ .04). Both heritability and the shared environmental estimate were moder- ated, however, by level of parental education. Specifically, among more highly educated families, the average h2 ϭ .74 (SE ϭ .10) and the average c2 ϭ .00 (SE ϭ .05). Conversely, among less well-educated families, herita- bility decreased and shared environmental influences increased, yielding similar proportions of variance ex- plained by genetic and environmental factors, average h2 ϭ .26 (SE ϭ .15), and average c2 ϭ .23 (SE ϭ .07).

INTRODUCTION adolescent adoption study (Scarr & Weinberg, 1978), the midparent-child correlation for IQ was .68 for bi- Genetically influenced traits are sometimes expressed ological offspring, but only .13 for adoptive off- more in certain environments than in others. For ex- spring. This pattern of parent-child associations in- ample, in bacteria the presence of particular nutrients dicated a substantial genetic influence on variation in the environment can regulate gene expression. in IQ (i.e., heritability) and only weak shared envi- Likewise, experiments on in rats have shown ronmental influences. Moreover, the same study also that enriched rearing can improve the performance of examined shared environmental effects via correla- genetically dull animals to about the level of geneti- tions between biologically unrelated siblings. These cally bright ones (Cooper & Zubek, 1958). These two correlations were close to zero for adolescent and young cases illustrate how the environment may moderate adult unrelated siblings, suggesting that a shared the genetic expression of certain traits. One way of family rearing environment had little impact on sib- testing this in humans is to examine whether the pro- ling similarity. On the basis of these findings and portion of variance in phenotypes due to genetic dif- other data on IQ, Scarr (1992) argued for no differen- ferences among individuals (i.e., the heritability) tial effect on IQ between average and above average varies across different environmental contexts. family environments. Previous research suggests that variance in IQ is Nonetheless, Scarr expected environmental influ- primarily attributable to genetic influences and not to ences to be of greater importance at the harmful end of environmental influences that are shared by children an environmental continuum, which would be rarely in the same family. By definition, common, or shared, represented in adoptive families, stating, “Environ- environmental influences are those influences that are ments that fall outside of the species normal range will shared among family members and serve to make not promote normal developmental patterns” (1992, family members similar to one another. Shared envi- p. 5). Scarr’s theory implies that the environment ronmental influences often include factors such as may moderate the relative genetic and environmental family structure and (SES). In influences on variation in development, because her- contrast, nonshared environmental influences are fac- itability may decrease and shared environmental ef- tors that are not shared among family members and fects may increase if a sample includes individuals contribute to differences among them. Examples of from these more harmful family environments. nonshared environmental influences among children Likewise, in their bioecological theory of nature- and adolescents are peer groups, perinatal traumas or nurture effects, Bronfenbrenner and Ceci (1994) ar- birth defects, and differential parental treatment. Scarr gued against the concept of a “single” heritability. presented evidence that genetic influences on chil- Their argument was that proximal processes, mean- dren’s IQ were stronger than common family envi- ronmental influences when families were working © 1999 by the Society for Research in , Inc. class to affluent (Scarr, 1992, 1993). Likewise, in her All rights reserved. 0009-3920/99/7005-0008

1152 Child Development ing the long-term interactions between children and number of pairs in the lower social class was quite their environments, are a necessary condition for the small (n ϭ 14 MZ and n ϭ 24 DZ pairs). Given the expression of any genetic trait. As they suggested, small sample size, it is possible that the correlations traits do not spring forth from fully found for the lower social class in this particular formed—in the developmental course of any single study are unreliable. individual, they must be nurtured. They further rea- Other studies have found evidence for secular soned that heritability is an estimate of the expression changes in heritability estimates. For example, the her- of genetic potential for a given trait. Thus, one specific itability of the age of first sexual intercourse has been prediction from their bioecological theory was that found to depend on birth cohort (Dunne et al., 1997). better environments (i.e., enhanced proximal influ- Specifically, the heritability of sexual intercourse on- ences) should increase heritability because genetic set was considerably lower among Australian men potential would be more fully realized. At the same and women older than 40 years than among younger time, however, they took the view that environmental Australian men and women. Thus, the shared envi- influences do not diminish in strength as environ- ronment effect (i.e., birth cohort) moderated genetic ments improve. They stated, influence on age of first intercourse. One explanation is that strict social norms may have prevented the ex- We take issue . . . with the prevailing conception of pression of any genetic propensity toward early sex- the reaction range simply as a curved plane, simi- ual intercourse in the older cohorts, but not in the lar to a bent piece of chicken wire that quickly younger ones. Two additional studies found some ev- straightens out to become horizontal . . . This rep- idence that the heritability of educational attainment resentation reflects the commonly held position (Heath et al., 1985) and (Sundet, Tambs, among behavioral geneticists that environment ex- Magnus, & Berg, 1988) increased with more recent co- erts an important influence only in severely de- horts. Again, these studies suggest that the greater prived environments . . . (p. 571) equality of educational opportunities available in more This last argument presents something of a para- recent cohorts has augmented the proportion of indi- dox, however. Given that the overall phenotypic vari- vidual differences that is due to genetic variation. Al- ation is the sum of genetic variation and environmen- though these examples suggest that the heritability of tal variation, if heritability increases in enriched certain phenotypes may vary across environments, environments, then by definition, less of the remain- these exceptions to the general rule are rare. ing variability would be environmental in origin. This One for the general lack of evidence for difference aside, Scarr (1992) and Bronfenbrenner and moderating effects of the environment on estimates of Ceci (1994) are in agreement that environments may heritability and shared environmental effects is that moderate the expression of genetic dispositions. the standard behavioral genetic model uses latent, Despite the conceptual allure of an environmental that is, unmeasured, variables as indicators of envi- context by heritability interaction, only a handful of ronmental influences (Anastasi, 1958). As the above studies have found that heritability is moderated by examples serve to illustrate, measured environmental context. Two early studies found support for the hy- variables increase statistical power to detect moderat- pothesis that the heritability of IQ would be greater in ing effects of the environment. Only a measured envi- more advantaged environments, and that shared en- ronment can specify the circumstances in which ge- vironmental influences would be greater in less advan- netic and environmental effects may change. Thus, taged environments (Fischbein, 1980; Scarr-Salapatek, the prior absence of findings of environmental mod- 1971). These studies, however, had important method- eration may be partly a function of limitations of the ological limitations. Specifically, in the sample of twins research design. from the Philadelphia area used by Scarr-Salapatek A different possibility, however, is suggested by an (1971), the zygosity of twins was not measured. In- alternative interpretation of genetic effects. As argued stead, Scarr-Salapatek calculated the correlations for by many behavior geneticists, prominently by Scarr identical (MZ) and fraternal (DZ) twins based on the and McCartney (1983), many genetic effects may also estimated proportions of same-sex MZ and DZ twins. occur through gene → environment (G→E) correla- In addition, social class was measured at the census- tions, a theory that “genes drive experience.” Such tract level, not through actual measurements of fam- correlations lead to particular genotypes being non- ily economic background. Although data from the randomly distributed across different environments. Swedish twins used by Fischbein (1980) did include In active G→E correlations, individuals may “pick” measured zygosity and family-level social class, social (consciously or unconsciously) those environments class was used as a categorical variable. Thus, the that most reinforce and sustain their genetic poten-

Rowe, Jacobson, and Van den Oord 1153 tials. For example, bright children might enjoy read- all high schools in the that had an 11th ing more than dull children. This process is often re- grade and had an enrollment of at least 30 students. A ferred to as a “self-selection” effect. In the reactive random sample of 80 high schools was selected from form of G→E correlation, social reactions to peoples’ this sampling frame, taking into consideration enroll- phenotypes may create an association between geno- ment size, region, school type, ethnicity, and urbanic- types and environments. For example, teachers might ity. The largest feeder school for each high school also recognize a greater potential among bright children was included in the sample. Seventy-nine percent of and may encourage them more strongly in their the schools initially contacted agreed to participate. studies. Both of these G→E correlations serve to re- Schools that refused to participate were replaced by duce the frequency of organism-environment “mis- another school in the same sampling stratum, result- matches,” such as bright children failing to read ing in a final sample of 134 schools. Within these books. If bright versus dull children could be ran- schools, 90,118 of 119,233 eligible students (75.6%) in domly assigned to high versus low book reading con- grades 7 to 12 completed a self-administered instru- ditions, moderating effects of the environment (in this ment for optical scanning. example, level of book reading) on IQ might emerge. Using both school roster information and informa- In real life, there are too few children in two of the tion provided by adolescents during the school inter- four possible experimental conditions: bright chil- view, a random sample of 15,243 adolescents also was dren with little reading experience, and dull children selected for a detailed home interview. This sample with voracious reading habits. In this way, the exist- was stratified by gender and age. The in-home inter- ence of a G→E correlation may greatly reduce the view was completed by 12,188 (79.5%) of these adoles- power to detect environmental moderation. In the stan- cents. In addition to this core sample, a number of dard behavioral genetic model G→E correlations are subsamples also was selected for the home interview. estimated as part of the genetic variation (Plomin, These subsamples included samples of disabled ado- DeFries, McClearn, & Rutter, 1997). lescents, adolescents from well-educated African Amer- In this article, we explore whether level of parental ican families, adolescents from typically understud- education (measured as a continuous variable) mod- ied racial and ethnic groups, and a special sibling pairs erates the genetic and environmental contributions to sample. Overall, 20,745 adolescents completed the in- variation in verbal IQ among adolescents. Although home interview. Details on the Add Health study have parental education may be related to children’s IQ via been reported elsewhere (Resnick et al., 1997). shared genes, parental education can also be consid- Included in the 20,745 adolescents is the Add ered a measure of “environmental quality,” because it Health pairs sample (N ϭ 3,139 sibling pairs), which is associated with the availability of intellectual stim- was selected using information from the in-school ulation and financial resources within the family. questionnaires and school rosters. Specifically, all ad- Moreover, because parental education is the same for olescents who were identified as pairs, half sib- siblings within a family, it is typically considered to lings, or unrelated siblings raised together were se- be a shared environmental influence. Our expectation lected for the home interview. Home interview data was that heritability estimates would be greater for were obtained from both the target adolescent and adolescents from families with more parental educa- his/her sibling. The sibling pairs were selected re- tion than from families with less parental education, gardless of whether they were present on the day of because the former may typically provide a satisfac- the school interview, and regardless of whether the tory context for intellectual development. Conversely, siblings attended the same school. A probability sam- shared environmental influences should be greater ple of full siblings also was drawn. The unrelated sib- among families with less parental education. ling pairs also included cousins who resided in the same household at the time of the survey. Thus, the pairs sample consists of all genetically informative METHOD sibling pairs. The average age in both the full sample and the sibling pairs sample is approximately 16 Sample and Procedure years (SD ϭ 1.7). The National Longitudinal Study of Adolescent The present sample contains 204 opposite-sex Health (Add Health) was designed to assess the dizygotic twins. The majority of same-sex twins were health status of adolescents and explore the causes of diagnosed as either monozygotic (MZ, n ϭ 247 pairs) adolescent health-related behaviors. Add Health or dizygotic (DZ, n ϭ 200) on the basis of their self- began with a total sample of over 90,000 adolescents reports of confusability of appearance. Confusability surveyed in school. The primary sampling frame was of appearance scales have been found to have greater

1154 Child Development than .90 agreement with zygosity determination based Table 1 Ethnic and Racial Characteristics of the Add Health on DNA evidence (Spitz et al., 1996). Eighty-nine Samples twin pairs of uncertain diagnosis were further classi- Sibling fied by their match on DNA genetic markers (47 DZ, Pairs 42 MZ). Twins were diagnosed as MZ if they were the Variable Full Sample Sample same for five or more genetic makers (error rate ap- proximately 4/1,000 or less) and DZ if they were dif- N 20, 745 6,278 Mean agea 16.1 (1.7) 16.0 (1.7) ferent at one or more markers. Zygosity determina- Ethnicity tion could not be ascertained for an additional 43 twin Non-Hispanic Whitesb 50.2 49.1 pairs. For the present analyses, these undecided (UD) African Americanc 20.8 22.8 twin pairs were combined with the DZ twins. Given Hispanicd 8.9 9.0 e that the UD group most likely contains a proportion Asian 2.3 2.0 Filipino 2.6 2.9 of true MZ twin pairs, this may slightly inflate our es- Cuban 2.2 1.0 timate of shared environment. (All analyses were re- Native American .5 .5 done without the UD twin pairs, and results did not Central/South American 1.4 1.3 differ from those presented here.) Although the major- Puerto Rican 2.1 1.7 ity of households had just two siblings enrolled in the Other 7.2 7.2 Missing 1.7 2.4 study, in households with multiple siblings, all possi- ble pairs were made (e.g., siblings A, B, and C form a Standard deviations are in parentheses. pairings AB, AC, and BC). This use of all pairs makes b Percentages of Whites are somewhat lower than national aver- the significance tests reported later slightly liberal. ages due to the over-sampling of racial and ethnic minorities in The data reported here come from the in-home the Add Health Study. c Wave I questionnaire. This in-home interview was Percentages of African Americans are greater than national aver- ages, due primarily to the over-sampling of African American ad- completed between May and December 1995. All re- olescents from college-educated families. spondents were given the same interview, which took d Includes Mexican Americans and Chicanos. from one to two hours to complete. All data were re- e Includes Chinese, Japanese, Korean, and Vietnamese adolescents. corded on laptop computers. Sensitive questions were asked via audio files drawn off the hard disk to coincide with presented questions. In addition to ican group, a decision was made to exclude these in- maintaining data security, this minimized the poten- dividuals. Nonetheless, over two thirds (67%) of the tial for interviewer bias. pairs sample were either White or Black. Table 1 presents the percentages of adolescents in Some pairs were excluded for special . various racial and ethnic groups. Because variance Twenty-two pairs were deleted because their relation- heterogeneity increases with the large number of ra- ship was not exactly a “sibling” relationship. Specifi- cial and ethnic groups included in the Add Health cally, these adolescents were identified as either aunt/ sample, the present analyses are restricted to adoles- uncle-nephew-niece pairs, boyfriend-girlfriend living cents pairs who belonged to one of the two largest ra- together, or unrelated adolescents living in a group cial and ethnic groups: non-Hispanic Whites (hereaf- home. We reasoned that these individuals would be ter, Whites; n ϭ 1,457 pairs, 46% of the sibling pairs less likely to experience similar family environments; sample) and African Americans (hereafter, Blacks; n ϭ thus they were deleted from the present analyses. 661 pairs, 21% of the sibling pairs sample). We consid- Further, 153 pairs were missing IQ data from one or ered including siblings from the third-largest group, both siblings. Thirteen pairs were outliers (i.e., one Hispanic Americans (n ϭ 227 pairs, 7.2% of the sam- or both sibling scored less than 50 on the verbal IQ ple); however, detailed analyses revealed that both scale). They were also deleted. Finally, a small number siblings indicated that in only 44% of the Hispanic of pairs (n ϭ 23) was deleted because of missing data American families was English the primary language regarding parental education. In sum, the present anal- spoken at home. A one-way analysis of variance yses are conducted on 1909 sibling pairs (1,322 White (ANOVA) revealed that Hispanic Americans from pairs, 587 Black pairs; 60.8% of the pairs sample; 90.1% non-English-speaking homes scored significantly lower of pairs who were identified as Black or White). on the Verbal IQ test than adolescents from English- Adolescents were from a variety of different living speaking home (M IQ ϭ 88.5, SD ϭ 14.2; M IQ ϭ 93.8, arrangements, with nearly one half (45.8%) living in SD ϭ 11.0, respectively), F(1, 394) ϭ 16.5, p Ͻ .01. Be- two-parent, biological families. Almost one quarter cause this systematic variation meant that there was (23.7%) lived in single-parent households (88% with considerable heterogeneity within the Hispanic Amer- biological mothers), and an additional 21.2% lived in

Rowe, Jacobson, and Van den Oord 1155 stepfamilies (78% with biological mothers and step- Table 2 Educational Levels of Residential Mothers and Resi- fathers). Just over 2% (2.1%) lived in adoptive homes dential Fathers, by Racial Group (with over 75% in two-parent adoptive homes), and Mother Father less than 1% (.4%) were in foster homes. Finally, an additional 6.8% lived in “other” arrangements, most White Black White Black often with other relatives (as in the case of cousins). N 2,415 1,039 2,004 443 No school .0 .1 .0 0 Measures Eighth grade or less 1.6 2.4 2.5 5.6 More than eighth grade, but Parental education. Adolescents were asked to indi- less than high school 8.9 14.0 7.2 7.9 cate the highest level of education completed by their High school equivalenta 38.6 39.7 35.5 41.1 residential mother (including biological mother, step- Some college 21.4 21.8 19.7 19.6 mother, adoptive mother, foster mother, and aunt) 4-Year college degree 21.7 17.3 24.4 19.9 Graduate or professional training 7.7 4.6 10.8 5.9 and residential father (including biological father, stepfather, adoptive father, foster father, and uncle). a Includes high school graduate, GED, and business or trade Of the 59% of individuals who reported information school instead of high school. for both mother and father (n ϭ 2,251), the correlation between residential mother and residential father ed- ucation was .53, p Ͻ .001. Thus, reports of mother and bal IQ test, and raw scores have been standardized by father education were averaged for each sibling (with age. Whites had a higher IQ mean than Blacks. This adolescents in single-parent families given the score racial difference would tend to increase sibling corre- for a particular parent). Likewise, the reliability of the lations in all sibling groups, because siblings usually educational reports is indicated by the high correla- share the same racial status. For this reason, the IQ tion between the siblings’ independent ratings of pa- scores were corrected for their association with racial rental education, r(1741) ϭ .81, p Ͻ .001. Thus, sibling group by using residuals from the regression of IQ on reports of parental education were averaged to create a dummy variable representing racial group (0 ϭ a single composite score for family educational level. White, 1 ϭ Black). The subsequent regression analy- The possible educational categories were: 0 ϭ never ses used these residual scores. went to school; 1 ϭ eighth grade or less; 2 ϭ more than eighth grade, but did not graduate high school; RESULTS 3 ϭ high school graduate or completed a GED or went to business, trade, or vocational school instead Table 3 presents the sibling correlations for the full of high school; 4 ϭ business, trade, or vocational sample of Black and White adolescents. In general, school after high school or some college; 5 ϭ gradu- the correlations order as expected by genetic theory. ated from college or university; and 6 ϭ professional In other words, sibling correlations increase accord- training beyond 4-year college. Educational levels for ing to the level of genetic relatedness, from .07 for bi- both mothers and fathers by race are reported in Table ologically unrelated sibling pairs to .73 for MZ twins. 2. The modal level of education for both residential There is no suggestion of a specific twin effect because mothers and residential fathers was 3, which corre- DZ twins and full-siblings were equally alike in IQ. sponds to a high school level education, and the mean Some correlations, however, suggested shared envi- level of average parental education was 3.7 (SD ϭ ronmental effects. In particular, the half sibling and 1.1), representing some post-high school education. cousin correlations are greater than their theoretical Add Health Picture Vocabulary Test (verbal IQ). This test is an abridged version of the Peabody Picture Vo- cabulary Test–Revised (PPVT-R). It was administered Table 3 Sibling Correlations for Peabody Picture Vocabulary IQ at the beginning of the in-home interview. This test of Sibling Group No. of Pairs r vocabulary involved the interviewer reading a word aloud. The respondent then selected the illustration MZ twins 176 .73*** that best fit the word. Each word had four, simple, DZ twins 347 .39*** black-and-white illustrations arranged in a multiple- Full siblings 795 .39*** choice format from which the respondent indicated Half siblings 269 .35*** Cousins 118 .21* his or her choice. For example, the word “furry” had Unrelated siblings 204 .07*** illustrations of a parrot, a dolphin, a frog, and a cat from which to choose. There were 78 items on the ver- * p Ͻ .05; *** p Ͻ .001.

1156 Child Development expectations based on a solely genetic model. Specifi- 103.0, SD ϭ 13.0; low education, M IQ ϭ 95.4, SD ϭ cally, if genetic factors were the only influences on 12.6, F(1, 3816) ϭ 257.6, p Ͻ .001. sibling similarity, then the half sibling correlation Table 4 summarizes the heritability and shared en- should be one half the full sibling correlation, and the vironment variance estimates obtained from the DF cousin correlation should be one quarter the full sib- models. In the full sample, the heritability of IQ was ling correlation ( rfull sibling ϭ .39). As can be seen from .57, and the shared environmental effect was .13. Esti- Table 2, the half sibling correlation of .35 and the mates of shared environmental (c2) and genetic effects cousin correlation of .21 exceeded the .195 and .098, (h2) were similar for Black and White adolescents (.13 respectively, that would be expected under a purely versus .12, and .57 versus .58, respectively). The ad- genetic model. Moreover, the twin groups indicated a justed R2 from this model was .17, F(3, 3814) ϭ 257.0, small shared environmental effect (using the common p Ͻ .001. The similarity of the variance estimates for formula for shared environment, 2rDZ Ϫ rMZ ϭ .05). Blacks and Whites indicates that the two groups can To explore the moderating effect of parental educa- be safely combined for further analyses. In the regres- tion, a DeFries-Fulker (DF) regression equation was sions that divided adolescents on the basis of their used to estimate genetic and environmental parame- parents’ education, we found a powerful moderation ters (DeFries & Fulker, 1985). The DF equation is an effect for both genetic and shared environmental in- unstandardized regression equation. The phenotype fluences. Heritability was greater for adolescents in of sibling 1 (P1) is predicted from the phenotype of homes with more educated parents (.74 versus .26), sibling 2 (P2), the coefficient of genetic relatedness (R; whereas the shared environmental estimate was lower with R ϭ 1.0 for MZ twins, R ϭ .5 for DZ twins and (.00 versus .23). full siblings, R ϭ .25 for half-siblings, R ϭ .125 for To test whether these differences were significant, cousins, and R ϭ .0 for unrelated siblings), and the in- we used an extension of the DF analysis that added teraction of the latter two variables (R ϫ P2). In this further interaction terms (for details, see LaBuda & form, the unstandardized regression coefficient on P2 DeFries, 1990). The full, extended equation is: is the estimate of the proportion of variance due to P ϭ b ϩ b P ϩ b R ϩ b ED ϩ (b RиP ) ϩ shared environmental effects (c2) and the regression 1 o 1 2 2 3 4 2 (b RиED) ϩ (b P иED) ϩ (b P иRиED) coefficient on the interaction term is the heritability 5 6 2 7 2 coefficient (h2; for details of how the unstandardized where P is verbal IQ (the subscripts 1 and 2 designate estimates translate into c2 and h2, see Cherny, DeFries, each sibling), R is the coefficient of genetic related- & Fulker, 1992; Rodgers & McGue, 1994). The DF ness, and ED represents level of parental education, model was run for the full sample, for Whites only, for coded as a continuous variable. If level of parental ed- Blacks only, and then for low and high parental edu- ucation moderates the c2 estimate, then the unstand- cation groups using high school graduate as the cut- ardized regression coefficient on the interaction term off (i.e, low education ϭ high school education or less, P2иED (i.e., the b6 coefficient) would be significant. high education ϭ more than a high school education). Similarly, the unstandardized regression coefficient The parental educational groups differed in average for the P2иRиED term (b7) is the test of whether the her- adolescent (unadjusted) IQs: high education, M IQ ϭ itability of verbal IQ differs across education levels.

Table 4 DeFries-Fulker Regression Equation Heritability and Shared Environmental Variance Estimates

2 a 2 a b No. of Pairs c (␴SE) t Value h (␴SE) t Value

Full sample 1,909 .1 .04 3.25*** .5 .08 7.13*** 3 7 Racial groups Whites 1,322 .1 .05 2.40** .5 .10 5.80*** 2 8 Blacks 587 .1 .07 1.86* .5 .15 3.80*** 3 7 Educational groups Low education 753 .2 .07 3.29*** .2 .15 1.73* 3 6 High education 1,156 .0 .05 .00 .7 .10 7.40*** 0 4

2 2 Note: c , shared environmental effects; h , heritability; ␴SE ϭ standard error. High education reflects greater than a high school education; low education reflects a high school education or less. a Standard errors corrected for double-entry of data. b Significance of t tests are based on one-tailed tests. * p Ͻ .05; ** p Ͻ .01; *** p Ͻ .001. Rowe, Jacobson, and Van den Oord 1157

Table 5 Results from the Extended DF Regression to Test The precise form of the interaction effect can be es- 2 2 Whether Parental Education Moderates c and h timated from formulas that combine the main effect a b and interaction regression coefficients. For the esti- B ␴SE t Value mation of the shared environmental effect, the for-

Intercept (b0) Ϫ9.78 1.87 Ϫ5.23*** mula is: 2 c c (b1) .50 .15 3.33*** 2 R (b2) 4.79 3.90 1.23 c ϭ .50 Ϫ .12(ED), ED (b3) 2.73 .49 5.57*** 2 c where ED is level of parental education and .50 and h (b4) Ϫ.12 .30 Ϫ.40 RиED (b5) Ϫ1.25 .99 Ϫ1.26 Ϫ.12 are the unstandardized regression coefficients 2 2 c иED (b6) Ϫ.12 .04 Ϫ3.00*** for the estimate of c when parental education is zero h2иED (b ) .19 .08 2.38*** 2 7 (b1) and the interaction of c with level of parental ed- b 2 2 ucation ( 6), respectively (see Table 5). Similarly, the Note: c , shared environmental effects; h , heritability; ␴SE ϭ stan- dard error; R, coefficient of genetic relatedness. corresponding formula for heritability is: a Standard errors corrected for double-entry of data. h2 ϭ Ϫ.12 ϩ .19(ED), b Significance is based on one-tailed t tests. c The estimates of c2 and h2 in these equations represent the shared where Ϫ.12 is the unstandardized regression coeffi- environmental effects and heritability when ED ϭ 0, that is, when cient for the estimate of heritability when parental ed- parents received no formal education. *** p Ͻ .001. ucation is zero (b4), and .19 is the unstandardized re- gression coefficient for the heritability by parental education interaction (b7; see Table 5). Table 5 shows that, as expected, both environ- Figure 1 takes the above equations and plots the mental (c2) and genetic (h2) interaction effects were values of c2 and h2 across varying levels of parental statistically significant: b6 ϭ Ϫ.12, adjusted SE ϭ .04, education. As can be seen from this figure, when par- t(3810) ϭ Ϫ3.00, p Ͻ .001; and b7 ϭ .19, adjusted SE ϭ ents have had at least a high school education (on .08, t(3810) ϭ 2.38, p Ͻ .001, respectively. The negative average; 60.6% of the sample), shared environmental sign on the b6 term indicates that shared environmen- effects were always below 20% of the total variance tal effects decrease with increasing levels of parent and genetic influences came to predominate. The con- education. Likewise, the positive sign of the b7 term verse held when the average parental education was indicates that as parental education rises, the herita- less than a high school degree. In this range, the influ- bility of verbal IQ also rises. Moreover, the fact that ence of shared environment was actually greater than 2 the coefficient for c was significant, b1 ϭ .50, adjusted that of genes. SE ϭ .15, t(3810) ϭ 3.33, p Ͻ .001, yet the coefficient At the extremes of parental education, variance es- 2 for h was not, b4 ϭ Ϫ.12, adjusted SE ϭ .30, t(3810) ϭ timates become less accurate because fewer families Ϫ0.40, p Ͼ .50, indicates that at the very extreme level are found (less than 1.1% at the lower extreme and of parental education (i.e., when parental education ϭ less than 10% at the upper extreme) and because the 0, representing no formal education), sibling resem- function form in Figure 1 was necessarily linear, blance for verbal IQ would be due entirely to shared which prevented the function from tapering off at the family environments. Finally, the adjusted R2 for this extremes. This explains why the estimate of c2 for model was .21, compared to the R2 ϭ .17 obtained those adolescents whose parents were college gradu- from the model without interactions, indicating that ates was less than zero, and why the estimate for h2 the full model explained an additional 4% of the vari- was less than zero for adolescents whose parents re- ance. Because we suspect that the group of cousins ceived no formal schooling. living in the same household could have been living The moderating effect of the environment on ge- there for only a short period, and because they were netic and shared environmental influences on verbal selected to be in the same household (typically with IQ also can be seen in the pattern of sibling correla- an aunt or uncle) by an unknown process, this DF re- tions when adolescents are divided into two groups gression was repeated with the cousin group omitted. on the basis of parental education (see Table 6). To re- In the new DF regression, the interaction terms were duce sampling variation, the less related sibling groups again statistically significant and the parameter esti- were combined into a single group (unrelated sib- mates were similar to those shown in Table 5 (results lings, cousins, and half siblings), as were the moder- not shown). Moreover, the pattern of regression coeffi- ately related sibling groups (DZ twins and full sib- cients also held when the DF equation was repeated for lings). A third group consisted of the MZ twins only. Whites and Blacks separately (results not shown). Thus, As shown in Table 6, the correlational patterns are the findings reported in Table 5 appear to be robust. consistent with the interactions found by the DF anal- 1158 Child Development

Figure 1 Relations between genetic and shared environmental variance components by level of parental education. Standard errors at the means and the extremes are indicated by vertical lines. Arrows point to the estimates of heritability and shared en- vironmental influences at the sample mean. yses. In the low relatedness group, the sibling correla- p Ͻ .05. Although the sibling correlations in Table 6 tion was greater in the offspring of low educated par- are given for illustrative purposes, they represent an ents (.32) than those offspring of highly educated arbitrary slice of the sample, whereas the DF analy- parents (.10), indicating greater shared environmen- sis estimated c2 and h2 using education as a continu- tal influences on sibling similarity in less educated ous variable. families, Zdifference ϭ 2.79, p Ͻ .01. In contrast, weaker genetic effects in less educated families were indi- DISCUSSION cated by a smaller MZ twin correlation in the low ed- ucation group: .75 in the high education group ver- Similar to results from two earlier studies (i.e., Fisch- sus .55 in the low education group, Zdifference ϭ 2.09, bein, 1980; Scarr-Salapatek, 1971), the present study adds support to the hypothesis that the heritability of IQ varies across social class levels. This study found Table 6 Sibling Correlations by Level of Parental Education that level of parental education moderated both the Low Education High Education genetic and shared environmental influences on Pea- body Picture Vocabulary IQ. Based on significant inter- No. of No. of action terms in an extended DeFries-Fulker (DF) model, Pairs r Pairs r heritability increased from about 25% in the offspring of parents with less than 12 years of education to about High relatedness 52 .55*** 124 .75*** Moderate relatedness 424 .33*** 718 .37*** 74% in the offspring of parents with greater than a high Low relatedness 277 .32*** 314 .10# school education. In contrast, the shared environmental effect decreased from 20–40% of IQ variation in the off- Note: High education reflects greater than a high school educa- spring of parents without high school degrees to about tion; low education reflects a high school education or less. High zero for more-educated parents. relatedness ϭ MZ twins; moderate relatedness ϭ DZ twins and full siblings; low relatedness ϭ half siblings, cousins, and unre- Our findings confirmed hypotheses stemming lated siblings. from bioecological theory of Bronfenbrenner and Ceci # p Ͻ .10; *** p Ͻ .001. (1994) and the theory of Scarr (1992), namely, that trait Rowe, Jacobson, and Van den Oord 1159 will increase with improvement in envi- less responsive to variation in family environments ronmental conditions. When the Add Health sample because shared environmental effect was weaker. was split between parents with a high school diploma Overall, the reaction range of IQ had the property en- (or less) versus some education beyond high school, visioned by Scarr (1992, 1993)—increasing environ- the heritability of verbal IQ was much greater for the mental effects in more deprived environments. children of more educated parents (h2 ϭ .74 versus Although encouraging, these results bear further .26). Thus, the genetic potential for learning vocabu- replication. For example, Capron and Duyme (1989) lary was expressed more fully when parents were bet- lacked the sample size to investigate any moderat- ter educated. ing effects of parental social class on the magnitude Although no exact threshold may exist for shared of genetic and shared environmental influences. In family environmental effects, it is clear that shared en- another study, moderating effects of the environ- vironmental effects became significant only among ment on children’s academic abilities were not found relatively uneducated parents. This finding corre- (Van den Oord & Rowe, 1998). In this study of chil- sponds to the hypothesis that environmental effects dren of mothers in the National Longitudinal Sur- on IQ may be nonlinear. According to this theoretical vey of Youth (NLSY), family environmental quality model, when family environments are average or bet- failed to moderate estimates of either the shared en- ter, the “reaction surface” (Turkheimer & Gottesman, vironmental effect or the genetic effect. Our failure 1991) of the IQ phenotype to environmental variation to replicate these earlier results was unexpected, given is fairly flat. In the region of poor quality family envi- that the NLSY, just as the Add Health study, was a ronments, however, the IQ phenotype could be highly large sample of children from a broad range of socio- responsive to environmental variation. economic statuses. One important implication from the present study One explanation may lie in the age difference is that results bear on the responsiveness of IQ to vari- among the two samples: The NLSY children averaged ation in normal family environments. In the group of about 9 years of age, whereas the average age in the parents with higher levels of education, family envi- Add Health sample was approximately 16, with 80% ronmental influences were not contributing to indi- of adolescents between 13.5 and 18 years old. If harm- vidual variation in IQ. They may, however, have ful environmental effects accumulate over time, then raised the mean IQ of the whole group. Shared envi- perhaps the moderating effect of environment may ronmental effects were most evident when the par- be more detectable in a sample of older children. Al- ents had less than a high school education. The chil- though the age ranges in the two samples do not dren of these parents also had lower IQs. Thus, the overlap, thus no direct comparison can be made, we provision of better family environments could raise did examine whether moderating effects could be IQs, and should decrease the IQ gap between socio- found in a subsample of adolescents where both sib- economic groups. lings were younger than 14 years old (results not The magnitude of IQ gain would depend on the shown). Although the sample size was quite small number of children who could be influenced and (n ϭ 336 pairs), the moderating effects of parent ed- the strength of any likely change. For example, at one ucation on heritability and the shared environmen- extreme, in 8.5% of the sample neither parent had tal influence were in the same direction as results more than an 11th grade education. For this level of from the whole sample, and were statistically sig- parental education, the estimate of shared environ- nificant. Thus, moderating effects of parental edu- mental effect was approximately .40. The average IQ cation were found even among the youngest adoles- (unresidualized on race) was 90.8, with a standard de- cents in our sample. viation of 11.9. Hence, a 1 SD improvement of their A more likely explanation lies in an important de- family environments could increase their IQs from sign difference between the studies. Specifically, the the 90.8 to approximately 98, that is, a gain of (.40 ϫ NLSY sample had only three levels of genetic kinship: 11.92),5 which equals 7.5. A 2 SD change could yield a full siblings, half siblings, and cousins, whereas this gain of approximately 15 IQ points. These values lie study also included twins and biologically unrelated within the range observed in a cross-fostering study siblings. To evaluate the effect of restricting our data of IQ done in France, where children born into poor analyses to the former groups, the DeFries-Fulker re- families were adopted by affluent families (Capron & gression was repeated omitting the twins and unre- Duyme, 1989; McGue, 1989). The children’s IQ gain lated siblings (a loss of almost 40% of the sample). Al- was 12 points (with a large standard error, because of though the pattern of results remained the same (i.e., the small number of children sampled). At higher a negative coefficient for the c2 ϫ ED interaction and a levels of parental education, however, IQ should be positive coefficient for the h2 ϫ ED interaction), the 1160 Child Development deletion of the twins and unrelated siblings in our for children with disparate intellectual abilities. Fur- sample caused the interaction terms to become non- thermore, intellectual simulation within families can significant. Thus, it is possible that large, multiple be compensated for, or complimented by, stimulation kinship samples are necessary to have sufficient sta- in schools and in peer groups. With their active G→E tistical power to detect moderating effects of the fam- effects, bright children can accelerate intellectual de- ily environment. velopment, regardless of their immediate family cir- Moreover, it is possible that our results reflect cumstances. In contrast, children with a potential for something other than a moderating effect of parental higher IQ in lower SES environments may not be able education on the genetic and environmental contri- to find the intellectual stimulation they need, given butions to sibling similarity in IQ. For example, heri- the lower quality of schooling often found in less tability may increase or decrease as phenotypic scores well-educated communities. Thus, the apparent mod- move toward an extreme. In the present study, ado- erating effect found in the present study may repre- lescents from the least educated families also have the sent a G→E correlation. lowest IQs, and adolescents from the most educated Furthermore, level of parental education, in and of homes have the highest IQs, F(2, 3815) ϭ 97.5, p Ͻ itself, is also likely to be influenced by genetic factors. .001. The effect of extreme scores on heritability has Thus, more highly educated parents are more likely been investigated using squared-phenotype terms in to pass on genes related to higher IQ as well as pro- DF regression analyses, or by regressing within-sibling vide more intellectually stimulating environments. pair phenotypic differences on phenotypic sums (Bailey There is evidence, at least among infants and chil- & Revelle, 1991; Detterman, Thompson, & Plomin, dren, that part of the correlation between the home 1990). Some twin studies have found no evidence of environment and offspring intelligence is mediated differential heritability (on infants 1–3 years, Cherny, genetically (Braungart, Fulker, & Plomin, 1992; Coon, Cardon, Fulker, & DeFries, 1992; on older adoles- Fulker, DeFries, & Plomin, 1990). Therefore, contin- cents, Sundet, Eilertsen, Tambs, & Magnus, 1994). Two ued investigations of the genetic and environmental twin studies were supportive of our findings, how- influences on the relationship between family envi- ever. One found greater IQ heritability among high ronment and IQ are warranted. IQ twin children, who were most likely the children Finally, this study had several limitations. One was of better educated parents (Bailey & Revelle, 1991). the use of just a single measure of IQ, verbal IQ, from The other study, of sixth grade twins in Cleveland, the Peabody Picture Vocabulary test. Because Add Ohio, found interaction effects most comparable to Health did not include a measure of nonverbal IQ, or ours: an increase of shared environmental influences nonvocabulary, verbal IQ, we cannot demonstrate when children had lower IQs and a trend toward conclusively that the genetic and environmental etiol- greater heritability when they had higher IQs (Thomp- ogy of general intelligence (g), rather than just vocab- son, Detterman, & Plomin, 1993). It should be noted, ulary, is moderated by the environment. Vocabulary however, that in an earlier report using fewer twin IQ, however, loads strongly on g in factor-analytic pairs from the same Western Reserve Twin Sample, studies of IQ, which makes the moderating effects Detterman et al. (1990) found the reverse: the herita- more plausible. An additional limitation is that our bility of IQ was greater among lower IQ children. findings were restricted to Non-Hispanic Whites and Findings for such interaction effects might become Blacks, because the other ethnic and racial groups more consistent if studies systematically over-sampled within Add Health were too few in number (or too individuals at IQ extremes. heterogeneous) to permit separate analyses. Although The present study found that for the offspring of in many cases racial and ethnic groups show similar parents with more than a high school diploma, verbal developmental processes (Rowe, Vazsonyi, & Flannery, IQ was highly heritable. It is not the case, however, 1994), the results presented here may not generalize that all working-class to affluent families provide the to other racial or ethnic groups. The DF regression same degree of intellectual stimulation. To explain a also assumed that the equal environments assump- high heritability of IQ despite unequal environments tion was not violated in any sibling group. This as- across families, a process of G→E correlation can be sumption is that the environmental influences that af- invoked. Brighter children may extract information fect a phenotype correlate equally between siblings from their environments more rapidly, handle com- for each type of sibling pair. In other words, it is as- plex information better, and expose themselves to in- sumed that MZ twins do not experience more similar formation at a greater rate than do less able children environments than DZ twins, full siblings, or any (Carroll, 1997; Gottfredson 1997). The same objective other sibling group and/or that more similar treat- environment is, in actuality, different environments ment does not systematically predict sibling similar- Rowe, Jacobson, and Van den Oord 1161 ity. Data on this assumption for twins is presented REFERENCES by Rowe (1994; also, Hettema, Neale, & Kendler, Anastasi, A. (1958). , environment, and the ques- 1995), but this assumption was not tested explicitly tion “How?” Psychological Review, 65, 197–208. in the present study. Finally, our model incorporated Bailey, M. J., & Revelle, W. (1991). Increased heritability for neither assortative mating, which could increase the lower IQ levels. Behavior , 21, 397–404. correlation of all siblings except MZ twins, nor genetic Braungart, J., Fulker, D. F., & Plomin, R. (1992). Genetic me- dominance, which could increase the correlation of DZ diation of the home environment during infancy: A sib- twins and biological full siblings, but not of the other ling adoption study of the HOME. Developmental Psy- sibling types. Although our simpler model is a limita- chology, 28, 1048–1055. tion, it probably captures the major sources of vari- Bronfenbrenner, U., & Ceci, S. J. (1994). Nature-nurture re- ance, which were additive genetic influences and conceptualized in developmental perspective: A bio- ecological model. Psychological Review, 101, 568–586. shared environmental influences, satisfactorily. Capron, C., & Duyme, M. (1989). Assessment of the effects of socio-economic status on IQ in a full cross-fostering ACKNOWLEDGMENTS study. Nature, 340, 552–554. Carroll, J. B. (1997). , intelligence, and public This research is based on data from the Add Health perception. Intelligence, 24, 25–52. project, a program project designed by J. Richard Cherny, S. S., Cardon, L. R., Fulker, D. W., & DeFries, J. C. Udry (Principal Investigator) and Peter Bearman, and (1992). Differential heritability across levels of cognitive funded by grant P01-HD31921 from the National Insti- ability. Behavior Genetics, 22, 153–162. tute of Child Health and Human Development to the Cherny, S. S., DeFries, J. C., & Fulker, D. W. (1992). Multiple regression of twin data: A model fitting approach. Behav- Carolina Population Center, University of North Caro- ior Genetics, 22, 489–497. lina at Chapel Hill; with cooperative funding partic- Coon, H., Fulker, D. W., DeFries, J. C., & Plomin, R. (1990). ipation by the National Cancer Institute; the Na- Home environment and cognitive ability of 7-year-old tional Institute of Alcohol Abuse and Alcoholism; children in the Colorado Adoption Project: Genetic and the National Institute on Deafness and Other Com- environmental etiologies. Developmental Psychology, 26, munication Disorders; the National Institute of Drug 459–468. Abuse; the National Institute of General Medical Cooper, R. M., & Zubek, J. P. (1958). Effects of enriched and Sciences; the National Institute of Mental Health; restricted early environments on the learning ability of the National Institute of Nursing Research; the Of- bright and dull rats. Canadian Journal of Psychology, 12, fice of AIDS Research, NIH; the Office of Behavior 159–164. DeFries, J. C., & Fulker, D. W. (1985). Multiple regression and Social Science Research, NIH; the Office of the analysis of twin data. Behavior Genetics, 15, 467–473. Director, NIH; the Office of Research on Women’s Detterman, D. K., Thompson, L. A., & Plomin, R. (1990). Dif- Health, NIH; the Office of Population Affairs, HHS; ferences in heritability across groups differing in ability. the National Center for Health Statistics, Centers for Behavior Genetics, 20, 369–384. Disease Control and Prevention, HHS; the Office of Dunne, M. P., Martin, N. G., Statham, D. J., Slutske, W. S., Minority Health, Centers for Disease Control and Dinwiddie, S. H., Bucholz, K. K., Madden, P. A. F., & Prevention, HHS; the Office of Minority Health, Of- Heath, A. C. (1997). Genetic and environmental contri- fice of the Assistant Secretary for Health, HHS; the butions to variance in age at first sexual intercourse. Psy- Office of the Assistant Secretary for Planning and chological Science, 8, 211–216. Fischbein, S. (1980). IQ and social class. Intelligence, 4, 51–63. Evaluation, HHS; and the National Science Founda- Gottfredson, L. S. (1997). Why g matters: The complexity of tion. Persons interested in obtaining data files from everyday life. Intelligence, 24, 79–132. The National Longitudinal Study of Adolescent Heath, A. C., Berg, K., Eaves, L. J., Solaas, M. H., Corey, L. Health should contact Jo Jones, Carolina Population A., Sundet, J., Magnus, P., & Nance, W. E. (1985). Educa- Center, 123 West Franklin Street, Chapel Hill, NC tional policy and the heritability of educational attain- 27516-3997; e-mail: [email protected]. ment. Nature, 314, 734–736. Hettema, J. M., Neale, M. C., & Kendler, K. S. (1995). Physi- cal similarity and the equal environments assumption in ADDRESSES AND AFFILIATIONS twin studies of psychiatric disorders. Behavior Genetics, 25, 327–335. Corresponding author: David C. Rowe, Campus Box LaBuda, M. C., & DeFries, J. C. (1990). Genetic etiology of 210033, University of Arizona, Tucson, AZ 85721-0033; reading disability: Evidence from a . In G. T. e-mail: [email protected]. Kristen C. Jacobson is Pavlidis (Ed.), Perspectives on dyslexia (Vol. 1, pp. 47–76). at Pennsylvania State University, University Park, New York: Wiley. PA; and Edwin J. C. G. Van den Oord is at Utrecht McGue, M. (1989). Nature-nurture and intelligence. Nature, University, Utrecht, the Netherlands. 340, 507–508. 1162 Child Development

Plomin, R., DeFries, J. C., McClearn, G. E., & Rutter, M. Scarr, S., & Weinberg, R. A. (1978). The influence of “family (1997). Behavioral genetics (3rd Ed.). New York: W. H. background” on intellectual attainment. American Socio- Freeman. logical Review, 43, 674–692. Resnick, M. D., Bearman, P. S., Blum, R. W., Bauman, K. E., Scarr-Salapatek, S. (1971). Race, social class, and IQ. Science, Harris, K. M., Jones, J., Tabor, J., Beuhring, T., Sieving, R. E., 174, 1285–1295. Shew, M., Ireland, M., Bearinger, L. H., & Udry, J. R. Spitz, E., Moutier, R., Reed, T., Busnel, M. C., Marchaland, (1997). Protecting adolescents from harm: Findings from C., Roubertoux, P. L., & Carlier, M. (1996). Comparative the National Longitudinal Study of Adolescent Health. diagnoses of twin zygosity by SSLP variant analysis, Journal of the American Medical Association, 278, 823–832. questionnaire, and dermatoglyphic analysis. Behavior Rodgers, J. L., & McGue, M. (1994). A simple algebraic dem- Genetics, 26, 55–63. onstration of the validity of DeFries-Fulker analysis in Sundet, J. M., Eilertsen, D. E., Tambs, K., & Magnus, P. (1994). unselected samples with multiple kinship levels. Behav- No differential heritability of intelligence test scores across ior Genetics, 24, 259–262. ability levels in Norway. Behavior Genetics, 24, 337–339. Rowe, D. C. (1994). The limits of family influence: Genes, expe- Sundet, J. M., Tambs, K., Magnus, P., & Berg, K. (1988). On rience, and behavior. New York: Guilford Press. the question of secular trends in the heritability of intel- Rowe, D. C., Vazsonyi, A. T., & Flannery, D. J. (1994). No ligence test scores: A study of Norwegian twins. Intelli- more than skin deep: Ethnic and racial similarity in de- gence, 12, 47–59. velopmental process. Psychological Review, 101, 396–413. Thompson, L. E., Detterman, D. K., & Plomin, R. (1993). Dif- Scarr, S. (1992). Developmental theories for the 1990s: De- ferences in heritability across groups differing in ability, velopment and individual differences. Child Develop- revisited. Behavior Genetics, 23, 331–336. ment, 63, 1–19. Turkheimer, E., & Gottesman, I. I. (1991). Individual differ- Scarr, S. (1993). Reply: Biological and cultural diversity: The ences and the canalization of human behavior. Develop- legacy of Darwin for development. Child Development, mental Psychology, 27, 18–22. 64, 1333–1353. Van den Oord, E. J. C. G, & Rowe, D. C. (1998). An examina- Scarr, S., & McCartney, K. (1983). How people make their tion of -environment interaction for academic own environments: A theory of genotype→environment achievement in an U.S. National Longitudinal Survey. effects. Child Development, 54, 424–435. Intelligence, 25, 205–228.