Intergenerational transmission of body mass index and associations with educational attainment

Participants were registered with the Netherlands Twin Register (NTR). The NTR collects longitudinal data on health, personality and lifestyle from twin families in the Netherlands. These families were recruited by NTR across the Netherlands through city councils, newsletters, and the NTR website. The present study is based on data collected between 2009 and 2012. Participants first received a written invitation including a link to a webpage, where they can log on to a web-based version of the survey with a unique, personal login name and password. If subjects did not access the web-based survey within the 6 weeks after the invitation, they received a paper version of the survey. Between 3–9 months after the paper versions of the survey were sent, subjects who had not responded received a reminder by post, or a reminder by email, if an email address was available. Several groups of non-responders (e.g. twins from incomplete twin pairs) were reminded by phone call [32]. We selected nuclear family members (parents and offspring), resulting in a sample of 19,135 participants. The survey was collected in two versions, a long and a short version. The short version, filled out by 3,421 participants, did not include questions concerning EA. Our sample comprises twins registered with NTR and their parents and siblings. We refer to the twins and their siblings as the offspring generation. There are 12,234 offspring aged between 16 and 97 years and 6,901 parents (2,817 fathers and 4,084 mothers) aged between 21 and 94 years. Note that the age distributions overlap, as twins registered with the NTR include younger twins, often registered with their parents and older twins, usually registered without their parents. A flowchart of the selection process can be found in supplementary figure S1.

The total number of families was 8,866 with an average of 1.4 offspring per family, where 18% of the families included only members of the parental generation. In addition, 45% of families included twins without their parents; 18% included twins and 1 parent; and 19% included twins and 2 parents (see supplementary table S1). There is information for both parent and offspring generation on BMI in 3,214 families (with N = 10,650), and on EA in 1,089 families (with N = 3,563) (see supplementary tables S2 and S3). EA from participants below the age of 25 years was not included, as they may not have yet reached their highest educational level, resulting in a sample size of 12,078 for EA. The data were checked for outliers and 5 individuals were excluded due to extreme values for height or weight.

BMI was based on self-reported height and weight, and was analyzed as a continuous variable. EA was self-reported and categorized according to the highest achieved educational level. In comparison to the one-dimensional progression that characterizes most secondary and tertiary educational systems around the world, the Dutch educational system is relatively complex, making the number of years of schooling alone insufficient to accurately capture educational level. The system is composed of two fundamental paths, namely a more vocational track (oriented towards manual occupation) and more general track. We classified the reported educational levels and ordered them according to academic performance requirements, associated SES, and projected income and occupational prestige associated with each level [33]. The resulting continuous EA scale has 8 levels (see Table 1). EA and BMI were adjusted for age of participant at the time of completing the survey. Age adjusted measures were used in the analyses. Outlying BMI data were checked against measurements obtained for the same subject from other surveys in the NTR database. Participants were grouped by sex in the analysis, and separate means were estimated for fathers, mothers, sons and daughters.

The steps included in SEM first involve model specification, i.e., specifying the hypothesized relationships between the variables, as outlined in Fig. 1. We chose full information (a.k.a. raw data) maximum likelihood (FIML) estimation to obtain estimates of the parameters (and standard errors), and the overall fit (i.e., χ², RMSEA) (Fan et al., 2016) of the model. The parameter estimates lead to an expected covariance matrix that is the most likely one given the observed data. The overall model fit measures provide information on how well our theoretical model fits the data.

We applied SEM to study the phenotypic transmission of BMI and EA from fathers and mothers to their offspring in a large population cohort from the Netherlands. Our model allows us to take into account the correlation of BMI and EA within and between parents, while examining the simultaneous transmission of BMI and EA. The latter includes the prediction of offspring BMI and EA from parental BMI and EA within (e.g., BMI to BMI) and across traits (e.g., BMI to EA). We tested whether the strength of the associations depend on the sex of the parent and the offspring. We examined direct effects, defined as the transmission from parent to offspring within or across traits, and indirect effects, resulting from 1) the correlations between EA and BMI within each parent and 2) the spousal correlations between parents for these traits. For instance, maternal EA has a direct effect path on offspring BMI, an indirect effect path through maternal BMI proportional to the EA-BMI correlation in the mother, and two indirect effect paths through paternal EA and BMI, given non-zero spousal correlations. We included a maximum of 6 offspring (3 males and 3 females) per family. We decided to limit the analysis to a maximum of 6 offspring because of computational considerations. This did not result in an appreciable loss of data: we analyzed 99.32% of the original study sample.

Model specifications: Paternal and maternal transmissions to sons and daughters were specified separately, i.e., we estimated 4 sets (mother-daughter, mother-son, father-daughter, and father-son) of 4 coefficients (EA to EA, BMI to BMI and the cross-transmission coefficients). The following parameters were specified, with estimates for offspring of the same gender (with a maximum of 3 sons and 3 daughters per family) constrained to be equal:

We estimated all intercepts for EA and BMI (father, mother, son, daughter) and the within and across-trait correlations for all residuals, i.e. the part in offspring BMI and EA that cannot be ascribed to parental transmission. All correlations were estimated without constraints and thus included 15 within-trait sibling correlations for EA and for BMI and 30 EA/BMI cross-trait correlations, as we allowed for 6 offspring per family.

Thus, in the full model (Fig. 1), all coefficients were estimated separately for sons and daughters and for mothers and fathers. Subsequently we imposed equality constraints on transmission parameters across genders of parents and offspring to test the differences in influence across various parent–offspring gender combinations. For these tests, we applied a Bonferroni correction. Given a family-wise α of 0.01 and 8 tests, the test-wise alpha was 0.01/8 = 0.00125. As mentioned above, we used FIML estimation to fit the model to the data. An advantage of FIML is that it handles missing data optimally: it exploits all available data and is more efficient than list- or pairwise deletion. Frequency tables and descriptive statistics were obtained from IBM SPSS (version 26). We carried out the structural equation modeling in the Lavaan package (version 0.6–6) in R (version 3.6.1).

Mean age of fathers was slightly higher than mothers (56 vs. 53 years), while offspring of both genders were both around 32 years. In both generations, the average male BMI was 0.5 points higher than the average female BMI. Mean BMI in parental generation was 26.1 for fathers and 25.6 for mothers, and in offspring generation it was 23.5 for sons and 23.1 for daughters. Correcting for age reduced the differences in BMI between the two generations (i.e., males: 2.6, females: 2.5) to 0.07 BMI points for males and 0.5 for females. Supplementary Figure S2 shows BMI distribution across age for males and females, with higher BMI in older individuals, and consistently higher male BMI in all age groups. EA levels were higher in males than in females, although the difference was smaller in the offspring generation. Figure 2 shows unadjusted mean BMI for each EA level clustered by sex, with higher EA levels associated with lower BMI, especially among females. Supplementary Figure S3 shows EA distribution across age for males and females, with higher levels of education and a narrowing gender gap for younger individuals. The variance of BMI (age adjusted) was larger in females than in males in the parental generation (19.63, CI 18.78–20.48 vs 10.49, CI 9.95–11.03), and in the offspring generation (13.27, CI 12.85–13.68 vs 9.30, CI 8.88–9.71) (Table 2). The variance of EA (age adjusted) was slightly larger in males than in females in the parental generation (3.35, CI 3.15–3.55 vs 2.59, CI 2.46–2.72), and in the offspring generation (2.83, CI 2.65–3.01 vs 2.33, CI 2.23–2.43). Negative within person EA-BMI correlations were observed for all family members (fathers, -0.102; mothers, -0.147; sons, -0.154; daughters, -0.173). Unadjusted correlations (see supplementary table S4) tended to be higher for most measures compared to age adjusted correlations.

Table 3 summarizes results for the full model, which are summarized in Fig. 3. The overall goodness of fit of the full model was acceptable (see supplementary table S7). A direct effect was defined as the transmission coefficient in the regression of parental BMI/EA on offspring BMI/EA. An indirect effect was calculated as the sum of three paths through correlations of one parental measures with the remaining three. For example, if we consider paternal BMI, there is one direct path and three indirect paths going through maternal BMI, paternal EA, and maternal EA. Total effects were calculated by adding up direct and indirect effects. In the full model, within-trait direct effects (parental BMI on offspring BMI and parental EA on offspring EA) were positive and significant for both parents, while cross-trait direct effects were insignificant, i.e., parental BMI did not have significant direct effects on offspring EA, and vice versa. Indirect effects were larger than direct effects for cross-trait transmission, due to significant cross-trait correlations at the parental level (see Table 2). For example, the standardized direct effect of maternal EA on male offspring BMI was -0.030, while the indirect effect was -0.072, adding to a total effect of -0.101). Residual within person EA—BMI correlations in the offspring generation remained significant after accounting for parental effects, which indicates EA and BMI are still associated after controlling for parental influence. To check if extreme BMI values (due to disorders such as anorexia nervosa or monogenic causes of morbid obesity) influenced the results, the analysis was repeated excluding 25 subjects with BMI values less than 15 or higher than 45. Parameter estimates obtained in the reduced sample hardly differed from those obtained in the full sample.

To assess gender differences in transmission from fathers and mothers to sons and daughters, we imposed equality constraints on transmission coefficients, and tested these constraints by likelihood ratio tests. We found significant differences only for maternal BMI on offspring BMI (supplementary table S5), where the direct effect was almost double in females (males: b = 0.105 CI 0.061, 0.148; females: b = 0.202 CI 0.161, 0.243). Similar analyses for parental gender showed no significant differences between paternal and maternal transmission coefficients (supplementary table S6), i.e., transmission from fathers and mothers were of equal magnitude. Based on the results of these tests, we arrived at model 2. Parameter estimates for this model, shown in Table 4, indicate small, but significant, direct effects for cross-trait transmission from parents to offspring. Standardized within-trait transmission coefficients were similar for both EA and BMI, and generally did not depend on the gender of the parent or offspring, with the exception of maternal BMI on offspring BMI. Cross-trait transmission coefficients were larger for offspring BMI than offspring EA, i.e., parental EA had a larger effect on offspring BMI than parental BMI had on offspring EA. Finally, the within trait spousal correlations were 0.228 for BMI, and 0.512 for EA. Cross-trait spousal correlations were -0.112 for maternal EA/paternal BMI, and -0.173 for paternal EA/maternal BMI. Both models had good model fit measures, with model 2 scoring slightly higher than the full model (e.g. x²: 59.5 vs 52.1; degrees of freedom: 51 vs 44) with a likelihood ratio test p-value = 0.39, indicating no significant difference between the two models. Model fit measures are presented in supplementary table S7.