Table 2 shows descriptive statistics for the raw parental variables contributing to this article. Both cohort and age effects are visible in both studies, as well as differences between them. Parents of younger and more recently born twins were younger on average, and German parents tended to be older than Minnesotan parents, matched for twin ages. This likely reflected trends toward later parenting (OECD 2017), as the German parents tended to have been born about a generation later. Parents of younger and more recently born twins were also taller on average, reflecting world-wide general height trends over the past century (Clark 2008). German parents tended to be taller than Minnesota parents, likely also reflecting their more recent birth. Parents of younger twins tended to weigh less on average, consistent with generally observed age trends in weight in adulthood, and German parents tended to weigh noticeably less than Minnesotan parents, matched for twin ages, despite Minnesota parents being assessed on average about 20 years previously. This was reflected in the BMIs as well. German parents varied more in height, but Minnesotan parents varied more in weight and BMI. In both countries, fathers’ heights varied more than mothers’, but mothers’ weights, and especially BMIs, varied more than fathers’.
The raw twin descriptive statistics contributing to BMI are shown in Table 3. At age 11, Minnesota twins were taller, but this had reversed by age 17, suggesting earlier puberty-related growth spurts in Minnesota, especially for girls. Minnesota twins weighed on average as much as 1 standard deviation more at all ages; this was consequently reflected in higher BMIs. At age 11, Minnesota girls were slightly taller and weighed more than boys (consistent with girls’ earlier puberty), but they were evenly matched in Germany, and boys were taller and weighed considerably more in both countries by age 17 (slightly taller and heavier at age 5 too). Variances in height were greatest at age 11, reflecting variance in timing of the pubertal growth spurt, but variances in weight and BMI increased with age in both sexes and countries. Like their parents, German twins varied more in height than did Minnesota twins, but Minnesota twins varied more in weight and BMI. Except at age 11, boys resembled their fathers in varying more in height than did girls. Minnesota girls resembled their mothers in varying more in weight and BMI than boys at all ages, but German girls only showed greater variance in BMI in Cohort 4 (ages 23–24). Figure 2 summarizes these comparisons in graphical form. The means were consistent with data from other, larger sources, but these sources do not usually address variances. Differences in variance may indicate factors relevant to mean level trends, however, such as society-level disparities in economic and community resources.
Rates of overweight and obesity
Table 4 shows rates of overweight and obesity in parents and twins, based on the standard BMI cut-offs of 25 and 30 for adults and both CDC and WHO standards for those under age 18. Parental rates were generally consistent with WHO and CDC statistics in magnitude, with Minnesotan rates considerably higher than German rates, especially among mothers. In contrast to WHO (2017) data, however, overweight rates were lower in mothers than in fathers in both countries, and obesity rates in Germany were lower too, though not in Minnesota. This could reflect the slightly elevated overall levels of education in both samples, especially if associated with greater attention to maintenance of muscle mass (physical fitness) in adulthood. MTFS age-24 and TwinLife Cohort 4 (ages 23–24) rates ran well behind their parents’, but were considerably higher than those of the younger twins (including, in MTFS, themselves at younger ages). The rates in younger twins showed likely impact of greater weight of muscle than fat mass in young males, which distorts BMI as a measure of adiposity. Overall, however, the rates were very consistent with concerns about indications that childhood overweight and obesity are emerging at younger ages and to increasing degrees over time, more so in the United States than in Germany.
Tables 5 and 6 show the phenotypic correlations among study variables, Table 5 for TwinLife and Table 6 for MTFS. Reflecting accumulation of overweight with age, correlations between height and weight were generally larger in younger groups and small in adulthood. Parental assortment for height was generally moderate, but that for weight and BMI small. There was enough parental assortment on BMI, however, to indicate that, on average, DZ twins would share 55% of their segregating genes rather than 50% when parental assortment is random (parental correlations 0.23 in TwinLife, 0.22 in MTFS). We adapted our modelling assumptions to reflect this. SES was generally associated positively with height and negatively with weight and BMI to small degrees, in both parents and twins in all age groups. One exception to this was TwinLife Cohort 3 (age 17) boys, for whom the SES–BMI correlation was positive. Because the phenotypic SES–BMI correlations were fundamental to our SES-moderation hypotheses, we highlight them in Fig. 3. Parent–offspring correlations reflected both genetic and family lifestyle transmission. Height correlations were moderate, and tended to be somewhat higher at older than younger ages and perhaps higher with the parent of the same sex than the opposite sex. Parent–offspring weight and BMI correlations were generally small to moderate and showed the same age pattern, but no particular pattern regarding same- and opposite-sex. Mid-parent–offspring correlations (not shown) were higher, but still ran about 0.10 lower than DZ correlations. These patterns were rather consistent in the two samples. They likely reflect combinations of gene-environment correlation for tendency to weight gain and lifestyle, generation-specific lifestyle effects, and development processes.
In MTFS, cross–time correlations of height were 0.60–0.80 before age 17, and well above 0.90 after that. Weight and BMI cross–time correlations were similar at the younger ages, but not as high after age 17. Girls tended to be slightly more consistent across time than boys.
Univariate variance decompositions
The twin correlations are shown in Table 7. Presence of genetic influence on height, weight, and BMI was clear throughout. Except at age 11 in MTFS girls and in TwinLife Cohort 4 (ages 23–24) men, all indicated important shared environmental influences on height as well. At the younger ages, weight and BMI also showed important shared environmental influences, but these eroded in the older ages. We fit standard univariate twin models to these data to produce the estimates of proportions of variance in BMI attributable to genetic and shared and non-shared environmental influences shown in Fig. 4. At all ages in both sexes in both countries, the majority of the variance could be attributed to genetic influence.
We next fit the SES moderation models in each TwinLife cohort and at each assessment age in MTFS, separately in females and males. The second column of Table 8 indicates the best-fitting model in each (fit statistics leading to selection of best-fitting models and confidence intervals for their moderating parameters are provided in online supplementary information). With the exception of men in TwinLife Cohort 4 (ages 23–24), there were significant negative main effects of SES on BMI in all groups in both sexes and both countries, so that BMIs tended to be lower in youth of higher SES-of-origin. The effect was significantly positive in this one older male German group, however. [But it was TwinLife male Cohort 3 (age 17) that showed the lone positive phenotypic correlation.] In most groups, SES moderated the genetic influences on BMI, so that there was less genetic variance and heritability in BMI at higher levels of SES, consistent with the adult studies based on attained SES reviewed above. TwinLife Cohort 1 (age 5) girls and boys and TwinLife Cohort 3 boys (age 17) were exceptions to this. In these groups, there was less variance in BMI at higher levels of SES too, but instead of moderating genetic influences, SES moderated shared environmental influences. SES also moderated non-shared environmental influences in many groups. Usually this also meant less non-shared environmental variance at higher levels of SES, but it meant more non-shared environmental variance in MTFS males at ages 17 and 23, and TwinLife males in Cohorts 2 (age 11) and 4 (ages 23–24).
We compared results in analogous age groups across countries in two ways. Our strictest test of differences was to fit each age–sex group’s data to the parameters generated in the analogous model of the other country’s data [e.g., by fixing the moderating and main-effect parameters to those generated by the girls’ age-11 MTFS data to TwinLife Cohort 2 (age 11) girls’ data]. In doing this, we allowed the A, C, and E parameters that reflect the magnitudes of the total variances to remain free, as it was plain that total variance in BMI was greater in MTFS than in TwinLife at all ages and in their parents as well. The extent to which we were able to do this depended not just on how different the parameters were, but on the relative sizes of the samples. In most groups, the power advantage went to MTFS, so there were several German groups in which we were able to do this without loss of model fit, but could not do it in the analogous Minnesotan groups.
To address which specific differences were most important, we also tested whether each significant moderating parameter could be constrained equal across countries in each group. This was possible without loss of model fit in girls at age 17 and in males at ages 11 and 24. In the other groups, some moderating parameters could be constrained equal, but not all. Maintaining equality constraints on the moderating parameters meriting it, we added equality constraints on the main effect parameters to each age–sex group’s two-country model. It was possible to do this in the three groups noted above without loss of fit, but not in the other three. Columns 3–5 in Table 8 summarize these results.
There was thus considerable consistency in results across age groups, countries, and sexes, as well as with observations in adults based on attained SES. Figure 5 shows the most typical patterns in the best-fitting models in each sex. At the same time, there was evidence of specific differences that suggested developmental patterns, sex differences, and country-specific contexts. Patterns in the female data were more consistent across both ages and countries. There was evidence of C moderation in TwinLife but not in MTFS, and it was consistent in direction with the observed A moderation wherever it appeared. Power to distinguish A variance from C variance is always relatively limited, and this is especially the case when the assumption that they are independent is violated. In German girls, evidence of C moderation was confined to the two younger cohorts, suggesting gene-environment correlation involving SES, genetic tendency to accumulate weight, and family lifestyles involving higher caloric intake, less exercise, and/or metabolic responses to stress in lower-SES groups. The parental lifestyle influence appeared to dissipate with age in girls, likely as they started to make more of their own diet and exercise choices, as well as choices involving experienced stress. In German boys, the C moderation appeared at ages 5 and 17. This suggested that parental influence might dissipate more rapidly in boys, but be replaced by within-pair peer- and school-related lifestyle choices such as similar levels of sports participation that affect muscle and thus weight development (without indicating obesity) to greater degrees in young males than in young females. The trends suggesting dissipation of parental influence were stronger at higher levels of SES.
Different emphasis on and opportunities for youth sports participation in Germany and the US, and their relations with SES may well have contributed to the TwinLife/MTFS differences observed. In addition to community-based and commercial sports programs, public schools often feature sports participation prominently in secondary school in the US. Participation is generally nominal in cost, though parents often must supply necessary equipment and after-school practices may involve transportation costs. Though these programs are especially common and tend to be of higher quality in better-funded school districts that tend to have students from higher-SES families, they often offer opportunities to students from lower-SES families within those districts to obtain access to university educations through sports scholarships that would otherwise be beyond their families’ means. This is very different from the situation in Germany, where community-based and commercial sports clubs for young people are completely separate from the school system. They are often inexpensive, but inevitably involve equipment purchase and transport to some location different from the participant’s school. Supporting the idea that access to, and thus participation in, sports might be greater in the United States, Physical Activity Council data (2017) indicated that healthy physical activity levels were maintained in relevant age groups by 46–49% of the population, while WHO (2012) data indicated that 27% of German children maintained healthy levels of physical activity. Other sources indicated different levels and trends, however, and definitions of ‘healthy’ activity levels and sampling may vary considerably among organizations collecting data.
Muscle development in males as they reach adulthood also seems likely involved in the positive main effects of SES on BMI and greater E variance at higher levels of SES at the older ages in both countries. The positive main effects suggest that young males of higher SES-of-origin tended to be more likely to invest actively in athletic development. This would be consistent with greater access to opportunities to do so and greater awareness of the health and status benefits of doing so. At the same time, the greater E variance at higher levels of SES suggest that seizing these opportunities is far from uniform, and, though some co-twins may choose to participate in physical activities (either mutually supportively or competitively), with resulting gene-shared environment correlation and interaction impact on their phenotypes, other co-twins may very intentionally choose different levels of physical activity, with correspondingly different gene-non-shared environment correlation and interaction impact on their phenotypes. Either way, higher SES tends to offer more opportunities.
Recall, however, that main effects and covariance were inevitably confounded in these models because twins shared SES-of-origin. What was measured as a uniformly applicable main effect in each model was actually the coefficient of a regression through the origin (including no intercept term) of the relation between SES and BMI within their shared variance, and this shared variance was omitted from the estimated variance components (Purcell 2002). The twins’ SES-of-origin was the parents’ attained SES and its covariance with their own BMIs was passed both genetically and environmentally (through gene-environment interaction, correlation, and direct effects) to the twins. This means that intergenerationally transmitted covariance was almost entirely the covariance inevitably excluded from our models. This missing covariance likely offered the most important hints about intergenerational transmission of tendencies toward obesity so we believed it important to try to characterize it.
Follow-up interpretative analyses
We suspect that this covariance consisted of personal characteristics that contribute to SES such as cognitive abilities, self-discipline, and working and playing well with others, which parents transmit both genetically and environmentally to their children. These also contribute to maintaining lifestyles involving diet, ample exercise, and minimal exposure to uncontrollable stress that facilitate maintaining healthy weight. This in turn can minimize expression of genetic vulnerabilities to accumulating excess weight. Such a process would at least be consistent with the observations we report here, as well as with those from the reviewed adult studies based on attained SES. Varying extents to which it is present could also contribute to differences in patterns among time and place cohorts. It is also exactly the kind of process that creates gene-environment Simpson’s paradoxes such as the contrast between Armour and Haynie’s (2007) phenotypic association between earlier age of first sexual intercourse and higher level of delinquent behavior and Harden et al.’s (2008) opposite association after controlling genetic and family influences. If within most countries the association between SES and overweight and obesity rates substantively involves, for example, relative economic and social standing as well as ability to maintain healthy lifestyle, this kind of process could also explain the Simpson’s Paradox observation that less ‘economically developed’ countries tend to have lower rates of overweight and obesity than do those more ‘economically developed’, but the reverse tends to be the case for individuals within countries.
But it does not explain the increases in obesity rates over time termed ‘epidemic’. Emerging environmental conditions would have to accentuate the general patterns of gene–environment correlation apparently affecting genetic expression observed here and in adult studies in ways that contribute to the population-level increases in overweight and obesity that have been observed over the past 40 years or more.
If so, we reasoned that we ought to be able to see analogous patterns of SES-of-origin moderating twin BMI if we controlled both twin BMI and SES for age-adjusted mid-parent BMI. This effectively removed the inter-generationally transmitted covariance that the moderation models confound with main effects. That is, any moderating effects of the resulting SES residual on the twin BMI residual should reflect whatever it is about parental SES that is independent of parental BMI yet acts on offspring BMI to accelerate it from one generation to the next. Because twins share this residual SES too, this moderating model still omits any covariance between the twins’ levels of whatever parental characteristics contributed to their own attained SES and the twins’ BMIs. Most of the relevant SES residual–twin BMI residual correlations were very small and not significant, however, so there was likely little such covariance. The model-indicated main effects should thus reflect mostly whatever direct residual-SES-of-origin effects applied uniformly to everyone. These would directly drive the increases over time in obesity levels that have been observed.
Results of fitting such models were consistent with this. The variance moderation patterns were the same as those using full SES and full BMI, but weaker, with not all reaching significance. This was consistent with presence of environmental changes that often accelerate the kind of hypothesized gene-environment interplay process described above. Similarly, modeled main effects were almost completely consistent, though weaker. All were still significant. This was consistent with presence of uniform main effects of SES on obesity levels. These would likely be increases over time in the kinds of SES-related environmental disparities involved in that hypothesized process.
The remaining moderating effects may have reflected stronger links between SES-of-origin and especially genetic vulnerabilities to weight gain in the offspring than the parental generation. If so, they would have contributed to the increases over time in obesity rates that have been observed. Such links would likely include decreasing job opportunities for people of limited educational attainment, increasingly stressful technological and economic environments that make healthy lifestyles increasingly difficult for those with lower levels of education and greater financial pressures, and increasing social and political challenges involving massive population migrations that strain public resources in ways that tend to impact those of lower SES more.
We suggest that the extent to which the main effects were weaker may have reflected environmental circumstances that foster development of overweight and obesity that affect everyone to much the same degree that have remained rather constant over the last 40 years or so. High social emphasis on educational attainment to access good job opportunities, high social and financial rewards to those who do relative to those who do not, strong correlation between house prices and school quality, ready availability and high advertising of junk foods and their relatively low cost, lower access to fresh fruits and vegetables in lower-SES neighborhoods, high proportions of sedentary jobs, and accessibility of socially and financially viable lifestyles involving little or no physical exertion, etc., are plausible examples. We also suggest that the extents to which there remained apparent main effects may have reflected increases in such circumstances over that period.
There were three exceptions to the consistency in main-effect patterns with removal of parental transmission. They were in the older male age groups in which the full-model main effects in both countries were particularly weak. In MTFS males at 17 and 24 and TwinLife Cohort 4 (ages 23–24) the directions of main effects in these groups were reversed. We suggest that this may have reflected environmental circumstances that have emerged between the parent and offspring generations that affect young men’s greater tendencies to gain muscle mass rather than adiposity in higher SES environments than in lower. Increasing popularity of endurance and extreme sports and body-building seems a likely possible contributor.
Figure 6 compares the phenotypic correlations between BMI and full SES-of-origin, the model-indicated main effects in their full covariances, and the model-indicated main effects in the residual SES-residual BMI covariances in each sex, age group, and country. We highlight this comparison rather than the moderating effects because we suspect that these main-effect indications are the more powerful drivers of the increasing rates of overweight and obesity and their associations with SES. The phenotypic correlations summarize the associations between the variables across the full range of SES as if they were constant in magnitude. This confounds direct effects of SES on BMI that apply uniformly to everyone with sometimes offsetting, sometimes enhancing moderating effects on genetic and environmental components as well as gene-environment covariance. This likely understates direct uniform main effects. The model-indicated main effects in the full models treat all the covariances as if they were only direct main effects, likely overstating direct uniform main effects. The model-indicated main effects in the parental BMI–SES-residualized covariances do the same, but the extent of covariance there was considerably reduced.
Again with the exception of the older male groups, the model-indicated main effects from the full models were more strongly negative than the phenotypic correlations, consistent with the former overstating actual direct uniform main effects and the latter understanding them. In girls, differences were especially pronounced at the younger ages, suggesting greater underlying gene-environment correlation. This would be consistent with more home-based lives in ways that matter for BMI at younger ages. The model-indicated main effects from the full models were stronger in the younger female TwinLife cohorts than the older, but stronger at age 24 in MTFS, suggesting some country-level differences. The younger boys from both countries showed similar patterns to the females, but to less striking degrees. As discussed above, the model-indicated main effects from the residualized models were the weakest of the three in all these groups. And we have also addressed the rather disparate patterns among the older male groups Summary and conclusions.