Parental body mass index and offspring childhood body size and eating behaviour: causal inference via parental comparisons and extended children of twins structural equation modelling

Introduction

The positive observational association between parental body mass index (BMI) and offspring adiposity in childhood is well replicated (1), but the mechanisms driving this association remain unknown. If greater maternal or paternal BMI causes greater offspring BMI via prenatal or intrauterine developmental mechanisms, a vicious cycle could amplify BMI through successive generations and be a major driver of the obesity epidemic (2). It is therefore crucial to establish why parental BMI is associated with offspring childhood BMI.

Several mechanisms are plausible (Figure 1a). Higher parental BMI could cause higher offspring adiposity through pre-conception and/or intrauterine developmental mechanisms (the developmental overnutrition hypothesis) (3-6), with some authors advocating that interventions to maintain women ‘s preconception BMI at a healthy level be used as a means to reduce offspring adiposity (1, 3, 7). Because adiposity is highly heritable across the life course (8), genetic confounding (via the inheritance of parental genetic alleles by the offspring) could result in intergenerational BMI associations. Non-genetic (environmental) confounding, for example via shared familial socioeconomic position or parental influences on offspring postnatal food intake and physical activity behaviours, could also contribute to these associations.

• Download figure
• Open in new tab

Figure 1:

(a) Directed acyclic graph (DAG) showing three plausible mechanisms for associations between parental BMI and offspring weight, BMI and appetite traits. *We define non-genetic (environmental) confounding as confounding that does not involve the offspring ‘s own genotype (non-genetic confounding could therefore still involve parental genetic effects, i.e. effects of parental genotype on offspring outcomes independently of offspring genotype, via the offspring ‘s environment). Examples of non-genetic confounding include shared familial socioeconomic position or parental influences on offspring postnatal food intake and physical activity behaviours. To the extent that postnatal parental BMI per se causally influences offspring outcomes after birth (e.g. by influencing feeding- and exercise-related parenting practices), this would constitute a postnatal causal effect rather than non-genetic confounding. (b) Conceptual diagram showing putative biological mechanisms by which parental BMI could have intrauterine or periconceptional causal effects on offspring outcomes. Arrows denote potential causal effects, bold outlined boxes denote variables analysed in the present study. The intent is to non-exhaustively show some key variables and relationships that have been hypothesized in the literature

Numerous animal studies purport to provide evidence in favour of the developmental overnutrition hypothesis (9). Potential biological mechanisms have been elucidated (Figure 1b) (5, 9), including a putatively key role for the programming of offspring appetite via energy homeostasis brain networks (10). In humans, child appetite traits are associated with the child ‘s own BMI (11), and with maternal overweight/obesity (12). However, whether developmental programming of adiposity and appetite occurs in humans remains unclear. Mendelian randomization (13), sibling studies (14, 15) and paternal negative exposure control studies (16, 17) suggest that familial confounding (either genetic or non-genetic) may be an important cause of parent-child BMI associations. However, such associations are generally unchanged on adjustment for measured variables (18), leaving the specific confounders unidentified.

We aimed to establish whether associations between peri-pregnancy parental BMI and offspring birth weight, childhood BMI and appetite-related eating behaviours are due to genetic confounding. We first compared the strength of the associations of maternal versus paternal BMI with offspring outcomes, which are likely to be similarly strong if they are primarily due to genetic confounding. We then applied a genetically informed structural equation model (SEM) to a population-based sample of twins, siblings and half siblings, and their children, to quantify the relative importance of genetic confounding versus other mechanisms in underpinning intergenerational associations. Based on prior evidence (13-19) we hypothesized that genetic confounding would not explain the associations of parental BMI with offspring birth weight, but would be a major driver of associations with offspring childhood BMI.

Methods

Results

Statistics are for the sample used for linear regression analyses of birth weight (n = 85,866), aside from the other outcome variables, for which statistics are from the corresponding linear regression samples. Equivalent data for the 8-year BMI sample (n = 46,620) are presented in Supplementary information S12. SD: standard deviations, WHO: World Health Organization The number of offspring included in linear regression analyses varied by outcome, from 85,866 (74.9% of recruited sample) to 30,904 (27.0% of recruited sample) for analyses of birth weight and 2-year BMI respectively. Supplementary information S1 shows the proportion of the sample with non-missing data for each analysis, and Table 1 shows the characteristics of the study participants. There was statistical evidence for selective attrition, such that the sample used for analyses of 8-year BMI (n = 46,620) was more highly educated and had lower obesity prevalence and greater maternal age versus the baseline sample, but the magnitude of such differences was relatively small (Supplementary information S12).

Linear regression analyses provided strong statistical evidence that the association of maternal BMI with offspring birth weight is stronger than that of paternal BMI with offspring birth weight (Figure 2a). However, after birth the associations with offspring BMI converged, and the associations of maternal and paternal BMI with offspring 2–5- year BMI were similar. Although for 8-year BMI there was statistical evidence that the paternal association was slightly weaker, the difference was not large, and when we used unstandardized parental variables the paternal association was actually slightly stronger than the maternal association (Supplementary table S1). These results were not markedly different when using offspring BMI predicted from a modelled growth curve (Supplementary information S13), when substituting birth weight for ponderal index/BMI at birth and substituting child BMI for weight, when stratifying by offspring sex, or when additionally adjusting for gestational age (Supplementary table S1). With respect to eating behaviour outcomes, both maternal and paternal BMI were positively associated with offspring food responsiveness and emotional overeating, and negatively associated with emotional undereating. Only paternal BMI was associated (negatively) with offspring satiety responsiveness and slow eating. Offspring 8- year BMI was associated with all eating behaviour outcomes except for emotional undereating, in the directions that would be expected from the behavioral susceptibility theory of obesity (22) (Supplementary information S14). We did not observe large departures from log-linear relationships (Supplementary information S7), and statistical interaction between maternal and paternal BMI was at most minor (Supplementary information S15).

• Download figure
• Open in new tab

Figure 2:

Linear associations of maternal and paternal BMI with offspring birth weight, child BMI and 8-year eating behaviours Linear associations of maternal and paternal BMI with offspring outcomes. a: offspring birth weight and child BMI, b: offspring 8- year old CEBQ eating behaviour traits, P: P-value testing the null hypothesis that regression coefficient is zero, P_dif: P-value testing the null hypothesis that maternal and paternal regression coefficients are equal

Table 2 shows the sample size available for MCoTS analyses, stratified by maternal and offspring relationship. The MCoTS results indicated that the positive phenotypic covariance between maternal BMI and offspring birth weight was not explained by genetic confounding, with genetic covariance estimates that were statistically indistinguishable from zero (Figure 3). The weak positive phenotypic covariance between paternal BMI and offspring birth weight was also not explained by genetic confounding. Surprisingly, there was statistical evidence for a small negative genetic covariance between paternal BMI and offspring birth weight, but this attenuated and became statistically indistinguishable from zero when, in exploratory analyses, we adjusted exposures and outcomes for potential confounders (including maternal BMI, paternal age and paternal income), suggesting that bias due to uncontrolled confounding may be the explanation (Supplementary information S16).

• Download figure
• Open in new tab

Figure 3:

MCoTS SEM estimates of phenotypic and genetic covariance of parental BMI with offspring birth weight and 8yr BMI Phenotypic covariance denotes the overall covariance between the exposure and outcome, Genetic covariance denotes the part of the phenotypic covariance that is due to genetic confounding, P: P-value for phenotypic covariance calculated via a z-test using the standard error from bootstrapping the MCoTS model, P-value for genetic covariance calculated via a chi squared test for deterioration of model fit on deletion of the a1’ path coefficient, n: number of offspring with outcome data available

From age 6 months onwards, genetic covariance estimates became positive and increased in magnitude, such that for offspring 8-year BMI, genetic confounding explained 79% (0.19 / 0.24 * 100) (95% CI: 62%, 95%) of the covariance with maternal BMI and 94% (0.20 / 0.21 * 100) (95% CI: 72%, 113%) of the covariance with paternal BMI (Figure 3, Supplementary information S17). Genetic confounding explained a high and stable proportion of the phenotypic covariance with predicted BMI from age 1 to 8 years (Figure 4, Supplementary information S18). Results did not appreciably differ when we fitted ACE or AE models instead of the primary ADE model in the parent generation (Supplementary Table 2), when birth weight was substituted for ponderal index/BMI at birth and child BMI was substituted for weight (Supplementary Table 2, Supplementary information S19), or when BW models were refit without offspring-generation twins (results available from the authors on request). Although sex stratified models were underpowered, they provided no evidence for large sex differences in estimates (Supplementary Table 2). MCoTS models for eating behaviour outcomes were underpowered and uninformative (Supplementary Table 2, Supplementary information S20). Full MCoTS results including model fit statistics and estimated variance components for parental and offspring phenotypes are presented in Supplementary Table 2.

• Download figure
• Open in new tab

Figure 4:

MCoTS SEM estimates of phenotypic and genetic covariance of parental BMI with offspring predicted BMIPhenotypic covariance denotes the overall covariance between the exposure and outcome, Genetic covariance denotes the part of the phenotypic covariance that is due to genetic confounding, P: P-value for phenotypic covariance calculated via a z-test using the standard error from bootstrapping the MCoTS model, P-value for genetic covariance calculated via a chi squared test for deterioration of model fit on deletion of the a1’ path coefficient, n: number of offspring with outcome data available

Discussion

We triangulated evidence from two analytic approaches applied to a large European birth cohort, to infer the mechanisms underlying associations between parental BMI and offspring BW, BMI until 8 years and 8-year eating behaviour. There were not large differences in the magnitude of the associations of maternal and paternal BMI with offspring BMI beyond early childhood, suggesting confounding within families as the most parsimonious explanation for such associations. This was confirmed by our MCoTS analyses, which indicated that the covariance between parental BMI and offspring BMI from age 6 months to 8 years is primarily due to genetic confounding.

We have previously used genomic restricted maximum likelihood (GCTA-GREML) to explore whether intergenerational BMI associations could be due to genetic confounding (18). Those analyses investigated whether mother-offspring BMI covariance was explained by a set of ∼8 million imputed common genetic variants, in nominally unrelated individuals from 3 European birth cohorts. In contrast, the present MCoTS analyses investigated confounding involving all genetic variants, using expected relatedness inferred via quantitative genetic theory in a pedigree of close relatives. GCTA-GREML indicated that for maternal BMI, imputed variants explained 43% (95% CI: 15%, 72%) of the covariance with offspring 15-year BMI, with a similar estimate for 10-year BMI. This is highly consistent with the present MCoTS estimates: 79% (95% CI: 62%, 95%) of the covariance between maternal BMI and offspring 8-year BMI was explained by genetic confounding. MCoTS estimates were somewhat higher than those from GCTA-GREML, which is expected given that GCTA-GREML uses a set of measured genetic variants whereas MCoTS infers the effects of all variants. Furthermore, both GCTA-GREML and MCoTS indicated that the covariance between maternal BMI and offspring birth weight was not due to genetic confounding. The high concordance of results between the two methods, which make different assumptions and were applied to different cohorts, provides strong evidence that genetic confounding is a major driver of associations between parental BMI and offspring adiposity in late childhood.

Our results are consistent with previous maternal-paternal comparison, MR and sibling studies (several of which share authors with the present study), which have not supported a large causal effect of maternal BMI on offspring childhood adiposity (13, 14, 16, 17). However, sibling studies do provide a degree of support for potential causal effects of more extreme maternal metabolic dysregulation (for example maternal diabetes and severe obesity) on offspring adiposity beyond birth (2, 14). Taken together, this suggests that putative causal effects observed in animal models of developmental overnutrition (5, 9) should be interpreted cautiously and do not necessarily occur in humans. The association between maternal BMI and offspring size/adiposity at birth has been invoked in the literature to argue that maternal adiposity has a causal effect on offspring adiposity beyond birth (28). Indeed, MR and sibling studies (13, 19, 29, 30) support a causal effect of maternal BMI on offspring birth size. However, the present results, alongside previous studies (13, 17-19, 29), strongly suggest that the causes of weight at birth are somewhat different from the causes of BMI in later childhood.

We observed some associations between parental BMI and offspring obesity-related eating behaviours assessed via the CEBQ questionnaire. Previous studies in smaller samples have found similar associations, albeit somewhat inconsistently (12, 31). In particular, we found that greater maternal and paternal BMI were associated with increased scores on the CEBQ food responsiveness and emotional overeating scales, and a reduced score on the emotional undereating scale. Offspring satiety responsiveness and slow eating were only associated with paternal BMI; however, we cannot exclude measurement error as an explanation for the absent maternal associations because the CEBQ questionnaire was completed by mothers and is inherently subjective. Given our maternal- paternal comparison results it is plausible that child appetite-related eating behaviours mediate the genetically driven association between parent and child BMI. However, it was not possible to confirm this because our MCoTS analyses were underpowered for eating behaviours.

Our results have important public health implications, when considered alongside prior evidence. Maternal BMI is unlikely to have a large causal effect on child BMI beyond birth, although a small causal effect remains plausible, potentially mediated via maternal glycaemia during pregnancy. Any causal effect of paternal BMI on offspring childhood BMI is likely to be similar to or smaller than that of maternal BMI. Consequently, reductions in the BMI of either parent before pregnancy are unlikely to cause large reductions in childhood adiposity. It is possible that interventions targeting parental BMI reduction could influence childhood adiposity via parental lifestyle changes that persist after birth and affect the offspring ‘s environment. However, whether preconceptional interventions are the optimal approach for preventing childhood obesity requires further evaluation in light of evidence from the present study. Importantly, women considering pregnancy should still be advised and supported to maintain a healthy weight, because there is good evidence that maternal obesity in pregnancy causes adverse perinatal outcomes in the mother and offspring (32).

Our study has important strengths. We have analysed data from a large prospective birth cohort, enabling precise estimation of the associations between parental BMI assessed during pregnancy and offspring outcomes in mid childhood. We leveraged a pedigree involving twin and sibling relationships in both the parental and offspring generations, to partition parent-offspring phenotypic covariance via an MCoTS SEM that is rooted in quantitative genetics theory. We gained increased power over a conventional children of twins model by including non-twin siblings in the parent and offspring generations, and up to two children for each parent. We also acknowledge potential limitations of this study. First, because the MCoTS model cannot simultaneously estimate dominance genetic effects and common environmental effects, we fitted an ADE model in the parent generation, which assumes that common environmental effects are absent. This assumption is supported by classic and extended twin studies of adult BMI (8) and the similarity of our findings when we used ACE or AE models in the parent generation (Supplementary table 2). Second, our MCoTS analyses assume that phenotypic associations between log parent BMI and offspring outcomes are linear. In our data there were only mild deviations from log-linearity which would be unlikely to meaningfully alter our conclusions (Supplementary information S7). Third, our MCoTS analyses did not account for assortative mating. However, we do not expect this to have had a large impact on our results because spousal phenotypic correlations for BMI are relatively weak (13, 33), and spousal correlations at BMI associated loci are weaker still (13). Fourth, the MCoTS model does not account for gene by environment interaction, and if interactions exist between the additive genetic and common environmental variance components our results could overestimate genetic confounding. We believe any such bias is likely to be small though because extended twin family design data suggest common environmental effects are negligible for adult BMI (8). Fifth, the residual covariance estimated by the MCoTS model will not be indicative of the true causal effect of parental BMI, to the extent that residual confounding affects associations between parental BMI and offspring outcomes. It is likely that the negative residual covariance estimate for paternal BMI and offspring birth weight reflects residual confounding, particularly as this estimate attenuated on adjustment for potential confounders (Supplementary information S15). Sixth, our maternal-paternal comparisons did not account for non-paternity, which could weaken paternal associations. However, a previous simulation study showed that for a maternal-paternal comparison analysis using MoBa data with follow up to age three years, results would have changed little with non-paternity rates of up to 10% (17). Seventh, BMI is an imprecise proxy measure for adiposity. Despite this, BMI is highly correlated with more direct adiposity measures in childhood (34). Eighth, MoBa has a participation rate of 41% and there has been attrition over follow up (20). Although we cannot exclude an effect of selection bias on our results, we believe it unlikely that this would be of sufficient magnitude to alter our conclusions. Lastly, we have studied a Norwegian population which has relatively high obesity prevalence and income per capita in international terms and it would be beneficial to replicate our analyses in other settings.

In summary, we have shown that in a Norwegian population the linear association between parental BMI around the time of pregnancy and offspring BMI from age 6 months to 8 years is primarily due to genetic confounding. Our results suggest that neither the mothers ‘nor fathers’ pre-pregnancy BMI has a large causal effect on childhood BMI. This implies that any hypothetical intervention that successfully reduced parental BMI before pregnancy, without altering the offspring ‘s postnatal environment, would be insufficient to achieve large reductions in the offspring ‘s childhood obesity risk. Our results suggest that in the studied population, maintaining healthy parental BMI before conception is unlikely to be a promising target for childhood obesity prevention interventions.

Contributors

Tom A Bond: conceptualisation, data analysis, software, data interpretation, literature search, writing- original draft Tom A McAdams: funding acquisition, data curation, methodology, software, data interpretation, writing- review and editing

Nicole M Warrington: methodology, software, data interpretation, writing- review and editing

Laurie J Hannigan: methodology, data curation, software, data interpretation, writing- review and editing Espen Moen Eilertsen: methodology, data curation, software, data interpretation, writing- review and editing Fartein A Torvik: methodology, data curation, software, data interpretation, writing- review and editing

Ziada Ayorech: data curation, software, data interpretation, writing- review and editing George Davey Smith: data interpretation, writing- review and editing

Deborah A Lawlor: data interpretation, writing- review and editing

Eivind Ystrom: funding acquisition, methodology, software, data interpretation, writing- review and editing Alexandra Havdahl: funding acquisition, conceptualisation, supervision, data interpretation, writing- review and editing

David M Evans: funding acquisition, conceptualisation, supervision, data interpretation, writing- review and editing

Data sharing

Data from the Norwegian Mother, Father and Child Cohort Study and the Medical Birth Registry of Norway used in this study are managed by the national health register holders in Norway (Norwegian Institute of Public Health) and can be made available to researchers, subject to approval from the Regional Committees for Medical and Health Research Ethics (REC), compliance with the EU General Data Protection Regulation (GDPR) and approval from the data owners. The consent given by the participants does not cover storage of data on an individual level in repositories. Researchers who want access to data sets for replication should apply through helsedata.no. Access to data sets requires approval from The Regional Committee for Medical and Health Research Ethics in Norway and an agreement with MoBa.

Declaration of interests

DAL received support from Medtronic Ltd and Roche Diagnostics for research unrelated to that presented here. All other authors report no conflict of interest.