Figure 1 summarizes all stages of the study design. We used the Avon Longitudinal Study of Parents and Children (ALSPAC)^17,18 as the discovery cohort, and explored potential effects of a wide range of maternal, family, environment, and child exposures, including air pollution, area-level deprivation index, individual level socioeconomic position, health, and behaviors such as smoking, detailed dietary factors and physical activity. Only prospective associations were explored (ie, with exposure assessed before the timing of the BMI measure). In total 2523 exposure outcome associations were included in the study with numbers being smallest (427) in relation to BMI in the first 12 months of life and highest in the 13-18 years old analyses (842). We conducted unadjusted analysis and selected exposures passing Bonferroni-corrected statistical significance to be taken forward for confounder adjusted multivariable regression in ALSPAC. For the exposures that remained after confounder adjustment, we sought replication in pooled analyses from eight independent European cohorts and/or MR evidence. Given the hypothesis free approach of this study and the likelihood that there would be few or no genome-wide analyses of many of the exposures assessed in childhood or of BMI in the age groups used here, we a priori decided to review the literature for any relevant published MR studies and to undertake one-sample MR where possible in ALSPAC.

Details of exposure and confounder selection and of how these were assessed in the cohorts are described in relevant sections below. Here we described the place, timing and recruitment of cohort participants.

We used linear regression to estimate the unadjusted and confounder-adjusted associations of each exposure with BMI. For unadjusted analyses, we used a Bonferroni corrected P-value threshold, which took into account all association tests across all age groups, with the threshold being 1.98*10⁻⁵ (ie, 0.05 divided by the number of tests 2523), to select which of the unadjusted associations would be taken forward for confounder adjusted analyses. We compared the adjusted with unadjusted associations in ALSPAC and defined those that attenuated by more than 50% as possibly being explained by confounding. This approach did not rely on the conventional P-value of 0.05 and took account of differences in statistical power between models of different sample sizes and covariables.

For every cohort that was able to contribute to a specific exposure age period analysis (ie, had complete data on the specific exposure, BMI in the age category and all confounders) (data-catalogue.molgeniscloud.org/catalogue), multivariable linear regression was undertaken in each cohort using the same adjustments as in the discovery ExWAS. The measures for exposure, outcome and confounders have been harmonized in different cohorts. Results were then pooled using random-effects meta-analysis. The analysis in each cohort was undertaken using the DataSHIELD platform (https://www.datashield.org/),²⁶ which supports federated analyses (ie, individual participant data remained on cohort-specific servers²⁷). We took into account of the sample size differences between the adjusted results in ALSPAC and the replication cohorts (Supplemental Table 4), and defined results as replicated with the following criteria:

For all the confounder-adjusted exposure-BMI associations in any age category that were taken forward into follow-up (Figure 1), we explored the possibility of obtaining relevant MR results. As most MR studies to date have been based on genome-wide association studies (GWAS) in adults, we did not anticipate many published GWAS of the exposures in earlier ages.^28,29 Therefore, we decided to take two approaches to explore MR evidence for follow-up associations.

First, we searched the literature for any published (one- or two-sample) MR studies on the potential effects of follow-up exposures on childhood BMI (See Supplemental Table 5 for search terms used for each follow-up exposure). In this search we included studies of childhood BMI of any age up to 18 years, with the idea that we would then consider their relevance to the specific age group. Second, we explored the potential of undertaking one-sample MR in ALSPAC, where we could identify GWAS of any of the relevant exposures. Where such a GWAS was identified we generated a weighted genetic risk score (GRS) for the relevant exposures using single nucleotide polymorphisms (SNPs) that were independent of each other (r² <0.001) and either associated with the exposure at genome-wide significance (P < 5 × 10⁻⁸) or a less stringent cut-off (P < 5 × 10⁻⁶) if there were fewer than 3 SNPs available at P < 5 × 10⁻⁸, in order to mitigate against weak instrument bias. The GRS was calculated by weighting each SNP by the per allele effect on the exposure and then summing the alleles across SNPs. The details are shown in Supplemental Methods. We compared MR results to the confounder-adjusted multivariable regression findings focusing on directional consistency.

Supplemental Table 1 describes the characteristics of the ALSPAC participants. The sample size for infancy, early-, middle-childhood and adolescence was 12 761, 12 138, 9105, and 7379, the median BMI was 17.2, 16.3, 17.9, and 21.4 kg/m², and the number of exposures was 427, 594, 660 and 842, respectively.

Of the total 2523 exposure-BMI associations explored across all age groups, 172 reached the Bonferroni corrected P-value threshold in unadjusted discovery ExWAS analyses, 8 of these were in infancy, 16 in early-childhood, 66 in middle-childhood and 82 in adolescence (Figure 1). The unadjusted associations for all exposures are shown in Figure 2 and Supplemental Table 3. These included previously reported exposure associations, such as maternal BMI during pregnancy, which associated with offspring BMI in all 4 age groups, and offspring prior BMI, which associated with higher BMI at early-, middle-childhood and adolescence. We also identified several potentially novel risk factors, such as children’s temperament associated with BMI in middle-childhood. For example, child’s shyness and child crying easily at age three were associated with lower BMI in middle-childhood, child being easy-going, higher sociability score, and child preferring to play with others than alone at age three were associated with higher BMI in middle-childhood. Of the 172 associations passing the Bonferroni correction which included 115 exposures (ie, some exposures were associated with BMI at several ages), 3 exposures associated with BMI at all 4 ages, 3 with three, 42 with 2 and 67 with just one of the age groups.

The 172 associations included 8, 16, 66, 82 associations in infancy, early-childhood, middle-childhood and adolescence, respectively, of these, 8, 11, 42, and 26 remained after controlling for confounders (Figure 3). These included previously reported associations such as maternal BMI and offspring prior BMI (Figure 3A–D), and maternal smoking (including smoking pre-pregnancy, during the first three months of pregnancy and the last two weeks of pregnancy) associated with higher BMI at early-, middle-childhood and adolescence (Figure 3B–D). Some potentially novel factors identified in the unadjusted analysis, such as children’s temperament, also remained in the confounder-adjusted analysis. Of the 20 derived variables from the Emotionality Activity Sociability (EAS) scale measured when children were aged three, seven were associated with childhood BMI, including whether children cry easily, whether children are always on the go, shyness (complete cases and prorated), EAS sociability score (complete cases and prorated), and whether child prefer to play with others than alone. For example, children who were described as being always on the go had a 0.16 kg/m² (95% confidence interval (CI) 0.07-0.25) higher BMI compared with those who are not. Children with a mother-reported higher shyness score, those reported to cry easily, and those who scored lower in the sociability score had lower BMI (Figure 3C). Additional adjustment for children’s age and sex were consistent with the confounder adjusted results without children’s age and sex, this additional adjustment had no effect on the selected potentially causal exposures (Supplemental Table 6).

Supplemental Table 2 describes the characteristics of each of the eight cohorts. The sample size of these cohorts ranges from 1344 (EDEN) to 39 835 (DNBC). The pooled sample size for replication ranged from 2631 to 71 056 depending on the specific exposure (Supplemental Table 4). Of the associations that remained after confounder adjustment, it was possible to attempt replication in independent cohort(s) for 5, 6, 11 and 12 exposures in infancy, early-childhood, mid-childhood and adolescence, respectively (Figure 4). We were not able to attempt replication of the other exposures retained in the confounder adjusted analyses. In particular, none of the replication cohorts had measures of the EAS scale meaning that we were unable to explore replication of the temperament and sociability associations. Of these exposures able to attempt replication, 4 (80%), 5 (83%), 7 (63%) and 9 (75%) replicated in infancy, early-childhood, mid-childhood and adolescence, respectively. These included maternal exposures, such as maternal BMI, for which we observed replication at all four ages, maternal smoking in pregnancy with BMI in early-, middle-childhood and adolescence, maternal saturated fat intake during pregnancy with BMI in middle-childhood, and children’s exposures, such as birth weight with BMI at all ages and pre-school TV viewing time with BMI in middle-childhood (Figure 4). For example, BMI in middle-childhood was on average 0.17 kg/m² (95% CI 0.10–0.24) higher per 1 hour/day higher pre-school TV viewing time in the replication pooled samples, compared with 0.21 (0.14, 0.28) in confounder adjusted analyses in ALSPAC. Of the 25 replicating exposure-BMI associations across all age groups, 12 showed low between-cohort heterogeneity (I² ≤50%) (Supplemental Figures 1–4), such as maternal BMI, maternal smoking pre-pregnancy, maternal total fat intake and saturated fat intake during pregnancy, birth weight and children’s TV viewing time. For those with large between-cohort heterogeneity, it largely reflected differences in magnitude, with most study results being directionally consistent.

Our search identified 2 relevant publications^28,30 (see Supplemental Table 7 for details). Both explored the effect of maternal pregnancy BMI on offspring birth weight, BMI and fat mass and both found causal evidence for an effect of higher maternal pregnancy BMI on higher offspring birth weight, but not subsequent BMI. The largest and most recently published included a total of 9339 infants with numbers varying from 9339 at birth, 8659 at age 1 to 4112 at age 15.

Excluding maternal pregnancy BMI, of the 63 exposures that were confounder-adjusted associated with childhood BMI at any age we identified GWAS that could provide genetic instruments for 16 (Supplemental Table 8) and conducted MR in ALSPAC. The strength of the GRS instruments varied, with F-statistics ranging from 0.01, for maternal pregnancy reported vitamin B6 intake, to 215.4 for birth weight (Figure 5). Sample size ranged from 4023, for the association between physical activity and BMI in adolescence to 6781 for birthweight and BMI in infancy. We found directional consistency in MR results between maternal pregnancy smoking and BMI in adolescence compared with confounder-adjusted results, though with wider confidence intervals in the MR (Figure 5): difference in mean BMI in adolescence comparing any vs no pregnancy smoking: 4.48 (95% CI 0.40-8.56) in MR vs 0.65 (0.34-0.95) in confounder adjusted regression. Similarly, we found directional consistency in MR results between child passive smoking and BMI in middle-childhood and adolescence, as well as TV viewing time and BMI in middle-childhood: 0.17 (95% CI 0.01-0.33) in MR vs 0.03 (95% CI 0.02 to 0.04) in confounder adjusted regression. Sensitivity analyses adjusting for fetal genotype for maternal smoking and TV viewing time were broadly consistent, but gave wider confidence intervals, eg, maternal smoking and BMI in adolescence: 2.33 (95% CI −2.62 to 7.28) in sensitivity analysis vs 4.48 (95% CI 0.40 to 8.56) in main analysis, TV viewing time and BMI in middle-childhood: 0.37 (95% CI −0.10 to 0.84) in sensitivity analysis vs 0.17 (95% CI 0.01-0.33) in main analysis (Supplemental Table 9).

Our study has several strengths. First, it included a relatively large sample size, with a wide range of exposures across childhood. As discussed in a previous ExWAS,¹³ “different exposures may act during different developmental periods, making it important to cover multiple periods”, our study covering different ages filled this research gap and provided valuable information for future interventional studies. Second, we conducted replication analyses in the EU Child Cohort Network, with harmonized methods for data management and analyses.²¹ We performed the replication analyses using DataSHIELD, which is a safe and robust data analysis platform to perform individual-level participant data meta-analyses, without physically transferring data.²¹ Third, we used MR to minimize confounding, which was promoted in a recent commentary on the ExWAS to strengthen causal inference.¹⁶

Our study also reflected several challenges in using ExWAS to identify modifiable targets. Of the 2523 exposure-BMI associations, only ∼7% associations passed Bonferroni-correction, and ∼3.4% of the total remained associated after controlling for confounders. Given the hypothesis free nature of ExWAS, we felt it was appropriate to use the conservative Bonferroni-corrected P-values to decide which exposure associations to take forward for confounder adjustment. However, this means we could have missed some potentially important effects due to a type 1 error. This is the nature of not attempting to take forward all associations explored in any hypothesis-free study. However, by providing all of these results, Supplemental Table 3 is a useful resource for future research, including by other researchers, that might explore the potential effects of some of the exposures that are close to the Bonferroni-corrected threshold in larger samples. Of those associations that passed the Bonferroni-corrected P-value threshold and remained after adjustment for confounders, for more than half (58%) were not able to seek replication in other cohorts, or to assess using MR studies. Again, this is a common challenge of any hypothesis-free study. As noted in the introduction, the tendency for large numbers of studies to repeatedly assess the established exposures, such as physical activity, reflects factors that are commonly measured because they are hypothesized. On the other hand, our hope is that this study will stimulate more studies to assess the potential novel risk factors that we could not explore beyond confounder adjustment.

Beyond these challenges there are some specific limitations. Although we used multivariable regression to control for pre-specified confounders, residual confounding cannot be ruled out as an explanation for some of our potential risk factors, particularly those for which replication and MR could not be undertaken. We used complete case analysis, which is generally unbiased when the chance of having complete data does not depend on the outcome conditional on all covariables in the main analyses.⁴⁰ It is not possible to test this assumption but bias from selection bias due to missing data is less likely where results were replicated as there were differences between the discovery and replication cohorts in amounts of missing data. Most exposures were parental- or child-reported, and these might be misreported. We would expect this to be random and not related to variation in BMI (outcome) which was always assessed after the exposure. Such measurement error would be expected to bias results towards the null in the unadjusted analyses and could bias in either direction in confounder adjusted analyses, depending on the direction and size of association of the error with confounders as well as the outcome.⁴¹ Due to the limited sample size in ALSPAC, and small number of genetic instruments for some exposures, MR results that we undertook in ALSPAC might be biased towards the confounder adjusted associations due to weak instrument bias for some exposures, such as maternal vitamin B6 intake. It is acknowledged that methods for exposome-wide associations are very varied, and that it covers a variety of very different aims, types of exposures and statistical methods.^42,43 Some exposure wide-association studies mutually adjust all of the exposures for each other.⁴² That can be useful if the aim is to identify a group of exposures that might accurately predict an outcome. The aim of this study is to identify causal risk factors for variation in childhood BMI. For causal analyses it is important to adjust for prespecified confounders (ie, factors that are known or plausible causes of the exposure and outcome,⁴⁴ as we did, as well as exploring replication and/or MR. When multiple variables are mutually adjusted in a model, the results do not provide unbiased causal estimates as many of the exposures will be adjusted for variables that are not confounders, and some will potentially be on the causal path and we would not want to control for them.^44,45 It is also likely there will be collider bias and the results from such a model are often not reproducible and difficult to interpret.⁴⁵ Similarly presenting the percent variation of a set of variables does not provide any evidence of causality, is not reproducible and is difficult to interpret.⁴⁶ Thus, for this study we have not mutually adjusted exposures and nor do we present percent variation explained by the exposures.

In conclusion, we found support for the effects of several hypothesized risk factors on BMI, such as maternal higher pregnancy BMI on higher infant BMI, and more offspring TV viewing time in early-childhood on higher subsequent (mid-childhood) BMI, and identified several potential novel risk factors related to childhood BMI, such as child emotion, sociability, and peer problems, which require further exploration. This study highlights the importance of replication and exploring causality in ExWAS.