Genetic instruments for childhood and adult body size were derived from a large-scale GWAS in the UKB conducted previously [7]. Full details of the GWAS protocol can be found in Supplementary Note. Linkage disequilibrium (LD) clumping was applied to identify our instruments using parameters of P < 5 × 10− 08 and r2 < 0.001 based on a reference panel based on 10,000 unrelated participants of European descent from UKB [8]. The final sets of genetic instruments can be found in Supplementary Table 2. These instruments have been validated in three independent populations which demonstrate their capability to reliably separate measured body mass index from childhood and adult timepoints as discussed in Supplementary Note. Furthermore, a recent study has found that the childhood genetic instruments have a much stronger effect on DXA-derived fat mass in early life compared to DXA-derived lean mass [9].

Reported illnesses of mothers (field 20110) and fathers (field 20107) were recorded in the UKB study by the majority of participants (n = 492,986 for maternal history and n = 488,077 for paternal history). Amongst these endpoints were; bowel cancer, breast cancer (mothers only), diabetes, heart disease, high blood pressure, lung cancer, prostate cancer (fathers only) and stroke. All outcomes were coded as 0 = neither parent with reported disease, 1 = one parent with disease and 2 = both parents with disease, with the exception of breast cancer and prostate cancer which was encoded as binary outcomes depending on whether mothers or fathers respectively had reportedly had these diseases. These fields in the UKB study were for blood relatives only as adopted mothers and fathers had separate fields for reported disease history (fields 20112 and 20113). If participants were unsure about any answers they were encouraged to respond with ‘do not know’. A summary of final sample sizes can be found in Supplementary Table 1. GWAS were applied to these outcome variables using the same protocol found in Supplementary Note to derive estimates for subsequent MR analyses.

Univariable MR analyses were initially undertaken to systematically estimate the total effect of genetically predicted exposures on each parentally proxied disease endpoint in turn. This was firstly conducted using the inverse variance weighted (IVW) method, which takes the SNP-outcome estimates and regresses them on those for the SNP-exposure associations. We subsequently applied the weighted median and MR-Egger methods which are more robust to horizontal pleiotropy than the IVW approach [2].

We next conducted multivariable MR to estimate the direct and indirect effects of exposures on disease endpoints which provided evidence of an effect based on FDR < 5% from IVW univariable analyses. Multivariable MR involves obtaining estimates for all instruments on each exposure being evaluated, thus allowing each estimated effect to take into account the effect of all other exposures in the model. Although this approach has been conventionally applied to analyse different risk factors as exposures (where estimates are typically interpreted as ‘lifelong effects’), the novelty of analysing the same exposure measured at different timepoints throughout the lifecourse (e.g. at age 10 and age 55 as conducted here) can facilitate inference in a lifecourse epidemiology setting. All analyses in this study were undertaken using R (version 3.5.1).