Abstract Higher childhood adiposity has been reported to lower breast cancer risk across multiple lines of evidence, including lifecourse Mendelian randomization (MR). These MR results show that the apparent univariable effect of higher adulthood adiposity on lower risk of breast cancer is almost entirely accounted for by the childhood adiposity effect when using multivariable analysis. It has been suggested, however, that this protective effect may arise from collider stratification bias, particularly in selective cohorts such as the UK Biobank. To examine the plausibility of this claim we first simulated data assuming no true causal effect of adiposity on breast cancer risk to test whether selection bias alone could generate multivariable MR results of the same magnitude as those obtained empirically. Selection into the analytic sample was modelled with log-additive effects of childhood body size and breast cancer status, together with their interaction. Log-additive selection on body size or cancer status alone produced minimal bias. In contrast, interaction-dependent selection generated appreciable downward bias, with the strongest three-way scenario producing the largest spurious protective effect. However, even under these severe scenarios, the bias fell short of the effect observed in earlier MR analyses. Second, we examined if selection bias could induce the joint patterns of univariable and multivariable results observed. We found that across 135k possible models, none strongly overlapped with the observed results, and only 0.7% overlapped with an agreement probability >10%, which mostly relied on a large three-way interaction term between cancer, adulthood body size and childhood body size influencing selection. To conclude, while selection bias, particularly that which is interaction-driven, can generate spurious protective effects, it is unlikely to fully explain the protective effect of childhood adiposity on breast cancer identified in MR studies.
Power et al. (Fri,) studied this question.