Sensitivity to unmeasured confounding
Using published studies (see ESM –Supplemental results: sensitivity analysis, Figure S2, S3, S4), we estimated association between theSLCO1B3 c.334T>G TT/TG genotype (vs. GG genotype, total N=241 vs. 400) and steady-state MPA AUC0-12 as ROM=1.136 (95%CI 0.949-1.361), and association between the UGT1A9 c.98T>C genotype TC (vs. TT, total N=25 vs. 593) and the outcome as ROM=1.098 (95%CI 0.548-2.198), with similar estimates in patients co-treated with CsA or tacrolimus/sirolimus in both cases; estimated prevalence of TT/TG (4 cohorts, total N=1192) and TC subjects (9 cohorts, total N=1827) was 31.1% and 3.9% respectively. Although there is no reason to expect higher prevalence of TT/TG (SLCO1B3 ) or TC (UGT1A9 ) patients among ABCG2 c.421C>Avariant carriers than among wt controls, we assumed scenarios with high chance imbalances (see ESM – Supplemental results: sensitivity analysis, Table S1, Table S2) and TT/TG and TC effects much higher than estimated. Even under such conditions, bias-corrected estimate of theABCG2 c.421C>A variant effect on exposure to MPA is still higher than the conventional limit of equivalent exposure (Figure 3). E-values suggested that the cumulative effect of unmeasured confounders would have to be strong, i.e., GMR=1.63 and GMR=1.59 (frequentist and Bayesian, respectively) in order to at least partly explain-away the observed, i.e., to “push” the observed GMR point-estimates (1.40) to 1.20.