Background Valproate inhibits clearance of lamotrigine and greatly increases its concentrations. We assessed whether this effect was moderated by a polymorphism (ABCG2 c.421C>A) of the breast cancer resistance protein. Methods In two consecutive independent studies in adults with epilepsy on lamotrigine monotherapy or cotreated with valproate: (i) Exposure to valproate was considered treatment, (ii) dose‐adjusted lamotrigine troughs at steady state were the outcome, and (iii) ABCG2 c.421C>A genotype (wild‐type wt homozygosity or variant carriage) was the tested moderator. We used entropy balancing (primary analysis) and exact/optimal full matching (secondary analysis) to control for confounding, including polymorphisms (and linked polymorphisms) suggested to affect exposure to lamotrigine (UGT1A4*3 c.142T>G, rs2011425; UGT2B7–161C>T, rs7668258; ABCB1 1236C>T, rs1128503) to generate frequentist and Bayesian estimates of valproate effects (geometric means ratios GMR). Results The two studies yielded consistent results (replicated); hence, we analyzed combined data (total N = 471, 140 treated, 331 controls, 378 ABCG2 c.421C>A wt subjects, 93 variant carriers). Primary analysis: in variant carriers, valproate effect (GMR) on lamotrigine (treated, n = 21 vs. controls, n = 72) was around 60% higher than in wt subjects (treated, n = 119 vs. controls, n = 259)—ratio of GMRs 1.61 (95%CI 1.23–2.11) (frequentist) and 1.63 (95%CrI 1.26–2.10) (Bayes). Similar differences in valproate effects between variant carriers and wt subjects were found in the secondary analysis (valproate troughs up to 364 μmol/L vs. no valproate; or valproate ≥364 μmol/L vs. no valproate). Susceptibility of the estimates to unmeasured confounding was low. Conclusion Data suggest that polymorphism rs2231142 moderates the effect of valproate on exposure to lamotrigine.