Risk differences are clinically relevant measures of treatment effect. But because of extreme baseline risk-dependent heterogeneity of risk differences, risk differences should be covariate-specific and not averaged. — Incorrect Covariate Adjustment May Be More Correct than Adjusted Marginal Estimates