Confounding & Causal Inference

Summary

Develops confounding through its three formal conditions and the critical distinction between confounders, which should be adjusted, and mediators, which should not when the total effect is the target. Contrasts the classical change-in-estimate detection rule with its limitations including non-collapsibility against the directed-acyclic-graph backdoor criterion from Pearl, and walks through five control strategies of restriction, matching with overmatching hazards, stratification through Mantel-Haenszel, multivariable regression, and propensity scores. Closes with marginal structural models for time-varying confounding in HIV care, sensitivity analyses with E-values for unmeasured confounders, and a theoretical critique of treating race, gender, and socioeconomic status as ordinary covariates.