Summary
The episode works through five questions that students commonly find difficult in the study of confounding and statistical inference. It shows with a hypothetical coffee and lung cancer example why stratified and crude estimates diverge under confounding, and why the classical criteria cannot identify a confounder without causal knowledge. Sarah works a hypothetical Simpson's paradox aloud, averages the stratum percentages without weighting them, and corrects the error with the counts, after which the hosts discuss the 1986 kidney stone comparison. A harder example combines power, alpha and the share of true hypotheses to show how many significant findings in a low-powered field may be false positives, and a missing data example shows how the same nonresponse can be missing not at random or missing at random depending on what was measured. The hosts close with a debate on adjusting for income and education in studies of racial disparities in maternal mortality.