Summary
Distinguishes selection bias as a structural problem in data generation from confounding as a variable problem, using Berkson's 1946 demonstration that hospital-based studies make independent conditions look correlated and the healthy-worker effect's standardized mortality ratios below 1.0 in asbestos cohorts. Covers attrition in Framingham, nonresponse in NHANES, and the three Rubin missingness categories of MCAR, MAR, and MNAR, then introduces survival filters including prevalence-incidence bias. Closes by stepping out to external validity, asking whether internally unbiased estimates apply to anyone outside the sample.