Summary
The episode works through four questions on design-specific and temporal biases. It explains why a properly randomized trial can still be biased after randomization, works a contamination example, and presents a disagreement between the hosts on intention-to-treat and per-protocol analysis that ends in a shared practical rule. It then works through a hypothetical cohort in which a drug with no effect appears protective because of immortal time, a screening example in which lead time and overdiagnosis raise five-year survival while deaths stay the same, and a case-control example in which unequal record lengths produce a false protective odds ratio. It closes with an illustration of why age, period and cohort effects cannot be separated without an assumption that must be stated.