Experimentation

Peeking, multiple testing, sample ratio mismatch, and novelty effects — the four standard ways a correctly-computed p-value still lies to you, and the checks that catch each one

Why observed associations routinely reverse under intervention — confounders, Simpson's paradox with real numbers, and the ladder of methods (randomization first) for earning a causal claim

Power, MDE, and the sample-size calculation: $n \approx 16\sigma^2/\delta^2$ per arm — how to size an A/B test before launch instead of explaining an inconclusive one after