When you cannot assign,
what may you still say?
/reg said randomisation balances even what you did not measure; /stat said adjustment removes only what you did. These twenty lessons are the question that follows.
You cannot assign someone to smoke. You cannot toss a coin for where they live. And yet Snow stopped a cholera outbreak without knowing what caused it, and smoking and lung cancer were settled without a trial. How that was possible — and exactly where observation misleads. Python runs right inside this page.
How you'll learn
Not invented examples. Snow's 315 against 37, fifty years of 34,439 British doctors, the 16,608 women of the WHI — only verified figures.
Confounding, selection, information, immortal time, collider. Five whole lessons — and two of them you avoid simply by knowing the name.
Start from data where the effect is exactly zero, then switch each bias on and watch a relationship appear out of nothing.
Unmeasured confounding cannot be removed. So you measure how strong it would have to be to overturn the result — the E-value.