Selection Bias Beyond Confounding

The one idea: Confounding is about why someone got treated. Selection bias is broader and meaner — it's about who ends up in your dataset at all, or in the slice of it you're analyzing. You can close every backdoor path in your DAG and still get a wrong answer, because the sample itself was never a fair draw from the population you think you're describing.

⚠️ The trap: "I controlled for every confounder I could think of, so my comparison is clean." That feels right and it's wrong. A comparison can have zero confounding and still be worthless if the sample that reached your dataset — the survivors, the volunteers, the people who didn't drop out — was filtered by a process related to the outcome. No control variable fixes a sample that was never representative in the first place.

The last few lessons treated bias as a story about who got treated: a manager targets the shaky accounts, a hotel cuts prices in the off-season, a school buys tablets because it's rich. Confounding is one instance of a bigger problem. Selection bias is the broader failure: your data is systematically filtered — by survival, by self-selection into the sample, by who stuck around long enough to be measured — in a way correlated with the outcome you're studying. Confounding is selection into treatment. Selection bias also covers selection into the dataset, which can bite you even when nobody chose who got treated at all.

The distinguishing question is different from "who chose the groups?" (the confounding audit from an earlier lesson). Here it's: **"How did a unit end up in front of me to be measured,…