Lesson 2 · Causal Inference · Step 5 of 6
Closing backdoor paths
A simulated dataset makes the arithmetic visible, because the true effect is 1.5 by construction. The naive comparison returns 2.07: too high, because a confounder is doing part of the work. Holding the confounder constant returns 1.47, close to the truth. Adding a collider to the same model moves the estimate away again. That is the lesson worth carrying: more control is not better control. A variable included for the sake of thoroughness can introduce exactly the bias you were trying to remove.
Also covered: regression adjustment · confounding · collider bias · overadjustment
Alongside this step
- Slides and transcript — the whole video
- Read the chapter — Causal Inference
Lesson 2 · 6 steps
The Validity Toolkit · lesson 2 of 4