Interactive demo · Probabilistic Machine Learning: Advanced Topics · Chapter 36

Adjusting for confounders

One confounder X pushes people into treatment and also moves the outcome, so the raw difference in means is wrong. Five estimators of the average treatment effect (true value 1) use a model for the outcome, a model for the treatment, or both. Make each model wrong in turn and see who survives: the outcome regression needs a right outcome model, inverse weighting a right propensity model, and the augmented estimator needs only one of the two.

Population

Models fitted

outcome model

propensity model

Try these

Model: X ~ N(0,1); treatment P(A=1 | x) = sigmoid(c x + d (x² − 1)); outcome Y = A + x + k (x² − 1) + noise with standard normal noise, so the true average effect is exactly 1. “Linear” fits 1, x; “quadratic” fits 1, x, x². Standard errors: Welch for the naive contrast, least squares for the regression, the influence-curve estimate for the three weighting estimators. The bars on the right panel are 95% intervals for this sample; after “Repeat”, the faint bar is the 5th to 95th percentile of the estimate over 300 new samples and the white tick is its mean.

In the lecture notes