Causal Effects via Propensity Scores

How to estimate effects from observational data

This article is the 2nd post in a series on causal effects. In the previous post, we laid a theoretical foundation for causal effects, but there were some lingering practical concerns. Namely, how can we compute causal effects from observational data? Here, I will discuss a set of techniques that do exactly this using something called a propensity score. The discussion will be followed up with example Python code of using these techniques with real-world data.

Key points:

1. Propensity scores estimate the probability of treatment
2. Subjects with similar propensity scores have similar baseline covariates
3. 3 popular propensity score techniques are matching, stratification, and inverse probability of treatment weighting

To estimate causal effects we need data. More specifically, we need data that contains outcomes, treatments, and covariates (defined previously). However, not all data sources are equal. Before discussing the propensity score, we must draw a clear distinction between two ways data are obtained.