Propensity Score Matching Methods
You'll learn how to match treated and control units using propensity scores through nearest-neighbor, caliper, and kernel matching techniques.
The Core Idea
Once you've estimated propensity scores (the probability of receiving treatment given covariates), you need to actually match treated units with similar control units. Think of it like finding "twins" in your data—people who looked equally likely to receive treatment, but one did and one didn't.
Three Main Matching Techniques
Nearest-Neighbor Matching pairs each treated unit with the control unit(s) having the closest propensity score. Imagine lining everyone up by their score and picking the nearest neighbor. You can use 1:1 matching (one control per treated) or 1:k matching (multiple controls). It's simple but can produce poor matches if the nearest neighbor is still quite far away.
Caliper Matching adds a safety rule: only match if the propensity scores are within a maximum distance (the "caliper"). Think of it as saying "I'll only accept a match if they're this close." Common caliper widths are 0.1 or 0.25 standard deviations of the propensity score. This prevents bad matches but may leave some treated units unmatched.
Kernel Matching uses a weighted average of all control units, with weights based on distance from each treated unit's propensity score. Closer controls get higher weights. It's like asking "what would happen if this treated person were a blend of all similar controls?" This uses more data but is more complex computationally.
Choosing Your Method
Nearest-neighbor is fast and intuitive. Caliper adds quality control. Kernel maximizes data usage but requires more assumptions about the weighting function.
Key Takeaway: Propensity score matching methods—nearest-neighbor, caliper, and kernel—offer different trade-offs between match quality, sample retention, and computational complexity when finding comparable treated and control units.