Both Inverse Probability Weighting (IPW) and Propensity Score Matching (PSM) are methods for addressing confounding in observational studies by using propensity scores. However, they differ in how they use these scores to create balance between treatment and control groups.
IPW assigns weights to individuals based on the inverse of their propensity score (PS).
Ensures that treatment groups are reweighted to look like a randomized experiment.
Inverse Probability Weighting (IPW)
logit treat x1 x2 x3
predict pscore
This generates the propensity score (pscore), i.e., the probability of receiving the treatment.
Alternative IPW Approach Using teffects ipw
teffects ipw (y) (treat x1 x2 x3) Automatically estimates the Average Treatment Effect (ATE).
Strengths of IPW :
-
Retains all observations (unlike PSM).
- Less sensitive to poor matches
Propensity Score Matching (PSM)
- PSM pairs individuals in the treatment and control groups based on similar propensity scores.
- After matching, treatment effects are estimated using only matched data.
logit treat x1 x2 x3
predict pscore
Strengths of PSM:
- Ensures high comparability between treated and control groups.
- Does not rely on extrapolation, reducing model dependence.
Key Differences: IPW vs. PSM
| Feature | IPW | PSM |
|---|---|---|
| Retains all observations? | ✅ Yes | ❌ No (drops unmatched cases) |
| Estimates ATE? | ✅ Yes | ❌ No (estimates ATT) |
| Sensitive to Poor Matches? | ❌ No | ✅ Yes |
| Sensitive to Extreme Weights? | ✅ Yes | ❌ No |
| More Computationally Expensive? | ❌ No | ✅ Yes |
| Better for Small Samples? | ❌ No (weights can be unstable) | ✅ Yes |
SUMMERY:
Inverse Probability Weighting (IPW) and Propensity Score Matching (PSM) are both methods for addressing confounding in observational studies using propensity scores. IPW assigns weights to individuals based on the inverse of their probability of receiving treatment, ensuring all observations are retained and enabling the estimation of the Average Treatment Effect (ATE). However, it can suffer from instability due to extreme weights. In contrast, PSM matches treated and control units based on similar propensity scores, improving comparability but discarding unmatched observations, which can reduce sample size. While IPW is better for retaining data and handling high-dimensional confounders, PSM is more intuitive and less dependent on model assumptions. The choice depends on study goals—IPW for ATE estimation with full data and PSM for better-matched groups, though a combination of both can enhance robustness.
Final Recommendation:
- Use IPW if you want to retain all observations and estimate ATE.
- Use PSM if you want a well-matched treatment/control group with fewer model assumptions.
- Consider combining both for robustness.
