Compare users exposed to certain factors (e.g., new dataset release) with similar users who were not.
Propensity Score Matching (PSM) is a statistical technique used to estimate causal effects in observational studies where randomization is not feasible. The goals of PSM are to:
1. **Reduce Selection Bias**: Match treated and untreated groups based on similar characteristics to ensure comparability, mimicking the balance achieved in randomized experiments.
2. **Control for Confounding Variables**: Adjust for confounding variables that might affect both the treatment and the outcome.
3. **Estimate Causal Effects**: Infer the causal effect of a treatment (or intervention) by comparing outcomes between treated and matched untreated groups.
### Why use Propensity Score Matching?
- **Non-Randomized Data**: In real-world settings, treatments or interventions are often not randomly assigned, leading to confounding.
- **Balance Covariates**: PSM creates comparable groups by balancing covariates (background variables) that might bias the results.
- **Clear Causal Interpretation**: By reducing confounding, PSM allows for a more accurate estimation of the treatment's effect.
- **Transparency**: The method is intuitive and allows for diagnostics to assess the quality of matching.
### How Propensity Score Matching works
1. **Calculate Propensity Scores**:
- Propensity scores are the probability of receiving the treatment, given a set of observed covariates.
- For example, using logistic regression: $P(Treatment = 1 | Covariates) = \text{logistic}(\beta_0 + \beta_1 X_1 + \beta_2 X_2 + \ldots + \beta_k X_k)$
- Covariates can include age, income, education, etc., that may influence the likelihood of receiving the treatment.
2. **Match Treated and Untreated Units**:
- Match each treated individual with one or more untreated individuals with similar propensity scores.
- Common matching methods include:
- **Nearest Neighbor Matching**: Pair treated individuals with the closest untreated individuals based on propensity scores.
- **Caliper Matching**: Only match if the propensity score difference is below a threshold.
- **Kernel Matching**: Weight untreated observations based on their distance to treated observations.
3. **Check Balance**:
- Evaluate whether covariates are balanced between treated and untreated groups after matching (e.g., through standardized mean differences or visualizations).
4. **Estimate Treatment Effect**:
- Compare outcomes (e.g., mean differences) between treated and matched untreated groups to estimate the Average Treatment Effect on the Treated (ATT).
### Example: Evaluating a training program's effect on income
**Objective**: Determine whether attending a job training program (treatment) increases participants' income (outcome).
**Step 1: Calculate Propensity Scores**
- Covariates: Age, education level, prior income, work experience.
- Use logistic regression to estimate the probability of attending the program (propensity score) for each individual.
**Step 2: Match Participants**
- Match program attendees (treated) with non-attendees (untreated) based on similar propensity scores. For example:
- Treated individual with a propensity score of 0.75 is matched with an untreated individual with a score of 0.74.
**Step 3: Check Balance**
- Confirm that covariates (e.g., education level, prior income) are similar between the two groups after matching.
**Step 4: Estimate Treatment Effect**
- Calculate the difference in average income between the treated and matched untreated groups:
- Average income of treated group: $50,000.
- Average income of matched untreated group: $45,000.
- Estimated causal effect (ATT): $50,000 - $45,000 = **$5,000**.
## Benefits and Limitations of PSM
### Benefits
- Reduces bias by controlling for observed confounders.
- Works with observational data.
- Clear and interpretable results.
### Limitations
- Unmeasured Confounders: Cannot adjust for variables that are not observed or included in the model.
- Data Loss: May exclude unmatched individuals, reducing sample size.
- Propensity Score Overlap: Requires sufficient overlap in propensity scores between treated and untreated groups.