Propensity Score

The validity of propensity score methods rests on the strong ignorability or unconfoundedness assumption. This requires two conditions:

• Conditional Independence: The potential outcomes (Y(1), Y(0)) are independent of treatment assignment given the covariates: (Y(1), Y(0)) ⟂ T | X.
• Positivity: Every unit has a non-zero probability of receiving either treatment: 0 < P(T=1 | X) < 1 for all X. If these hold, conditioning on the propensity score e(X) is sufficient to satisfy unconfoundedness, allowing for unbiased estimation of the Average Treatment Effect (ATE).

Propensity scores are not directly observed and must be estimated from data, typically using a binary classification model.

• Logistic Regression: The most common method, modeling log-odds of treatment as a linear function of X.
• Machine Learning Classifiers: Methods like boosted trees (e.g., XGBoost) or random forests can better capture non-linearities and interactions in high-dimensional settings, though they require careful cross-fitting to avoid overfitting bias.
• Regularization (L1/L2): Used in high-dimensional settings to prevent overfitting. The goal is not prediction accuracy per se, but achieving covariate balance in the resulting matched or weighted sample.

Estimated propensity scores are used in several key adjustment techniques:

• Propensity Score Matching: Pairs treated units with control units that have similar propensity scores, creating a balanced sample for effect estimation.
• Inverse Probability of Treatment Weighting (IPTW): Creates a pseudo-population by weighting each unit by the inverse of its probability of receiving the treatment it actually received. Weights are 1/e(X) for treated units and 1/(1-e(X)) for control units.
• Stratification (Subclassification): Units are grouped into strata (e.g., quintiles) based on their propensity score, and treatment effects are estimated within each stratum before being aggregated.

After estimation, diagnostics are essential to validate the propensity score model's performance.

• Covariate Balance: The primary diagnostic. Standardized mean differences, variance ratios, and empirical quantile-quantile plots for key covariates should be compared between treated and control groups after matching or weighting. Imbalance should be minimal (< 0.1 standard deviations).
• Overlap/Common Support: Visualizing the distribution of propensity scores for treated and control units reveals the region of common support. Estimation should be limited to this region where both groups have substantial density, as extrapolation outside this region is unreliable.

Inverse Probability Weighting (IPW) is a primary application of propensity scores. It creates a pseudo-population where treatment assignment is independent of observed covariates by weighting each unit by the inverse of its probability of receiving the treatment it actually received.

• Formula: Weight for a treated unit = 1 / PS. Weight for a control unit = 1 / (1 - PS).
• Effect: Units with a low probability of receiving their actual treatment (e.g., a treated unit with a low PS) are up-weighted, as they are rare and informative.
• Outcome: The weighted sample allows for an unbiased estimate of the ATE using a simple weighted average of outcomes.

The Average Treatment Effect (ATE) is the target causal quantity often estimated using propensity scores. It is the expected difference in an outcome if every unit in the population received the treatment versus if none did.

• Definition: ATE = E[Y(1) - Y(0)], where Y(1) is the potential outcome under treatment and Y(0) is the potential outcome under control.
• The Fundamental Problem of Causal Inference: We never observe both Y(1) and Y(0) for the same unit. Propensity score methods help overcome this by creating comparable groups.
• Interpretation: A positive ATE indicates the treatment, on average, improves the outcome across the population.

Propensity score matching is a specific method that uses the estimated propensity score to pair each treated unit with one or more "similar" control units, based on their PS. The matched control group serves as the counterfactual.

• Process: For each treated unit, find control unit(s) with the closest propensity score (e.g., using nearest-neighbor matching).
• Outcome: The average outcome difference within each matched pair is computed, then averaged across all pairs to estimate the Average Treatment Effect on the Treated (ATT).
• Assumption: Relies on the common support condition, meaning there must be overlap in PS distributions between groups.

The seminal paper "The Central Role of the Propensity Score in Observational Studies for Causal Effects" by Paul Rosenbaum and Donald Rubin formally defined the propensity score and established its key theoretical properties.

• Key Theorem: They proved that if treatment assignment is strongly ignorable given observed covariates, then it is also strongly ignorable given just the propensity score. This reduces the dimensionality of the confounding adjustment problem.
• Impact: This work provided the mathematical foundation for all modern propensity score methods, transforming the analysis of observational data in fields from economics to medicine.
• Legacy: The paper introduced core concepts like balance checking and the propensity score theorem.