Propensity Score Matching

The propensity score is the conditional probability of receiving a treatment given a vector of observed baseline covariates. It is typically estimated using a logistic regression model where treatment assignment is the dependent variable and all measured confounders are independent variables.

• Reduces a high-dimensional set of covariates to a single scalar value between 0 and 1.
• Satisfies the balancing property: at any given score value, the distribution of covariates is identical between treated and control groups.
• Enables the identification of comparable subjects who would have been equally likely to receive the intervention.

Once propensity scores are estimated, subjects are paired using specific algorithms to create balanced groups. The goal is to find a control subject whose score is identical or very close to that of a treated subject.

• Nearest Neighbor Matching: Pairs each treated unit with the control unit possessing the closest propensity score, often within a specified caliper width (maximum allowable distance).
• Kernel Matching: Uses a weighted average of all controls, with weights inversely proportional to the distance in scores.
• Stratification: Divides subjects into blocks (quintiles) based on the propensity score, ensuring balance within each stratum.

The common support (or overlap) is the range of propensity scores where both treated and control observations exist. PSM requires sufficient overlap to make valid comparisons.

• Observations falling outside this region are typically discarded to avoid extrapolation bias.
• A visual inspection of the propensity score distribution histograms is a critical diagnostic step.
• Violations of the overlap assumption indicate that the treated and control populations are fundamentally too different to compare reliably.

After matching, the researcher must verify that the matching procedure successfully eliminated systematic differences in covariates. This is not assumed; it is tested.

• Standardized Mean Difference (SMD): The absolute difference in means divided by the pooled standard deviation. An SMD less than 0.1 is generally considered adequate balance.
• Variance Ratios: Compares the variance of covariates between groups; a ratio near 1.0 indicates good balance.
• Love Plots: Graphical tools that visualize the reduction in covariate imbalance before and after matching.

PSM only controls for observed confounders. Sensitivity analysis tests how strongly an unmeasured confounder would need to influence selection to nullify the estimated treatment effect.

• Rosenbaum Bounds: Quantifies the magnitude of hidden bias required to alter the inference for a statistically significant result.
• Gamma (Γ): A parameter representing the odds ratio of differential assignment due to an unobserved covariate. If the p-value remains significant at high gamma values, the result is robust.
• Essential for establishing credibility in clinical research where randomization is absent.

The primary estimand in PSM is usually the Average Treatment Effect on the Treated (ATT). This measures the effect of the intervention specifically for those who actually received it.

• Calculated as the mean difference in outcomes between the treated group and their matched controls.
• Contrasts with the ATE (Average Treatment Effect), which estimates the effect if the entire population were treated.
• Standard errors must be adjusted to account for the uncertainty in the propensity score estimation step, typically using bootstrapping.

A confounding variable is an extraneous covariate that correlates with both the treatment assignment and the outcome of interest. In clinical studies, confounders like disease severity or socioeconomic status can create spurious associations or mask true causal relationships. Propensity score matching explicitly balances these variables across treatment and control groups to break the backdoor path and isolate the treatment effect.

IPTW is an alternative to matching that uses the propensity score to create a pseudo-population where treatment assignment is independent of observed covariates. Instead of pairing subjects, IPTW assigns weights equal to 1/propensity score for treated subjects and 1/(1-propensity score) for controls. This method preserves the full sample size but is more sensitive to extreme propensity scores near 0 or 1.

After matching, researchers must verify that covariates are balanced between groups using standardized mean differences (SMD). An SMD below 0.1 (10%) is generally considered adequate balance. Additional diagnostics include:

• Love plots to visualize pre- and post-matching balance
• Variance ratios to check distributional similarity
• Kolmogorov-Smirnov tests for continuous variables

The ATT is the primary estimand in propensity score matching, measuring the effect of treatment specifically on those who actually received it. Unlike the ATE (Average Treatment Effect), which estimates the effect across the entire population, the ATT answers: What was the benefit for treated patients compared to what would have happened had they not been treated? This is often the most policy-relevant quantity.

Rosenbaum bounds sensitivity analysis tests how strongly an unmeasured confounder would need to influence treatment assignment to overturn the study's conclusions. The key parameter, Gamma (Γ), represents the odds ratio of treatment assignment between matched pairs due to hidden bias. If results remain significant at Γ = 2, the findings are considered robust to moderate unobserved confounding.

Caliper matching restricts matches to pairs whose propensity scores fall within a pre-specified distance threshold, typically 0.2 standard deviations of the logit of the propensity score. This prevents poor matches where treated and control subjects have substantially different covariate profiles. The trade-off is that some treated subjects may remain unmatched, reducing sample size but improving covariate balance.