Propensity Score Weighting Methods

In the quest to uncover cause-and-effect relationships from observational data, researchers face a fundamental challenge: treated and untreated groups are often not directly comparable due to confounding variables. Propensity score weighting, specifically inverse probability of treatment weighting (IPTW), elegantly solves this by re-weighting the observed data to create a pseudo-population where treatment assignment is independent of the measured confounders. This allows for the unbiased estimation of causal effects like the average treatment effect (ATE), provided key assumptions are met.

Understanding the Propensity Score and IPTW

The propensity score, denoted as $e(X)$, is the conditional probability of receiving a treatment given a set of observed covariates $X$. Formally, $e(X)=P(T=1|X)$, where $T=1$ indicates treatment. It is typically estimated using logistic regression. The core idea of IPTW is to give each unit a weight that is the inverse of the probability of receiving the treatment they actually received. This effectively creates a balanced synthetic sample.

For a treated unit ($T=1$), the weight is $w=1/e(X)$. For a control unit ($T=0$), the weight is $w=1/(1-e(X))$. In this re-weighted population, the covariates $X$ are no longer predictive of treatment status. The ATE can then be estimated as a simple weighted difference in means. For example, if $Y$ is the outcome, the weighted mean for the treated group is $\frac{\sum (T_i\cdot Y_i\cdot w_i)}{\sum (T_i\cdot w_i)}$, and similarly for controls. The ATE is the difference between these two weighted averages.

Stabilized Weights and Trimming Extreme Values

Basic IPTW weights can be highly variable, especially when propensity scores are very close to 0 or 1. This increases the variance of the effect estimate. A common refinement is to use stabilized weights. For treated units, the stabilized weight is $w^{stab}=P(T=1)/e(X)$, and for control units, it's $w^{stab}=P(T=0)/(1-e(X))$. Here, $P(T=1)$ is the marginal probability of treatment in the sample (e.g., the proportion treated). These weights have a mean of approximately 1, which dramatically reduces variability without introducing bias.

When propensity scores are extremely low or high, the corresponding IPTW weights become enormous, giving a few units disproportionate influence and destabilizing the analysis. Trimming is a practical solution. Common approaches include truncating weights at a specified percentile (e.g., the 1st and 99th) or discarding units with a propensity score outside a range like [0.1, 0.9]. While trimming reduces variance, it does so by changing the estimand from the ATE in the full population to the ATE in a subpopulation with overlap, which must be clearly communicated.

Doubly Robust Estimation

A major advancement that combines the strengths of modeling the treatment and the outcome is doubly robust estimation. This method, often implemented via augmented inverse probability weighting (AIPW), integrates propensity score weighting with outcome regression modeling.

The procedure is twofold: First, you model the outcome using regression (e.g., $Y$ on $T$ and $X$). Second, you calculate IPTW using the propensity score model. The doubly robust estimator combines these. Its powerful advantage is that it yields an unbiased estimate of the ATE if either the propensity score model or the outcome regression model is correctly specified—not necessarily both. This "two chances to get it right" property makes it a preferred and more robust approach in many applied settings, guarding against model misspecification in one of the two key components.

Comparing IPTW with Matching and Stratification

Propensity score analysis isn't limited to weighting; matching and stratification (or subclassification) are two other primary techniques. The choice among them depends on your data characteristics and research goals.

IPTW uses all available data, which is efficient, but it can be sensitive to extreme weights. It directly targets the ATE for the entire population. Matching (e.g., 1:1 nearest-neighbor) creates a matched sample of treated and control units with similar propensity scores, discarding unmatched units. This often improves covariate balance intuitively and is less sensitive to extreme propensity scores, but it changes the estimand to the average treatment effect on the treated (ATT) and reduces statistical power by discarding data. Stratification divides the sample into strata (e.g., quintiles) based on the propensity score and calculates effects within each stratum before pooling. It is simple and less model-dependent than IPTW but can suffer from residual imbalance within strata.

In practice, IPTW is often favored when estimating the ATE with large datasets, especially when combined with stabilization and trimming. Matching is a strong choice when the ATT is the parameter of interest or when visual verification of balance is a priority. Stratification offers a straightforward, less parametric alternative.

Common Pitfalls

1. Ignoring Extreme Propensity Scores: Applying IPTW without checking the distribution of weights is a critical error. Extreme weights indicate a lack of overlap between treatment groups (violation of the positivity assumption), and the resulting estimate will be unstable and potentially biased. Correction: Always visualize the propensity score distribution by treatment group, calculate weight statistics (mean, max), and implement trimming or consider changing the target population.
2. Failing to Assess Balance: A key step after weighting is to check if the method successfully balanced the covariates. Correction: After applying weights, re-calculate standardized mean differences for all confounders $X$. Successful balancing should reduce these differences to near zero (e.g., |SMD| < 0.1). If balance is poor, the propensity score model may need to be re-specified.
3. Misinterpreting the Estimand After Trimming/Dropping Units: Trimming weights or using matching discards data, which changes the population for which you are estimating an effect. Correction: Clearly define and report your final target population. For instance, "This analysis estimates the average treatment effect for individuals with a propensity score between 0.2 and 0.8."
4. Assuming Unconfoundedness is Proven: Propensity score methods only adjust for observed confounders. The critical unconfoundedness assumption (that all confounders are measured) cannot be tested with the data. Correction: This is a fundamental limitation of observational studies. Sensitivity analyses, like assessing how strong an unmeasured confounder would need to be to nullify the result, are essential for robust interpretation.

Summary

• Inverse Probability of Treatment Weighting (IPTW) uses propensity scores to weight observations, creating a pseudo-population where treatment is independent of confounders, enabling estimation of the Average Treatment Effect (ATE).
• Stabilized weights ($P(T)/e(X)$) reduce the variance of IPTW estimates, while trimming extreme weights addresses violations of the positivity assumption and improves stability.
• Doubly robust estimation (e.g., AIPW) combines outcome regression with IPTW, providing a valid estimate if either the propensity score model or the outcome model is correct, offering superior protection against model misspecification.
• Choosing between IPTW, matching, and stratification involves trade-offs: IPTW is efficient for the ATE but sensitive to model specification; matching is intuitive for the ATT but discards data; stratification is simple but may leave residual imbalance.
• A valid analysis requires careful diagnostics: always check for overlap in propensity scores, assess covariate balance after weighting, and transparently acknowledge the untestable assumption of no unmeasured confounding.