Imagine running a randomized controlled trial (RCT) or an A/B test where a fair coin determines whether a participant receives a treatment. Because you designed the experiment, you know the exact probability of treatment assignment for every individual is precisely 50%.
When it comes time to estimate the average treatment effect using inverse probability weighting (IPW), intuition suggests you should plug this true, known probability ($0.5$) into your estimator. After all, why estimate something you already know with absolute certainty?
Yet, a foundational insight in causal inference reveals that this intuition is wrong. Estimating the propensity score from the observed sample data—even when the true population score is perfectly known—yields causal estimates with lower variance and higher statistical efficiency.
The Intuition of Sample-Specific Balance
To understand why this paradox occurs, we have to look at the difference between a population and a finite sample.
The true propensity score reflects the probability of treatment assignment in the limit, across an infinite population. However, any real experiment relies on a finite sample. By pure stochastic chance, the coin flips will result in slight, accidental correlations between baseline patient covariates and the treatment assignment. For instance, the treatment group might end up being slightly older or sicker than the control group.
If you use the true propensity score ($0.5$), your estimator is completely blind to these accidental correlations. The baseline imbalance leaks directly into your final outcome analysis, inflating the variance of your treatment effect estimate.
True Propensity Score (Population Level) ──> Ignores finite sample fluctuations ──> Higher Variance
Estimated Propensity Score (Sample Level) ──> Captures and adjusts for chance imbalances ──> Lower Variance
An estimated propensity score model, by contrast, looks directly at the sample. It asks: "Given the data we actually observed, what characteristics predict treatment?" If older patients accidentally received treatment more often, the empirical model will yield a slightly higher propensity score for older individuals. When these estimated scores are used for weighting or matching, they actively adjust for the sample-specific chance imbalances, effectively neutralizing experimental noise.
As Hui Guo et al. demonstrate in Sufficient covariates and linear propensity analysis, the estimation process acts as an information-preserving mechanism. By capturing and adjusting for these stochastic correlations specific to the observed sample, empirical estimation reduces the overall variance of the final causal estimate.
Universal Efficiency Gains Across Frameworks
This phenomenon is not a niche quirk of standard weighting; it is a structural property that spans across diverse causal inference and machine learning paradigms.
Policy Learning and Bandits
In off-policy evaluation for multi-armed bandits, researchers use historical data logged by a known algorithm to evaluate a new policy. Even though the logging policy's probabilities are known exactly, Yusuke Narita et al. show in Efficient Counterfactual Learning from Bandit Feedback that estimating these probabilities from the data using machine learning models minimizes estimation variance, allowing the estimator to achieve the theoretical limit of semiparametric efficiency.
Optimal Treatment Rules
When the goal is to discover personalized medicine rules in high-dimensional settings, Chengchun Shi et al. (Robust learning for optimal treatment decision with NP-dimensionality) find that estimating propensity scores borrows vital information from observed covariates. This empirical adjustment leads to more precise decision boundaries and more accurate treatment rules than fixed assignment probabilities.
Propensity Score Matching
The benefits carry over to matching architectures. In survival analysis, Tongrong Wang et al. (Propensity score matching for estimation of pairwise marginal hazard ratios) prove that matching based on an estimated propensity score is statistically more efficient than matching on the true score, because the estimation step utilizes internal treatment assignment information to lower the overall asymptotic variance.
Data Integration
Even when transporting clinical trial findings by integrating them with external observational data, Wenjie Hu et al. (Identification and Estimation of Treatment Effects on Long-Term Outcomes in Clinical Trials With External Observational Data) confirm that IPW estimators are systematically enhanced when true scores are replaced by their estimated counterparts.
Designing the Optimal Propensity Model
If estimating the propensity score is inherently superior to using the true score, the next logical question is: how should we construct the estimation model to maximize this efficiency? Recent literature focuses on tuning the propensity model specifically to optimize downstream causal precision.
Targeted Covariate Balancing
Rather than relying on standard maximum likelihood estimation (like standard logistic regression), Jianqing Fan et al. introduce an optimal Covariate Balancing Propensity Score (oCBPS) framework in Optimal Covariate Balancing Conditions in Propensity Score Estimation. This method explicitly forces the propensity model to prioritize sample-specific balance, allowing the estimator to hit theoretical efficiency limits faster, even if the underlying statistical models are slightly misspecified.
Outcome-Adaptive Estimation
Traditional modeling dictates that a propensity score should only include variables that predict treatment. However, Jiacong Du et al. (Outcome adaptive propensity score methods for handling censoring and high-dimensionality: Application to insurance claims) demonstrate that incorporating information about the outcome into the propensity model significantly enhances efficiency. By injecting the predicted probability of the outcome as a covariate into the treatment model, the estimator focuses its balancing power on the covariates that matter most for reducing bias and variance.
Collaborative Tuning
In a similar vein, Kara E. Rudolph et al. (Improving efficiency in transporting average treatment effects) propose a collaborative estimator that tunes the propensity score model based directly on the Conditional Average Treatment Effect (CATE). This approach optimizes the transportability of treatment effects across populations without requiring prior knowledge of effect modifiers.
High-Dimensional Regularization
When dealing with a massive number of patient characteristics, Xin He et al. (Nonparametric augmented probability weighting with sparsity) combine nonparametric imputation with a penalized propensity score estimator using a group lasso penalty. Operating within an Augmented Inverse Probability Weighting (AIPW) framework, this approach maintains low bias and standard error even in sparse, high-dimensional settings.
Stabilizing Weights for Practical Power
While empirical propensity scores reduce variance by capturing sample-specific fluctuations, they can occasionally produce extreme weights if the sample contains regions with low overlap (where certain individuals have a near-zero or near-one probability of treatment). To prevent these extreme weights from destabilizing the estimator, researchers have developed smoothing and alternative weighting architectures.
For example, Fujin Zhu et al. (Pareto-smoothed inverse propensity weighing for causal inference) introduce a Pareto-smoothing technique that stabilizes extreme propensity weights. This method yields a higher effective sample size and enhances statistical power compared to raw IPW or standard normalization.
Similarly, when evaluating disparities across multiple groups, Jiajun Liu et al. (Assessing racial disparities in healthcare expenditure using generalized propensity score weighting) utilize a weighting scheme based on the harmonic mean of propensity scores. Instead of excluding or heavily truncating individuals with low overlap, this approach smoothly down-weights them, preserving vital sample information and maximizing the statistical power of the analysis.
Ultimately, the shift from true to estimated propensity scores represents a fundamental conceptual realization: causal inference is an exercise in balancing the specific groups under study. By treating the propensity score as an empirical target to be estimated rather than a fixed population parameter, researchers can actively anchor their analysis to the unique composition of their sample, maximizing both precision and robustness.
Key Foundational Work
Sufficient covariates and linear propensity analysis (2010) — contributed 3 concept elements to this direction.
Robust learning for optimal treatment decision with NP-dimensionality (2016) — contributed 2 concept elements to this direction.