Abstract
Inverse probability of treatment weighting (IPTW) is a popular method for estimating the average treatment effect (ATE). However, empirical studies show that the IPTW estimators can be sensitive to the misspecification of the propensity score model. To address this problem, researchers have proposed to estimate propensity score by directly optimizing the balance of pre-treatment covariates. While these methods appear to empirically perform well, little is known about how the choice of balancing conditions affects their theoretical properties. To fill this gap, we first characterize the asymptotic bias and efficiency of the IPTW estimator based on the Covariate Balancing Propensity Score (CBPS) methodology under local model misspecification. Based on this analysis, we show how to optimally choose the covariate balancing functions and propose an optimal CBPS-based IPTW estimator. This estimator is doubly robust; it is consistent for the ATE if either the propensity score model or the outcome model is correct. In addition, the proposed estimator is locally semiparametric efficient when both models are correctly specified. To further relax the parametric assumptions, we extend our method by using a sieve estimation approach. We show that the resulting estimator is globally efficient under a set of much weaker assumptions and has a smaller asymptotic bias than the existing estimators. Finally, we evaluate the finite sample performance of the proposed estimators via simulation and empirical studies. An open-source software package is available for implementing the proposed methods.
Introduction
In causal inference, estimating the Average Treatment Effect (ATE) is a fundamental goal. For instance, determining if a job training program increases earnings requires comparing outcomes between treated and control groups. In observational studies, where treatment assignment is not random, researchers often use Inverse Probability of Treatment Weighting (IPTW) based on propensity scores—the probability of receiving treatment given observed covariates.
However, IPTW is notoriously sensitive to model misspecification. If the propensity score model is even slightly incorrect, the resulting ATE estimate can be significantly biased. To mitigate this, the Covariate Balancing Propensity Score (CBPS) was developed to prioritize covariate balance over model fit. Yet, a critical theoretical question remained: which specific functions of the covariates should be balanced to achieve optimal results? This research introduces the optimal Covariate Balancing Propensity Score (oCBPS), providing a rigorous mathematical foundation for selecting balancing conditions.
Methodology: The oCBPS Estimator
The authors derive the oCBPS estimator by analyzing the asymptotic behavior of IPTW under local model misspecification. Key methodological contributions include:
- Optimal Balancing Functions: The research identifies that the optimal functions to balance are those related to the outcome model—specifically, the conditional mean of the outcomes. By prioritizing the balance of variables that actually drive the outcome, the estimator achieves higher precision.
- Double Robustness: The oCBPS estimator is "doubly robust," meaning it provides consistent ATE estimates if either the propensity score model or the outcome model is correctly specified.
- Sieve Estimation: To move beyond rigid parametric assumptions, the authors utilize a nonparametric sieve approach. This uses flexible basis functions (like splines or wavelets) that increase in complexity with the sample size, allowing the model to capture complex non-linear relationships without assuming a specific functional form.
Key Insights and Theoretical Contributions
The paper provides several surprising theoretical insights:
- Efficiency Paradox: Estimating the propensity score by forcing covariate balance can actually result in a more efficient ATE estimator than using the "true" propensity score. This challenges the intuition that knowing the true treatment assignment mechanism is always best.
- Superiority over AIPW: Under certain misspecification scenarios, oCBPS converges to the true effect faster than Augmented Inverse Probability Weighting (AIPW), which has long been considered the gold standard for efficiency.
- Relaxed Assumptions: The nonparametric sieve approach requires significantly less "smoothness" in the data (measured as the ratio of smoothness $s$ to dimension $d$, $s/d > 3/4$) compared to previous state-of-the-art methods that required much stricter assumptions ($s/d > 7$).
Empirical Validation
The authors evaluated oCBPS using both synthetic simulations and the classic LaLonde (1986) dataset.
Simulation Results
Across various scenarios involving linear and non-linear outcomes, oCBPS consistently outperformed standard Generalized Linear Models (GLM) and the original CBPS in terms of Root Mean Square Error (RMSE) and confidence interval coverage.
The LaLonde Benchmark
The LaLonde dataset is a standard test for causal estimators, comparing observational data to an experimental benchmark ($886 earnings increase). Under a linear specification, oCBPS produced an estimate of -$306, which is significantly closer to the benchmark than the -$1190 produced by traditional GLM methods.
Conclusion and Future Directions
The oCBPS framework provides a principled way to select balancing conditions in propensity score estimation. By achieving semiparametric efficiency and offering double robustness, it provides a more reliable tool for causal inference in observational studies. An open-source software package is available for implementing the proposed methods.
While oCBPS is a powerful step forward, the authors note some areas for growth:
- Basis Selection: While the method is optimal given a set of basis functions, the researcher still has to choose those functions. Future work could automate this selection process.
- High-Dimensional Data: In settings where the number of variables exceeds the number of observations, further specialized regularization techniques are required.
- Matching Estimators: Extending these optimal conditions to matching-based methods is a promising next step.
Summary updated