Abstract
Evaluating treatments received by one population for application to a different target population is a central problem in causal inference. We study the minimax linear estimator of the treatment-specific mean outcome on a target population and provide a theoretical basis for inference. Focusing on the case where the unknown outcome function belongs to the unit ball of a reproducing kernel Hilbert space (RKHS), we show that the resulting linear estimator is asymptotically optimal. We justify the common practice of ignoring bias when building confidence intervals and establish bounds attesting to the estimator's good finite sample properties. Extensive simulations demonstrate promising performance across various sample sizes and overlap levels.
Introduction: The Challenge of Transportability
In causal inference, a recurring challenge is "transportability": observing treatment effects in one population (the treated) and estimating the effect on a different target population. Differences in characteristics between these groups—known as covariate shift—mean that simple averages are often biased.
While "balancing weights" are commonly used to adjust for these shifts, two primary issues remain:
- Inference Challenges: Estimators often retain residual bias. Practitioners typically ignore this bias when constructing confidence intervals, but theoretical justification for this practice has been lacking.
- Model Selection: Determining the "smoothness" of the underlying regression function is difficult, complicating the choice of an appropriate statistical model.
Methodology: Minimax Linear Estimators
The authors propose Minimax Linear Estimators ($\hat{\psi}_{ML}$), which identify weights that minimize the worst-case mean squared error over a class of regression functions.
Key technical contributions include:
- RKHS Framework: Analyzing the estimator within the "unit ball" of a Reproducing Kernel Hilbert Space (RKHS) allows for modeling complex, non-linear relationships.
- Duality: The authors prove that this minimax weighting approach is mathematically equivalent to a "plug-in" estimator using kernel ridge regression.
- Asymptotic Analysis: They derive the influence function, showing the estimator is asymptotically normal and efficient.
- Translation Invariance: The introduction of $\hat{\psi}_{MLt}$ ensures the estimator is invariant to shifts in the outcome scale, enhancing practical stability.
Experimental Setup
The estimator was evaluated using two benchmark simulations:
- Hainmueller (2012): A complex setup with six covariates and a quadratic outcome model to test handling of non-linearities.
- Kang and Schafer (2007): A "stress test" with poor group overlap and intentionally misspecified models to evaluate robustness.
Key Results
The simulation results highlight the superior performance of the translation-invariant minimax estimator ($ML_t$) compared to traditional methods like Inverse Probability Weighting (IPW) and Ordinary Least Squares (OLS).
- Superior RMSE: In the Hainmueller simulations at $n=50$, $ML_t$ achieved an RMSE of 3.30, significantly lower than IPW (4.37) or OLS (4.78).
- Robust Coverage: In the Kang-Schafer test, $ML_t$ maintained a 95% confidence interval coverage between 0.93 and 0.99. Conversely, OLS coverage dropped to 0.46 at $n=1000$ due to its inability to account for covariate shift bias.
- Efficiency: The estimator reaches the semi-parametric efficiency bound, utilizing data optimally under the given assumptions.
Discussion and Insights
- Justifying "Bias-Unaware" Intervals: The study shows that in correctly specified RKHS models, the minimax estimator's bias is "second-order," disappearing faster than variance as sample size grows. This provides theoretical grounding for using standard confidence intervals without explicit bias correction.
- Adaptivity: The estimator automatically adapts to the smoothness of the true underlying function, improving performance without requiring manual model adjustments.
- Stability: By linking balancing weights to kernel ridge regression, the authors demonstrate why these estimators are more stable than IPW, which can fail when group overlap is poor.
Limitations and Future Work
While powerful, the choice of kernel (e.g., Matérn or Gaussian) still requires domain knowledge or cross-validation. Additionally, solving the quadratic optimization for weights can be computationally intensive for very large datasets. Finally, the method assumes "strong ignorability" (no unobserved confounders); extending this to settings with unobserved confounding is a key area for future research.
Summary updated