Calibrated and Conformal Propensity Scores for Causal Effect Estimation.

Area of Science:

• Causal inference
• Statistical modeling
• Bioinformatics

Background:

• Propensity scores are widely used in observational studies to estimate treatment effects.
• The probabilistic output of propensity score models is often uncalibrated, potentially leading to biased estimates.
• Calibration ensures predicted probabilities reflect true event rates.

Purpose of the Study:

• To propose and validate calibration techniques for learned propensity score models.
• To demonstrate the necessity of calibration for unbiased treatment effect estimation.
• To improve the accuracy and efficiency of causal inference methods.

Main Methods:

• Developed simple recalibration techniques for probabilistic propensity score models.
• Proved calibration as a necessary condition for unbiased estimation with inverse propensity weighting and doubly robust estimators.
• Derived error bounds relating causal effect estimates to model uncertainty and calibration quality.

Main Results:

• Calibration strictly improves error bounds on causal effect estimates.
• Calibrated propensity scores avoid extreme propensity weights, enhancing stability.
• Demonstrated improved causal effect estimation in high-dimensional image and GWAS data.
• Achieved over two-fold speedup in GWAS analysis using calibrated propensity scores.

Conclusions:

• Calibration of probabilistic propensity score models is essential for reliable causal effect estimation.
• Calibrated propensity scores offer a pathway to more accurate and efficient causal inference.
• The proposed methods enhance the utility of propensity scores in complex, high-dimensional data analyses, including GWAS.