Nonparametric Estimation for Propensity Scores With Misclassified Treatments.

Area of Science:

• Causal Inference
• Statistical Modeling

Background:

• Estimating average treatment effect (ATE) is crucial in causal inference.
• Propensity score methods are widely used but assume precise binary treatment measurement.
• Existing methods fail with measurement error in treatments and nonlinear confounder relationships, leading to biased results.

Purpose of the Study:

• To analytically examine the impact of measurement error on ATE estimation.
• To develop a valid method for addressing binary treatments with misclassification.
• To improve the accuracy of ATE estimation in the presence of treatment measurement error.

Main Methods:

• Analytical examination of bias in ATE estimators due to treatment measurement error.
• Development of a correction method for misclassified binary treatments.
• Utilizing random forest for propensity score estimation with nonlinear confounders.
• Derivation of an error-eliminated ATE estimator with established asymptotic properties.

Main Results:

• Quantified the bias introduced by measurement error in binary treatments on ATE estimation.
• Proposed a novel method to correct for misclassified treatments.
• The proposed estimator demonstrates improved finite sample performance in numerical studies.

Conclusions:

• Measurement error in binary treatments significantly biases ATE estimation.
• The developed method effectively corrects for misclassification and nonlinear confounder effects.
• Correcting for measurement error is essential for valid and reliable causal inference.