Estimating weighted quantile treatment effects with missing outcome data by double sampling.

Area of Science:

• Causal inference
• Biostatistics
• Health data analysis

Background:

• Standard causal inference methods focus on average effects, limiting analysis of extreme outcomes.
• Estimating causal weighted quantile treatment effects (WQTEs) is crucial for understanding tail distributions.
• Real-world data, like electronic health records (EHRs), often have missing-not-at-random (MNAR) data, biasing WQTE estimation.

Purpose of the Study:

• To develop a method for estimating causal WQTEs in the presence of MNAR data.
• To mitigate bias in WQTE estimation using double sampling strategies.
• To provide robust causal inference for tail counterfactual distributions using real-world data.

Main Methods:

• Proposed a double sampling strategy to ascertain missing data on a sub-sample.
• Developed a novel inverse-probability weighted estimator with derived asymptotic properties.
• Introduced a bootstrap method for both pointwise and uniform inference, estimating propensity scores and double-sampling probabilities.

Main Results:

• The proposed method identifies causal WQTEs without requiring missingness assumptions on the original data.
• Asymptotic properties were derived for the novel estimator, supporting pointwise and uniform inference.
• Simulation studies demonstrated the finite sample performance of the proposed estimators.

Conclusions:

• Double sampling effectively mitigates bias from MNAR data in causal WQTE estimation.
• The novel estimator and bootstrap inference provide reliable tools for analyzing tail treatment effects.
• The method was successfully illustrated using EHR data on bariatric surgery outcomes.