Confounding-adjustment methods for the causal difference in medians.

Area of Science:

• Causal inference
• Biostatistics
• Epidemiology

Background:

• Traditional average causal effect estimation using population means is problematic with skewed continuous outcomes.
• Existing methods like outcome transformation or using population means may not be satisfactory for skewed data.
• Methods for estimating the causal difference in medians, particularly with confounding adjustment, are less discussed.

Purpose of the Study:

• To describe and compare confounding-adjustment methods for estimating the causal difference in medians.
• To address the gap in understanding how to handle skewed outcome data in causal effect estimation.

Main Methods:

• Evaluated multivariable quantile regression, inverse probability weighted (IPW) estimators, weighted quantile regression, and g-computation.
• Assessed methods via a simulation study with varying outcome skewness.
• Applied methods to an empirical dataset from the Longitudinal Study of Australian Children.

Main Results:

• Inverse probability weighted (IPW) estimators, weighted quantile regression, and g-computation minimized bias when models were correctly specified.
• G-computation additionally minimized variance.
• Multivariable quantile regression produced biased results due to its constant-effect assumption.

Conclusions:

• The inverse probability weighted (IPW) and g-computation methods offer effective strategies for estimating the causal difference in medians.
• These methods are valuable for handling skewed outcome data in causal analyses.
• The study highlights practical applications and the utility of these advanced statistical techniques.