High-dimensional generalized propensity score with application to omics data | JoVE Visualize

Area of Science:

Statistics
Biostatistics
Epidemiology

Background:

Propensity score (PS) methods are crucial for causal effect estimation in non-randomized studies, relying on the untestable unconfoundedness assumption.
Including numerous covariates strengthens the unconfoundedness assumption but can introduce bias and reduce efficiency in PS models.
High-dimensional data, such as omics, present challenges for variable selection in causal inference, with existing methods primarily focusing on binary treatments.

Purpose of the Study:

To propose a novel variable selection method, the generalized outcome-adaptive LASSO (GOAL), for continuous treatments in high-dimensional causal inference.
To address the limitations of existing methods by developing a technique robust to model misspecification and suitable for complex datasets.
To evaluate the performance of GOAL in accurately and efficiently identifying relevant covariates for causal effect estimation.

Main Methods:

Developed the generalized outcome-adaptive LASSO (GOAL) algorithm for selecting covariates in the context of continuous treatments.
Conducted simulation studies to assess GOAL's performance under linear outcome models, comparing its accuracy and precision against ideal estimates.
Applied GOAL to analyze seven DNA methylation datasets from the Gene Expression Omnibus (GEO) database, focusing on Alzheimer's disease incidence.

Main Results:

Simulation studies demonstrated that GOAL effectively identifies true confounders and outcome predictors while excluding irrelevant covariates.
GOAL achieved high accuracy and precision in causal effect estimation, closely approximating ideal results in simulations.
The GOAL method proved robust to potential outcome model misspecification, maintaining reliable performance.

Conclusions:

GOAL offers an effective and robust approach for variable selection in causal inference with continuous treatments, particularly in high-dimensional settings.
The method enhances the reliability of causal effect estimates by appropriately selecting relevant covariates.
The application to DNA methylation data demonstrates GOAL's utility in real-world biomedical research, specifically for investigating epigenetic aging and Alzheimer's disease.