Selective inference for effect modification via the lasso.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Effect modification is crucial for decision-making but challenging with numerous covariates.
• Selecting appropriate models for effect modification is essential for valid statistical inference.
• Existing methods may lack interpretability or lead to excessive false discoveries.

Purpose of the Study:

• To propose a novel two-stage procedure for selecting parsimonious effect modification models.
• To enable valid statistical inference on selected effect modification models.
• To enhance the interpretability and reduce false discoveries in effect modification analysis.

Main Methods:

• Utilizes Robinson's transformation to isolate treatment effects from nuisance parameters.
• Employs machine learning algorithms for estimating nuisance parameters.
• Applies the least absolute shrinkage and selection operator (LASSO) for parsimonious model selection.

Main Results:

• The proposed method yields a more interpretable model compared to full covariate models.
• Significantly reduces false discoveries compared to univariate subgroup analyses.
• Demonstrates asymptotic validity of conditional selective inference under specified assumptions.

Conclusions:

• The two-stage procedure effectively identifies parsimonious effect modification models.
• The method offers a robust approach for statistical inference in high-dimensional settings.
• Validated through simulations and an epidemiological application, highlighting practical utility.