Upscaling the uncertainty-based approach to very large predictor sets

Investigate how to upscale the Fisher’s unit information matrix–based uncertainty calculation for individual-level predictions to settings with a very large number of predictors and develop practical methodology to maintain valid uncertainty quantification in high-dimensional models.

Background

The proposed method computes individual-level uncertainty using a decomposition involving Fisher’s unit information matrix for logistic regression with a specified core set of predictors. While effective for modest numbers of predictors, the authors note practical and methodological challenges in extending the approach to very high-dimensional settings.

They explicitly call for further research to generalize and scale their uncertainty-based sample size methodology to models with many predictors.