PPubMed3 Nov 2020Cited 3×

Variable selection methods for identifying predictor interactions in data with repeatedly measured binary outcomes

Bethany J. Wolf, Yunyun Jiang, Sylvia H. Wilson, Jim C. Oates

Abstract

It has been hypothesized that many common diseases result from interactions among genetic, clinical, and environmental factors . Identifying predictors of patients’ disease status may necessitate modeling interactions among predictors . Clinical studies also often evaluate patient outcomes over time requiring methods that can account for within-subject correlation. Generalized linear mixed model (GLMM) and generalized estimating equations (GEEs) address the correlation between repeated measures collected on a patient within a linear model framework . GLMMs incorporate a subject-specific random

A figure from Variable selection methods for identifying predictor interactions in data with repeatedly measured binary outcomes — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

Penalized and boosted approaches are effective for variable selection in data with clustered binary outcomes.

What’s new

i.Evaluated stepwise selection, penalized GLMM, boosted GLMM, and boosted GEE for variable selection with repeatedly measured binary outcomes
ii.Demonstrated that penalized and boosted methods recover true predictors and interactions more often than stepwise selection
iii.Introduced and assessed a two‑stage bias‑reduction approach applicable to penalized and boosting methods
iv.Provided practical guidance for choosing among glmmLasso, GMMBoost, and GEEBoost in real applications

How they did it

The authors conducted simulation studies comparing stepwise selection, penalized GLMM (glmmLasso), boosted GLMM (GMMBoost), and boosted GEE (GEEBoost) for identifying main effects and two‑way interactions in clustered binary data. They also evaluated a two‑stage approach that refits a traditional GLMM using predictors selected in the first stage to reduce bias. The methods were applied to two real clinical data sets: hypothermia during surgery and treatment response in lupus nephritis.

What they found

glmmLasso selected 18 predictors, GMMBoost 4, and GEEBoost 6 in the hypothermia data set
False discovery rate (FDR) for glmmLasso ranged 65‑81% and for GEEBoost 68‑88% across simulations
GMMBoost achieved FDR = 0 for large sample sizes
Boosted GEE was the fastest computationally among the three sparse methods
Two‑stage bias correction reduced bias and error for all penalized and boosting methods

Caveats

✕glmmLasso models often included many spurious predictors, leading to high false‑discovery rates
✕GMMBoost required substantially longer computation time, preventing simulations at n=500 with 10 observations per subject
✕Penalized and boosting approaches produced regression estimates biased toward zero
✕Stepwise selection performed poorly in recovering true predictors and interactions

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Variable selection methods for identifying predictor interactions in data with repeatedly measured binary outcomes

The paper, in five minutes

Ask a question about this paper

Similar papers in all four archives

Similar papers in all four archives

The paper, in five minutes