Online Program

Logistic Ridge Regression for Domain Importance Assessment in Patient-Reported Outcomes Studies

*Tolulope T Sajobi, University of Calgary 
Pooneh Pordeli, University of Calgary 
Wilfrid Kouokam, Université de Bretagne Sud, France 
Lisa M Lix, University of Manitoba 

Keywords: variable importance, logistic ridge regression, patient-reported outcomes, inflammatory bowel disease

Variable importance measures based on logistic regression (LR) model have been proposed to rank order patient-reported outcome (PRO) domains according to their ability to discriminate between independent groups (e.g., treatment and control). But these measures are sensitive to strong domain correlations, which may result in incorrect ranking of PRO domains. This study investigates importance measures based on logistic ridge regression (LRR) for ranking correlated PRO domains. The investigated importance measures include standardized logistic regression coefficients, Pratt’s index, dominance analysis, and relative weight analysis. The accuracy of measures based on LRR and LR models were compared using Monte Carlo methods. Variable importance measures based on LRR resulted in more than a 10% increase in any-variable and per-variable correct ranking percentages than measures based on LR for moderately/strongly correlated PROs. The Short Form (36) Health Survey (SF-36) data from the Manitoba Inflammatory Bowel Disease (IBD) study, a longitudinal PRO study of persons diagnosed with IBD, is used to demonstrate the recommended methods.