A Comparison of Model-Building Strategies for Lower Respiratory Tract Infection in Long-Term Care

Abstract

Five strategies for creating predictive models of lower respiratory tract infection in residents of long-term care facilities were compared. A linear judgment model was derived by administering clinical vignettes to physicians who indicated the risk of infection based on the presence or absence of five predictor variables. A model based on physician consensus was created using the same variables. Three models based on empirical data (logistic regression, proportional hazards, and recursive partitioning) were created from a “derivation” sample of data from a cohort study of lower respiratory tract infections in nursing homes using the five predictor variables. All models were applied to a validation set and compared using receiver operating characteristic (ROC) curves. The data-derived and consensus models showed the highest discriminative ability while the linear judgment model showed inferior performance.