mwheymans/psfmi: Predictor selection and model performance evaluation in multiple imputed datasets

Package contains functions to apply pooling or backward selection for logistic or Cox regression prediction models. Backward selection is done from the pooled model on basis of the Wald p-value. The model can contain continuous, dichotomous, categorical predictors, interaction terms between all type of these predictors. Continuous predictors can also be introduced as restricted cubic spline coefficients. For categorical and spline predictors, the overall p-value is obtained by pooling the total covariance matrix (D1 method), pooling Chi-square values (D2 method), pooling Likelihood ratio statistics (method of Meng and Rubin) or pooling the median p-values (MPR rule). For continuous and dichotomous variables Rubin's rules are used as the basic pooling method, although one of the four methods can also be chosen. The package also contains functions to generate apparent model performance measures over imputed datasets as ROC/AUC, Nagelkerke R-squares, Hosmer & Lemeshow test values and calibration plots. A wrapper function over Frank Harrell's validate function is used for that. Bootstrap internal validation is performed in each imputed dataset results are pooled. Backward selection as part of internal validation is optional and recommended.

Getting started

Package details

LicenseGPL (>= 2)
Package repositoryView on GitHub
Installation Install the latest version of this package by entering the following in R:
mwheymans/psfmi documentation built on Aug. 18, 2018, 7:38 p.m.