psfmi: Prediction Model Selection and Performance Evaluation in Multiple Imputed Datasets

Provides functions to apply pooling or backward selection of logistic, Cox regression and Multilevel (mixed models) prediction models in multiply imputed datasets. Backward selection can be done from the pooled model using Rubin's Rules (RR), the D1, D2, D3 and promising median p-values method. The model can contain continuous, dichotomous, categorical predictors and interaction terms between all these type of predictors. Continuous predictors can also be introduced as restricted cubic spline coefficients. It is also possible to force (spline) predictors or interaction terms in the model during predictor selection. The package includes a function to evaluate the stability of the models using bootstrapping and cluster bootstrapping. The package further contains functions to generate pooled model performance measures in multiply imputed datasets as ROC/AUC, R-squares, Brier score, fit test values and calibration plots for logistic regression models. A function to apply Bootstrap internal validation is also available where two methods can be used to combine bootstrapping and multiple imputation. One method, boot_MI, first draws bootstrap samples and subsequently performs multiple imputation and with the other method, MI_boot, first bootstrap samples are drawn from each imputed dataset before results are combined. The adjusted intercept after shrinkage of the pooled regression coefficients can be subsequently obtained. Backward selection as part of internal validation is also an option. Also a function to externally validate logistic prediction models in multiple imputed datasets is available. Eekhout (2017) <doi:10.1186/s12874-017-0404-7>. Wiel (2009) <doi:10.1093/biostatistics/kxp011>. Marshall (2009) <doi:10.1186/1471-2288-9-57>.

Downloads:

Linking:

Please use the canonical form https://CRAN.R-project.org/package=psfmi to link to this page.