xv != xi: Cross-validation in Stata

Evaluating the out-of-sample properties of statistical models is important, especially for predictive modeling/analytics. Although Stata currently implements cross-validation methods natively for some model-fitting commands—dslogit, dspoisson, dsregress, elasticnet, lasso, poivregress, pologit, popoisson, poregress, sqrtlasso, xpoivregress, xpologit, xpopoisson, and xporegress—broader use of cross-validation is not natively supported. At last year’s conference, a user explained the challenges that students and new users face when trying to use cross-validation procedures in Stata. While it is possible to implement the four-step process of splitting the sample, fitting the model to the training sample, predicting outcomes on the validation/test sample, and computing metrics related to the fit, doing so is tedious and time consuming. Developing a program that implements the four-step process above is not a trivial task, despite what one of the authors initially thought. In this talk, we present xv, an extensible prefix command implementing cross-validation for Stata estimation commands.