UC Davis

Missing data is a ubiquitous problem in predictive modeling, but the majority of current methods are designed for statistical estimation rather than prediction, leaving the comparative behavior of imputation-based and non-imputational strategies in prediction-oriented settings poorly understood. This paper introduces 𝚖𝚟𝙿𝚛𝚎𝚍, an open-source R program for prediction under missing data in supervised regression. The package includes three complementary paradigms: imputation-based prefilling (𝚕𝚖_𝚙𝚛𝚎𝚏𝚒𝚕𝚕), pairwise available-case estimation (𝚕𝚖_𝚊𝚌) and the Tower technique (𝚕𝚖_𝚝𝚘𝚠𝚎𝚛), which in turn is based on the Tower Property of conditional expectation and aims to predict accurately. 𝚕𝚖_𝚙𝚛𝚎𝚏𝚒𝚕𝚕 supports four techniques, including complete-case analysis, MICE, Amelia and the missForest algorithm, ranging from list-wise elimination to model-based and nonparametric imputation approaches. We perform a rigorous empirical evaluation on five benchmark datasets with different sample sizes, missingness rates and properties of the response variable, using 5-fold cross-validation and four conventional predictive metrics. Our evaluation shows that the performance of a method depends heavily on the dataset characteristics and no single approach universally dominates all dataset combinations, thus motivating the necessity for a uniform benchmarking methodology. Practical guidelines for selecting a method depending on dataset features are also presented.The package is publicly available at https://github.com/matloff/mvPred.