Bootstrapping and sample splitting for high-dimensional, assumption-lean inference (Q2284380)

In recent times, there has been a crisis in the sciences because too many research results are found to lack replicability and reproducibility. Some of this crisis has been attributed to a failure of statistical methods to account for data dependent extraction and modeling that precedes statistical inference [\textit{R. Berk}et al., ``Assumption lean regression'', Preprint, \url{arXiv:1806.09014}]. Valid inference after model selection is performing. An older method is sample splitting: use part of the data for model selection and the rest for inference. A new method is proposed for sample splitting combined with the bootstrap (or the normal approximation). New parameters that measure variable importance are defined and that can be inferred with greater accuracy than the usual regression coefficient.