Blind computational reproduction of published statistical analyses with MerQur — Study 4: previously untested analyses and re-testing of corrected errors on independent data (preregistration)

> This document preregisters Study 4 of a blinded computational reproduction programme for MerQur, a free > statistical software package. Part A tests analyses not used in Studies 1-3 (tree-based learning, other > classifiers and performance measures, Bayesian regression, multiple imputation and bootstrap intervals, > multiple-response sets, complex survey totals and age standardisation). Part B re-tests, on independent data, > errors confirmed in the first round of Studies 1-2 and corrected in MerQur 2.3.2, using articles whose analyses > exercise the corrected code paths; Part B is reported separately and not pooled. Twenty-seven articles with > openly licensed raw data (CC0 or CC BY 4.0) were selected before registration. Their analyses will be re-run, > blind to the published results, by three independent analysts using MerQur 2.4.1 (frozen at code commit > b35ff79da354f93fb2b0d64f9ed6778ff4eb3e62), R and SAS, and every reported value will be classified with a > predefined decision tree. Study 4 follows the design and rules of Studies 1-3 (Zenodo > https://doi.org/10.5281/zenodo.24021910, https://doi.org/10.5281/zenodo.24021656, > https://doi.org/10.5281/zenodo.24021061). No analysis had been run on the study data at the time of this > deposit; SHA-256 hashes of the frozen analysis specifications are given in Appendix A.