We would like to bring to your attention a new R software package available from CRAN for doing Bayesian analyses that are free from making parametric assumptions about measurement error. This package is called DFBA, which stands for Distribution-Free Bayesian Analyses, and it was developed by Dan Barch and Rich Chechile. The package provides a set of functions for performing distribution-free Bayesian analyses. The term distribution-free is deliberately used to differentiate the Bayesian analyses in the DFBA package from the more complex modeling that has come to be called Bayesian nonparametric models. The distribution-free methods focus on the rank and categorical information in the data. The package contains some novel procedures that emerged from the relatively recent statistical literature as well as some more well-known Bayesian methods. Included in the DFBA package are Bayesian analogues to the frequentist Mann-Whitney U test, the Wilcoxon signed-ranks test, the Kendall tau rank correlation, the Goodman-Kruskal gamma, the McNemar test, the binomial test, the sign test, the median test, as well as distribution-free methods for testing contrasts among conditions and for computing Bayes factors for hypotheses. The package also includes procedures to estimate the power of distribution-free Bayesian tests based on data simulations from various probability models for the data. The DFBA functions provide users with a versatile and readily understandable set of Bayesian procedures that avoid making parametric assumptions about measurement error and are robust to the problem of extreme outlier scores.
The package website can be found at the following link:
The website also provides information about a set of 14 vignettes that discuss the various functions in the package as well as a manual of the R documentation. The vignettes and documentation are designed to be informative for both advanced statistical users as well as for scientists who are unfamiliar with the Bayesian approach to statistical inference and how it differs from the orthodox relative-frequency perspective. The package is designed to be easy to use and easy to interpret regardless of the level of the user’s statistical experience. In most cases the users simply can enter their data and get posterior estimates of the unknown population parameter under investigation along with the corresponding probabilities and Bayes factors for the pertinent hypotheses. The prior distribution for most analyses is set to a noninformative default, but that prior can be easily adjusted by the user to their preferred prior. The users are also provided with other options so as to better tailor the analyses to fit their needs.
Richard Chechile and Daniel Barch
Tufts University