Comments on Pardoe and Weitner: sentencing convicted felons in the United States: a Bayesian analysis using multilevel covariates

Abstract

This paper presents a regression analysis of the factors related to sentencing, which is obviously a very complicated and important problem. Iwillmake somevery general remarks about the paper, because I consider it to be a representative example of a general approach to statistical analysis. As an example, it looks quite good, although I am not an expert in sentencing. But neither are the authors, I think. Most important, from my point of view, is that the general approach is deeply flawed. From the methodological point of view, the paper uses techniques which differ from logistic regression analysis popular in criminology. It replaces logistic regression by multilevel logistic regression, and, as a further elaboration, it replaces the conventional multilevel analysis ofWong and Mason (1985) by a fully Bayesian approach. The question we have to answer is simple: does this analysis improve on standard logistic regression analysis? And can we expect the results of the analysis to be more reliable and interesting? Unfortunately, there are several problems with the paper, or more precisely with this type of statistical analysis, that make it difficult to give a positive answer to these questions. Why do Pardoe et al. (and many other social and behavioral scientists with similar data structures) argue that it makes sense to use multilevel analysis in this context? If we look in the multilevel literature for a general answer, we are in for some disappointment. As in various other social science statistics contexts, arguments are often replaced by references to presumably unassailable expert sources. We see this in an early discussion of factor analysis, in the literature around LISREL and structural equation models, and now again in multilevel analysis. Pardoe et al. maintain, for instance, that conventional logistic regression