Multinomial Regression Versus Quantile

Abstract

With multinomial regression, instead of overall functions of all predictors, the predictors are restructured into stepped binary predictors. This produces a lot of significant p-values and the investigators may rapidly be at a loss to know how to interpret them. Moreover, some readers may reject the multiple p-value approach and consider it a case of plenty type I errors of finding differences, where there are none. However, currently, test statistics in regression analyses are no longer usually adjusted for multiple testing, because they are assumed to stem from a family of null hypotheses with many interactions within a single experiment. Therefore, adjustment as though they were entirely independent null hypotheses does not seem right. The p-values are not null hypotheses here, but rather kind of goodness of fit tests, something that makes you happy. We should add, that various closely related methodologies are available for analyzing categorical predictor variables. Multinomial regression is fine for predictors without a stepping function,

1 Supplementary Electronic Material (S)

chap17ordinalregression (SAV 3 kb)