useR! 2024

In this presentation we show how to model data from the closed unit interval [0, 1] using extended-support beta regression and heteroscedastic two-limit tobit models. In contrast to zero- and/or one-inflated beta regression, both approaches only require estimation of a single latent process that captures both the distribution of the inner observations and the point masses for observations on the boundaries at 0 and/or 1. The heteroscedastic two-limit tobit model does so by fitting a Gaussian distribution censored at 0 and 1 which is conveniently available in the R package "crch". Extended-support beta regression has recently been proposed and implemented in the development version of the "betareg" package. It contains both classic beta regression and heteroscedastic two-limit tobit as special cases, shifting between the two with just one additional parameter. Both approaches are illustrated by modeling reading accuracy scores of children and investments in an economic loss aversion experiment, respectively, discussing the models' relative (dis)advantages.