Evaluating Logistic Mixed-Effects Models of Corpus-Linguistic Data in Light of Lexical Diffusion

Abstract

We explore methods for evaluating logistic mixed-effects models of both corpus and experimental data types through simulations. We suggest that the fit of the model should be evaluated by examining the variance explained by the fixed effects alone, rather than both fixed and random effects put together. Nonetheless, for corpus data, in which frequent items contribute more observations, coefficient estimates for fixed effects should be derived from a model that includes the random effects. Including random effects in the model with such datasets allows for better estimates of the fixed-effects predictor coefficients. Not having random effects in the model can cause fixed-effects coefficients to be overly influenced by frequent items, which are often exceptional in linguistic data due to lexical diffusion of ongoing changes.