8 - Comparing Bayesian and Frequentist Models of Language Variation

Summary

This chapter compares standard frequentist and more recent Bayesian approaches to logistic regression analyses. Starting out from a multifactorial case study of the verb help complemented by either the bare infinitive or the to-infinitive, the key components and the main conceptual differences of frequentist and Bayesian inference are discussed. Conceptually, the Bayesian rationale of directly testing hypotheses on the effects of multiple factors on an outcome variable is argued to be preferable and more sensitive than the conventional approach of testing null hypotheses. On the practical side, Bayesian statistics enables the researcher to recycle and integrate the results of previous analyses based on different datasets as informative priors, which can help improve and stabilize statistical modelling. Recourse to prior research can thus produce synergies and reduce data preparation expense. In cases of data sparsity, it can by the same token enable researchers to analyse small samples. Bayesian methods are thus put forward as powerful tools for overcoming the limitations of isolated corpus studies and for promoting synergies between data collected by individual researchers.