The most under-used statistical method in corpus linguistics: multi-level (and mixed-effects) models

Abstract

Much statistical analysis of psycholinguistic data is now being done with so-called mixed-effects regression models. This development was spearheaded by a few highly influential introductory articles that (i) showed how these regression models are superior to what was the previous gold standard and, perhaps even more importantly, (ii) showed how these models are used practically. Corpus linguistics can benefit from mixed-effects/multi-level models for the same reason that psycholinguistics can – because, for example, speaker-specific and lexically specific idiosyncrasies can be accounted for elegantly; but, in fact, corpus linguistics needs them even more because (i) corpus-linguistic data are observational and, thus, usually unbalanced and messy/noisy, and (ii) most widely used corpora come with a hierarchical structure that corpus linguists routinely fail to consider. Unlike nearly all overviews of mixed-effects/multi-level modelling, this paper is specifically written for corpus linguists to get more of them to start using these techniques more. After a short methodological history, I provide a non-technical introduction to mixed-effects models and then discuss in detail one example – particle placement in English – to show how mixed-effects/multi-level modelling results can be obtained and how they are far superior to those of traditional regression modelling.