Does language-as-used fit a self-paced reading paradigm? (The answer may well depend on the statistical model you use).

We report on a self-paced reading experiment that was run to ascertain whether the effect of differential tense, aspect and mood (henceforth TAM) marking on verbs would affect processing. TAM properties were identified as the strongest predictors for the choice between 6 near synonyms meaning TRY in Russian on the basis of regression models fit to manually annotated corpus data (Divjak 2010, Divjak & Arppe 2013). We will discuss how we used a Generalized Linear Mixed Model to account for the fact that we deviated from the traditional set-up for self-paced reading in two ways: we used an imbalanced design and ran the task with actually attested sentences rather than artificially created ones. These deviations were motivated by the need to accommodate the natural restrictions on TAM combinations and to respect the lack of a strict word order, which are both typical for Russian. We will also describe how we used a Generalized Additive Model to handle the non-linearities that we encountered in the reading times data.