Not so SUBTLE(X): Word frequency estimates and their fit to sentential reading times in interaction with predictability

Frequency and predictability are two prominent psycholinguistic variables that determine the ease of word comprehension and have informed models of language processing. Here, we pooled the data from five self-paced reading studies to investigate (1) the usefulness of three well-known frequency databases of German in accounting for word reading times in context (i.e., the SUBTLEX-DE, CELEX, and dlexDB databases), and (2) whether frequency and predictability have additive or interactive effects on lexical processing. Regarding (1), goodness of fit comparisons between the three frequency measures showed that, in the majority of models, dlexDB frequencies performed best (in contrast to earlier investigations recommending to use SUBTLEX), even though nearly all frequency effects were statistically invariant and dwarfed by the contributions of other more potent variables such as predictability or trial number. Regarding (2), we found that, even though predictability influenced reading times, there was no evidence for interactive effects of frequency and predictability. Our results call into question the current default practice in many psycholinguistic studies to rely on subtitle norms when it comes to estimating lexical frequencies, but they also suggest that frequency effects may be negligible in paradigms which promote contextual word-by-word reading. Our findings are more in line with modular models of language comprehension in which lexical access operates independently from contextual predictability.