Individual differences in spillover during naturalistic reading

Spillover---the influence of previous words on the processing of current words---is widely reported in reading studies, yet population-level findings are mixed. We argue that spillover is neither noise nor a uniform phenomenon; it may reflect systematic individual differences in how processing difficulty carries over across words. We analyze six naturalistic eye-tracking-while-reading datasets and estimate individual-level spillover effects of preceding-word length, frequency, and surprisal using Bayesian hierarchical models. Because studying individual differences requires reliable measurement, we first assess the cross-context reliability of these effects across experimental sessions and reading tasks, operationalized as correlations between individual-level estimates from two measurement contexts. The measured reliability differs across sources and depends on experimental design and signal-to-noise ratio, being strongest for word length and weaker for lexical frequency and surprisal. Crucially, we find that individual differences in spillover vary systematically with reading fluency. Low-fluency readers exhibit greater spillover from low-level lexical properties, such as word length and frequency, whereas high-fluency readers show more pronounced spillover from higher-level contextual difficulty, such as surprisal, which may help explain why population-level spillover findings are mixed. Finally, spillover is related to skipping and regression rates, indicating that some spillover effects may reflect idiosyncratic patterns of gaze behavior. Overall, our findings suggest that population-level effects can obscure divergent reader profiles, motivating theories and models of reading beyond the "average reader".