Emotional valence and arousal in spoken word recognition: a hierarchical drift diffusion reanalysis of MALD

Abstract

Emotional valence and arousal in spoken word recognition: a hierarchical drift diffusion reanalysis of MALD This project asks whether a spoken word’s emotional valence and arousal affect how quickly listeners accumulate evidence that it is a word. We reanalyse the Massive Auditory Lexical Decision database (MALD; Tucker et al., 2019) for the 9,572 words studied by Guo et al. (2022), plus all nonwords, from 440 participants. Guo et al. fitted generalised additive models to mean log response times per word, with smooth terms for each predictor and tensor-product interactions between valence, arousal and frequency. Each word contributed one averaged observation, so their analysis generalises across words but averages over participants, response accuracy and the shape of the response-time distribution. We analyse every trial instead, with a hierarchical drift diffusion model that models each response time and choice jointly. It separates the rate of evidence accumulation (drift) from response caution, response bias and non-decision time. Drift is predicted by Guo et al.’s variables: valence and arousal (Warriner et al., 2013), spoken frequency (CELEX), bigram frequency, concreteness, contextual diversity and age of acquisition residualised on frequency, two principal components of length, duration and neighbourhood measures, part of speech, and initial phoneme voicing. Words and participants are crossed random factors, so conclusions generalise across both. Hypotheses. Pleasant words (H1a) and arousing words (H1b) are recognised with higher drift; valence and arousal interact (H2); and their interaction depends on frequency (H3, an extension, since Guo et al. modelled only two-way interactions). Established lexical effects should appear in their usual directions as a check on the model. Expected outcomes. Small but credible effects of valence and arousal on drift, much smaller than the frequency effect; no directional prediction for H3. Our predictors are linear, so non-linear effects detected with Guo et al.’s smooth terms cannot be reproduced. Where results differ, we discuss whether the difference reflects the model (drift versus mean RT), the unit of analysis (trials versus item means), or linearity. Transparency. The data already exist. Before registering, we had fitted a simpler model without item effects and observed the direction of the valence, arousal and valence × arousal effects; H1 and H2 therefore test the robustness of those effects to modelling item variability. Results from the primary model and from all model comparisons were not inspected before registration. The registration pre-specifies the primary model, inference criteria, sensitivity analyses, and a decision tree for convergence failures and model misfit. All analysis code is archived with checksums.