Pupil data analysis

Baseline pupil diameter in each trial was calculated as averaged pupil trace 1s before the sentence onset. The pupil diameter measured from the sentence onset to the repeat prompt was subtracted from that baseline level to obtain relative pupil diameter changes elicited by the task. In Additional analyses, to examine whether the impact of luminance was independent of the PPD baseline-correction method, we attempted a posteriori two alternative ways of calculating PPD, as a percentage of the baseline, or as a percentage of the dynamic range. In all these methods, the search for PPD was restricted to a window starting from sentence onset and ending at the verbal response prompt (thereby excluding any confound with the pupil’s arousal induced by verbally repeating the sentence). Note that all sentences were left unprocessed in duration to avoid unnatural acoustic manipulation (compression or stretching of original sentences). This procedure was common in listening effort studies due to varied length of standardized sentences and the variability in sentence duration could be well controlled by consistent trace alignment (either by sentence onset or offset) [33]. Additional analysis took into account of this variability in sentence duration and alignment method and explored whether they affected our results.

A linear mixed-effect model (LME] was fitted on three dependent variables successively: baseline, PPD amplitude, and PPD latency. Each of these metrics was extracted on the trace averaged over a block (20 sentences), so there was only one observation per condition limiting the model complexity to random intercepts at most. Thus, just like for the behavioral data, we considered a model with one fixed factor in Exp.1 (SNR, as categorical variable) or two fixed factors in Exp.2 (SNR and luminance, both as categorical) and random intercepts by subject. Further analyses are presented (S1 Appendix) where we considered a trial-based approach that allowed us to examine more complex models with by-subject random slopes for each fixed factor and look specifically at the effect of position within a block. Chi-squared tests were used to examine main effects and possible interactions.