Applying linear additive models to isolate component processes in task-evoked pupil responses

Abstract

The diameter of the eye’s pupil is constantly in flux. Decades of research have established a strong relationship between modulations of pupil diameter and internal mental processes. One of the major challenges in applied pupillometry research has been delineating the various factors that influence pupil diameter and isolating distinct mappings between mental processes and time-varying features of the task-evoked pupil response (TEPR). Here we applied a linear additive modeling (LAM) framework that posits the TEPR as a temporal sequence of transitions in arousal state reflecting task-dependent changes in attention and cognitive processes. We evaluated and compared LAM models on a longitudinal repeated-measures data set in which 26 subjects performed a 10-min psychomotor vigilance task (PVT) biweekly for 16 weeks (8 sessions per subject). PVT performance was captured by mean response time (rt) and lapse rate (proportion of rt’s>0.5 sec). The mean TEPR had a biphasic shape that was well-described by a LAM consisting of an early response (peaking between 0.5–1.0 sec post-stimulus-onset) and a broader late response (peaking between 1.25–3.0 sec), likely reflecting rapid orienting of attention followed by allocation of attentional resources to support decision making, respectively. There was, however, substantial variability in TEPR shape from subject-to-subject and session-to-session. We used multilevel linear models to examine the relationship between model-fitted TEPRs and performance measures at multiple levels including the group-level, subject-level, and individual-trial-level. We discovered two parameters of the fitted LAMs that explained unique variance in performance across all three levels including 1) peak amplitude of the early (orienting) response and 2) latency of the later (executive arousal) response. These two components represent a scale-invariant signature of performance. This work highlights the value of applying LAMs to identify component processes in TEPRs, and we explore applicability to other task domains.