Evidence accumulation in a complex task: Making choices about concurrent multiattribute stimuli under time pressure.

Abstract

Evidence accumulation models transform observed choices and associated response times into psychologically meaningful constructs such as the strength of evidence and the degree of caution. Standard versions of these models were developed for rapid (∼1 s) choices about simple stimuli, and have recently been elaborated to some degree to address more complex stimuli and response methods. However, these elaborations can be difficult to use with designs and measurements typically encountered in complex applied settings. We test the applicability of 2 standard accumulation models-the diffusion (Ratcliff & McKoon, 2008) and the linear ballistic accumulation (LBA) (Brown & Heathcote, 2008)-to data from a task representative of many applied situations: the detection of heterogeneous multiattribute targets in a simulated unmanned aerial vehicle (UAV) operator task. Despite responses taking more than 2 s and complications added by realistic features, such as a complex target classification rule, interruptions from a simultaneous UAV navigation task, and time pressured choices about several concurrently present potential targets, these models performed well descriptively. They also provided a coherent psychological explanation of the effects of decision uncertainty and workload manipulations. Our results support the wider application of standard evidence accumulation models to applied decision-making settings.

Adapting to workload in complex environments: Insights from evidence accumulation models of decision making

“…Our manipulation of the number of stimuli, although analogous to Palada et al (2016), used trials with a predetermined number of ships, rather than a probabilistic manipulation to draw the number of ships presented in a 3-minute period. Regardless of this difference, like Palada et al (2016), we found that participants did not respond to the number of stimuli by changing their rate of information processing. Rather, the number of stimuli selectively influenced threshold; participants strategically lowered their threshold when more ships were present, prioritizing the speed over the accuracy of their choices to reduce non-responses.…”

“…They also observed that non-decision time decreased when participants had tighter deadlines. Palada et al (2016) extended this approach to a more complex task, testing whether the LBA was capable of accounting for performance on a simulation involving a navigation and target classification task that is more representative of the types of tasks found in applied settings. They found that the LBA was able to provide an accurate characterization of all aspects of performance even though multiple stimuli were simultaneously present in the display, and participants were free to decide how they allocated their attention amongst tasks and stimuli.…”

“…A second limitation is that Palada et al (2016) used a range of randomly varying stimulus configurations that differed markedly with respect to the difficulty of the target vs. nontarget classification. The hypothesis that the heterogeneity of stimulus difficulty would be accounted for by the rate variability parameter was supported, with target variability being greater than nontarget variability.…”

“…The simplest model was fit first and thereafter, the best-fitting values informs the optimization algorithm start-point for the next model variant where one extra factor is introduced. This method continues up until the top model is fitted, with more complex models fit from many different start points, and reduces the chance of the fitting procedure finding sub-optimal solutions (Donkin et al, 2011; for examples, see Heathcote & Love, 2012;Palada et al, 2016;Rae et al, 2014).…”

“…We also assumed that non-decision times did not differ between stimulus types, corresponding with the assumption that encoding and response production times do not differ between targets and nontargets. Further, we assumed that time pressure manipulations would not influence rate variability (sv), though we did allow rate variability to be a function of the match factor (for examples using the same approach see Palada et al, 2016;Strickland, Heathcote, Remington, & Loft, 2017). We estimated a single value for the start point parameter (A) for all conditions.…”