Prof. Ian Krajbich
Value-based decision making can be complex because many factors influence choices. For example, one might choose Item A over Item B because As are generally better, because As come with a bonus, or because you get more of A than of B. These factors would all manifest as increased choice probabilities for As over Bs despite arising from different considerations. Here, we use the diffusion model (HDDM) to identify patterns associated with these different kinds of biases, namely response, stimulus, and magnitude biases.In two experiments, subjects (n_1=67, n_2=89) first rated how much they would like to eat various snack foods, then made binary choices between those foods. We randomly designated options on one side of the screen as the targets. Our conditions were as follows:1) Proportion – the target side had higher value foods a majority of the time2) Quarter – the target side came with a 25-cent bonus3) Double – the target side had twice as much food4) Demand – we told subjects that we were testing if people favor options on the target sideOur results revealed an overall response bias in the Proportion condition, but not in the Demand condition; a stimulus bias in the Quarter condition; and a magnitude bias in the Double condition. We also investigated how response biases develop and disappear over time. In conditions where the target side was consistently better, starting-point biases increased over time; in conditions where the target and non-target side were equally attractive, starting-point biases decreased over time.
Eddy J. Davelaar
Dr. Brandon Turner
For a long time, choice and response time (choice-RT) have been the central behavioral measures used to explore the mechanism of choice. In addition to choice-RT, confidence has also been considered as a measure of behavioral judgment. Recently, many researchers have attempted to integrate those behavioral measures (choice, response time, and confidence) into the unified modeling framework (Poisson Race Model: Merkle and Van Zandt, 2006; RTCON: Ratcliff and Starns, 2009; 2DSD: Pleskac and Busemeyer, 2010). Here, we propose a way of modeling confidence with the leaky, competing, accumulator (LCA; Usher and McClelland, 2001). To do so, we rely on a simple solution to mapping the continuous states of evidence in the LCA with the relative balance of evidence hypothesis (Vickers, 1979). The competitive nature of the accumulation process in the LCA framework can produce continuous decision states, as an asymptotic accumulation and have different effects on accuracy, RT, and of course, confidence. In this study, we will investigate how the LCA can be accompanied by confidence and how the dynamic competitions between the accumulators with information leakage and lateral inhibition can affect confidence as a continuous state of the evidence. Simulation results show that the LCA can successfully account for all the main benchmarks of confidence modeling (Pleskac and Busemeyer, 2010).
Prof. Ian Krajbich
When making decisions, how people allocate their attention influences their choices. One empirical finding is that people are more likely to choose the option that they have looked at more. This relation has been formalized with the attentional drift-diffusion model (aDDM; Krajbich et al., 2010). However, options often have multiple attributes. Attention is also thought to govern the relative weighting of those attributes (Roe et al. 2001). However, little is known about how these two distinct features of the choice process interact; we still lack a model (and tests of that model) that incorporate both option and attribute-wise attention. Here, we propose a multi-attribute attentional drift-diffusion model to account for attentional discount factors on both options and attributes. We then use five eye-tracking datasets (two-alternative, two-attribute preferential tasks) from different choice domains to test the model. We find very stable option-level and attribute-level attentional discount factors across datasets, though non-fixated options are consistently discounted more than non-fixated attributes. Additionally, we find that people generally discount the non-fixated attribute of the non-fixated option in a multiplicative way, and so that feature is consistently discounted the most. Finally, we also find that gaze allocation reflects attribute weights, with more gaze to higher-weighted attributes. In summary, our work uncovers an intricate interplay between attribute weights, gaze processes, and preferential choice.
Stephanie M. Smith
Prof. Ian Krajbich
All else being equal, preference-based choices become faster as the values of the alternatives increase, a phenomenon we refer to as the overall-value effect. There are many competing explanations for the overall-value effect, with normative, mechanistic, and artefactual origins. Here, we examined these potential explanations, applying diffusion modeling to choice and response-time (RT) data from three difference choice domains (snack foods, abstract art, and conditioned stimuli). Within each experiment, we manipulated whether subjects knew the range of overall values in the upcoming block of trials. In each experiment, when subjects did not know the value of the upcoming trials, we observed the overall-value effect. The presence of the overall-value effect in the conditioned-stimulus experiment indicates that it is not due to artefacts from how we measure value, nor from familiarity with high-value items. Moreover, in most cases we also observed accuracy increasing with overall value, which rules out many of the mechanistic, noise-based explanations. When subjects knew the value of the upcoming trials, their response caution (i.e., boundary separation in the diffusion model) increased for high-value trials compared to middle-value trials, reflecting longer RT. This indicates that the overall-value effect is not normative since value-informed subjects display a reverse overall-value effect. Our results indicate that the overall-value effect is a robust phenomenon and is best explained as stronger evidence (i.e. higher drift rates) for higher value items.
In a repeated decisions task, subjects were asked to choose between minimizing their chances of losses and maximizing their chances of gains. After each trial, both losses and gains were simultaneously reported (even if 0). The probabilities of losses and gains were manipulated by the choices of the participants such that the expected value of net earnings was 0 regardless of choice, but negative if participants failed to make a choice in time. Multiple iterations of the study were conducted, manipulating factors such as payoff magnitudes, probabilities, and time limit. In addition to the primary task, additional measures were included as covariates such as for loss-aversion and behavioral approach/inhibition(BIS/BAS). Participants were compensated monetarily for their time by an amount partially dictated by their task earnings. The talk will overview the changing patterns of responses over the course of the task for the different conditions and their connection to the measured covariates.