Portal de Periódicos da CAPES

Sequential effects: A Bayesian analysis of prior bias on reaction time and behavioral choice

2014; Wiley; Volume: 36; Issue: 36 Linguagem: Inglês

1551-6709

Shunan Zhang, He Huang, Angela J. Yu,

Behavioral Health and Interventions

Sequential effects: A Bayesian analysis of prior bias on reaction time and behavioral choice Shunan Zhang He Crane Huang Angela J. Yu (s6zhang, heh001, ajyu@ucsd.edu) Department of Cognitive Science, University of California, San Diego 9500 Gilman Drive, La Jolla, CA 92093-0515 Abstract Human subjects exhibit “sequential effects” in many psychological experiments, in which they respond more rapidly and accurately to a stimulus when it reinforces a local pattern in stimulus history, compared to when it violates such a pattern. This is often the case even if the local pattern arises by chance, such that stimu- lus history has no real predictive power, and therefore any behavioral adjustment based on these erroneous pre- dictions essentially amounts to superstition. Earlier, we proposed a normative Bayesian learning model, the Dy- namic Belief Model (DBM), to demonstrate that such be- havior reflects the engagement of mechanisms that iden- tify and adapt to changing patterns in the environment (Yu & Cohen, 2009). In that earlier work, we assumed a monotonic relationship between prior bias and response time (bias toward a stimulus was assumed to result in faster reaction time when that was the actual stimu- lus; conversely, when the other stimulus was present, it was assumed to result in a slower response). Here, we present a more detailed and quantitative analysis of the relationship between prior bias and behavioral out- come, in terms of response time and choice accuracy. We also present novel behavioral data, along with a frame- work for jointly identifying subject-specific parameters of the trial-by-trial learning (Dynamic Belief Model, DBM) and within-trial sensory processing and decision-making (Drift-Diffusion Model, DDM) based on the behavioral data. Our results provide strong evidence for DBM, and reveal potential individual differences, in their differen- tial beliefs about the timescale at which local patterns persist in sequential data. Keywords: Perceptual Decision Making; Se- quential Effects; Bayesian Model; Drift-Diffusion Model Introduction In a variety of behavioral experiments, human subjects display “sequential effects”, a modulation of response time and/or accuracy by recent trial history (e.g. Ber- telson, 1961; Laming, 1968; Kornblum, 1973; Soetens, Boer, & Hueting, 1985; Cho et al., 2002; Jones, Curran, Mozer, & Wilder, 2013). For example, in two-alternative forced choice experiments, in which subjects discriminate between two types of stimuli (A or B), subjects respond more accurately and rapidly if a trial is consistent with the recent pattern (e.g. AAAAA followed by A, ABABA followed by B), than if it is inconsistent (e.g. AAAAA followed by B, ABABA followed by A). This sequen- tial effect depends on the length of the run (Cho et al., 2002). For instance, an alternation following four repe- titions affects responding more than one following only two repetitions. Figure 1 illustrates a robust finding of the dependence of RT and error rate on recent trial his- tory, both being largest when a relative long run of rep- etitions or alternations are broken by the current obser- vation (middle two trial types), and smallest when such runs are extended (left and right end). Previously, we proposed a Bayesian learning model, the Dynamic Belief Model (DBM), to account for sequential effects, via a human learning mechanism that assumes the potential for discrete, un-signaled changes in the environ- ment. Consequently, DBM repeatedly modifies internal estimates of the relative probability of one stimulus type versus another occurring, based on recent stimulus his- tory (Yu & Cohen, 2009). By assuming reaction time and error rate to be monotonically and inversely corre- lated with the estimated prior probability of observing the actual stimulus prior to stimulus onset, DBM can qualitatively reproduce the empirically observed sequen- tial effects shown by Cho et al. (2002). In this work, we give a more precise and quantitative treatment of the influence of prior expectations on sen- sory processing and decision-making within a trial, by as- suming an evidence-integration-to-bound process (Gold, 2002), which is formally similar to the Drift-Diffusion Model (DDM) (e.g. Bogacz, Brown, Moehlis, Holmes, & Cohen, 2006) and appears to explain activities of pari- etal cortical neurons during primate perceptual decision- making (Roitman & Shadlen, 2014). We present a Bayesian method for simultaneously identifying subject- specific parameters of DBM and DDM based on an in- dividual’s choice accuracy and reaction times, and apply it to behavioral data collected in a simple 2-alternative choice perceptual discrimination task. Using this quan- titative method, we will compute the relative support, measured in Bayes factors, the data lend to DBM versus a competing model, the Fixed Belief Model (FBM) (Yu & Cohen, 2009), which assumes that human subjects do not believe the task statistics to be changeable over time. We will also characterize the population distributions of subject-specific Bayesian model parameters, which cor- respond to semantically readily interpretable variables, such as subjects’ beliefs about the rate of change in the environmental statistics, the overall relative frequency of repetition and alternation trials, and the subjective diffi-