Learning to Make Selection and Detection Decisions: The Roles of Base Rate and Feedback

The effects of type of feedback and base rate on threshold learning in a multiple‐cue decision task were examined. In most such decision experiments, participants receive feedback after every trial (full feedback), and a single base rate (usually 0.5) is used. Our experiment explored conditional feedback (feedback only after positive decisions) representing common selection and detection tasks (such as hiring), where the decision maker receives no feedback unless the decision is positive (e.g., hire the applicant). We used three base rates (0.2, 0.5, and 0.8). As expected, performance was best in full feedback, but after 300 learning trials, the difference was small. Conditional feedback generally resulted in fewer positive decisions than full feedback, but this difference was not found in the low (0.2) base rate condition. There were interactions between base rates and types of feedback. Results provide partial support for the constructivist encoding hypothesis of Elwin and colleagues. Simulation results suggest that our results may reflect overconfidence when feedback is not given. With respect to rate of learning, when the base rate was 0.2, conditional feedback participants reached approximately the same selection rate but did so more slowly than the full feedback participants. Partial feedback participants learned slower and appeared to be still learning after 500 trials. When the base rate was 0.5 or 0.8, partial feedback was nearly as good as full feedback, but conditional feedback resulted in a systematically lower rate of positive decisions. Copyright © 2011 John Wiley & Sons, Ltd.     

An extension to the constructivist coding hypothesis as a learning model for selective feedback when the base rate is high

Elwin, Juslin, Olsson, and Enkvist (2007) and Henriksson, Elwin, and Juslin (2010) offered the constructivist coding hypothesis to describe how people code the outcomes of their decisions when availability of feedback is conditional on the decision. They provided empirical evidence only for the .5 base rate condition. This commentary argues that the constructivist coding hypothesis imposes an ever-declining selection rate and overestimates base rate bias for high base rate conditions. We provide support based on a simulation model of learning under selective feedback with different base rates. Then we discuss possible extensions to constructivist coding that can help overcome the problem.     

Paradoxical effects of base rates and representation in category learning

The inverse base rate effect (IBRE) continues to be a puzzling case of decision making on the basis of conflicting information in human category learning. After being trained via feedback over trials to assign combinations of cues to high- and low-frequency categories, participants tend to respond with the low-frequency category to an otherwise perfectly conflicting pair of test cues, contrary to the category base rates. Our Experiment 1 demonstrated that decision making on the basis of an explicit summary of the cue-outcome and outcome base rate information from the standard learning task does not result in the effect. The remaining experimental conditions evaluated the necessary and sufficient conditions for the effect by systematically exploring experimental deviations between the standard learning task and the pure decision-making task. In partial disagreement with both recent theoretical accounts of the effect, these experiments indicate that asymmetric outcome representation and profound base rate neglect are individually necessary and jointly sufficient conditions. Broader theoretical implications are discussed.     

The decision to check in multialternative choices and limited sensitivity to default

Given the rapid proliferation of smartphone applications and data aggregation websites, in many situations people can use decision aids to guide their choices. For example, they may consider whether to use a navigation device to check the fastest route or whether to use a price comparison website to find the cheapest deal. In what circumstances will subjects use a costly comparison decision aid (which I refer to as “checking”) to choose for them? In six studies, I investigate the impact of the number of available alternatives and checking's attractiveness on the decision to check. While at first increasing the attractiveness of checking led to higher checking rates, a further increase in the number of available alternatives (and thus checking's attractiveness) did not increase the checking rate. Surprisingly, even when checking had a 40% higher expected value compared with not checking, the observed checking rate was below 45%, contrary to risk and ambiguity aversion predictions. Furthermore, labeling the checking alternative as the default had no impact on its choice rate. I find large individual differences in decisions to check. Surprisingly, subjects' initial decisions had high predictive power over their subsequent checking rates, even after 100 trials with full feedback. I propose two simple learning models that capture well the aggregated results.     

Affective state influences perception by affecting decision parameters underlying bias and sensitivity

Studies of the effect of affect on perception often show consistent directional effects of a person's affective state on perception. Unpleasant emotions have been associated with a "locally focused" style of stimulus evaluation, and positive emotions with a "globally focused" style. Typically, however, studies of affect and perception have not been conducted under the conditions of perceptual uncertainty and behavioral risk inherent to perceptual judgments outside the laboratory. We investigated the influence of perceivers' experienced affect (valence and arousal) on the utility of social threat perception by combining signal detection theory and behavioral economics. We compared 3 perceptual decision environments that systematically differed with respect to factors that underlie uncertainty and risk: the base rate of threat, the costs of incorrect identification threat, and the perceptual similarity of threats and nonthreats. We found that no single affective state yielded the best performance on the threat perception task across the 3 environments. Unpleasant valence promoted calibration of response bias to base rate and costs, high arousal promoted calibration of perceptual sensitivity to perceptual similarity, and low arousal was associated with an optimal adjustment of bias to sensitivity. However, the strength of these associations was conditional upon the difficulty of attaining optimal bias and high sensitivity, such that the effect of the perceiver's affective state on perception differed with the cause and/or level of uncertainty and risk.     

The Impact of Asymmetric Partial Feedback on Response‐Bias

For many decisions we make, we are more likely to receive feedback about some outcomes than about others, and thus we must operate under conditions of asymmetric partial feedback. In this paper, we examine the effects of asymmetric partial feedback on decision making across three experiments. Experiment 1 examines the effects of four partial feedback schedules on decision making in a single‐attribute auditory detection task. Experiment 2 examines these same schedules in the context of a simulated experiment in which virtual participants perform a multi‐attribute choice task. Experiment 3 then examines effects of one partial feedback schedule representative of a real‐world decision‐making domain (personnel selection) using an ecologically informed multi‐attribute choice task. All three experiments reveal that asymmetric partial feedback primarily impacts on response‐bias, a parameter of signal detection theory. These findings can be explained in terms of decision makers' attempts to maximize accuracy and the fact that shifts in response‐bias will, under partial feedback, produce misleading impressions of greater accuracy. Copyright © 2013 John Wiley & Sons, Ltd.     

What is learned from experience in a probabilistic environment?

Three experiments explored what is learned from experience in a probabilistic environment. The task was a simulated medical decision‐making task with each patient having one of two test results and one of two diseases. The test result was highly predictive of the disease for all participants. The base rate of the test result was varied between participants to produce different inverse conditional probabilities of the test result given the disease across conditions. Participants trained using feedback to predict a patient's disease from a test result showed the classic confusion of the inverse error, substituting the forward conditional probability for the inverse conditional probability when tested on it. Additional training on the base rate of the test result did little to improve performance. Training on the joint probabilities, however, produced good performance on either conditional probability. The pattern of results demonstrated that experience with the environment is not always sufficient for good performance. That natural sampling leads to good performance was not supported. Further, because participants not trained on joint probabilities did, however, know them but still committed the confusion of the inverse error, the hypothesis that having joint probabilities would facilitate performance was not supported. The pattern of results supported the conclusion that people learn all the necessary information from experience in a probabilistic environment, but depending upon what the experience was, it may interfere with their ability to recall to memory the appropriate sample set necessary for estimating or using the inverse conditional probability. Copyright © 2004 John Wiley & Sons, Ltd.     

The assumed effects of positive feedback paired with success on motivation to do a task: The cases of characters with high and low initial levels of interest

This study investigated the effects of positive feedback paired with success on motivation to do a task, and examined whether this differed according to whether the person performing the task had a high or low initial level of interest in that task. Subjects (448 junior college students) were asked to rate (from 0–10) the level of motivation of characters in four different stories. Subjects believed that positive feedback would encourage the character to do a task regardless of initial level of interest. However, when positive feedback could not be expected, ratings for motivation fell. Ratings for motivation were further analyzed when no positive feedback could be expected, either after an earlier experience of positive feedback paired with success or after success only (no feedback). In the case of a low initial level of interest, it was assumed that motivation would be increased after positive feedback, but in the case of a high initial level of interest motivation was assumed to be decreased by an earlier experience of positive feedback paired with success. These results suggest that the effect of positive feedback paired with success will differ according to the level of a person's initial interest.     

Optimal classifier feedback improves cost-benefit but not base-rate decision criterion learning in perceptual categorization

Unequal payoffs engender separate reward- and accuracy-maximizing decision criteria; unequal base rates do not. When payoffs are unequal, observers place greater emphasis on accuracy than is optimal. This study compares objective classifier (the objectively correct response) with optimal classifier feedback (the optimal classifier's response) when payoffs or base rates are unequal. It provides a critical test of Maddox and Bohil's (1998) competition between reward and accuracy maximization (COBRA) hypothesis, comparing it with a competition between reward and probability matching (COBRM) and a competition between reward and equal response frequencies (COBRE) hypothesis. The COBRA prediction that optimal classifier feedback leads to better decision criterion leaning relative to objective classifier feedback when payoffs are unequal, but not when base rates are unequal, was supported. Model-based analyses suggested that the weight placed on accuracy was reduced for optimal classifier feedback relative to objective classifier feedback. In addition, delayed feedback affected learning of the reward-maximizing decision criterion.     

Alternative antecedents, probabilities, and the suppression of fallacies in Wason's selection task

Three experiments examined the influence of a second rule on the pattern of card selections on Wason's selection task. In Experiment 1 participants received a version of the task with a single test rule or one of two versions of the task with the same original test rule together with a second rule. The probability of q was manipulated in the two-rules conditions by varying the size of the antecedent set in the second rule. The results showed a significant suppression of q card and not-p card selections in the alternative-rule conditions, but no difference as a function of antecedent set size. In Experiment 2 the size of the antecendent set in the two-rules conditions was manipulated using the context of a computer printing double-sided cards. The results showed a significant reduction of q card selections in the two-rules conditions, but no effect of p set size. In Experiment 3 the scenario accompanying the rule was manipulated, and it specified a single alternative antecedent or a number of alternative antecedents. The q card selection rates were not affected by the scenario manipulation but again were suppressed by the presence of a second rule. Our results suggest that people make inferences about the unseen side of the cards when engaging with the task and that these inferences are systematically influenced by the presence of a second rule, but are not influenced by the probabilistic characteristics of this rule. These findings are discussed in the context of decision theoretic views of selection task performance (Oaksford & Chater, 1994).     

Rapid-response versus free-time selection tasks using different logical connectives

When reasoning with conditional statements (i.e., if [not] p then [not] q), people tend to display matching bias: Options that match the entities named in the rule are selected even when logically inappropriate. Three different Wason selection tasks were administered under free-time and rapid-response formats. For the latter, individual cards were presented for one second, and required a response within a further one second. Previous research using these formats (Roberts & Newton, 2001) has shown that this increases matching bias, in line with the action of preconscious heuristic processes which direct attention towards relevant aspects of a problem, but whose action can be overturned if there is sufficient time to apply analytic reasoning processes. The selection tasks administered included a standard abstract conditional task, a disjunctive version (i.e., either [not] p or [not] q), and a conditional task in which the cards showed explicitly negated values. Both conditional tasks demonstrated matching bias, but under rapid-response presentation, matching bias only increased for the standard conditional and disjunctive tasks. Overall, the data support Evans’ (e.g., 2006) heuristic-analytic framework albeit with some caveats, and it is suggested that the broad question, of whether individual selection task formats show or do not show matching bias, requires more detailed investigation.     

Responses to false physiological feedback in individuals with panic attacks and elevated anxiety sensitivity

Participants with elevated anxiety sensitivity and a history of panic attacks were compared to a low anxiety comparison group with respect to physiological and subjective reactivity to false heart-rate feedback and reactivity to a priming procedure. Whereas accurate heart-rate feedback elicited minimal responses, participants across groups showed significant physiological and subjective responses to false feedback. High risk and low risk participants did not differ in heart-rate responses to false feedback, though panic attack frequency did predict physiological and subjective reactions to false feedback in the high risk group. Self-reported nonspecific anxiety was significantly higher in high risk female participants than in low risk female participants, while males did not different in general subjective anxiety. However, high risk participants reported more panic-specific symptoms during the false feedback task than low risk participants, regardless of the sex of the participant. Therefore, although the experimental paradigm appeared to trigger nonspecific anxiety in high risk female participants, panic attack symptoms in reaction to the task were specific to risk group, not sex, and consistent with hypotheses. Surprisingly, the priming procedure did not influence physiological or subjective responses to false feedback in either group. These results raise additional questions regarding the process and impact of interoception in individuals with panic attacks, and suggest that false perception of internal changes may contribute to risk for panic disorder when exposed to believable cues.     

Feedback on health status and peer popularity influences social risk decision-making processes in adolescents and young adults

Cognitive and dual-processes models, involving cognitive and socio-emotional components, for adolescents’ risky behaviour have been proposed. This study tested their predictions by manipulating the presence or absence of feedback about gains and losses in health and peer popularity in a decision-making task with peers. Risky (e.g., taking or refusing a drug) and ambiguous decisions (e.g., eating hamburger or hotdog) were examined in 256 adolescents (aged 13–14; 15–16; 17–18) and young adults (aged 19–20). Participants made more risky choices and required less decision time when receiving feedback about the negative consequences on health and positive consequences on peer popularity. Decision times were comparatively larger for risky than for safe options in late adolescence and young adulthood. Results supported the dual-processes model showing that, though late cognitive changes were observed that could eventually lead to the selection of safe options, feedback gains in peer popularity outweighed unhealthy consequences leading to risky decisions.     

Decision support for real-time, dynamic decision-making tasks

By definition, dynamic decision making dictates that multiple and interrelated decisions be made in a continuously changing environment. Such decision making is difficult and often taxes individuals’ cognitive resources. Here I investigated ways in which to support decision making in these environments. I evaluated three forms of decision support: outcome feedback, cognitive feedback, and feedforward that incorporated (to varying degrees) common features of learning theories associated with dynamic tasks. Participants in a laboratory experiment performed a real-time, dynamic decision-making task while receiving one of the different types of decision support. During the first 2 days, individuals received one type of decision support, but on the third day they performed the task without this support. Participants who received feedforward improved their performance considerably and continued to exhibit improved performance even after discontinuation of the decision support on the third day. Neither outcome feedback nor cognitive feedback resulted in improved performance. More research is necessary to conclusively identify the best forms of dynamic decision-making support and their durability when transferred to new tasks.     

Set size,assertion form,thematic content and sampling in the selection task

Participants attempted to solve a modified version of Wason's (1966) selection task. Variables were: (1) sizes of the sets referenced by a specified assertion, (2) form of the assertion (conditional or categorical), (3) thematic content of the assertion (letters and numbers or makes and owners of taxicabs), and (4) the need for sampling or not. In Experiment 1, participants were given enough information to determine the truth or falsity of the specified assertion with certainty; in Experiment 2, they had to rely on sampling and could not determine the assertion's truth or falsity with certainty. Performance was better in Experiment 1 than in Experiment 2, but in both cases much better than what is typically obtained with the conventional selection task. The results support the hypothesis that performance of the selection task is sensitive to the sizes of the sets involved, add credence to the conclusion that framing the task in a thematically meaningful way can facilitate performance, and demonstrate that facilitation does not require deontic reasoning or a well-known convention being involved. The relationship between expressed confidence and level of performance is consistent with other studies showing the former to be a good predictor of the latter.     

Learning From Peers’ Eye Movements in the Absence of Expert Guidance: A Proof of Concept Using Laboratory Stock Trading,Eye Tracking,and Machine Learning

Existing research shows that people can improve their decision skills by learning what experts paid attention to when faced with the same problem. However, in domains like financial education, effective instruction requires frequent, personalized feedback given at the point of decision, which makes it time‐consuming for experts to provide and thus, prohibitively costly. We address this by demonstrating an automated feedback mechanism that allows amateur decision‐makers to learn what information to attend to from one another, rather than from an expert. In the first experiment, eye movements of N = 100 subjects were recorded while they repeatedly performed a standard behavioral finance investment task. Consistent with previous studies, we found that a significant proportion of subjects were affected by decision bias. In the second experiment, a different group of N = 100 subjects faced the same task but, after each choice, they received individual, machine learning‐generated feedback on whether their pre‐decision eye movements resembled those made by Experiment 1 subjects prior to good decisions. As a result, Experiment 2 subjects learned to analyze information similarly to their successful peers, which in turn reduced their decision bias. Furthermore, subjects with low Cognitive Reflection Test scores gained more from the proposed form of process feedback than from standard behavioral feedback based on decision outcomes.     

Decision making under risk condition in patients with Parkinson's disease: a behavioural and fMRI study

We aimed to study whether previously described impairment in decision making under risky conditions in patients with Parkinson's disease (PD) is affected by deficits in using information about potential incentives or by processing feedback (in terms of fictitious gains and losses following each decision). Additionally, we studied whether the neural correlates of using explicit information in decision making under risk differ between PD patients and healthy subjects. We investigated ten cognitively intact PD patients and twelve healthy subjects with the Game of Dice Task (GDT) to assess risky decision making, and with an fMRI paradigm to analyse the neural correlates of information integration in the deliberative decision phase. Behaviourally, PD patients showed selective impairment in the GDT but not on the fMRI task that did not include a feedback component. Healthy subjects exhibited lateral prefrontal, anterior cingulate and parietal activations when integrating decision-relevant information. Despite similar behavioural patterns on the fMRI task, patients exhibited reduced parietal activation. Behavioural results suggest that PD patients' deficits in risky decision making are dominated by impaired feedback utilization not compensable by intact cognitive functions. Our fMRI results suggest similarities but also differences in neural correlates when using explicit information for the decision process, potentially indicating different strategy application even if the interfering feedback component is excluded.     

Selection of strategies for narrowing choice options: Antecedents and consequences

This study investigates the antecedents (task and decision maker characteristics) and consequences (set size and decision quality) of prescreening strategy selection. In Experiment 1 we investigated which strategy, inclusion or exclusion, is more natural for narrowing choices in tasks with a single correct answer; about 70% of the participants selected exclusion. Experiment 2 directly contrasted correct answer tasks with personal judgment tasks, using the same foils for the two tasks. Participants were more likely to use exclusion for items with a correct answer than for personal judgments. In Experiment 3, participants could choose different strategies for different items and rated the difficulty of each item. The greater the perceived difficulty of an item, the more likely participants were to choose an exclusion strategy. In all three experiments exclusion led to larger set sizes, across task type and experimental design. There were no differences in decision quality as a function of strategy selection after correcting for set size. Individual differences based on personality inventories were not found to be good predictors of strategy selection, but had moderate effects on set size for personal judgment tasks. Results are discussed in terms of a status quo bias for adding or deleting options from an initial reference frame.     

“先扬后抑”vs.“先抑后扬”:反馈顺序对决策信心建构的影响

本研究采用2×3被试内设计,通过操纵不同性质信息的反馈比例和反馈顺序,重点考察了反馈顺序对决策信心动态建构的影响。结果表明:(1)在不同反馈比例条件下,当个体接收到的正性反馈多于负性反馈时,个体的决策信心上升,反之则下降;当个体接收到的正性反馈等于负性反馈时,个体的决策信心下降,表现出"负性偏向";(2)"先扬后抑"与"先抑后扬"的信息反馈顺序对决策信心的动态建构产生了不同影响,表现出类似于"近因效应"的现象,"惊讶假设(surprise hypothesis)"可为此提供解释。本研究表明,在决策信心的动态建构过程中,信息的反馈顺序发挥着重要作用。