Elsevier

Consciousness and Cognition

Volume 27, July 2014, Pages 246-253
Consciousness and Cognition

Variance misperception explains illusions of confidence in simple perceptual decisions

https://doi.org/10.1016/j.concog.2014.05.012Get rights and content

Highlights

  • Confidence must depend on the strength and the reliability of the evidence.

  • Human observers ignore the reliability of evidence in a perceptual decision.

  • Observers are overconfident for stimuli with a low reliability.

  • Overconfidence in real-life decisions can now be studied with perceptual tasks.

Abstract

Confidence in a perceptual decision is a judgment about the quality of the sensory evidence. The quality of the evidence depends not only on its strength (‘signal’) but critically on its reliability (‘noise’), but the separate contribution of these quantities to the formation of confidence judgments has not been investigated before in the context of perceptual decisions. We studied subjective confidence reports in a multi-element perceptual task where evidence strength and reliability could be manipulated independently. Our results reveal a confidence paradox: confidence is higher for stimuli of lower reliability that are associated with a lower accuracy. We show that the subjects’ overconfidence in trials with unreliable evidence is caused by a reduced sensitivity to stimulus variability. Our results bridge between the investigation of miss-attributions of confidence in behavioral economics and the domain of simple perceptual decisions amenable to neuroscience research.

Introduction

Subjective confidence is used ubiquitously in approximate everyday expressions such as – “I think…”, “Maybe…”, “I am sure that”. In signal detection theory, confidence has a precise mathematical definition, indexing the probability that the decision is actually correct (Kepecs & Mainen, 2012). A large corpus of data suggests that the brain can be close to optimal when performing perceptual inferences under uncertainty, integrating multiple sources of evidence weighted by their reliability (Knill and Pouget, 2004, Körding and Wolpert, 2006). The view that confidence is an accurate index of decision uncertainty has been a conceptual anchor for neurophysiological and psychophysical studies of confidence (Bach & Dolan, 2012). Indeed, recent studies have demonstrated how cortical circuits can represent uncertainty in perceptual decisions and how these representations guide action (Kepecs et al., 2008, Kiani and Shadlen, 2009).

The emerging picture of probabilistic inference in neuroscience, however, is in striking tension with the principles of behavioral economics. Decades of experimentation with “real-life” decision problems have shown that humans are sub-optimal decision makers and, specifically, that their estimation of confidence deviates from the predictions of probability theory (Griffin and Tversky, 1992, Lichtenstein et al., 1977, Slovic, 1972, Slovic and Lichtenstein, 1968, Tversky and Kahneman, 1974, Tversky and Kahneman, 1981). In economics, humans exhibit reliable inconsistencies: they rely on sub-samples of the data, focus on tokens (representative exemplars), ignore the variance (or reliability) of the distribution, and over-weight evidence confirming their previous commitments and choices (Griffin and Tversky, 1992, Tversky and Kahneman, 1973, Tversky and Kahneman, 1974, Tversky and Kahneman, 1981). Consider the following example of variance neglect from behavioral economics (Kahneman, 2011). When people are presented with the sentence “In a telephone poll of 30 seniors, 60% support the president”, many conclude that “most seniors support the president”. This leap of confidence ignores the fact that the reliability of a poll depends on the number of participants.

The present work is an effort to reconcile theories of probabilistic inference in neuroscience with behavioral economics (Griffin and Tversky, 1992, Jarvstad et al., 2013, Wu et al., 2009). We investigate if subjects are able to take the variance of visual stimuli into account and if conditions exist where the perceptual choices of subjects deviate systematically from their confidence reports.

Section snippets

Results

We studied sensory decision making with a field of 60 oriented line segments. Orientations were sampled from a uniform distribution of mean orientation μ and width φ. We controlled orientation jitter across the bars (exemplars) by varying φ, which was randomly selected from a set of three values (φi, i = 1, 2, 3) (Fig. 1a). Participants reported with a single (four-choice) response whether the average orientation was right (clockwise, CW) or left (counter-clockwise, CCW) relative to the vertical

Discussion

Confidence in a perceptual decision represents an estimate of the quality of an internal representation of sensory evidence, and should depend on the sensory evidence (‘signal’) and its reliability (‘noise’) (Yeung & Summerfield, 2012). Here we showed that the confidence of subjects in their perceptual decision deviates systematically from the veridical signal-to-noise ratio. Paradoxically, subjects assign higher confidence to stimuli with a larger variability that is associated with poorer

Experiments

Six subjects aged 18–26 participated in experiment one after giving informed consent. All had normal or corrected-to-normal vision and where naive to the purpose of the experiment. Each participant performed 6 blocks of 300 trials each.

Trials began with the presentation of a white fixation point (0.27°) on a black background. After 0.3 s, the oriented bars were presented at the center of the screen for 0.3 s. After stimulus offset, a blank screen was presented until the participant made a

Acknowledgments

This work was funded by CONICET, the Netherlands Organization for Scientific Research, the Human Frontier Science Program, and by the James McDonnell Foundation.

References (29)

  • Andreas Jarvstad et al.

    Perceptuo-motor, cognitive, and description-based decision-making seem equally good

    Proceedings of the National Academy of Sciences

    (2013)
  • P. Juslin et al.

    Thurstonian and Brunswikian origins of uncertainty in judgment: A sampling model of confidence in sensory discrimination

    Psychological Review

    (1997)
  • Daniel Kahneman

    Thinking, fast and slow

    (2011)
  • A. Kepecs et al.

    A computational framework for the study of confidence in humans and animals

    Philosophical Transactions of the Royal Society B: Biological Sciences

    (2012)
  • Cited by (0)

    View full text