Limits...
Weighting mean and variability during confidence judgments.

de Gardelle V, Mamassian P - PLoS ONE (2015)

Bottom Line: In a motion categorization task on moving dots, we manipulated the mean and the variance of the motion directions, to obtain a low-mean low-variance condition and a high-mean high-variance condition with matched performances.Critically, in terms of confidence, observers were not indifferent between these two conditions.Observers exhibited marked preferences, which were heterogeneous across individuals, but stable within each observer when assessed one week later.

View Article: PubMed Central - PubMed

Affiliation: Laboratoire de Psychologie de la Perception, CNRS & Université Paris Descartes, Paris, France; Centre d'Economie de la Sorbonne, CNRS & Université Paris 1, Paris, France; Paris School of Economics, Paris, France.

ABSTRACT
Humans can not only perform some visual tasks with great precision, they can also judge how good they are in these tasks. However, it remains unclear how observers produce such metacognitive evaluations, and how these evaluations might be dissociated from the performance in the visual task. Here, we hypothesized that some stimulus variables could affect confidence judgments above and beyond their impact on performance. In a motion categorization task on moving dots, we manipulated the mean and the variance of the motion directions, to obtain a low-mean low-variance condition and a high-mean high-variance condition with matched performances. Critically, in terms of confidence, observers were not indifferent between these two conditions. Observers exhibited marked preferences, which were heterogeneous across individuals, but stable within each observer when assessed one week later. Thus, confidence and performance are dissociable and observers' confidence judgments put different weights on the stimulus variables that limit performance.

No MeSH data available.


Related in: MedlinePlus

Logistic regression for confidence.(A) The metacognitive regression curve for one participant, plotting confidence comparison responses (in probability, y-axis) against differences in signal-to-noise ratios between the two trials to be compared (threshold units, x-axis). Large dots are real data, binned along the x-axis variable. Aligned small dots are regression fits. Trial pairs in which the low variance stimulus was presented first (resp. second) are presented in black (resp. gray). (B) Parameter estimates for the logistic regression (EQ1). Small dots represent for each participant the average estimate of the parameter across the two sessions. Big dots and error bars represent the mean and SEM at the group level. (C) Stability of parameter estimates across the two sessions, one week apart. For each predicting variable, we plotted the beta coefficient measured on the second session against the value on the first session, and report the correlation between the two sessions. Error bars represent the standard errors for each parameter estimate. Cross-sessions correlations are reported along with their significance (** p<0.01, *** p<0.001).
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4368758&req=5

pone.0120870.g004: Logistic regression for confidence.(A) The metacognitive regression curve for one participant, plotting confidence comparison responses (in probability, y-axis) against differences in signal-to-noise ratios between the two trials to be compared (threshold units, x-axis). Large dots are real data, binned along the x-axis variable. Aligned small dots are regression fits. Trial pairs in which the low variance stimulus was presented first (resp. second) are presented in black (resp. gray). (B) Parameter estimates for the logistic regression (EQ1). Small dots represent for each participant the average estimate of the parameter across the two sessions. Big dots and error bars represent the mean and SEM at the group level. (C) Stability of parameter estimates across the two sessions, one week apart. For each predicting variable, we plotted the beta coefficient measured on the second session against the value on the first session, and report the correlation between the two sessions. Error bars represent the standard errors for each parameter estimate. Cross-sessions correlations are reported along with their significance (** p<0.01, *** p<0.001).

Mentions: To take a different angle on these results, we carried out a logistic regression (EQ3) in which the probability that the observer reported greater confidence in the first trial of the pair, P(C1>C2), was predicted from three quantities. First, a constant term with weight β1 measured the overall tendency of the observer in this confidence comparison task. Second, we quantified using threshold units (EQ2) the signal-to-noise ratio of the stimuli in the two trials, noted S1 and S2, and we used their difference as a predictor with weight β2 in the regression. Third, to capture more formally the variance-related bias reported above, we included a predictor with weight β3, that coded whether the low variance or the high variance stimulus was presented first. This logistic regression was estimated for each participant and session. We report for each parameter the mean estimate across the two sessions, and the correlation between the two sessions to evaluate stability in time of the parameter estimates. The results of this analysis are presented Fig. 4.


Weighting mean and variability during confidence judgments.

de Gardelle V, Mamassian P - PLoS ONE (2015)

Logistic regression for confidence.(A) The metacognitive regression curve for one participant, plotting confidence comparison responses (in probability, y-axis) against differences in signal-to-noise ratios between the two trials to be compared (threshold units, x-axis). Large dots are real data, binned along the x-axis variable. Aligned small dots are regression fits. Trial pairs in which the low variance stimulus was presented first (resp. second) are presented in black (resp. gray). (B) Parameter estimates for the logistic regression (EQ1). Small dots represent for each participant the average estimate of the parameter across the two sessions. Big dots and error bars represent the mean and SEM at the group level. (C) Stability of parameter estimates across the two sessions, one week apart. For each predicting variable, we plotted the beta coefficient measured on the second session against the value on the first session, and report the correlation between the two sessions. Error bars represent the standard errors for each parameter estimate. Cross-sessions correlations are reported along with their significance (** p<0.01, *** p<0.001).
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4368758&req=5

pone.0120870.g004: Logistic regression for confidence.(A) The metacognitive regression curve for one participant, plotting confidence comparison responses (in probability, y-axis) against differences in signal-to-noise ratios between the two trials to be compared (threshold units, x-axis). Large dots are real data, binned along the x-axis variable. Aligned small dots are regression fits. Trial pairs in which the low variance stimulus was presented first (resp. second) are presented in black (resp. gray). (B) Parameter estimates for the logistic regression (EQ1). Small dots represent for each participant the average estimate of the parameter across the two sessions. Big dots and error bars represent the mean and SEM at the group level. (C) Stability of parameter estimates across the two sessions, one week apart. For each predicting variable, we plotted the beta coefficient measured on the second session against the value on the first session, and report the correlation between the two sessions. Error bars represent the standard errors for each parameter estimate. Cross-sessions correlations are reported along with their significance (** p<0.01, *** p<0.001).
Mentions: To take a different angle on these results, we carried out a logistic regression (EQ3) in which the probability that the observer reported greater confidence in the first trial of the pair, P(C1>C2), was predicted from three quantities. First, a constant term with weight β1 measured the overall tendency of the observer in this confidence comparison task. Second, we quantified using threshold units (EQ2) the signal-to-noise ratio of the stimuli in the two trials, noted S1 and S2, and we used their difference as a predictor with weight β2 in the regression. Third, to capture more formally the variance-related bias reported above, we included a predictor with weight β3, that coded whether the low variance or the high variance stimulus was presented first. This logistic regression was estimated for each participant and session. We report for each parameter the mean estimate across the two sessions, and the correlation between the two sessions to evaluate stability in time of the parameter estimates. The results of this analysis are presented Fig. 4.

Bottom Line: In a motion categorization task on moving dots, we manipulated the mean and the variance of the motion directions, to obtain a low-mean low-variance condition and a high-mean high-variance condition with matched performances.Critically, in terms of confidence, observers were not indifferent between these two conditions.Observers exhibited marked preferences, which were heterogeneous across individuals, but stable within each observer when assessed one week later.

View Article: PubMed Central - PubMed

Affiliation: Laboratoire de Psychologie de la Perception, CNRS & Université Paris Descartes, Paris, France; Centre d'Economie de la Sorbonne, CNRS & Université Paris 1, Paris, France; Paris School of Economics, Paris, France.

ABSTRACT
Humans can not only perform some visual tasks with great precision, they can also judge how good they are in these tasks. However, it remains unclear how observers produce such metacognitive evaluations, and how these evaluations might be dissociated from the performance in the visual task. Here, we hypothesized that some stimulus variables could affect confidence judgments above and beyond their impact on performance. In a motion categorization task on moving dots, we manipulated the mean and the variance of the motion directions, to obtain a low-mean low-variance condition and a high-mean high-variance condition with matched performances. Critically, in terms of confidence, observers were not indifferent between these two conditions. Observers exhibited marked preferences, which were heterogeneous across individuals, but stable within each observer when assessed one week later. Thus, confidence and performance are dissociable and observers' confidence judgments put different weights on the stimulus variables that limit performance.

No MeSH data available.


Related in: MedlinePlus