Development of Metacognition in Probabilistic Decision Making

• In Event: Metacognition

Abstract

Metacognitive awareness of our own uncertainty is central to many human judgements and decisions involving alternative possible outcomes, from basic perceptual detection to predicting social behaviour (Dunlosky & Metcalfe, 2008). Yet investigations of the developmental origins of metacognitive awareness and its influence on decision making have been hindered by measurement difficulties, often resulting in the use of simplistic, binary measures. The present study introduced a novel computerised paradigm and continuous measure to investigate 6- to 9-year-old children’s (N = 108) implicit and explicit confidence when making probabilistic decisions.
During training, participants were introduced to the “Randomiser”: a digitised lottery machine that randomly selected one box from among multiple boxes containing coloured balls. Participants were also introduced to a novel continuous scale in the form of a cup, which allowed them to rate how confident they were that a particular-coloured ball would exit the Randomiser (see Figure 1). In phase 1, participants saw two boxes containing red and/or blue balls enter the Randomiser. Children selected whether they thought the ball to come out of the Randomiser would be red or blue, and subsequently rated their confidence on the continuous scale. In phase 2, participants saw four boxes go into the Randomiser, with probabilities of the highest-likelihood option varying from 50%, to 75%, to 100%. In phase 3, participants again saw four boxes go into the Randomiser, although this time some boxes were occluded by a lid that prevented children from knowing the contents (either red or blue). On each trial of each phase, participants were scored for whether they selected the highest-probability option, as well as for implicit confidence (response latency) and explicit confidence (ratings on the continuous scale).
Results indicated that in phases 2 and 3, probability was a significant predictor of accuracy, χ2(1, 108) = 41.66, p < .001, and χ2(1, 108) = 25.54, p < .001, respectively. That is, as the high-probability option approached certainty (100%), participants became increasingly likely to predict that this option would exit the Randomiser. Supporting the validity of the measure, both response latency and explicit confidence varied linearly as a function of the high-probability option (see Figure 2). Response latency was also a significant unique predictor of explicit confidence in all phases, F(1, 108) = 12.30, p = .001, F(1, 108) = 28.50, p < .001, and F(1, 108) = 30.15, p < .001, such that as response latency decreased, explicit confidence increased.
In phase 3, the number of closed boxes was an additional significant predictor of accuracy, χ2(1, 108) = 48.54, p < .001, such that participants were less accurate as the number of closed boxes increased. Intriguingly, this effect was more pronounced in older children than younger children, χ2(1,108) = 27.69, p < .001. This pattern indicates that older children in particular are less inclined to calibrate decisions in line with known probability parameters as the quantity of unknown information increases. One possibility is that, with increasing age, children become more likely to form misguided metacognitive evaluations about unknown probabilistic information.

Authors

• Cate MacColl, University of Queensland
  Presenting Author

• Jonathan Redshaw, University of Queensland