Confidence ratings are better predictors of future performance than delayed judgments of learning

ABSTRACT

What is the best way to predict future memory performance? The intuitive answer is through judgments of learning (JOLs), in which people estimate how likely they are to remember something in the future. Recent theory, however, suggests that a retrospective confidence rating made just after a retrieval attempt might be a better predictor in some situations. In three preregistered experiments, we compared delayed JOLs to confidence ratings. People studied paired associates (E1) or psychology vocabulary terms (E2 & E3), then took a practice cued-recall test in which they made either a JOL or confidence rating after each response. They then took a final test. In Experiment 1, confidence ratings offered higher resolution (metacognitive accuracy) of memory for paired associates than did JOLs, but in Experiments 2 and 3, the advantage of confidence ratings was much smaller. A mini meta-analysis indicated that confidence ratings have a small advantage in predicting future performance over delayed JOLs. We argue that the two judgments rely on similar cues, and that even though JOLs explicitly ask people to predict future performance, doing so does not enhance prediction accuracy. Rather, the presence of a retention interval in the JOL cue adds variability to the judgment process.

KEYWORDS:

• Metacognition
• judgments of learning
• retrospective confidence judgments
• paired associates
• self-consistency

Acknowledgements

We thank the Carleton College Psychology Department and Furman University for funding this research. We also thank Grace Gilmore and Peter Dehkes for their essential work on Experiments 1 and 2A and Jae Eun Lee, Lucia Ray, and Hailey Rinella for their assistance.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 We conducted analyses using the data from all subjects. An analysis with just the preregistered subjects—described in the supplemental materials—yielded the same outcomes.

2 Thanks to an anonymous reviewer for suggesting this alternative computation for gamma.

3 Resolution for first test recall is reported in the supplemental materials.

4 Analyses using Kendall’s Tau, p = .004, d = 0.80, and G_trap (a measure of gamma based on ROC curves), p < .001, d = 0.61, yielded similar outcomes.

5 We originally preregistered a one-tailed test (which was significant) but report the more conservative two-tailed test here. Also, note that indexing resolution with Kendall’s Tau, p = .009, d = 1.12, or G_trap, p = .040, d = 0.26, led to statistically significant differences with two-tailed tests.

6 Using G_trap led to an outcome similar to the analysis reported here with gamma, with p = .060 d = 0.29; see supplemental materials for details.

7 Using G_trap to measure resolution yielded identical outcomes, p = .498. See supplemental materials for details.

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.