A Statistical Explanation of the Dunning–Kruger Effect
Magnus, J. R. (Vrije Universiteit Amsterdam) & Peresetsky, A. A. (Higher School of Economics, Moscow), Frontiers in Psychology 13:840180, 25 March 2022. Open access.
The paper dunning-kruger-misread referred to and did not name. Both authors are econometricians, which is the first thing to know about it: the case against a psychological finding is being made from outside psychology, with the tools of censored regression.
The argument
Their claim is not that the effect fails to replicate. It is that the effect is real, reproducible, and empty of psychology:
There is thus no need for a psychological explanation of the DK effect: it is a statistical artifact.
And, more bluntly, in the conclusion: “there is an effect, but it does not reflect human nature.”
The model and the data
Data: 665 undergraduates at the International College of Economics and Finance, HSE Moscow, across four cohorts (2016–2019), each predicting their own grade on a statistics exam on a 0–100 scale. Mean actual grade 39.8; mean predicted 38.6.
Model: a three-parameter doubly-censored tobit. A prediction on a 0–100 scale cannot leave the interval, so the latent judgment gets censored at both ends. The three parameters are measurement noise plus a random lower and upper bound; fitted by nonlinear least squares, the conditional expectation lands almost on top of the nonparametric curve. The authors stress the fit is robust to the parameter values — several different settings give indistinguishable curves.
The mechanism in one sentence: a student who really is near the floor has nowhere to err but upward, and one near the ceiling has nowhere to err but downward, so the crossing pattern appears with no difference in self-knowledge anywhere in the sample.
Lineage. Krueger & Mueller (2002) opened this line with regression-plus-better-than-average; Burson, Larrick & Klayman (2006) formalized a noise-plus-bias model; Gignac & Zajenkowski (2020) had already argued in Intelligence that the effect is “(mostly) a statistical artefact.” This paper’s contribution is modelling the boundary constraints explicitly rather than invoking regression to the mean as a known fact — a move the authors argue most of the literature makes too casually.
Two limits, and the wiki can only see the second because it holds both papers
Stated by the authors: only 14% of observations exceed 60, so the model is fitted where the data are thin at the top.
Not stated, and larger: the paper never engages Study 4. Its entire case rests on fitting one cross-sectional dataset — grades predicted once, on one exam. kruger-dunning-1999 anticipated the statistical objection and built an experiment against it: train the bottom quartile in logic for ten minutes and their self-estimates fall sharply, while an untrained control group’s do not. A boundary artifact does not explain why a training packet delivered after the test moves self-assessment, and Magnus and Peresetsky neither model that result nor mention it. What they demonstrate is that the cross-sectional shape needs no psychology. Whether the experimental manipulation needs any is untouched.
They also get a fact about the target paper wrong. They write that top-performer underestimation was “not discussed in Kruger and Dunning (1999)” but is “often also associated with their names.” The 1999 paper devotes a titled section to it — The Burden of Expertise — attributes it to the false-consensus effect, and built the second phase of Study 3 to test that attribution. The error is small, and it points the same way as the omission above: this is a statistical re-analysis of a pattern, by authors reading the original for its shape rather than its experiments.
What this leaves standing
The tempting summary — “Dunning-Kruger has been debunked” — is not what either paper supports. The honest statement is narrower and more interesting, and dunning-kruger-effect carries it: the familiar percentile-crossing chart is fully explained by censoring, and the training experiment is not. Two primary sources, disagreeing, neither refuted.
Related
dunning-kruger-effect · kruger-dunning-1999 · dunning-kruger-misread · replication-crisis · jan-magnus · anatoly-peresetsky · heuristics-and-biases · synthesis