What your self-rating is worth

You are about to rate your own readiness. Before you do, the instrument needs a warning label.

Across 55 studies comparing self-evaluations of ability with performance measures, the mean correlation was r = .29 with a standard deviation of .25 (Mabe & West, 1982). Thirty-two years later, a synthesis of 22 meta-analyses across academic, intelligence, language, medical, sports and vocational domains produced r = .29 again, with individual effects ranging from .09 to .63 (Zell & Krizan, 2014). Two independent syntheses, three decades apart, on the same number.

An r of .29 means self-assessment shares roughly 8 per cent of its variance with actual ability. Not none. Not much.

The story you have probably heard about this is the Dunning-Kruger effect, and most of what is said about it is wrong in a way that matters here. The original studies are small, single-institution and student-based: 65 participants on humour, 45 on logical reasoning, 84 on grammar, 140 on a logic task. The bottom quartile in the grammar study sat at the 10th percentile and rated themselves at the 67th, which is the striking result everyone repeats.

The problem is the analysis. Plotting the difference between self-rating and score against the score itself generates that pattern from pure random numbers, which Nuhfer and colleagues demonstrated by simulation on a modelled dataset of 1,154 participants. Applied to 929 community participants with an objective ability measure, the two statistically valid tests of the hypothesis both came back null: no significant heteroscedasticity, and an essentially linear relationship between measured and self-assessed intelligence (Gignac & Zajenkowski, 2020). A double-censored model reproduced the entire pattern from scale boundary effects alone in 665 students predicting their own exam scores, because low performers cannot predict below zero and high performers cannot predict above 100 (Magnus & Peresetsky, 2022).

So the correct position is not that you are deluded about your abilities. It is narrower and more useful: your self-rating carries some signal, it carries much less than you experience it as carrying, and there are specific conditions that make it better.

Mabe & West identified them, and nine measurement conditions accounted for R = .64 of the variance in how valid a self-evaluation turned out to be. Four of those conditions are things you can arrange this week.

Self-evaluation was more accurate when the rater expected the self-rating to be compared against an objective criterion. When the rater had prior experience of self-evaluating. When anonymity was guaranteed. And when the instruction asked for comparison against peers rather than against an absolute standard.

The first of those conditions is the one that does the work, and it is why the Level 0 exercise asks a question that looks like bureaucracy: for each line of your readiness map, who could check this and how?

The mechanism is not mysterious. A rating you expect nobody to test is a statement about how you would like to be seen. A rating you expect somebody to open a document against is a prediction, and predictions get made more carefully. The difference is not honesty. It is that the second one has a cost attached to being wrong.

There is a second, harder implication that most readiness exercises avoid. If self-assessment correlates .29 with ability, then the readiness map you produce alone is worth about 8 per cent of the information you need, and the external check is not a nice-to-have step at the end. It is most of the exercise. This is the reason Exercise 0.B exists and the reason it is gated separately.

And there is a specific version of this that shows up later in the course, so it is worth naming now. Anxious candidates systematically believe they performed far worse in interviews than the interviewer thought. Interview anxiety correlates between −.15 and −.49 with self-rated interview performance, and between −.07 and −.28 with interviewer-rated performance (McCarthy & Goffin, 2004). In the largest field study of the question, 8,343 real applicants across 373 companies in 93 countries, the relationship between anxiety and interview performance was γ = −.03, statistically significant and practically negligible (McCarthy et al., 2021).