---
title: "Self-Ratings Track Tested Ability at .29. Comparing the Three Ways to Fill a Skills Inventory."
description: "Self-evaluations of ability correlate .29 with objective performance across 22 meta-analyses, and self- and supervisor ratings agree at .22. A comparison of self-reported, manager-rated and tested skills data, and a tiering rule that uses all three."
canonical: https://www.assessall.com/blog/self-ratings-track-tested-ability-at-29-comparing-the-three-ways-to-fill-a
published: 2026-10-28
updated: 2026-10-28
author: "AssessAll Editorial"
source: AssessAll
---

# Self-Ratings Track Tested Ability at .29. Comparing the Three Ways to Fill a Skills Inventory.

Self-evaluations of ability correlate .29 with objective performance across 22 meta-analyses, and self- and supervisor ratings agree at .22. A comparison of self-reported, manager-rated and tested skills data, and a tiering rule that uses all three.

A skills inventory can be filled three ways: ask people to rate themselves, ask their managers, or test them. Self-ratings track objectively measured ability at about r = .29 across 22 meta-analyses. Manager ratings agree with self-ratings at .22. Only the third method produces evidence that will hold up when the inventory starts making decisions.

That ordering is the whole design problem. Every skills-based workforce plan — internal mobility, redeployment, reskilling spend, succession slates — eventually reads from a skills inventory. How that inventory was populated determines what it can legitimately be used for, and most organisations populate it the cheapest way and then use it for the most consequential decisions.

## The evidence on self-reported skill

Zell and Krizan's metasynthesis in *Perspectives on Psychological Science* pooled 22 meta-analyses covering roughly 350,000 participants and found a mean correlation of **r = .29** (SD = .11) between people's self-evaluations of ability and their objectively measured performance. The authors' own summary: people have "only moderate insight into their abilities." See [Zell & Krizan (2014)](https://doi.org/10.1177/1745691613518075).

The domain breakdown matters more than the headline. Self-insight was strongest where feedback is constant and weakest where it is absent:

- **Academic ability** — .33
- **Sports** — .31
- **Vocational skill** — .26
- **Medical skill** — .22
- **Memory** — .15
- **Nonverbal skill** — .09

Two other moderators are directly actionable. Domain-specific self-ratings outperformed global ones (.28 vs .18), and self-insight fell as task complexity rose (.32 at low complexity, .20 at high). In other words: self-report degrades exactly where your skills inventory is most ambitious — complex, judgment-heavy, feedback-poor capabilities like stakeholder influence, commercial judgment or coaching.

For cognitive ability specifically, [Freund and Kasten's meta-analysis](https://doi.org/10.1037/a0026556) of 41 studies and 22,256 participants put the correlation between self-estimated and psychometrically measured ability at **r = .33**. A 2026 study of a UK working-adult cohort found weaker agreement still: self-estimated intelligence correlated **r = .12** (non-significant) with tested general ability in the subsample that took the ability test, while correlating .25 with conscientiousness, .25 with risk approach and .23 with curiosity ([Furnham & Semmelink, 2026](https://doi.org/10.1016/j.paid.2025.113457)). The tested subsample was small — under 200 people — so treat the exact figure as indicative. The pattern is not: self-ratings carry a large dose of disposition.

That is the single most useful thing to know about a self-reported skills inventory. It is not mostly noise. It is partly a measure of something real and partly a measure of confidence — and the two are inseparable in the data you receive.

## The evidence on manager-rated skill

The obvious fix is to have managers rate instead. The agreement data makes clear this is a different measurement, not a better version of the same one.

Heidemeier and Moser's meta-analysis of 115 samples and 37,752 employees found self- and supervisor ratings of performance correlated **r = .22** observed, **.34** corrected for unreliability, and that self-ratings ran higher than supervisor ratings by about a third of a standard deviation (d = 0.32 in Western samples). See [Heidemeier & Moser (2009)](https://doi.org/10.1037/0021-9010.94.2.353) in the *Journal of Applied Psychology*.

Manager ratings also move with what they are for. Jawahar and Williams's meta-analysis of 22 studies and 57,775 ratings found ratings collected for administrative purposes averaged **d = .32** higher than ratings collected for research or development — the same raters, the same people, roughly a third of a standard deviation of inflation attributable to consequence alone ([Butler University record](https://digitalcommons.butler.edu/cob_papers/2)). A skills inventory that feeds promotion slates is an administrative instrument whatever the intranet copy says.

And the third option now in wide use — inferring skills from job titles, CV text, project records and system activity — has an evidence problem of its own: published criterion-validity evidence linking inferred skill ratings to job performance is scarce. Inference is best understood as a very fast way to generate hypotheses about skill, not as a measurement. Even [O*NET](https://www.onetcenter.org/dataCollection.html), the taxonomy much of this tooling is built on, derives its descriptor ratings from surveys of job incumbents and occupational experts — self-report and expert judgment, at the base of the stack.

## The evidence on tested skill

Tested and demonstrated skill is the only one of the three with a long criterion-validity record. The revised estimates in Sackett and colleagues' correction of the selection literature put structured interviews at .42, job knowledge tests at .40, work sample tests at .33 and general mental ability at .31 for predicting job performance ([Sackett et al., 2022](https://doi.org/10.1037/apl0000994)).

Those coefficients are not dramatically larger than .29. The difference that matters is not size, it is defensibility: a test score comes with documentation — a specification, reliability evidence, a scoring rule, a validity argument — of the kind the [Standards for Educational and Psychological Testing](https://www.testingstandards.net/) and the [SIOP Principles](https://www.siop.org/principles/) expect before scores drive employment decisions. A self-rating has no such record, which is why it cannot survive a challenge even when it happens to be accurate.

The real cost of testing is coverage. You cannot test 4,000 people on 60 skills, and you should not try.

## What each method is good for

**Self-reported skill**

- Breadth and speed across a whole population
- Interest, aspiration and willingness to be deployed — which only the person holds
- Identifying candidates for assessment, not conclusions about them
- Development planning the individual owns

**Manager-rated skill**

- Observed on-the-job behaviour the person cannot see
- Context: what the role actually demanded
- Weakest where the manager has thin observation or a stake in the outcome

**Tested or demonstrated skill**

- Decisions with consequences: selection, promotion, certification, redeployment
- Comparability across sites, cohorts and years
- Defensible evidence when a decision is questioned
- Too costly to apply to everything, so it has to be targeted

## When self-report is the right call

It is the right call more often than the numbers suggest. The Dunning-Kruger framing that made self-assessment a joke has itself been challenged: Gignac and Zajenkowski showed the classic effect is largely a statistical artefact of regression and measurement error rather than evidence that low performers are uniquely deluded ([*Intelligence*, 2020](https://doi.org/10.1016/j.intell.2020.101449)). Self-ratings are a weak signal, not a worthless one.

Use self-report when the decision is low-stakes and reversible, when the construct is concrete and feedback-rich (software tools used last quarter, languages spoken, machines operated, certifications held), when you need coverage in weeks rather than quarters, and when what you actually need is the person's intent rather than their proficiency. Use it also as a triage layer — a cheap filter that decides who gets measured properly.

## A tiering rule that uses all three

1. **Sort your skills list by decision stakes, not by importance.** A skill that only ever informs a learning recommendation needs self-report. A skill that gates a role needs a test.
2. **Ask self-report questions at the task level, with a frequency or recency anchor.** "Which of these have you done unsupervised in the last six months" beats a 1–5 proficiency slider; specificity bought .28 against .18 in the Zell and Krizan moderators.
3. **Keep the inventory's purpose visible and separate.** If the same field feeds both development and promotion, expect the administrative inflation Jawahar and Williams measured.
4. **Test only the gating skills, and only the people who reach the gate.** Pay-as-you-go assessment credits exist for exactly this shape of demand — AssessAll prices individual assessments at ₹30 / US$0.50 a credit, so a tested layer over a self-reported base is a budgeting question rather than a procurement project.
5. **Record the provenance of every rating.** Each skill entry should carry whether it was self-reported, manager-rated, inferred or tested, and when. An inventory that cannot tell you this cannot tell you what it is safe to do.

## What this means for the inventory you already have

If your skills data was collected by survey, you do not have a capability measurement — you have a well-populated hypothesis set with a .29 relationship to reality, weaker in the complex capabilities you care most about, and partly reflecting confidence. That is genuinely useful for coverage, deployment preference and learning routing. It is not evidence for a promotion, a redundancy selection or a client capability claim.

The fix is not to re-run the survey more carefully. It is to decide which handful of skills actually gate decisions, test those, and let self-report do the job it is good at: telling you where to look.

---

Source: https://www.assessall.com/blog/self-ratings-track-tested-ability-at-29-comparing-the-three-ways-to-fill-a