Construct validity
Whether a measure actually captures the abstract thing it claims to.
What it means
Construct validity is the degree to which a test or measure genuinely assesses the theoretical concept it is intended to, such as intelligence, anxiety, or trust. Because such constructs are not directly observable, validity is built up from a web of evidence: theoretical justification, the measure's relationships with related and unrelated variables, and its behavior across known groups and conditions. Threats include construct underrepresentation, where the measure misses facets of the concept, and construct-irrelevant variance, where it picks up extraneous influences. It is widely regarded as the overarching form of validity, since reliable, well-controlled studies of the wrong construct still answer the wrong question.
How the evidence is assembled
In practice, construct validity is not one test but an accumulating case. Convergent evidence shows the measure correlating with things it should track; discriminant evidence shows it staying distinct from things it should not. Campbell and Fiske formalised both in the multitrait-multimethod matrix, which crosses several traits with several measurement methods so the trait signal can be separated from the method artefact. Known-groups validation checks that the measure distinguishes people who ought to differ, such as diagnosed patients from controls. Factor analysis probes whether items cluster in the pattern the theory predicts. No single result settles the question; each rules out a rival explanation, and the construct interpretation survives only as long as the whole pattern holds together.
The nomological network
Cronbach and Meehl's original move was to insist that an unobservable construct means nothing in isolation. It acquires meaning only through a nomological network, the web of lawful relations linking it to other constructs and to observable behaviour. Validating a measure therefore means testing that entire theory at once. This creates a well-known circularity: when a predicted relationship fails to appear, the fault may lie with the measure, with the theory, or with the study design, and the data alone rarely say which. Progress comes from many studies gradually tightening the network rather than from a single decisive experiment. It is why construct validation is described as never finished, only better or worse supported at a given moment.
Still argued over
Theorists disagree about what construct validity even is. Messick folded content, criterion and consequence into a single overarching construct validity, arguing that the social consequences of using a test belong inside its validity, not outside it. Borsboom and colleagues pushed the other way: a test is valid, they said, simply if the attribute exists and its variation causally produces variation in the scores, and everything else, from networks of correlations to fairness and consequences, is a separate question the field had wrongly absorbed into validity. The dispute is not academic hair-splitting. It changes what evidence you must collect before claiming a number measures what its label says, and how much of a test's real-world impact you are answerable for.
Where it shows up
The stakes are highest wherever a number stands in for something no one can see directly. Employment tests must defend, sometimes in court, that a score reflects job-relevant ability rather than schooling or fluency. Clinical scales claim to measure depression or pain, then drive treatment. Schools and health systems are ranked on metrics that quietly bundle in intake, wealth or catchment. The same trap runs through analytics, where a team names a dashboard number 'productivity' and manages toward it without ever checking what it captures. Flake and Fried document how routinely researchers build ad hoc scales, rename old ones, and report no validity evidence at all, a pattern that lets a plausible label pass for a measured construct.
Examples
A 'job-satisfaction' survey that mostly measures pay contentment lacks construct validity for the broader concept it names.
A product team measures 'engagement' by minutes spent in the app, then celebrates a rise that came from a slower search and a confusing new menu keeping people hunting.
Ranking hospitals on patient-satisfaction scores measures something real, but that something includes parking, food and wifi, not only whether the clinical care was any good.
A reading-comprehension exam timed so tightly that fast readers outscore deep ones is partly measuring reading speed, a construct-irrelevant influence on the score it reports as comprehension.
An AI 'reasoning' benchmark a model can ace by having memorised similar problems is measuring recall, not reasoning; the label holds only until someone tests it on genuinely novel items.
First described in Cronbach & Meehl (1955).
Key references
- Flake, J. K., & Fried, E. I. (2020). Measurement schmeasurement: Questionable measurement practices and how to avoid them. Advances in Methods and Practices in Psychological Science, 3(4), 456-465. doi.org/10.1177/2515245920952393
- Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2004). The concept of validity. Psychological Review, 111(4), 1061-1071. doi.org/10.1037/0033-295X.111.4.1061
- Messick, S. (1995). Validity of psychological assessment: Validation of inferences from persons' responses and performances as scientific inquiry into score meaning. American Psychologist, 50(9), 741-749. doi.org/10.1037/0003-066X.50.9.741
- Campbell, D. T., & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 56(2), 81-105. doi.org/10.1037/h0046016
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281-302. doi.org/10.1037/h0040957