Construct Validity

What it is

Lee Cronbach and Paul Meehl named it in 1955. The problem they were describing is that most of what psychology wants to measure cannot be observed directly. Nobody can see anxiety, intelligence or attachment security. What can be observed is a score, and the score is assumed to stand for the thing.

Three things have to hold for that assumption to be safe. The instrument has to agree with itself when nothing has changed. It has to agree with other instruments that claim to measure the same thing. And it has to predict what the underlying construct is supposed to predict, and not predict what it is not.

Failing the last one is called a specificity failure, and it is diagnostic. A measure that predicts everything equally is probably picking up something more general than its name suggests.

In effect

Attachment research supplies the clearest recent example of the second test failing.

The field has two adult instruments. The Adult Attachment Interview takes about an hour and scores how coherently a person narrates their childhood. Self-report questionnaires ask how a person feels about closeness now. Both produce a result called secure, and for two decades the field proceeded as though they were two windows onto one construct.

Glenn Roisman and colleagues pooled every study that had run both on the same people, nearly a thousand adults, then ran a fresh study with the best version of each. The overlap was, in their words, trivial to small. Under one part in a hundred.

And the third test failed too. When the Strange Situation was put through a preregistered predictive test in 2023, infant attachment forecast academic skills about as well as it forecast socioemotional outcomes. A measure of emotional development that predicts arithmetic equally well is telling you it is measuring something broader.

What it does not say

It does not mean either instrument is worthless. Both may be valid. They are not valid for the same thing, which is a different and more specific complaint.

It is not the same question as reliability. An instrument can agree with itself perfectly and still measure the wrong construct, and attachment's infant measure manages to fail both tests independently.

And it is not a reason to abandon measurement. The way out is to name what each instrument actually captures and stop borrowing authority across the gap between them.

---