Latent Variable Realism in Psychometrics

Abstract

This dissertation concerns the theoretical status of latent variables in psychometrics and the philosophical foundations of psychometrics. I work toward the construction of a philosophical framework for psychometrics by examining and refining fundamental psychometric concepts such as validity, and by proposing a theoretical interpretation of latent variables. I mine psychometric methods for tacit philosophical commitments, make them explicit, and evaluate them. With its philosophical presuppositions made explicit, I then articulate what are realistic epistemic aspirations for psychometrics.

Latent variables cannot be measured directly; they must be inferred from observable variables phenomena. For example, variability in scores on psychometrics tests, are "explained" by positing an unobservable source of the observed variability, i.e., a latent variable. In psychometrics, many central theoretical constructs have their provenance in latent structure analysis. I focus on a particular latent variable, g, the general factor of intelligence.

At the most general level, this dissertation will address the following questions:

1. Does psychometric practice require regarding abilities as real entities?

2. Is epistemic realism regarding g tenable?

I argue that the answer to (1) is "yes" irrespective of any one psychometrician's professed philosophical commitments. However, unlike the critics of psychometrics, I argue that this commitment is innocuous and productive.

The answer to (2) depends on what it would mean to be a realist about g. I argue that g is not a causally efficacious attribute of individuals. It makes sense only as a relation between individuals. Statistical models of g do not contain information about causal processes within individuals. While it may be wrong to conceive of g as causally efficacious, models of g do constrain theories of mental ability at the level of individuals.