ABSTRACT
Connections With Other Chapters
Chapter 3 introduced generalizability theory as a statistical model that views item scores as samples from a domain. In chapter 5, the associated conceptual model of domain scores was reviewed and analyzed. The chapter concluded that, despite appearances to the contrary, a minimal set of causal assumptions can hardly be avoided in the interpretation of the domain score model. In the present chapter, we discuss models that explicitly use causal relations between constructs and test scores. Two such models are of particular importance: the reflective model, in which test scores are modeled as effects of the construct, and the formative model, in which they are modeled as causes of the construct. These models are discussed in the light of validity theory. In addition we provide a brief review of alternative causal models.
In most cases of structured observation interpreted as measurement, there is an asymmetry between the measures and the attribute being measured. One uses a pointer reading of one’s scale to measure weight, but one does not use one’s weight to measure the pointer reading; one uses diagnostic interviews to assess psychiatric problems, not the other way around; and one uses personality test scores to assess personality dimensions, but personality dimensions never serve to assess test scores.