Dec 7, 2010, Bruce Ratner
The purpose of this article is to resurface two essential concepts of a predictive model: Reliability and validity. I reexamine these concepts so that model builders and marketing users of models know whether their predictive models are meritable.
Two essential characteristics of a predictive model are its reliability and validity, terms that are sometimes misknown, and thus misused interchangeably. Reliability refers to a model that yields consistent results. 1 For a predictive model (the statistical regression model, the primary, but not the only kind implied in this article; e.g., CHAID), the proverbial question is "How dependable is the model for predictive purposes?"
If a model is reliable, it is said to have internal consistency, which means the predictor variables are sufficiently contributing to the model predictiveness. Thus, the model is trustworthily producing good predictions. However, reliability is not a principle without a middle ground; it is a matter of degree. No model itself is perfectly reliable. It is quite normal for individuals to vary in performance, as chance influences are always in operation, but performance is expected to fall within narrow limits from one model implementation to another. In other words, predictions for the same individuals obtained, from repeated implementations of a reliable model, are expectedly to be moderately variable.