Is It You or Your Model Talking?
A Framework for Model Validation
This report lays out a conceptual framework for validation, arguing that some models can be validated and used to predict, while others cannot be validated and may only be put to nonpredictive uses. To be validatable, a model must be observable and measurable, must exhibit constancy of structure in time, must exhibit constancy across variations in conditions not specified in the model, and must permit the collection of ample data. Nonvalidatable models can be used as a bookkeeping device, as an aid in selling an idea of which the model is but an illustration, as a training aid to induce a particular behavior, as part of an automatic management system whose efficacy is not evaluated by using the model as if it were a true representation, as an aid to communication, as a vehicle for a fortiori arguments, and as an aid to thinking and hypothesizing. The report shows that the appropriate form of model quality assurance depends fundamentally on how the model is used.