Prediction (General)

Criteria

Predicting criteria on the basis of predictors is one of the most important activities in psychological assessment (cf. Wiggins, 1973). Criteria in psychological assessment are, in most cases, criteria of success which refer to future behaviour of the persons concerned in situations that are different from those in which the original assessment, i.e. the assessment of the predictors, takes place. Typical problems dealt with in psychological assessment are selecting treatments and interventions, recommending programmes of training, education and exercise. Parents may ask for help when selecting an educational career for their children, adolescents when trying to overcome difficulties at university. In each case, success is the crucial issue: success of a treatment or intervention; success of a training programme, an educational measure, or an exercise. Will the educational career of the child chosen by the parents be successful? Will the adolescent successfully overcome the difficulties at university when a certain kind of intervention is applied? etc.

The reliability of a criterion determines an upper bound to its predictability: an unreliable criterion cannot be predicted. Quite often, the instability of replications is due to a lack of replicability of the criterion.

Selection and Characterization of Predictors

The optimum is to derive predictors from a theory which links constructs on the predictor side to constructs on the criterion side – if such a theory exists. Given a set of predictors, reliability of each one is of utmost concern. Some models like linear regression assume that the predictors are measured without error. If the scale level of a variable is high, this variable in general is better suited as a predictor because the higher scale level indicates that the variable may provide more information. Then chances are better to have information available which is useful for prediction.

One reason why more than one predictor should be used is the possibly complex nature of the criterion. If it is composed out of several components, predictors may be related to different subsets of these components. This should lead to a relatively low correlation between predictors. On the other hand, within some prediction models, the predictors may serve like items in classical test theory to increase the reliability of the overall score.

...