Flexible regression-based norming of psychological tests

Psychological tests are often used in the diagnosis and selection of individuals. The test scores themselves are typically not informative, so they are usually interpreted by comparing them with the test scores of a reference population. For intelligence tests, for example, a test score is usually compared with those of the general population of the same age as the testee. The required normed scores are obtained by estimating a statistical norming model for a range of age values, in which the raw test score distribution is related to age. To arrive at a realistic norming model, it is required to use appropriate models.

In this thesis, I investigate challenges in realistic norming, such as model selection and possibly complex models with larger sampling fluctuations. I show that good model selection is possible with an automated model selection procedure. I also investigate the consequences of using a too strict versus a too flexible model, and reveal that a too flexible model is generally better. In addition, I demonstrate how uncertainty in normed scores due to sampling fluctuations can be expressed in confidence intervals. The sampling fluctuations can be decreased by increasing the normative sample size, but this is costly and not always possible in practice. I show that the sampling fluctuations can be decreased without increasing the sample size by using prior norm information in the estimation of new normed scores. This demonstrates that good model selection and efficient test norming can be used to improve psychological test score interpretation.