Center of Research on psychological and somatic disorders
Methodology & Statistics / Medical & Clinical Psychology
School of social and behavioral sciences
Latent variable prediction models in clinical and medical psychology
Traditional prediction in health research involves the use of unweighted sum scores to predict health outcomes. The use of sum scores in prediction is problematic for several reasons. First, this approach lacks a measurement model, so it does not take into account measurement errors and the attenuated predictions that they create. Second, the absence of a measurement model inhibits researchers from answering some important substantive questions, like whether the latent variable underlying tests scores is itself responsible for a particular prediction or whether any of the test items (e.g., reflecting narrow symptoms) show any additional direct paths to the predicted outcome. A third limitation of prediction based on sum scores is their inability to detect which items are more or less informative for certain regions of the latent scale. Failing to model such item characteristics might lead to non-linear test characteristic curves and hence to non-linearity and/or heteroscedasticity in regression analyses. Fourth, predictions based on sum scores do not allow for differences between unobserved (latent) subgroups in the prediction and/or measurement model.
These limitations can be solved by latent prediction models, a specific type of latent variable model involving both a measurement model and a prediction model. In each project we will focus on assumptions underlying such prediction models (e.g. linearity; multivariate normality) and use simulation studies to investigate how violations of these assumptions affect Type I & II errors and the accuracy of parameter estimates. By doing so, we not only aim to be psychometrically innovative, but we also strive to answer important substantive questions in medical psychology. We will specifically investigate the relationship between health and Type D personality. As this personality type can be modeled as a statistical interaction between its two subcomponents, our projects will involve the modeling of latent interaction effects.
Prof. Johan Denollet, Prof. Jelte Wicherts & Dr. Wilco Emons
4 January 2017 – 4 October 2020