Predictive Unfolding Models for Single-Peaked Items with Binary and Graded Response Data
The project will develop and extensively test new models for the prediction of individual attitudinal responses, preferences, emotional and behavioral tendencies as measured in questionnaires. A common basic assumption of unfolding models is that the probability of endorsement of an attitude item or the occurrence of a behavioral tendency is a single-peaked function of the underlying scale being measured.
When background variables for persons and/or design characteristics for items are available, we can incorporate them in the model and then predict and potentially explain the response of new persons with known background profiles and/or interpolate their response tendency to items not previously measured. The explanatory version of this approach represents an alternative to Structural Equation Modelling (SEM) in problems characterized by a unidimensional Latent Variable (LV) data structure.
This approach will improve the assessment of items in attitudinal and diagnostic assessment for item bank building and optimization, as items can be screened to meet ad-hoc specifications (such as a regular coverage of the LV dimension). In marketing and opinion survey research this approach can be used to rank items in terms of appeal to different types of audiences. Person response assessment is also improved by a more accurate and flexible specification of item features.
A common simulation framework for experimental testing is used for all models and estimation strategies, and the software developed for model estimation and hypothesis testing will be integrated in a user-friendly R-package.
Prof. M. De Rooij (Leiden University)
Prof. W.J. Heiser (Leiden University)
14 February 2015 – 14 February 2019