Optimal designs for Thurstonian IRT models

• Applicants: Dr. Heiko Grossmann; Professor Dr. Heinz Holling; Professor Dr. Rainer Schwabe
• Subject Area: Personality Psychology, Clinical and Medical Psychology, Methodology
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 538403948

This proposal is concerned with the development of optimal experimental designs for the estimation of personality traits (trait scores) based on Thurstonian IRT models. These models use forced-choice response formats in the form of paired comparisons or paired comparisons derived from choice sets. Compared to Likert scales, this leads to much greater robustness against biases caused by socially desirable response behaviour. In other words, faking is only possible to a much lesser extent. Previous simulation studies have consistently shown that in Thurstonian IRT models trait scores can only be estimated with satisfactory accuracy if the respondents are presented with item pairs or item sets that consist of mixed-keyed items, i.e. both positively keyed items and negatively keyed items. However, the use of mixed-keyed items has the serious disadvantage that the items in question may differ significantly in terms of social desirability and the measurement of trait scores can therefore be easily distorted, which contradicts the basic intention of using Thurstonian IRT models. It is therefore desirable to use items, which a) counteract potential faking and b) allow a sufficiently precise estimation of the trait scores. In order to meet requirement a), corresponding item pairs or larger choice sets should only contain equally keyed items. Optimal experimental design theory can be used to simultaneously achieve b). The aim of the present proposal is to develop optimal experimental designs for pairwise comparisons and larger choice sets that allow sufficiently good estimation of the trait scores in Thurstonian IRT models; whereby only items that are keyed in the same direction are permitted. The aim is to determine designs in which the number of paired comparisons or choice sets is not too large and which can be used in practice. Optimal experimental designs will be developed for Thurstonian IRT models with metric, binary and ordinal responses, and also for binary paired comparisons derived from rankings of three or more alternatives. In order that the knowledge gained from the planned project can be implemented in practice in a user-friendly way, a corresponding R program and Shiny app will be developed. Initial optimal experimental designs derived specifically for this application are promising that suitable designs for practical applications of Thurstonian IRT models can be developed.

• DFG Programme: Research Grants