Advances in Computerized Adaptive Measurement of personality

Resumen

Personality traits remain a primary focus of study in many psychological areas. Notwithstanding the advances achieved with the consolidation of the Big Five model as a common framework of study, personality assessment still presents some limitations that need to be addressed. First, traditional paper-and-pencil questionnaires are quite long for modern evaluation settings where several instruments are administered or testing time is very limited. Second, although some attempts have been made to measure personality more efficiently through computerized adaptive testing (CAT), they have completely ignored the hierarchical nature of domains and facets of personality traits. Third, most personality research and assessment relies on self-report measures, which is well known are sensitive to the influence of item wording effects that can distort research results. Accordingly, this dissertation sought to address these limitations by means of three studies. Study 1 presents the process of construction and calibration of a wide pool to measure the Big Five facets. Results from a post-hoc simulation study demonstrated that the adaptive administration of the items produced accurate facet scores using only a third of the total of the items in the pool. Study 2 goes one step further and illustrates the construction of a CAT based on the bifactor model, which allows to approach the study of the Big Five while considering its hierarchical nature. A post-hoc simulation study demonstrated that the CAT based on the bifactor model is more advantageous to assess the Big Five personality traits than other traditional competing approaches. Finally, Study 3 used Monte Carlo methods to evaluate the impact of three types of item wording effects (careless, item verification difficulty, and acquiescence) on person score estimates and other aspects (model fit, factor loadings, and structural validity) in the context of unidimensional fixed-length texts. Two models were evaluated to this end: the random intercept item factor analysis (RIIFA) model and the traditional model with one substantive factor (1F). Results revealed that, although the RIIFA model was consistently superior in terms of model fit to the 1F model, it was not able to better estimate the uncontaminated person scores and other parametes for any type of wording effect than the 1F model. In conclusion, the three studies included in this dissertation provided a series of tools to measure personality traits more efficiently and contributed to the advancement of knowledge in the area of wording effect measurement