Development and Extension of Item Response Theory Models for Extreme Response Style

In this proposal, three studies are proposed and the goal of the three studies aims to develop a new class of item response theory (IRT) models for extreme response style (ERS). Extreme response style is very common in the likert-type rating-scale items when attitude inventory and personality assessment are used to measure the degree of performance in respondents. Ignoring the possible extreme response style in the assessment can threaten test reliability and validity and results in the inferiority of scoring inference. The three studies are scheduled to carry out in the next three years because the three studies connect to each other coherently and each study will take a lot of time to deal with computational burden in the simulation designs. In the first, study, multiple tests measuring different latent traits are incorporated and the multidimensional IRT models for ERS are proposed. Four polytomous multidimensional IRT models are extended to incorporate the ERS tendency and thus the multiple latent traits and ERS tendency would be quantified. In the second study, a higher-order latent trait structure is considered and a new class of higher-order IRT models for ERS is developed because the hierarchical structure in latent traits and ERS phenomena may occur simultaneously. The author provides four higher-order IRT models for ERS and quantifies both the factor loadings of latent traits and ERS tendency. In the third study, a mixture latent distribution is incorporated and the mixture IRT models for ERS are proposed because not all independents will show ERS in the responses to test items. For the three studies, a series of comprehensive simulations will be conducted to assess the parameter recovery under diverse manipulations using Bayesian estimation. The pilot studies showed that each study is feasible because the parameter recovery was satisfactory in a small number of simulations. In addition, three empirical studies that are derived from the Survey Research Data Archive (SRDA) will be provided to demonstrate the applications of the three new models to real data analysis.