IRT Software for Health Outcomes and Behavioral Cancer Research
Type Research Project
Duration Sept. 30, 2008 - March 29, 2011
- Institute for Statistics and Mathematics IN (Details)
TagsPress 'enter' for creating the tag
- Mair, Patrick (Former researcher) Project Head
Health and behavioral cancer researchers require highly reliable, objective, and methodologically founded measurement instruments, accurate scaling of subjects on important health-related attributes, and statistically sound procedures for evaluating differences among groups (such as treatment vs control) and change across time. Item response theory (IRT) provides a powerful modeling framework for achieving these goals via the measurement of latent attributes that are only indirectly measured by observable data. Unfortunately, there is a massive lack of user-friendly IRT software that allows for a straightforward computation of a variety of IRT models in daily research applications.
The aim of this project is the development of flexible, user-friendly IRT software especially suited for researchers in the health sciences. This software will run on multiple platforms and cover a broad spectrum of IRT models such as classical binary models (Rasch, 1-PL, 2-PL, 3-PL), classical polytomous models (GRM, RSM, PCM, NRM), as well as up-to-date approaches such as models with covariates and multidimensional models that are highly relevant for health related research questions. The generalized models enable analysis of longitudinal and multilevel data, as well as examination of treatment group effects on a scale. Multidimensional models overcome the sometimes rather restrictive assumptions that require analysis of only one attribute at a time.
From a technical point of view, the program will offer numerous statistical estimation approaches for item and person parameters such as MML, nonparametric MML, fully nonparametric models, MCMC, Bayesian EAP, weighted likelihood, etc. Once the parameters are estimated, a researcher can evaluate the model by means of a large set of model tests and fit indices. Numerous interactive high-level plots will allow for a customizable visualization of the results, and an XML export will assure that tables and figures are publication quality.