Graded response model-based item selection for behavior and symptom identification

In measuring outcomes of health care, information is obtained from subjects employing instruments that often use Likert scales. These instruments are typically designed using classical testing theory which assumes the errors around the true scores are normally distributed and constant. Advances in psychometric practices through the use of item response theory (IRT) models have led to more flexibility in scale development and in data analyses. In this paper, we introduce statisticians and health services researchers to IRT models through a case-study of data collected to measure subjective distress. The data consist of self-reports of symptom and problem difficulty obtained from a sample of 2,656 patients discharged with a psychiatric disorder from 13 hospitals in the United States between May 2001 and April 2002. Dimensionality of the trait is assessed using principal factor analysis. Model assessment is made using χ2 statistics and residual analyses. We select items for the scale using the Fisher Information available at selected levels of the underlying trait. © Springer Science+Business Media, LLC 2006.