1078 — Differential Item Functioning in a Graded Response IRT Model: A Bayesian Approach to Item Discrimination

Seal P (Center for Health Quality, Outcomes, & Economic Research (CHQOER)), Eisen S (CHQOER)

Objectives:
Mental health survey instruments are often useful at diagnosing and summarizing the well-being of respondents. A typical survey involves respondents evaluating themselves on a number of items via a set of ordinal choices. If certain subgroups of patients with the same mental health status give systematically different responses to certain items, the instrument is said to exhibit differential item functioning (DIF). This talk develops a novel approach in diagnosing DIF from ordinal response data.

Methods:
Responses to mental health surveys are fit to item response theory models under a Bayesian framework. We focus in particular on Samejima's (1969) graded response model. Our approach involves positing two lack-of-fit diagnostics to examine departures from conditional independence, and then summarizing the posterior distribution of these diagnostics. The diagnostics are examined on a per-item basis, supplemented with graphical summaries.

Results:
We applied our approach to the analysis of responses by mental health patients from the BASIS-24, a widely-used self-report mental health assessment instrument, to study differences among Latino and white patients in item response patterns. We have found that DIF exists primarily in items relating to psychotic symptoms between Latino and white patients, as well as some suggestion of differences in the substance abuse domain.

Implications:
Our approach in estimating DIF through lack-of-fit diagnostics in a Bayesian framework allows for a flexible and fruitful way to identify differences between cultural and language groups in mental health self-report instruments.

Impacts:
The use of cutting edge diagnostic statistical tools can enhance the development of mental health self-report instruments for use within the VA health system.