Porthole
Item-Level Response Shift Detection using Item Response Theory Analyses with the Graded Response Model (307887)
*Olawale Fatai Fatai Ayilara, Department of Community Health Sciences, University of ManitobaLisa M Lix, University of Manitoba
Lara Russell, Centre for Health Evaluation and Outcome Sciences, University of British Columbia
Tolulope Sajobi, Department of Community Health Sciences, Cumming School of Medicine, University of Calgary
Richard Sawatzky, School of Nursing, Trinity Western University
Keywords: Response shift, likelihood ratio test, full information maximum likelihood
Analyses of longitudinal change in patient-reported outcomes (PROs) can be confounded by the presence of response shift (RS). Our objective was to demonstrate the longitudinal graded response model (GRM) to test for RS and estimate change in latent variable with ordinal items. PRO data were collected before and after surgery using the 12-item Short Form Survey, version 2; we focused on the 6-items that measure physical health (PH) for this demonstration. We compared two measurement models to detect RS using the likelihood ratio test (LRT). The parameters of the model were estimated using the full information maximum likelihood. The overall RS test was statistically significant (LRT = 3068.62, df = 24, p < 0.001), which indicates the presence of RS. The standardized estimated PH mean for the post-operative occasion (pre-operative occasion means were constrained to zero and variance of one) was 1.02, which coincides with the change in PH means before accounting for RS. Further analyses will test for reprioritization and recalibration RS. The proposed method allows for detection of RS at the item level and has the ability to produce unbiased estimates when missing data are ignorable.