Collapsing or not? A practical guide to handling sparse responses for polytomous items

In ordinal data analysis, category collapse is the process of combining adjacent response options to create fewer response categories than were originally measured. When collapsing response categories, researchers need to be aware of inducing data-model misfit and of obtaining biased parameter estimates. Through mathematical derivation we show that category collapse induces data-model misfit when using Generalized Partial Credit IRT model (GPCM) generated data. This data-model misfit is not present when using Graded Response IRT model (GRM) generated data. Using simulation studies, we found that category collapse can indicate better data-model fit in GRM- and GPCM-generated data. In the case of GPCM data, this result is spurious and can lead practitioners to draw conclusions from models that do not fit the data well. Recovered GPCM IRT item parameters were also significantly biased. Recommendations for practitioners who wish to collapse categories are provided. peerReviewed publishedVersion