Comparing Imputation Methods for Trait Estimation Using the Rating Scale Model.

This study examined the performance of four methods of handling missing data for discrete response options on a questionnaire: (1) ignoring the missingness (using only the observed items to estimate trait levels); (2) nearest-neighbor hot deck imputation; (3) multiple hot deck imputation; and (4) semi-parametric multiple imputation. A simulation study examining three questionnaire lengths (41-, 20-, and 10-item) crossed with three levels of missingness (10, 25, and 40 percent) was conducted to see which methods best recovered trait estimates when data were missing completely at random and the polytomous items were scored with Andrich's (1978) rating scale model. The results showed that ignoring the missingness and semi-parametric imputation best recovered known trait levels across all conditions, with the semi-parametric technique providing the most precise trait estimates. This study demonstrates the power of specific objectivity in Rasch measurement, as ignoring the missingness leads to generally unbiased trait estimates.