A Comparison of Rounding Methods for Categorical Data Using Multiple Imputation

• In Event: Poster Session 7
  In Poster Session: Division D Section 2 Poster Session

Abstract

When variables are categorical, multiple imputation based on multivariate normal model will replace the missingness with continuous values. Rounding methods aim to transform the continuous values onto ordinal scales. The current study investigated the performance of four rounding methods: crude rounding, coin-flip rounding, adaptive rounding and calibration-based rounding. We also examined at how many categories the rounding issues will not be a concern. A one-factor CFA model was employed for data generation and analysis. Missing values were created following missing at random. We varied the sample size (100, 300), item loading (.4, .8), number of categories (2, 4, 6, 8, 10), data distribution (symmetric, positively skewed) and percentages of missingness (25%, 50%). Parameter estimation bias and efficiency were evaluated.

Authors

• Yan Xia, Florida State University

• Yanyun Yang, Florida State University