What Do We Do When Data Are Missing on Multiple Variables?

• In Event: Roundtable Session 117
  In Roundtable Session: Research on Accommodating Missing Data

Abstract

Recently, analysts have become more concerned over handling missing data. However, current literature on imputation techniques is limited: studies only consider consequences with data missing on one variable. This study compares the relative bias of commonly used and recommended techniques (Listwise Deletion, Multiple Imputation) with data missing on multiple variables. Factors varied in this study are: type (MCAR, MAR) and degree of missingness (10%, 25%, and 50%), and the number of variables where missing occurs. This study uses real data, deleting cells to create realistic scenarios of missingness across variables. Parameter estimates from LD and MI are compared to estimates from the full dataset. Results showed that, overall, MI performed better than LD with less biased regression coefficients and R2.

Authors

• Jonathan D. Rubright, American Institute of Certified Public Accountants

• Ratna Nandakumar, University of Delaware

• Joseph J. Glutting, University of Delaware