Comparing the Performance of Different Missing Data Imputation Approaches in Discrete Outcome Modeling

Abstract

Although several approaches exist for data imputation, these approaches are not commonly applied in transportation. The current paper is geared toward assisting transportation researchers and practitioners in developing models using datasets with missing data. The study begins with a data simulation exercise evaluating different solutions implemented for missing data. The dimensions considered in our analysis include: the nature of independent variables, different types of missing variables, different shares of missing values, multiple data sample sizes, and evaluation of single imputation (SI), multiple imputation (MI) and complete case data (CCD) approach. The comparison is conducted by adopting the appropriate inference process for the MI approach with multiple realizations. From the simulation exercise, we find that the MI approach consistently performs better than the SI approach. Among various realizations, the MI approach with five realizations is selected based on our results. The MI approach with five realizations is compared with the CCD approach under different conditions using model fit measures and parameter marginal effects. In the presence of a small share of missing data, for larger datasets, the results suggest that it might be beneficial to develop a CCD model by dropping observations with missing values as opposed to developing imputation models. However, when the share of missing data warrants variable exclusion, it is important and even necessary that the MI approach be employed for model development. In the second part of the paper, based on our findings, we implemented the MI approach for real empirical datasets with missing values for four discrete outcome variables.