A Simulation Study Comparing Handling Missing Data Strategies

Missing data is a threat to the accurate reporting of substantive results within data analysis. While handling missing data strategies are widely available, many studies fail to account for missingness in their analysis. Those who do engage in handling missing data analysis sometimes engage in less than-gold-standard approaches. These gold-standard approaches: multiple imputation (MI) and full information maximum likelihood (FIML), are rarely compared with one another. This paper assess the efficiency of different handling missing data techniques and directly compares these gold-standard methods. A Monte Carlo simulation is performed to accomplish this task. Results confirm that under a missing at-random assumption, methods such as listwise deletion and single use imputation are inefficient at handling missing data. MI and FIML based approaches, when conducted correctly, provide equally compelling reductions in bias under a Missing at Random (MAR) mechanism. A discussion of statistical and time-based efficiency is also provided.