Date of Award


Document Type

Open Access Dissertation


Educational Studies

First Advisor

Bethany Bell

Second Advisor

Robert Johnson


Problems of missing data are pervasive in social science research. Because of this, researchers have begun to use techniques after data collection to deal with missing data, including traditional methods (i.e. listwise deletion, pairwise deletion, and single imputation procedures) and modern procedures (i.e. multiple imputation and full information maximum likelihood). In the past, several organizations and researchers have warned that traditional missing data techniques (MDTs) can introduce bias into parameter estimates, and can result in a loss of statistical power (e.g., Becker & Powers, 2001; Wilkinson & the APA Task Force on Statistical Inference, 1999). However, previous research has shown that using a traditional method does not necessarily reduce statistical power or bias parameter estimates (Roth & Swizer, 1995). Research using traditional regression techniques has shown that sample size, percent of missing data, and missing data mechanism are key characteristics in determining under what conditions each MDT should be used. To further complicate matters, the multilevel modeling (MLM) literature has largely ignored the impact of missing data. Thus, it is not known if the results from single-level missing data research apply to hierarchical data. The present simulation study compare the performance of multilevel multiple imputation (MLMI) and listwise deletion in the context of linear two-level organizational models with continuous predictors. Design factors of interest included missing data technique, missing data mechanism, level-1 sample size, level-2 sample size, level-1 percent of missing data, and level-2 percent of missing data totaling 2,000 conditions. Design factors were evaluated on four outcomes, including bias, Type I error, statistical power, and confidence interval (C.I.) coverage. Results from this study showed that listwise deletion performed well for bias, level-2 Type I error rates, level-1 power, and C.I. coverage. Listwise deletion did have minor problems with Type I error rates at level-1, and power at level-2. MLMI performed well for level-2 Type I error rates, level-1 power, and level-2 C.I. coverage, but had minor issues with level-1 Type I error rates and level-2 power, and major issues with bias at both levels, and C.I. coverage at level-1. Recommendations for applied researchers based upon these results are discussed.