| Christopher Swoboda
University of Wisconsin-Madison
A New Method for Multilevel Multiple Imputation: MLMI
FINAL REPORT:
Multilevel data with missing information across many variables is common in educational research. Inappropriate decisions about how to proceed with this missing data can lead to biased estimation, underestimated standard errors and, most importantly, incorrect conclusions. This paper presented a combination of strategies and improvements for adapting multilevel data analysis to a commonly used tool for missing data called multiple imputation. These ideas were combined into a method called multilevel multiple imputation (MLMI), which reduces the emphasis on eciency and parsimony in the imputation process and focuses instead on prediction and the multilevel structure of the data. A simulation study using real data from the National Educational Longitudinal Study of 1988 demonstrated substantial improvements over more common missing data options of listwise deletion and conventional multiple imputation (U.S. Department of Education, National Center for Education Statistics, 1990). These results illustrate the importance of embracing the multilevel structure of data when choosing a method for accommodating missing values in a multilevel data analysis.For researchers and policy makers, erroneous inferences can devastate time, resources and most importantly, the very people they are trying to help. With ubiquitous missing data in large-scale research and the increased emphasis on multilevel data collection and analysis, it is important for educational researchers to have access to accurate statistical tools. Understanding the relationships between important educational outcomes and characteristics and policies of students, teachers, schools, districts, states and countries is key to improving the educational system and giving students an opportunity for a better future. The MLMI procedure gives researchers a tool to provide more accurate assessments of these relationships when analyzing multilevel data with missing values.
Back to Funded Dissertation Grants Page
|