I imagine it is highly cited because (1) it articulates and deals with some very practical issues that are confronted by social scientists whenever they analyze data, and (2) it offers recommendations based on sound statistical thinking.

There are no new discoveries introduced in this paper. It describes missing data methodology, which is relatively new in the social sciences.

Two of us (Collins and Kam) are social scientists and one of us (Schafer) is a statistician. We work together at The Methodology Center; a research center devoted to social science methodology. Schafer is a noted expert on missing data. Collins and Kam were learning about missing data and wondered how you can compare multiple imputation and maximum likelihood approaches to missing data, and what are the benefits and drawbacks of each. All three of us had noticed that some social scientists were assuming that if they used a maximum likelihood approach their missing data problems were automatically solved, which of course is not necessarily true. In talking through these issues, we realized there was a need for an article clarifying these and other issues concerned with missing data.

When applied under similar circumstances, maximum likelihood and multiple imputation procedures for missing data analysis perform similarly. These two procedures differ most in how people tend to use them, specifically in whether people include auxiliary variables. Auxiliary variables are included in an analysis for the sole purpose of improving estimation related to missing data. The maximum likelihood procedure is usually associated with a restrictive approach to missing data, where very few, if any, auxiliary variables are included. By contrast, multiple imputation is associated with an inclusive approach to missing data, where many auxiliary variables are included. We showed that there are many benefits and very few costs associated with the inclusive approach, and recommended that social scientists adopt it. We also pointed out that there is nothing inherent in the maximum likelihood procedure that prevents the use of the inclusive approach; however, today's computer software for maximum likelihood applications make it difficult to include auxiliary variables, and documentation rarely alerts users to the benefits of doing so.

Linda M. Collins
Professor
The Methodology Center and Department of Human Development and Family Studies
Penn State
University Park, PA, USA