Chapter 12: Analyzing Structural Equation Models with Missing Data

Researchers have traditionally dealt with missing data by discarding incomplete cases or by filling in missing scores with a single set of replacement values (i.e., single imputation). These ad hoc methods have fallen out of favor in the methodological literature because of their tendency to produce biased parameter estimates (Wilkinson & Task Force on Statistical Inference, 1999). In the 1970s, statisticians developed the underpinnings of two state-of-the art approaches, maximum likelihood missing data handling and multiple imputation (Beale & Little, 1975; Dempster, Laird, & Rubin, 1977; Finkbeiner, 1979; Rubin, 1978, 1987). However, as late as the mid-1990s, a lack of software options made it difficult to implement these techniques. Muthén, Kaplan, and Hollis (1987) had earlier proposed a multiple group approach that produced maximum likelihood estimates with incomplete data, but their model required complicated parameter constraints that became unwieldy with more than a few missing data patterns. Fortunately, the last decade has seen an explosion in missing data handling technology. All of the major structural equation modeling programs now implement maximum likelihood solutions, and a few also offer multiple imputation facilities. The methodology has grown to the point where most complete-data analyses now have a straightforward missing-data analog.