Multiple Imputation

In the United States, the modern survey sampling revolution began largely at the U.S. Census Bureau. The process, so the story goes, took off in the late 1930s when Jerzy Neyman, at the invitation of W. Edwards Deming, came to Washington and lectured at the U.S. Department of Agriculture (USDA 1937; Duncan and Shelton 1978). The Census Bureau work that led to the two-volume master piece by Hansen, Hurwitz, and Madow (HHM, 1953) actually started earlier, as America tried to respond as a country to the Great Depression (Stephan 1949). In any case it is from the per spective of the HHM book that I begin my discussion of the seminal work on imputation that Don Rubin launched in 1977. Those of you who grew up in another data-collection environ ment, like experimental design, may object to the narrowness of my focus. After all, other statistical traditions also tackled miss ingness, albeit in ways different from those used in surveys (e.g., Little and Rubin 2002). The effort to impute, or as it was called then, allocate for missingness began in the 1940 Decennial Census and was in full swing by the 1960 Decennial. One of the related factors that contributed to this advance was that the Cen