An Example of Handling Missing Data (Correctly and Incorrectly)


  • Prof. Donald Rubin – Harvard University
  • Faculdade de Psicologia e de Ciências da Educação Universidade Lisboa – 15:00
  • Terça-feira, 4 de Junho de 2002
 Abstract: The correct way to handle missing data is not obvious. This talk will present a simple example of a randomized experiment in which there were missing outcomes, but more missing in one treatment group than the other. How these missing values were addressed made a big difference to the answers. In this real example, handling them the way a government agency (the US FDA) initially wanted led to the conclusion that the experimental treatment wasn’t any better than the control, and possibly worse with respect to negative side effects. Handling the missing values in a more scientific way led to the conclusion that the experimental treatment was much superior. The scientific analysis used multiple imputation. Creating multiple imputations for the missing values revealed uncertainty about what the missing outcomes really were. The agency approved the new treatment for use. This real example illustrates several important ideas about the right way to handle missing data.