Multiple Imputation for Missing Data: Motivation and Examples

 

  • Prof. Donald Rubin – Harvard University
  • Instituto Nacional de Estatística, Lisboa – 14:30-16:30; 17:00-19:00
  • Segunda-feira, 3 de Junho de 2002
 
 Abstract: In the first part, I will introduce the multiple imputation for missing data perspective using simple examples. This part will also show why adhoc methods, such as fill in the mean, complete case, available case, and single imputation do not work in any generality. That is, in general they all lead to invalid statistical inferences. Weighting also cannot work in general. Moreover, this first part will show how multiple imputation does work in general, and that software to implement multiple imputation is now available. The second part will discuss particular real examples of the use of multiple imputation to address real problems: First, the problem of dropout in a randomized trial of a surgical device used to prevent adhesions, where the use of multiple imputation was important to getting the device approved by the US FDA (1); second, the planned use of multiple imputation to address a variety of problems, including noncompliance and dropout, in CDC (2) randomized trials of anthrax vaccine. If possible, time will be made available to consider new problems from the audience. (1) FDA : Food and Drug Administration; (2) CDC: Centre for Disease Control