Title
|
Room
|
Multiple Imputation for Missing Data
|
M-International Salon E
|
Date / Time
|
Sponsor
|
Type
|
08/05/2001
8:00 AM
-
4:00 PM
|
ASA, Section on Statistics in Epidemiology*
|
Other
|
Organizer:
|
n/a
|
Chair:
|
n/a
|
Discussant:
|
|
CE Presenter
|
Donald Rubin
Donald Rubin
John Barnard
John Barnard
|
Description
Missing data are a pervasive problem in many scientific disciplines. Multiple imputation (MI) is a method of addressing such missing data. Multiple imputation replaces each missing value by a set of plausible values (for example, five values) that represent the uncertainty about the right value to impute (I.e., fill in). Special software is needed to create multiple imputations in large data sets, but such software is now becoming available. Multiply-imputated data sets are analyzed using (1) standard complete-data software (just as if there were no missing values) and (2) simple general-purpose macros that combine the results of complete-data analyses in essentially the same way, no matter which complete-data analysis is used. This process results in valid statistical inferences that properly reflect the loss of information due to the missing values: consistent estimates, valid p-values, and valid confidence intervals. Although originally designed two decades ago to tackle missing data problems in public-use sample surveys, MI is gaining popularity in many fields, particularly in clinical trials, because of its general validity, flexibility, and simplicity. In this course we will introduce multiple imputation, show why it is generally superior to ad-hoc methods, and describe and demonstrate how to generate multiple imputations.
|
|