Imputation

Imputation, also called ascription, is a statistical process that statisticians, survey researchers, and other scientists use to replace data that are missing from a data set due to item nonresponse. Researchers do imputation to improve the accuracy of their data sets.

Missing data are a common problem with most databases, and there are several approaches for handling this problem. Imputation fills in missing values, and the resultant completed data set is then analyzed as if it were complete. Multiple imputation is a method for reflecting the added uncertainty due to the fact that imputed values are not actual values, and yet still [Page 323]allow the idea of complete-data methods to analyze each data set completed by imputation. In general, multiple imputation can lead to valid inferences from imputed data. Valid inferences are those that satisfy three frequentist criteria:

• 1.
  Approximately unbiased estimates of population estimands (e.g. means, correlation coefficients)
• 2.
  Interval estimates with at least their nominal coverage (e.g. 95% intervals for a population mean should cover the true population mean at least 95% of the time)
• 3.
  Tests of significance that should reject at their nominal level or less frequently when the null hypothesis is true (e.g. a 5% test of a zero population correlation that should reject at most 5% of the time when the population correlation is zero)

Among valid procedures, those that give the shortest intervals or most powerful tests are preferable.