Nicklas Pettersson: Real donor imputation improved by kernel methods

Imputation is a general tool for dealing with missing data. Auxiliary information is used to predict the missing data, which is then filled in as to create seemingly ‘complete’ datasets.

Real donor imputation uses copies of values that have already been observed to ‘replace’ the missing values, as opposed to model donor imputation where the generated imputed values may be non-observable in real life.

Real donors can be used with any model but are often used in combination with nonparametric methods, typically hot-deck imputation. Such methods are usually less sensitive to model misspecifications, but make implicit assumptions through distance metrics and variable choices.

A benefit from using real donors is that natural possible values are imputed, but the method relies on having observed enough potential donor units, and has problems at boundaries of the data. We try to mitigate these weaknesses through several methods related to kernel estimation. In the seminar I will show how the methods behave under several scenarios.