Regression-based nearest neighbour hot decking (Q1584178)

Imputation is typically used when missing item values need to be substituted with certain fabricated values in surveys or censuses. Numerous alternative imputation methods are mentioned in the literature, but, as the author asserts, the most current classifications are causing confusions, because some methods are, in fact, variations of a more general methods family. The author classifies the imputation methods into four main categories: 1) use of complete cases, when any missing items have not been imputed; 2) deductive or logical imputation, there is a known function (identity equation) between certain observed values and missing values; 3) imputed values are derived from a (behavioral) model, that is, imputed values may be non-observable in real life world; 4) imputed values are derived from a set of observed values, from a real donor respondent. This paper considers a complex imputation case and compares the three imputation methods with each other. The first method is random hot decking which assumes that the overall response mechanism is ignorable. The second method is based on the following two steps: (a) using logistic regression to impute whether the missing value is greater than, or equal to, zero; (b) for previous cases, using the standard regression imputation method. The third method is called regression-based nearest neighbour hot decking (RBNNHD), since it exploits both the regression model and hot decking. This method is of the highest interest in this paper. The paper also discusses a diagnostic test for the quality of imputations. This test checks how many times the same donor is used in imputing missing values.