Missing Data: Comparison of Imputation Methods on the Example of RLMS

Current research states that in the case of various methods of data imputation it is crucial to compare them and apply for the elimination of biases, losses of representation, and quality losses of data analysis techniques. The main idea is to indicate the significance of imputation process and to compare methods within given data conditions. This study investigates the performance of data imputation using simple mean imputation, regression imputation, Expectation-Maximization (EM) algorithm, and Multiple Imputation by Chained Equations (MICE) method. These techniques are applied to generated samples in order to compare the effectiveness of chosen methods. The Russia Longitudinal Monitoring Survey – Higher School of Economics (RLMS-HSE) dataset is used for analysing and comparing the effect of working experience on earnings between samples with applying imputation techniques and without using them. It was shown that MICE and regression imputation are the most effective methods among chosen ones with the use of generated samples. However, regression imputation may inflate correlations between variables. There were almost no differences between the coefficients’ significance of full imputed results and imputed results of independent variables. Also, the outcomes of the four methods had no differences in the significance of the coefficients.