Comparison of EM Algorithm Imputation with Two Methods of Mean Imputation and New Samples Imputation in Panel Surveys

In economics and other social sciences, researchers are interested in modeling panel data where sample units iteratively are observed in different occasions. One application of panel data is the estimation of change mean rate of response variable over time. In all surveys especially panel surveys, nonresponse is a serious problem that mostly occurs in social science and medical data. This type of study usually encounters attrition in second wave and the waves after. Nonresponse causes bias and reduces efficiency of estimates. For adjusting this problem in panel survey, there are different “imputation” and “weighting” methods. One of the imputation methods is EM algorithm. In this article after introduction initial concept of panel survey, type of missingness in panel surveys and missing mechanisms, EM algorithm is introduced as a method of imputing missing data. Then by using of British Household Panel Survey data, EM algorithm imputation method is compared with two imputation methods in view of different criteria. Results show when correlation between two waves is high, EM algorithm has better performance than the others.