Abstract

Missing data are often a problem in social science data. Imputation methods fill in the missing responses and lead, under certain conditions, to valid inference. This article reviews several imputation methods used in the social sciences and discusses advantages and disadvantages of these methods in practice. Simpler imputation methods as well as more advanced methods, such as fractional and multiple imputation, are considered. The paper introduces the reader new to the imputation literature to key ideas and methods. For those already familiar with imputation methods the paper highlights some new developments and clarifies some recent misconceptions in the use of imputation methods. The emphasis is on efficient hot deck imputation methods, implemented in either multiple or fractional imputation approaches. Software packages for using imputation methods in practice are reviewed highlighting newer developments. The paper discusses an example from the social sciences in detail, applying several imputation methods to a missing earnings variable. The objective is to illustrate how to choose between methods in a real data example. A simulation study evaluates various imputation methods, including predictive mean matching, fractional and multiple imputation. Certain forms of fractional and multiple hot deck methods are found to perform well with regards to bias and efficiency of a point estimator and robustness against model misspecifications. Standard parametric imputation methods are not found adequate for the application considered.