The problem of counterfactual and control group is at the core of impact evaluation. Almost all existing methods aim to find the best control group to compare with the treated group. The aim of this study is to use imputation methods to estimate counterfactual and derive average treatment effect estimators from the data sets completed using the basic definition of treatment effect described in Rubin framework. The estimators obtained are called Imputation Based Treatment Effects estimators. A number of imputation methods are tested, among them there is Maximum likelihood, Multiple Imputation, Linear and Quantile regressions. Using simulations and bootstrap methodology, we found that the best imputation methods (data reconstruction) in the framework of impact evaluation are Quantile regression and Multiple Imputation. We also found that our estimators (taking average) obtained from data imputed are convergent and can perform as well as average treatment effects estimators obtained from classical methods such as Difference in Difference and Propensity Score Matching. Imputation Based Treatment Effect estimators are then tested on a program (Lalonde data) and the results show that they can perform as well as existing estimators and even better in certain cases especially when there is a shortage in data.
Mathematics Subject Classification: 62M10
Keywords: Average Bias, Estimator, Impact, Imputation, Treatment Effect