Performance of Imputation Methods towards Increasing Percentage of Missing Values

Abstract

The aim of this paper is to study the performance of eightdifferent existing imputation methodsused on simulatedand real dataset. The methods are compared in term of their ability to estimate the missing observationsand estimate some statistics(mean, standard deviation and coefficient of a regression) using the full data set completed by the imputation. The comparisons are made using root mean square error, mean absolute deviationand bias observed after estimation of statistics. Simulation results using specific simulated data and bootstrap show that Mean Imputation and Complete case analysisare the best method in completing the data set and in obtaining best estimators for statistics.However, the results are subject to major changes if parameters like sample size, number of replication and type of distribution chosen are modified. In short with real data, result will change depending on the structure of dataset to impute. For example, application of the simulation results to a Rwandan dataset on smallholder farmers revealed that k-NN is the best method in reconstructing and Multiple Imputation can be used as imputation method in case we are to estimate some statistics. Our final conclusion is that imputation methods cannot be compared since in most cases their performance is parametrically linked to the data. We finally proposed a methodology and a simulation protocol to identify for any data set which imputationmethod will give the best results and therefore should be applied in priority. Key words. Bias, Bootstrap, Imputation,Root Mean Squared Error, Mean Absolute error.