Comparison between imputation method for handling missing data / Ayunie Ezadin, Nur Izzaty Chumin and Siti Nur Izzatulnisa Salit

This paper presents imputation method for the National Institute of Diabetes and Digestive and Kidney Diseases data from Arizona, United States. Missing data occurs in this data for five variables which are plasma glucose concentration, diastolic blood pressure, triceps skin fold thickness, serum insulin intake and body mass index (BMI). Missing data leads to problem that can cause bias and invalid conclusions to be made. This research objectives are to improve the data by filling the missing value and to compare which imputation method is better to handle missing value in a data set. In this research, imputation method and evaluation of the performance are applied for this data using Rstudio software. Five imputation methods used in this paper are Mean imputation method, K-Nearest Neighbour (KNN) imputation method, Multiple imputation method, Hot-Deck imputation method and Regression imputation method. The performance of these methods are evaluated using statistical analysis, coefficient of determination (R2), mean-squared eror (MSE), root of mean square error (RMSE), mean absolute error (MAE), index of agreement (d) and bias (B). Based on the result obtained from this research, it can be concluded that K-Nearest Neighbour imputation method is the best method among the five methods that are applied to handle the missing value. Conclusions are made as K-Nearest Neighbour (KNN) imputation method shows the best performance and has the lowest error value compared to other methods.