Imputing Missing Values Using Support Variables with Application to Barley Grain Yield

In any data collection process, regardless of the sampling method, missing data values are encountered due to many different reasons. Depending on the amount of missing data the results to be obtained from the analysis of such data will somehow be affected. Therefore, starting from 1950s an increasing interest is shown by statisticians on one hand how to minimize the missing data values and also how to impute the missing values. In this thesis the theory and methods employed so far for the imputation of missing values in a data set are studied in detail. This is followed by the introduction of a new concept in the imputation of missing data using the support variables as part of multivariate regression process. Conversion of the units of support variables to that of the response variable is very important and is studied in detail via the imputation of missing values in a barley grain yield data set. Application results of the support variable concept is comepared with the results obtained from Markov Chain Monte Carlo (MCMC), Gaussian and Epanechnikov Kernel regression and found to be better performer in terms of lower error levels and in terms of robustness. The robustness of the results of all methods are checked using the Relative Aitchison Distance (RDA) concept. Keywords: Missing Value, Imputation, Support Variables, Mean Squared Error (MSE), Regression, Correlation, Kernel Regression.