Comparative performance of imputation methods for different proportions of missing data in classification of crop genotypes

Abstract

Most crop datasets contain missing values, a fact which can cause severe problems in the analysis and limit the utility of resulting inference. Classification techniques for grouping of crop genotypes are used when the data is complete. However, the presence of missing values limits the utility of these techniques and creates bias in the resulting inferences. In majority of the cases, missing values are handled by deleting the genotype or traits which contain missing values there by losing information on these genotypes. An interesting approach to handle this problem is to impute the missing values. In this paper, we provided some solutions to handle missing data in crop breeding experiments for classification of crop genotypes.The performance of the imputation techniques is assessed by using the hit ratio criteria computed through four different classifiers by using extensive simulation procedure. This paper has also attempted to provide a description of missing data mechanism in agricultural experiments and various imputation techniques for missing data analysis in classification problems. For lower proportions of missing data, all four of the imputation techniques provided satisfactory results for classification of crop genotypes. For moderate level of missingness in the data, regression and multiple imputation techniques provided same levels of precision for classification of crop genotypes. When there is a high proportion of missing data, multiple imputation technique outperformed all imputation techniques for classification of crop genotypes. Among the classifiers, k-th nearest neighbor is the best classification technique in missing data situations.