Yield Prediction Models For Brassica Juncea Based On Various Trait Selection Algorithms

Abstract

Residual analysis and various traits selection procedures are performed and compared on brassica juncea(mustard) data which consists of 33 genotypes and nine different morphological and yield related traits. In Residual analysis it was observed that genotype UCD-6/21and P53-126/40 have outlying morphological traits because of high values of DFI, DFC and PH, so it is suggested to brassica juncea breeders to control DFI and PH traits because these traits have negative impact and have insignificant effect towards oil content (OC%). Also genotypes UCD-10/15 and P118-R8 have outlying values of OC% and needed to genetically improvement and cannot be recommended for adoption at wide scale but in influential analysis it can be concluded that although genotype UCD-6/21and P53-126/40 are outlying with respect to morphological traits taken as fixed and genotypes UCD10/15 and P118-R8 are outlying w. r .t. OC% but none of them are influential observation in fitting of prediction models. Before developing prediction models based on various traits selection procedures, assumptions of linearity, normality of residuals, homoscedasticity of error variances, autocorrelation of residuals and multicollinearity were tested and found appropriate accept the multicollinearity. Due to the problem of multicollinearity in the data, ridge regression was applied as a remedy. Besides the development of prediction models, one of the major objectives was to point out morphological traits which effect negatively towards yield and it was observed that days to flowering initiation (DFI) has negative significant effect on OC% in all traits selection methods and it requires controlling it genetically by the breeders. Also days to flowering completion (DFC), YIELD and 100-SW have positive significant effect on OC%. In context of prediction models three models, based on Stepwise, PRESS combined with subset regression and ridge regression were developed, compared and finally one model was suggested on the basis of several measures of goodness of fit for prediction of OC % of brassica juncea genotypes. The effects of traits and the fitted models were verified by simulation (through program written in Minitab), in which 1000 random samples each of size 33 were generated using Normal, lognormal and exponential distribution and it was concluded that random samples generated through normal distribution gives close results to the original results as compared to lognormal and exponential distribution.