|KIMBENG, COLLINS - LSU Agcenter
|PONTIF, MICHAEL - LSU Agcenter
|SEXTON, DAVID - LSU Agcenter
Submitted to: American Society of Sugar Cane Technologists
Publication Type: Abstract Only
Publication Acceptance Date: 5/10/2018
Publication Date: 9/1/2018
Citation: Kimbeng, C., Todd, J.R., Pontif, M., Sexton, D., Dufrene Jr, E.O. 2018. Minimizing the influence of genotype by environment interaction on selection decisions in a sugarcane variety trial [abstract]. Journal of the American Society of Sugar Cane Technologists. 38:58.
Technical Abstract: The advanced stages of most plant breeding trials is usually comprised of a number of highly selected genotypes evaluated in a number of environments representing a much larger target population of environments. Genotype x environment (GxE) interaction is a widely recognized phenomenon in these types of trials. Genotype x environment interaction complicates selection decisions especially when it results in changes in the ranking of experimental genotypes in different environments. Sugarcane being a perennial crop, the environment main effect (E) in the GxE matrix is further complicated because it constitutes a combination of location, year and crop age (plant and several ratoon cane crops). The effects of year and crop are often confounded as each crop stage is grown in a different year and data in the first and second ratoon crops are taken (repeatedly) on the same plots as the plant-cane crop in order to assess ratooning ability. It is possible that several assumptions in the analysis of variance (ANOVA) are violated by this data structure, which if true, could obscure the true interpretation of results from the ANOVA. This paper presents an exploratory but systematic approach to analyze data from advanced stage selection trials in sugarcane. Data on sugar yield (sugar per acre, SPA) collected from 13 genotypes evaluated in three locations over 2 crop-years were used to demonstrate the concept. First the data were analyzed for normality using the Proc Univariate procedure in SAS and found not to follow a normal distribution (Pr > = ISI < 0.001). Pattern analyses, including factor, principal component and cluster techniques, which describe environmental similarities based on genotype response were used to analyze the data. These quasi non-parametric approaches, unlike the ANOVA, are less susceptible to the assumption of normality. The three locations and crops within each location formed three distinct groups or clusters, suggesting a strong location effects in this data set. As expected, analysis of the data using the Proc GLM procedure in SAS confirmed the location effect to be significant (P = 0.01) while the crop-year effect was not (P > 0.05). Next the data were analyzed using the Proc Mixed procedure in SAS with genotype (G) as fixed effect and environment main effects (location (L), crop-year (C)) and the G x E interactions (GL, GC, GLC) as random effects in the model. The G main effect was significant (P < 0.01) and the variance due to L was more than 2-fold larger than that attributed to GL, GC and GLC. We then fitted the model with location as a repeated measure allowing for heterogeneity of the location error variances. Fitting location as a repeated measure improved the goodness of fit of the model as the Akaike information criterion (1397 vs. 41.6) and Bayesian information criterion (1396 vs. 39.7) values reduced substantially and a chi-square test of the difference in the two -2 restricted log likelihood ratios was significant (P < 0.01), suggesting that the error terms are not homogeneous across locations. Based on this exploratory analyses, it is obvious that observations in this data set are not independent, normally distributed and the variances are not homogeneous which violates the basic assumptions of the ANOVA. In an attempt to reduce the influence of the location main effect (location was found to be the major cause of GxE in this data set), we analyzed each location separately and the mean, minimum, maximum and standard deviation (sx,s) values from each location were used to adjust each data point (Xi) using two methods. Each data point was normalized (SPAn) which scaled the data to a range of 0 to 1 and standardized (SPAz) which scaled the data to a mean of zero and a unit variance. A normality test using the Proc Univariate routine in SAS revealed that SPAz was normally distributed (Pr > = ISI < 0.