Brule-Babel: Advanced Statistical Analysis of Strip-Plot Canola Variety Trial Data and Comparison to Small-Plot Variety Trial Data
Date: March 2014
Term: 3 years
Status: Completed
Researcher(s): Anita Brule-Babel, Gary Crow, Lyle Friesen, University of Manitoba, Winnipeg MB
SaskCanola Investment: n/a
Total Project Cost: $54,995
Funding Partners: n/a
Project Summary
Every year in western Canada there is a large investment in testing of canola genotypes/varieties in both small-plot and larger-scale trials, or strip plot trials. In 2011, a three-year project was launched using a Mixed Model analysis to investigate the question of how well the small-plot and strip trial results correspond with respect to ranking of variety performance. Overall, the project showed that a Mixed Model analysis of canola small-plot and field-scale/strip trial data is appropriate, provides variety yield estimates that appear to be accurate even when data are limiting, and provides the opportunity to compare varieties that were not tested together.
Every year in western Canada there is a large investment in testing of canola genotypes/varieties in both small-plot and larger-scale trials, or strip plot trials. The intention of crop variety small-plot performance testing is to predict how the variety will perform in commercial fields. Researchers and industry wanted to know how well the small-plot and strip trial results correspond with respect to ranking of variety performance (yield).
In 2011, a three-year project was launched based on Canola Performance Trials (CPT) strip trial yield data to investigate this question. Researchers conducted a Mixed Model analysis on the data using the Mixed Model statistical computer software program ASReml, which has been designed/optimized to accommodate large unbalanced datasets. The objectives of the project were to: apply modern statistical techniques (Best Linear Unbiased Prediction methods (BLUP)) to variety performance data; explore statistical techniques for evaluating environmental (location/year) and genotype by environmental (GxE) effects, and assess the effect of the predictions over the current single-year, single check comparisons. Researchers also wanted to develop guidelines that the MCVET committee could use for future testing.
Researchers included datasets and results from other companion projects in the final report, in order to have the best dataset available for analysis. The first project was to subject small-plot data to the Mixed Model analysis before any comparisons could be made between strip trial and small-plot results. This second comprehensive strip trial project included and encompassed the small-plot canola data analysis and the comparison to the small-plot Mixed Model results. The small-plot and strip trial datasets were matched by variety prior to analysis; this resulted in relatively large datasets of 5,210 and 4,344 datalines, respectively. The variety-matched datasets were subjected to Mixed Model statistical analysis, and variety yield estimates (BLUP's) were compared. Researchers also invited a number of companies to submit several years of recent canola strip trial yield data, which was compared to small-plot data from a number of sources. Also included was a summary of a comparison of canola yield small-plot results to commercial field results (Manitoba Agricultural Services Corporation, MASC data).
From the study and comprehensive analysis, the results showed that a Mixed Model analysis of canola small-plot and field-scale/strip trial data is appropriate and provides variety yield estimates that appear to be accurate. Currently, the results of CPT yield data analysis are presented as arithmetic means (in the 'Seed Manitoba' publication). However, mathematical and statistical theory indicates that the Mixed Model analysis (or least-squares linear models) will always provide better or equal results to an arithmetic mean based approach. The advantage of Mixed Model analysis and adjusted 'means' (BLUP estimates) over arithmetic means becomes apparent where data is limiting and/or the year (growing season weather which influenced yield) was unusual as compared to a 10-year mean yield.
Based on the results of comparing canola yield small-plot results to larger scale growing environments (CPT strip trial, MASC commercial field), the results indicate that small-plot results are not a perfect predictor of variety performance under larger scale growing conditions. In both the CPT small-plot/strip trial and the small-plot/MASC commercial field comparisons, the performance (yield) for approximately one-fifth of the varieties differed by 5 per cent or more between the growing environments. This difference in performance may be related to growing conditions that are unique to small-plot trial environments. If small-plot production/agronomic/growing conditions can be altered to more closely reflect actual commercial field conditions, then the predictive accuracy of small-plot variety testing might be improved. (For more information: https://umanitoba.ca/faculties/afs/agronomists_conf/media/Brule-Babel_Pres_Dec_13_2012.pdf)
Overall, the project showed that a Mixed Model analysis of canola small-plot and field-scale/strip trial data is appropriate and provides variety yield estimates that appear to be accurate. The Mixed Model analysis provides the opportunity to compare varieties that were not tested together, and relative variety performance is better represented even when data are limiting. The model analysis of crop variety trial data also indicated that approximately five years of data seems to be optimum for analysis, including more than one year of testing for a number of varieties in the multi-year dataset. As a result of the study, Seed Manitoba began implementing the use of BLUP data in 2013 for the main variety comparison tables.