Predictive and Postdictive Analysis of Forage Yield Trials.

Field-based research programs are subject to numerous externalities that impact both precision and accuracy of phenotypic assessments made on cultivars, breeding lines, management systems, and evironmental factors.  Young researchers are faced with numerous critical decisions related to choice of equipment, test sites, experimental design, trial size, and trial duration, often without the benefit of experience, empirical results, or local knowledge. Many long-term research programs have access to huge stores of data that can be retrospectively queried to provide predictive answers to many of the critical questions related to these choices.  The purpose of this presentation is to illustrate a case-study of a 30-year research program, illustrating how postdictive queries and analyses can be employed to improve future precision and efficiency of treatment-mean estimation.  The study was based on 114 genetic experiments of 11 forage grass species conducted at Arlington, WI between 1981 and 2005.  Both incomplete block designs and postdictive spatial-autocorrelation analyses resulted in significant improvements to precision that were highly dependent on plot size, averaging 120% for 5.6-m² plots, 148% for 2.8-m² plots, and 226% for 1.4-m² plots.  The synergistic effects of reducing plot size and employing an incomplete block design or spatial analysis resulted in a complete reversal of the classic relationship between coefficient of variation and plot size: mean CV = 19, 13, and 11% for 5.6, 2.8, and 1.4-m² plots, respectively.  Power functions were found to be highly effective in predicting the number of replicates required to detect future differences in treatment means, particularly for highly sensitive and critical genetic experiments where small differences between treatments were expected.