Removing Spatial Variation from Wheat Yield Trials: A Comparison of Methods

Abstract

Most cultivar evaluation trials use blocked designs and are analyzed using classical analysis of variance; however, a standard analysis for blocked designs often does not adequately account for spatial variability. Recent advances in spatial statistics suggest that there are better alternatives. The primary objective of this research was to compare randomized complete block (RCB) analysis with two nearest neighbor adjustment (NNA) methods; a random field procedure was also examined, as another way to remove spatial variability. Yield data from three replicated breeding nurseries involving diverse adapted and unadapted germplasm, each grown at four locations in Nebraska during 1988–1989, were used for the comparisons. The NNA approach was superior for all nurseries at all locations, on the basis of lower coefficient of variation and greater ability to distinguish cultivar differences. For 9 of 12 trials, RCB cultivar means were highly correlated with NNA cultivar means, indicating that both procedures would identify similar elite lines. For three trials, however, the estimated cultivar yield and cultivar rank from RCB and NNA procedures were different, and so different lines would be selected. The random field analysis was applied to the trial for which the discrepancy between the RCB and NNA analysis was greatest; this yielded results similar to the nearest neighbor analysis. Our results suggest that spatial trends are common in Nebraska and compromise the accuracy and precision of standard analysis of blocked designs. Therefore, NNA or random field analysis should be used to improve the analysis of breeding trials.

Spatial statistical analysis and selection of genotypes in plant breeding

“…The residuals were then used to estimate the spatial correlation structure. This was done graphically by means of a so-called semivariogram or simply variogram (Stroup et al, 1994).…”

“…Under stationarity -spatial law unaffected by translation -the variogram has a direct and simple relation to the function of autocovariance C(h), that is: S(h) = σ 2 -C(h); in which σ 2 = C(h = 0) (Es & Es, 1993;Stroup et al, 1994;Pannatier, 1996). Thus, fitting a continuous model to the sample variogram, the corresponding spatial covariance function for this relation is obtained.…”

“…This increases the chance of violating the independence among observations assumed when using the OLS method, due to the likely similarity of observations of neighboring plots (Stroup et al, 1994). This phenomenon, referred to as spatial correlationalso called spatial dependence or autocorrelation -can seriously affect the comparison of treatments.…”

“…The traditional analysis of variance relies on randomization to neutralize the harmful effects of this type of correlation, but frequently this is not attained adequately (Stroup et al, 1994). For this reason, Kempton et al (1994) support a greater use of methods that consider some accounting for spatial dependence to improve the precision of variety trials.…”

Evaluation of experimental designs in durum wheat trials

“…In this design treatments must be arranged all together in each group, but the number of treatments usually cannot exceed 20, because heterogeneity increases within the block and consequently experimental error increases too (Stroup et al, 1994;Stringer et al, 2012). To overcome this problem, a new category of designs has been developed with each group divided into smaller more homogeneous subgroups that do not contain all treatments.…”