Statistical methods used for small plot analyses are unsuitable for large-scale on-farm experiments because they fail to take into account the spatial variability in treatment effects within paddocks. Several new methods have recently been proposed that are inspired by geostatistical analyses of spatially-varying treatment effects, which are typical for site-specific crop management trials with quantitative treatments. However, these methods do not address the objective of comparative experiments, where the overall assessment of treatments’ performance is of interest. Moreover, most biometricians, who routinely analyse data from field trials, are either unfamiliar with the new geostatistical techniques or reluctant to include these in their regular analytical toolkits due to the unavailability of easy-to-use software tools. The linear mixed model is widely used for analysing small plot field trials because it is extremely versatile in modelling spatial and extraneous variability and is accessible through user-friendly software implementation. Motivated by comparative experiments, conducted in large strip trials using qualitative treatment factors, and yield data obtained from harvest monitor, we propose a linear mixed effects model for determining the best treatment at both local and global spatial scales within a paddock, based on yield predictions and profit estimates. To account for the large spatial variation in on-farm strip trials, we divide the trial into smaller regions or pseudo-environments (PEs), each containing at least two replicates. We propose two approaches for creating these PEs. In the presence of appropriate spatial covariates, a clustering method is proposed; otherwise, the trial area is stratified into equal-sized rectangular blocks using a systematic partitioning scheme. PEs are used to estimate the treatment effects by incorporating treatment-by-PE interactions in linear mixed effects models. The optimum treatment within each PE is found by either comparing the best linear unbiased predictions solely or incorporating profit and comparing economic performance. To illustrate the applicability of our method, we have analysed two large strip trials conducted in Western Australia.

用于小地块分析的统计方法不适用于大规模的农场实验，因为它们没有考虑到围场内处理效果的空间变异性。最近提出了几种新方法，这些方法受到空间变化处理效果的地质统计分析的启发，这些方法对于具有定量处理的特定地点作物管理试验来说是典型的。然而，这些方法没有解决比较实验的目标，其中对治疗效果的整体评估很感兴趣。此外，大多数经常分析现场试验数据的生物识别学家要么不熟悉新的地质统计技术，要么由于没有易于使用的软件工具而不愿意将这些技术纳入他们的常规分析工具包中。线性混合模型广泛用于分析小地块田间试验，因为它在建模空间和外来变异性方面非常通用，并且可以通过用户友好的软件实现来访问。受比较实验的启发，使用定性处理因素在大型条带试验中进行，以及从收获监测器获得的产量数据，我们提出了一个线性混合效应模型，用于根据产量预测确定围场内局部和全球空间尺度的最佳处理和盈利预测。为了考虑到农场条带试验中的巨大空间变化，我们将试验划分为更小的区域或伪环境 (PE)，每个区域至少包含两个重复项。我们提出了两种创建这些 PE 的方法。在存在适当的空间协变量的情况下，提出了一种聚类方法；否则，试验区将使用系统分区方案分层为大小相等的矩形块。PE 用于通过将治疗与 PE 相互作用纳入线性混合效应模型来估计治疗效果。通过单独比较最佳线性无偏预测或合并利润并比较经济绩效，可以找到每个 PE 中的最佳处理方式。为了说明我们方法的适用性，我们分析了在西澳大利亚进行的两次大型带状试验。通过单独比较最佳线性无偏预测或合并利润并比较经济绩效，可以找到每个 PE 中的最佳处理方式。为了说明我们方法的适用性，我们分析了在西澳大利亚进行的两次大型带状试验。通过单独比较最佳线性无偏预测或合并利润并比较经济绩效，可以找到每个 PE 中的最佳处理方式。为了说明我们方法的适用性，我们分析了在西澳大利亚进行的两次大型带状试验。