Experimental Designs and Estimation Methods for On‐Farm Research: A Simulation Study of Corn Yields at Field Scale

• Published in: Agronomy journal 2019-11, Vol.111 (6), p.2724-2735
• Main Authors: Alesso, Carlos Agustín, Cipriotti, Pablo Ariel, Bollero, Germán Alberto, Martin, Nicolas Federico
• Format: Article
• Language: English
• ISSN: 0002-1962; 1435-0645
• DOI: 10.2134/agronj2019.03.0142

On‐farm experimentation using Precision Agriculture technology enables farmers to make decisions based on data from their fields. Results from on‐farm experiments depend on the experimental design and statistical analyses performed. Detailed information about the accuracy of the treatment effect estimates, and Type I error rates of hypothesis testing under different spatial structure scenarios attained by alternative experimental designs and analysis is required to improve on‐farm research experiments. Three thousand yield data sets were drawn from 15 random fields simulated by unconditional Gaussian geostatistical simulation technique and were modeled by applying 10 experimental designs and three estimation methods with experimental units ranging from 138 to 9969 m2. No effect of spatial structure, experimental design, and estimation methods was observed on overall mean yield and treatment bias. Unaddressed changes of nugget/sill ratio and range of variogram had a significant effect on estimator efficiency and accuracy with Type I error rates above the nominal rate, which increased with higher spatial autocorrelation. Spatial methods were robust to changes in spatial structure regardless of the design. Randomization of treatment increased the uncertainty of model estimators. In general, the accuracy of treatment effect estimates increased with the number of replications of smaller size. The opposite trend was observed between those estimates and the size of the plots. Analyses showed that the best designs for testing the overall treatment effect in two‐treatment experiments would be split‐planter, strip‐plots, and chessboard because of their size and number of experimental units. Core Ideas Spatial autocorrelation increases grand mean estimator variance in any design or method. Spatial autocorrelation reduces treatment effect estimator efficiency if not modeled. Spatial autocorrelation increases Type I error if not modeled. Designs with small experimental units (strip plots or chessboard) performed better.