Proposed method for statistical snalysis of on-farm single strip treatment trials

Authors: CHO, JASON - Cornell University; GUINNESS, JOSEPH - Cornell University; Kharel, Tulsi; MARESMA, ÁNGEL - Cornell University; CZYMMEK, KARL - Cornell University; VAN AARDT, JAN - Rochester Institute Of Technology; KETTERINGS, QUIRINE - Cornell University

Submitted to: Agronomy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/7/2021
Publication Date: 10/12/2021
Citation: Cho, J.B., Guinness, J., Kharel, T.P., Maresma, Á., Czymmek, K.J., Van Aardt, J., Ketterings, Q.M. 2021. Proposed method for statistical snalysis of on-farm single strip treatment trials. Agronomy. 11(10):2042. https://doi.org/10.3390/agronomy11102042.
DOI: https://doi.org/10.3390/agronomy11102042

Interpretive Summary: Randomized replicated field trials require careful planning and implementation of experiment for better statistical interpretation. Many growers hesitate to participate due to time and resource required to implement these trials during the peak farming season. Yield monitor generate intensive data points (every second) during crop harvest and can be a better solution to easily implement, record and analyze on-farm trials data. Hence, researchers from Cornell University, Ithaca, NY, Rochester Institute of Technology, and USDA-ARS Crop Production System Research Unit (CPSRU), Stoneville, MS have conducted 2 year (2018 and 2019) N-rich single strip trials on 6 corn fields (2 farms with 3 fields each) from central New York. Historic yield monitor data was used to develop yield stability zone for each of the field. N-rich strip crossed different stability zones within the field creating a N x Zone treatment combination. N-rich and control (outside the N-rich strip) treatments were evaluated per stability zone. The analysis showed that estimates of treatment effects using the Generalized Least Squares approach are unstable due to over-emphasis on certain data points, while assuming in-dependence leads to underestimation of standard errors. These results will help researcher to fine tune appropriate statistical analysis methods needed to analyze on-farm experimentation while providing timely recommendation to growers for their research question investigated on their own farm.

Technical Abstract: On-farm experimentation (OFE) allows farmers to improve crop management over time. The randomized complete blocks design (RCBD) with field length-strips as individual plots is commonly used but it requires advanced planning and has limited statistical power when only three to four replications are implemented. Harvester-mounted yield monitor systems generate high resolution data (1-s intervals), allowing for development of more meaningful, easily implementable OFE designs. Here we explored statistical frameworks to quantify the effect of a single treatment strip using georeferenced yield monitor data and yield stability-based management zones. Nitrogen-rich single treatment strips per field were implemented in 2018 and 2019 on three fields each on two farms in central New York. Least Squares and Generalized Least Squares approaches were evaluated for estimating treatment effects (assuming independence) versus spatial covariance for estimating standard errors. The analysis showed that estimates of treatment effects using the Generalized Least Squares approach are unstable due to over-emphasis on certain data points, while assuming in-dependence leads to underestimation of standard errors. We concluded the Least Squares approach should be used to estimate treatment effects, while spatial covariance should be assumed when estimating standard errors for evaluation of zone-based treatment effects using the single-strip spatial evaluation approach.