Traditional evaluation procedure in National Turfgrass Evaluation Program (NTEP) relies on visually assessing replicated turf plots at multiple testing locations. This process yields ordinal data; however, statistical models that falsely assume these to be interval or ratio data have almost exclusively been applied in the subsequent analysis. This practice raises concerns about procedural subjectivity, preventing objective comparisons of cultivars across different test locations. It may also lead to serious errors, such as increased false alarms, failures to detect effects, and even inversions of differences among groups.
We reviewed this problem, identified sources of subjectivity, and presented a model-based approach to minimize subjectivity, allowing objective comparisons of cultivars across different locations and better monitoring of the evaluation procedure. We demonstrate how to fit the described model in a Bayesian framework with Stan, using datasets on overall turf quality ratings from the 2017 NTEP Kentucky bluegrass trials at seven testing locations.
Compared with the existing method, ours allows the estimation of additional parameters, i.e., category thresholds, rating severity, and within-field spatial variations, and provides better separation of cultivar means and more realistic standard deviations.
To implement the proposed model, additional information on rater identification, trial layout, rating date is needed. Given the model assumptions, we recommend small trials to reduce rater fatigue. For large trials, ratings can be conducted for each replication on multiple occasions instead of all at once. To minimize subjectivity, multiple raters are required. We also proposed new ideas on temporal analysis, incorporating existing knowledge of turfgrass.