Plant disease severity assessment-how rater Bias, assessment method, and experimental design affect hypothesis testing and resource use efficiency

Abstract

The effect of rater bias and assessment method on hypothesis testing was studied for representative experimental designs for plant disease assessment using balanced and unbalanced data sets. Data sets with the same number of replicate estimates for each of two treatments are termed "balanced" and those with unequal numbers of replicate estimates are termed "unbalanced". The three assessment methods considered were nearest percent estimates (NPEs), an amended 10% incremental scale, and the Horsfall-Barratt (H-B) scale. Estimates of severity of Septoria leaf blotch on leaves of winter wheat were used to develop distributions for a simulation model. The experimental designs are presented here in the context of simulation experiments which consider the optimal design for the number of specimens (individual units sampled) and the number of replicate estimates per specimen for a fixed total number of observations (total sample size for the treatments being compared). The criterion used to gauge each method was the power of the hypothesis test. As expected, at a given fixed number of observations, the balanced experimental designs invariably resulted in a higher power compared with the unbalanced designs at different disease severity means, mean differences, and variances. Based on these results, with unbiased estimates using NPE, the recommended number of replicate estimates taken per specimen is 2 (from a sample of specimens of at least 30), because this conserves resources. Furthermore, for biased estimates, an apparent difference in the power of the hypothesis test was observed between assessment methods and between experimental designs. Results indicated that, regardless of experimental design or rater bias, an amended 10% incremental scale has slightly less power compared with NPEs, and that the H-B scale is more likely than the others to cause a type II error. These results suggest that choice of assessment method, optimizing sample number and number of replicate estimates, and using a balanced experimental design are important criteria to consider to maximize the power of hypothesis tests for comparing treatments using disease severity estimates.