Location: Fruit and Nut Research
Title: Some consequences of using the Horsfall-Barratt scale for estimating disease severity compared to nearest percent estimation) AuthorBock, Clive | |
Gottwald, Timothy | |
Parker, Paul | |
Ferrandino, Frank | |
Welham, Sue | |
Van Den Bosch, Frank | |
Parnell, Stephen |
Submitted to: Plant Disease Epidemiology International Workshop
Publication Type: Proceedings Publication Acceptance Date: 4/16/2009 Publication Date: 6/8/2009 Citation: Bock, C.H., Gottwald, T.R., Parker, P.E., Ferrandino, F., Welham, S., Van Den Bosch, F., Parnell, S. 2009. Proceedings of the 10th International Workshop on Plant Disease Epidemiology, Cornell, NY. June 7-12, 2009. Pp 20-22.2009. Interpretive Summary: The Horsfall-Barratt (H-B) scale has been widely used as a tool for disease severity assessment. However, it might detract from the quality of the assessments compared to other methods, for example, rating disease to the nearest percent. The objective of this work was to test whether use of the H-B scale systematically affects the outcome of hypothesis tests. Simulation modeling was used to sample from hypothetical disease severity distributions to compare data based on nearest percent estimates (NPEs) and the H-B scale. From the results the quality of the standard deviation of the mean estimates of disease suggested the H-B scale was less good than NPEs between ±20-50%. Assuming different severity means for hypothetical populations, increased sample size increases the probability to reject the null hypothesis when it was false, but NPEs of disease severity had a greater probability to correctly reject the null-hypothesis compared to H-B scale data, the effect being marked between 20 and 50% severity. Up to 50% more samples are needed when using the H-B scale. Thus there are situations where using H-B scale-based data compared to using NPEs for hypothesis testing can result in a greater risk of failing to reject null hypothesis when the null hypothesis is false. This is a type II error. Many aspects of plant pathology, including comparing the effects of fungicide treatments, yield loss, cultural, agronomic, and management factors, disease resistance and studying development of epidemics rely on hypothesis testing, and plant pathologists should consider the potential ramifications of applying the H-B scale for disease assessments when an hypothesis test is being made. Technical Abstract: The accuracy and precision of the Horsfall-Barratt (H-B) scale has been questioned, and some of the psychophysical law on which it is based found to be inappropriate. It has not been demonstrated whether use of the H-B scale systematically affects the outcome of hypothesis testing. A simulation modeling approach was used to sample from hypothetical disease severity distributions to compare data based on nearest percent estimates (NPEs) and the H-B scale. Hypothetical, normally distributed disease severity distributions were assumed. Treatment A has mean µA, and treatment B has mean µB = µA + µ', where µ' is the difference between the means and thus reflects the treatment difference. A generalized rater ability was developed to estimate disease at severities from 0 to 50% from these distributions. The aim was to investigate whether probability to reject H0 when H0 is false is influenced by using H-B midpoint values compared to the NPEs. A t-test was used to compare the populations for each sampling. The proportion of tests that H0 was rejected was plotted against various parameters including sample size. Increased sample size increased the probability to reject the null hypothesis (H0), when this hypothesis is false. But between mean disease 20% and 50%, NPEs have a greater probability to correctly reject H0 compared to H-B scale data. If mean disease severity was <20% there was little difference in the two assessment methods to reject H0. Where there is a difference between rating methods, up to a 50% larger sample size can be taken which allows a similar probability to NPEs. The larger sample size addresses the problems of relatively inaccurate individual H-B midpoint estimates that lead to greater variability in the sample mean estimate. Many aspects of plant pathology, including comparing the effects of fungicide treatments, yield loss, cultural, agronomic, and management factors, disease resistance and studying development of epidemics rely on hypothesis testing, and plant pathologists should consider the potential ramifications of applying the H-B scale for disease assessments when an hypothesis test is being made. |