Quantitative ordinal scale estimates of plant disease severity: Comparing treatments using a proportional odds model

Authors: CHIANG, KUO-SZU - National Chung-Hsing University; LIU, H - National Chung-Hsing University; CHEN, Y - National Chung-Hsing University; EL JARROUDI, MOUSSA - Universite De Liege; Bock, Clive

Submitted to: Phytopathology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/14/2019
Publication Date: 4/1/2020
Citation: Chiang, K., Liu, H.I., Chen, Y.L., El Jarroudi, M., Bock, C.H. 2020. Quantitative ordinal scale estimates of plant disease severity: Comparing treatments using a proportional odds model. Phytopathology. 110:734-743. https://doi.org/10.1094/PHYTO-10-18-0372-R.
DOI: https://doi.org/10.1094/PHYTO-10-18-0372-R

Interpretive Summary: Disease severity assessment is needed to measure disease in pecan and other crops and often it is based on use of an ordinal scale. Ordinal scales in plant pathology are generally based on the percent area with symptoms. We used a parametric proportional odds model to analyze directly the ratings obtained from disease scales, without converting ratings to percentages based on class midpoints of quantitative ordinal scales (currently a standard procedure). The purpose of this study is to evaluate the performance of the proportional odds model for the purpose of comparing treatments (e.g. varieties, fungicides, etc.) based on ordinal estimates of disease severity. A simulation method was implemented to perform the study. The proportional odds model was compared with the model using midpoint conversions of ordinal intervals. Our results show that the performance of the proportional odds model is never inferior to using the midpoint of the severity range at severity <40%. Especially at low disease severity (=10%), the proportional odds model is superior to the midpoint conversion of the interval method. Thus, for early onset of disease, or for comparing treatments that happen to share severities <40%, the proportional odds model is preferable for analyzing quantitative disease severity estimation data based on ordinal scales when comparing treatments, and at severities >40% is equivalent to other methods.

Technical Abstract: Studies in plant pathology and plant breeding requiring disease severity assessment often use a certain type of ordinal scale based on defined numeric ranges, which can be termed a quantitative ordinal scale – with plant disease this special form of the ordinal scale is generally based on the percent area with symptoms [e.g. the Horsfall-Barratt (HB) scale]. We used a parametric proportional odds model to analyze directly the ratings obtained from disease scales, without converting ratings to percentages based on class midpoints of quantitative ordinal scales (currently a standard procedure). This useful feature of the proportional odds model also renders it amenable to comparing estimates from studies using different response scales. The purpose of this study is to evaluate the performance of the proportional odds model for the purpose of comparing treatments (e.g. varieties, fungicides, etc.) based on ordinal estimates of disease severity. A simulation method was implemented to perform the study. The parameters of the simulation were estimated using actual disease severity data from the field. The proportional odds model was compared with the model using midpoint conversions of ordinal intervals. The criterion for comparison was the power of the hypothesis test. Our results show that the performance of the proportional odds model is never inferior to using the midpoint of the severity range at severity <40%. Especially at low disease severity (=10%), the proportional odds model is superior to the midpoint conversion of the interval method. Thus, for early onset of disease, or for comparing treatments that happen to share severities <40%, the proportional odds model is preferable for analyzing quantitative disease severity estimation data based on ordinal scales when comparing treatments, and at severities >40% is equivalent to other methods.