In postharvest disease (PHD) control experiments, treatments with the same observed values for all their replicates are frequently present, which leads to one or more treatments with zero observed variance—this type of data are referred to as data with zero-variance patterns. Such data patterns clearly violate the assumptions of classical parametric statistical models. These patterns are frequently overlooked in phytopathological studies, thus resulting in inadequate inference. In vivo experiments evaluating postharvest disease (PHD) control in fruits involve diverse response variables describing disease progress over time, including binary, discrete, and continuous data, which require different probability distributions for appropriate statistical modeling. These experiments typically follow a completely randomized design (CRD), with individual fruits serving as experimental units and being evaluated over a time interval (t, days) defined by the host–pathogen system. Disease progress is commonly quantified by incidence, severity (measured by discrete scales or mean lesion diameter, MLD), and the area under the disease progress curve (AUDPC). In PHD datasets, high/very low treatment efficacy often leads to zero/maximum values across all replications which produces zero-variance patterns, thus rendering the use of general linear models (GLM) unsuitable. Under such conditions, alternative nonparametric approaches or permutation tests are required, including Fisher’s exact test for incidence and tests for nonparametric contrasts for severity or AUDPC. Our objective is to contribute towards the adoption of more adequate statistical inferential methods for analyzing data from PHD experiments.
Maia et al. (Fri,) studied this question.