Alternatives for simulating and modeling simplified insect feeding eletropenetrography discrete data

The study of insect feeding behavior using electropenetrography (EPG) typically involves analyzing complex data. EPG data comprises a temporal sequence of behaviors summarized using a collection of counts, durations, and sequential variables. These variables can be counts, means, percentages, or linear combinations of behaviors. This results in numerous variables being correlated to a certain degree. Consequently, statistical analysis is rendered complex, particularly in terms of model fitting and selection. This study proposed a statistical approach to simulate overdispersed correlated count data based on a previous comparative experiment to monitor the feeding behavior of untreated Euschistus heros versus E. heros treated with an entomopathogen. The waveforms included non-feeding (Z), pathway (Eh1), laceration/maceration of endosperm tissue (Eh3a), short ingestion events of lacerated/macerated endosperm tissue (Eh3b), xylem sap ingestion (Eh2), and ingestion from an unknown location (Eh4). Simulated scenarios involved the creation of differences between groups of insects based on the total number of events or the proportion of events of Z. Several statistical models were then fitted to the simulated data and evaluated based on goodness-of-fit, type-I error rate, and power analysis. The multinomial model exhibited the lowest type-I error rate and was more sensitive in detecting higher (>1.35x) differences between groups. Only the multinomial model achieved a power greater than 0.8. Conversely, models such as the Poisson and normal models exhibited limitations such as inflated type-I error rates in the presence of overdispersion. Among the univariate models, the mixed model exhibited the best fit.

Keywords:
EPG; multinomial model; double-poisson