Regressions Fit for Purpose: Models of Locust Phase State Must Not Conflate Morphology With Behavior

Phenotypic plasticity often entails coordinated changes in multiple traits. The effects of two alternative environments on multiple phenotypic traits can be analyzed by multivariable binary logistic regression (LR). Locusts are grasshopper species (family Acrididae) with a capacity to transform between two distinct integrated phenotypes or “phases” in response to changes in population density: a solitarious phase, which occurs when densities are low, and a gregarious phase, which arises as a consequence of crowding and can form very large and economically damaging swarms. The two phases differ in behavior, physiology and morphology. A large body of work on the mechanistic basis of behavioral phase transitions has relied on LR models to estimate the probability of behavioral gregariousness from multiple behavioral variables. Mart́ın-Blázquez and Bakkali (2017; [10.1111/eea.12564]10.1111/eea.12564) have recently proposed standardized LR models for estimating an overall “gregariousness level” from a combination of behavioral and, unusually, morphometric variables. Here I develop a detailed argument to demonstrate that the premise of such an overall “gregariousness level” is fundamentally flawed, since locust phase transformations entail a decoupling of behavior and morphology. LR models that combine phenotypic traits with markedly different response times to environmental change are of very limited value for analyses of phase change in locusts, and of environmentally induced phenotypic transitions in general. I furthermore show why behavioral variables should not be adjusted by measures of body size that themselves differ between the two phases. I discuss the models fitted by Mart́ın-Blázquez and Bakkali (2017) to highlight potential pitfalls in statistical methodology that must be avoided when analysing associations between complex phenotypes and alternative environments. Finally, I reject the idea that “standardized models” provide a valid shortcut to estimating phase state across different developmental stages, strains or species. The points addressed here are pertinent to any research on transitions between complex phenotypes and behavioral syndromes.