Getting the most out of ecotoxicological field data with generalized linear and additive mixed models: a six-step analysis framework

Analysing datasets from ecotoxicological studies conducted under field or semi-field (i.e.,, enclosure) conditions to evaluate the risks posed by chemicals can be challenging due to the inherent variability of natural systems, complex interactions between environmental factors, presence of non-linear dynamics, and difficulties working with free-ranging wildlife. Regression-based statistical approaches, including generalized linear and additive effect models (GLM, GAM) and their "mixed" counterparts (GLMM, GAMM) have a long tradition in ecology to find signals in noisy data by disentangling the influence of multiple factors. They have gained attention in analysing ecotoxicological data for risk assessment of chemicals. Nevertheless, GLMM and GAMM are often perceived as complex, leading to hesitation to account for their results in regulatory evaluations. To enhance the understanding and uptake of GLMM and GAMM, we present a framework to demystify the development (i.e.,, calibration, internal validation, and selection) of GLMM and GAMMs within the context of ecotoxicological field study data. An initial data exploration; an evaluation of the smallest level of significant difference that the model can discriminate (Minimum Detectable Difference, MDD); and a final model interpretation and visualization complete the framework in a total of six steps which enable significance of treatment-related effects to be checked at two independent stages. The framework is exemplified with a case study on common voles exposed to a fungicide under field conditions. The case study demonstrated the advantages of GLMM and GAMM in obtaining most out of valuable ecotoxicological field data, namely, their flexibility to different data types (e.g.,, counts, proportions, continuous data) recorded as study endpoints, ability to incorporate all data within a single analysis while considering the repeated sampling within the same fields (i.e.,, avoiding pseudoreplication), the potential non-linear dynamics of the endpoints over time, and the multiple influencing factors of direct and indirect interest to the study interpretation.