Accurate modeling of animal activity patterns is important for biological assessment surveys, management plans, and fundamental understanding of how organisms respond to environmental and climatic conditions. Often for logistic reasons, researchers collect activity data of animals repeatedly from the same sites. The resulting longitudinal data has the added benefit of being able to distinguish between “cohort” and “age” effects. When analyzing longitudinal data it is important to account for the correlation associated with repeated sampling of the same sites to avoid pseudoreplication and violations of model assumptions. This is often accomplished using mixed-effects models, which are conditional (subject-specific). If the specific sites are not of interest, generalized estimating equations (GEE) are computationally simpler than mixed-effects models and provide marginal (population-level) estimates. We compare GEE and mixed-effects models for estimating seasonal activity of red-backed salamanders (Plethodon cinereus). We obtained counts of salamanders from nighttime visual encounter surveys throughout their activity season over four years. We used two modeling approaches to evaluate these data. First, we used the same fixed effects in all models to compare salamander responses to meteorological conditions. Second, we conducted independent model selection to determine the best predictive model of salamander surface activity for mixed-effects and GEE models.
Results/Conclusions
The explanatory model used in all modeling approaches produced estimates that were in the same direction and similar rank order for all mixed-effect and GEE models. Soil temperature had a significant quadratic effect with peak activity around 15 C. Rainfall amount and relative humidity had positive effects on salamander surface activity. Salamanders were most active in the spring. At higher temperatures rainfall had less effect on activity and wind speed has less effect on humid nights. However, the magnitude of the effects and the associated error differed among models. Linear mixed models (LME) on log-transformed count data and GEE resulted in similar estimates of the fixed effects. Generalized linear mixed models (GLMM) estimated steeper slopes (positive and negative) for nearly all variables compared with GEE and LME models. The second approach resulted in different models predicting salamander activity. All models included temperature, rainfall, relative humidity, and windspeed as important variables. The GAMM was the least complex because terms were absorbed by the smoothing term. The LME model was the next simplest, whereas the GLMM and GEE models included all potential variables and interactions. Despite differences, the overall predictions of the models for mean conditions throughout the year were fairly similar.