Evaluation of GLM and GAM for estimating population indices from fishery independent surveys

We evaluate the performance of generalized linear (GLM) and generalized additive (GAM) models for deriving population indices from fishery independent survey data. Six model types (3 GLMs and 3 GAMs) were formulated that differed in how spatial covariates were represented, with each using one of three alternative ways to include temporal covariates. The models were applied to summer and fall survey data on 127 species from fisheries-independent bottom-trawl surveys conducted by the Southeast Monitoring and Assessment Program (SEAMAP) in the northwest Gulf of Mexico. Three response variables were analyzed: occurrence, density, and abundance. The best model (from the alternative temporal representations) were identified for each response variable for each of the six model types and their performance analyzed in more detail. Model performance was based on residual autocorrelation (Moran's I), prediction error (AIC weights), and predication variance (based on simulated sampling). We examined for patterns in these metrics based on the magnitude of the response variables (i.e., by quartiles). Results suggest that sample size (indexed here by response value quartiles) could be useful for a priori consideration when choosing among GLM and GAM models. GAM models that use geoposition with smoothing as the spatial covariate performed comparable to some of the other models at low abundances and densities (lower quartiles), and significantly outperformed all of the other models at higher densities and abundances (quartiles 3 and 4). We discuss how our results provide guidance on selecting GLM and GAM models for deriving population indices from survey data.