MSGSS
Talk by Joseph Barss
A model for species abundance based on ecological survey data must properly account for the presence of spatial and temporal correlations in order to provide accurate abundance index estimates. Data on lobster abundance collected by trawl survey programs in the Bay of Fundy poses additional modelling complications, including the use of different types of sampling gear with unknown catchability coefficients, varying survey coverage by program and year, and a highly skewed data distribution with a high proportion of zero observations. The negative binomial geostatistical generalized additive mixed model fit to these lobster data includes Gaussian random fields for modelling spatial and spatiotemporal effects. The model is fit with the sdmTMB R package, which performs maximum likelihood estimation using the Laplace approximation to the marginal likelihood. The Gaussian random fields have a Matérn covariance function, and are fit as Gaussian Markov random fields using a triangular mesh laid across the study area. Model selection is informed by the use of residual analysis, Akaike’s Information Criterion, and spatial block cross-validation. The main results of interest include the estimates of the model parameters, as well as maps of the predicted abundance at all locations in the study area for every year, and an index whose value at any year is the total abundance in the study area for that year. Simulation studies are also performed to examine the accuracy of the model under different scenarios of reduced data quality or quantity.