Impacts of different types of data integration on the predictions of spatio-temporal models: A fishery application and simulation experiment

Abstract

Integrated spatio-temporal models, which enable the sharing of information across locations, time and data sources, are gaining traction for their potential to generate more precise and more accurate estimations compared to models fitted to single data sources. Standard integrated spatio-temporal models combine multiple data sources via a catchability factor. Recently, spatially varying catchability (SVC) integrated spatio-temporal models were developed to implement data integration via the estimation of an SVC term for the least reliable data sources. Expanded-domain integrated spatio-temporal models are models integrating data from different spatial areas. Spatio-temporal models can combine standard or SVC integrated modelling with expanded-domain integrated modelling. The above-mentioned types of data integration have never been evaluated through a comparative analysis. Here, we investigate the impacts of these different types of data integration on the predictions of spatio-temporal models, via an application to the southern hake (Merluccius australis) HAK4 stock, where the bottom trawl data collected within the New Zealand observer programme are integrated with data from five different bottom trawl research surveys, and a simulation experiment. In total, six models were compared in the present study, where the three last models constitute expanded-domain integrated models: (Model 1) a model fitted to observer-only data for HAK4; (Model 2) a standard integrated model fitted to both observer and survey data for HAK4; (Model 3) an SVC integrated model fitted to both observer and survey data for HAK4; (Model 4) a model fitted to observer-only data for HAK4 and the other New Zealand hake stocks (HAK1 and HAK7); (Model 5) a standard integrated model fitted to both observer and survey data for HAK4, HAK1 and HAK7; and (Model 6) an SVC integrated model fitted to both observer and survey data for HAK4, HAK1 and HAK7. For the simulation experiment, we produced simulated data from Model 5, fitted the six models to the simulated data, and evaluated the performance of the models by comparing their estimations to the simulated data. Overall, the indices obtained with the different types of integrated models outperformed the indices obtained with models using observer-only data: indices from the integrated models were more precise, better matched the traditional stratified random index and had less bias and a smaller root-mean-squared-error, yet characterised uncertainty less well. Moreover, expanded-domain integrated models outperformed other models regarding habitat assessments: (1) they provided insights into spatial density patterns for much larger regions and predicted these patterns more precisely for the area common to all models; and (2) models combining expanded-domain integrated modelling with standard or SVC integrated modelling predicted patterns of distribution shifts and range expansion/contraction more precisely than the expanded-domain integrated model employing observer-only data. Finally, expanded-domain integrated models did not outperform other integrated models regarding supporting stock assessments. The indices obtained with expanded-domain integrated models were more precise and had less bias and a smaller root-mean-squared-error, but agreed less with the traditional stratified random index and characterised uncertainty less well. This result suggests that standard or SVC integrated models should be preferred to produce indices for local areas.