Latent variable models handled with optimization aided by automatic differentiation; application to marine resource assessments

The use of models with latent variables is important in the management of marine resources. Catch-at-age models used in fisheries stock assessments may contain several hundered unknown parameters, but the are typically only a few parameter of real interes t (e.g. current stock size). We will develop method for estimating interest parameters by maximum likelihood, with the remaining unknown parameters integrated out of the likelihood function. Numerical integration in high dimensions is challenging, particularly in combination with optimization, and new efficient optimization algorithms are required. Estimation is only one part of statistical inference. We will also develop practical diagnostic tools for latent variable models, and we will develop simulation-based methods for estimating bias-corrected confidence distributions. Finally, we will communicate the results from the project to people working in assessment of marine resources. This is a network project with nodes at the Institute of Marine Research, The University of Bergen, and the Norwegian Computing Center, Oslo.