Bayesian inference with tmbstan for a state-space model with VAR(1) state equation (2101.05635v1)

Abstract: When using R package tmbstan for Bayesian inference, the built-in feature Laplace approximation to the marginal likelihood with random effects integrated out can be switched on and off. There exists no guideline on whether Laplace approximation should be used to achieve better efficiency especially when the statistical model for estimating selection is complicated. To answer this question, we conducted simulation studies under different scenarios with a state-space model employing a VAR(1) state equation. We found that turning on Laplace approximation in tmbstan would probably lower the computational efficiency, and only when there is a good amount of data, both tmbstan with and without Laplace approximation are worth trying since in this case, Laplace approximation is more likely to be accurate and may also lead to slightly higher computational efficiency. The transition parameters and scale parameters in a VAR(1) process are hard to be estimated accurately and increasing the sample size at each time point do not help in the estimation, only more time points in the data contain more information on these parameters and make the likelihood dominate the posterior likelihood, thus lead to accurate estimates for them.