Standard errors via the observed Fisher information
Produced in R version 4.4.0.
Fisher information is a common way to get standard errors in various settings, but is not so suitable for POMP models.
We often find ourselves working with complex models having some weakly identified parameters for which the asymptotic assumptions behind these standard errors are inadequate.
Further, the technique requires evaluation of second derivatives of the log likelihood; a numerically unstable problem when one has the capability to obtain only noisy estimates of the log likelihood.
We suppose that \(\theta\in\mathbb{R}^D\) and so we can write \(\theta=\theta_{1:D}\).
The Hessian matrix of a function is the matrix of its second partial derivatives. We write the Hessian matrix of the log likelihood function as \(\nabla^2\ell(\theta)\), a \(D\times D\) matrix whose \((i,j)\) element is \[ \big[\nabla^2\ell(\theta)\big]_{ij} = \frac{\partial^2}{\partial\theta_i\partial\theta_j}\ell(\theta).\]
The observed Fisher information is \[ I^* = - \nabla^2\ell(\theta^*).\]
A standard asymptotic approximation to the distribution of the MLE for large \(N\) is \[ \hat\theta(Y_{1:N}) \approx N[\theta, {I^*}^{-1}],\] where \(\theta\) is the true parameter value. This asserts that the MLE is asymptotically unbiased, with variance asymptotically attaining the Cramer-Rao lower bound. Thus, we say the MLE is asymptotically efficient.
A corresponding approximate 95% confidence interval for \(\theta_d\) is \[ \theta_d^* \pm 1.96 \big[{I^*}^{-1}\big]_{dd}^{1/2}.\]
We usually only have one time series, with some fixed \(N\), and so we cannot in practice take \(N\to\infty\). When our time series model is non-stationary it may not even be clear what it would mean to take \(N\to\infty\). These asymptotic results should be viewed as nice mathematical reasons to consider computing an MLE, but not a substitute for checking how the MLE behaves for our model and data.
