autoUpdate: Run MCMC iterations until full convergence
In thlytras/FluHMM: Hidden Markov Model for influenza sentinel surveillance

Description Usage Arguments Details Value

This function runs further MCMC iteration to extend the current posterior sample until convergence is reached for all parameters.

1	autoUpdate(x, iter = 20000, maxit = 2e+05, Rhat = 1.1, thin = 1)

`x`	An object of class ‘FluHMM’, from which to generate posterior samples.
`iter`	Number of iterations to run.
`maxit`	Maximum number of iterations to run before giving up.
`Rhat`	Gelman-Rubin diagnostic cutoff value to determine convergence (only for the sigma[1] parameter).
`thin`	Thinning interval. If the beta[2] and beta[3] parameters are very slow to mix, and your RAM memory is insufficient, it is advisable to use a larger thinning interval (e.g. 10 or 50), i.e. keep in the posterior sample every `thin` MCMC iteration.

Most parameters in the model converge fairly quickly, but mixing is slower for parameters beta[2] and beta[3] (i.e. the slopes of the epidemic growth and epidemic decline phases). Initial convergence of the model is judged by monitoring the sigma[1] parameter and is performed by the function autoInitConv, which is usually run in the FluHMM constructor. If all other parameters have converged but beta[2] and beta[3] have not, due to slow mixing, then it is advisable to run the MCMC chain for longer, until beta[2] and beta[3] cover their posterior distributions adequately. This can be done manually using update with argument enlarge=TRUE. Alternatively autoUpdate can be used, which extends the sample by iter iterations, checks the Gelman-Rubin diagnostic for all parameters, and if convergence has not been reached it extends again the sample until convergence or until maxit iterations.

In some cases (usually near the end of the influenza surveillance period, when all five model phases have occured) mixing is very slow and the chains need to be run for very long. This requires lots of RAM memory and a fair amount of CPU power to handle the very large posterior sample. In such cases it is advisable to thin the chain (i.e. by 10 or 20, i.e. keep every 10th or 20th element in the MCMC chain), to generate a smaller posterior sample with good coverage of the posterior distribution.

None. The function mutates its argument ‘x’ directly.

thlytras/FluHMM documentation built on May 31, 2019, 10:44 a.m.