Burn-in iterations: is there any way to change the proportion of burn-in iterations of MCMC sampling?

[RYZ0521]

The default proportion of burn-in is set to be 0.1 of the total number of MCMC samplings. Is there any way to change this to a larger proportion?

[alhauser]

Currently, it is not possible to change this proportion.

This could be enabled by adding a new entry to the model.args argument of CausalImpact(), and passing it down to GetPosteriorSamples() where the burn-in rate of 0.1 is currently hardcoded. This would require downstream changes in tests and documentation, too.

We currently don't have plans to add this option as the default burn-in rate typically works well, but feel free to create a pull request if you want to implement that functionality.

[steve-the-bayesian]

Note that the burn-in rate for bsts has not been a strict 10% of draws for a long time. Bsts looks at the log-likelihood associated with each parameter draw. Bsts looks at the last 10% of draws (assuming convergence has been achieved by then) to find an upper quantile of the log-likelihood distribution. Burn-in is declared the first time the MCMC sample path of log-likelihood exceeds this level. In non-Gaussian models bsts does not compute log likelihood (because it is not available in closed form), so it falls back to eliminating a fixed percentage of draws.

…

On Thu, Jul 15, 2021 at 9:14 AM Alain Hauser ***@***.***> wrote: Currently, it is not possible to change this proportion. This could be enabled by adding a new entry to the model.args argument of CausalImpact(), and passing it down to GetPosteriorSamples() where the burn-in rate of 0.1 is currently hardcoded <https://github.com/google/CausalImpact/blob/7e0f59f944b22bfb3186a726215f56919d25e37f/R/impact_inference.R#L31>. This would require downstream changes in tests and documentation, too. We currently don't have plans to add this option as the default burn-in rate typically works well, but feel free to create a pull request if you want to implement that functionality. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#43 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABMVDVJ7NPWKXISOOHF2E2TTX4COLANCNFSM4QNBYVKA> .