MCMC FAQs

Can I change the length of the burn-in without re-running the model?

Example question: I've run a model using MCMC, and when I looked at the trajectories for the parameters I realised that I had not set a long enough burn-in period. Can I change this retrospectively, or will I have to run the model again specifying the longer burn-in?

It is possible to get parameter estimates based on a longer burn-in after you have run a model. If you click on the trajectory graph for any parameter and answer Yes to the message Calculate MCMC diagnostics? then the MCMC diagnostics window will appear, and if you select the Diagnostic Settings button at the bottom of this window you will see a box at the bottom of the window that appears, with Increase burn-in length by next to it. Type in here the iteration number up to which you want the burn-in period to run (this will give a burn-in length of the number you type in plus the burn-in length you were already using to run MCMC) and select Apply and Done. The 'posterior mean' estimate in the MCMC diagnostics window is now the parameter estimate under this longer burn-in and the 'SD' is the standard error of the parameter estimate. You can now click on the trajectory graph for any other parameter and the 'posterior mean' and the 'SD' will be adjusted to take into account the longer burn in (i.e. you do not need to select the Diagnostic Settings button for all the other parameters). Note however that the estimates displayed above the trajectory graphs in the Trajectories window will not be updated, and nor will the estimates displayed in the Equations window, the estimates stored in c1096-c1099, or the stacked parameter chains in c1090. Note also that the length of the monitoring chain (the iterations after the burn-in period) will be reduced by the same amount that the burn-in is increased.

Other MCMC questions

(Back to top)