I understand the default in Mplus is to discard half of the iterations as burn-in. Could I specify something different, a specific number of iterations or a different ratio (e.g., 1/10)? I am working on a model that converges very fast, but some of the parameters have large auto-correlation, so I would like to use a big thinning factor and run long chains. This means the burn-in I need is only a small fraction of the number of actual iterations. Thanks!
Auto correlations are not a reason of concern in principle. I would recommend to not alter Mplus convergence decision and posterior distribution. If the auto correlations is extremely high the MCMC sequence can not possibly converge fast. I would recommend using thin as you have done but leave the rest as is.
You can save all parameters in all MCMC iterations with the BPARAMETERS and summarize the posteriors any way you want from that file.
I just wanted to share a bit more. I used four chains and set BCONVERGENCE = 0.001, and convergence happens within a few hundred iterations. This is very clear from the trace plots, where the four chains quickly converge and mixing is very good. Most of the parameters have low auto-correlation, except for binary variables' thresholds and regression coefficients relating these to a latent mediator. Auto-correlation is not a concern except that it reduces effective sample size, and I was trying to get to effective sample sizes that I feel comfortable with (a few thousand).
Yes, I have saved all the iterations and discarded only part of the first half. I was just wondering whether there was a way to specify a burn-in option so you don't have to do this manually.
PS: Sorry, I misspoke. In that model I have an observed continuous mediator, which is why it converges very fast. I have another model with a latent mediator underlying an ordinal variable; that one takes a longer time to converge.
on page 8, some text indicates that convergence is reached when PSR values for all parameters is less than 1 + e, where e = f*c. c is set by the user (via BCONVERGENCE = c), but what is f? How can we determine f?
Daniel Lee posted on Monday, January 14, 2019 - 9:33 am
Yanling Li posted on Friday, May 01, 2020 - 6:36 pm
I wonder if Mplus adapts the samplers during the burn-in phase. That is, can the first half samples be treated as independent samples drawn from the posterior distribution? I am running a long chain and do not want to discard the whole first half, so I keep some iterations in the first half for estimation, but if these are not independent samples, I may not use them for estimation.
In some estimations we adapt the sampler in the burn-in phase. This happens only up to iteration 1000. We do not recommend that you use the first half of the estimation as we can't guarantee that convergence has been achieved at that point and the estimates could still be moving towards the posterior distribution.