I am running MCMC for binary/categorical variables. I know that a decent method to check for convergence is to look at the traceplot of the variance, but I don't know a good citation for this. Does anyone know where I can find an article on this, or a better method for verifying that a binary/categorical variable has converged to a stationary distribution?
$\begingroup$
$\endgroup$
4
-
$\begingroup$ Do you think the standard convergence diagnostics such as Gelman, Rubin or Geweke would not be appropriate for binary/categorical variables? $\endgroup$– bdeonovicNov 2, 2015 at 21:58
-
1$\begingroup$ I think it would be appropriate for binary, but I question whether it is for categorical. The reason for this is because for multiple categories, moving from category 0 to 2 is penalized (with variance), much more than it is moving from 0 to 1, even though this should not be the case. For binary, we know only one movement, and thus tracking the variance (although "meaningless"), still gives us some information. $\endgroup$– RayVelcoroNov 2, 2015 at 23:10
-
$\begingroup$ I see, I agree. Although I have had some experience working with binary parameters I have not had any experience working with categorical parameters. I will think about this. $\endgroup$– bdeonovicNov 4, 2015 at 23:22
-
1$\begingroup$ I suppose that you could run many chains (with different starting points), and then check the mean/sd each of the plots of the indicators. Seems kinda heuristic though. $\endgroup$– RayVelcoroNov 5, 2015 at 3:42
Add a comment
|