[statnet_help] MCMC does not converge even after 50 iterations for STERGM

Vivek Kulkarni viveksck at gmail.com
Sun Aug 16 20:01:57 PDT 2020


Thanks a ton, Pavel for the great information! I will try these out.

Best,
Vivek.

On Tue, Aug 11, 2020 at 7:40 PM Pavel N. Krivitsky <p.krivitsky at unsw.edu.au>
wrote:


> Hi, Vivek,

>

> The first thing to try might be to reduce the dgwdsp decay to something

> less than .25, since you might be dealing with a phase transition (a model

> such that a small change in parameter value causeing a huge change in

> network structure).

>

> *Questions*:

> 1. In particular: how should I interpret the log output that output

> Estimating Function Values. For a model that is going to converge do these

> values get close to 0?

>

>

> Yes. At the MLE the expected network statistics equal to the observed, and

> so their difference as a function of the parameters is called an *estimating

> function*. Of course, when we use MCMC, we have to approximate this by

> taking the sample means of the simulated statistics and subtracting off the

> observed.

>

> 2. What are the scaling limits of STERGMS? How many nodes/edges per time

> point can they handle gracefully? Also is there a scalability limit on

> number of timepoints?

>

>

> Underneath it all, STERGM CMLE just calls ergm() with a specially

> formulated specification. So, STERGM CMLE will scale about as well as ergm

> itself.

>

> 3. In general, what diagnostics in the output log attached can I look at

> to get a sense of why my model is not converging?

>

>

> That the estimating function values are fluctuating like this suggests

> that you might have a phase transition, hence my advice above.

>

> If some of these statistics were going off to infinity, you would have a

> diagnostic of multicollinearity or related problems. (The current master

> branch on GitHub, if you feel up to installing it, tries to diagnose those

> problems before the estimation.)

>

> You can use the MCMC.runtime.traceplot= control parameter to have it make

> traceplots of MCMC runs to check for insufficient burn-in or a phase

> transition.

>

> I hope this helps,

> Pavel

>

>

>

>

> I would greatly appreciate any suggestions.

>

> Best,

> Vivek.

>

>

> _______________________________________________

>

> statnet_help mailing list

>

> statnet_help at u.washington.edu

>

>

> http://mailman13.u.washington.edu/mailman/listinfo/statnet_help

>

>

>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman13.u.washington.edu/pipermail/statnet_help/attachments/20200816/cd30bcb2/attachment.html>


More information about the statnet_help mailing list