search

statnet / lolog

Latent Order Logistic (LOLOG) Graph Models
Other
5 stars 1 forks source link

lolog-ergm vignette attempts to fit an ergm specification that exhibits a phase transition. #6

Closed krivit closed 8 months ago

krivit commented 2 years ago

In the estimation

https://github.com/statnet/lolog/blob/5fdd171bfed17be505ceb50e8f0f7aa4ee91456f/vignettes/lolog-ergm.Rmd#L116-L117

the choice of GWESP decay (0.25) exhibits a phase transition near the MLE. This can be seen in trace plots produced by

fitukErgm <- ergm(ukFaculty ~ edges() + nodematch("GroupC") + nodefactor("GroupC", base=3) + gwesp(decay=.25, fixed=TRUE),
                  control=control.ergm(MCMC.runtime.traceplot=TRUE))
ifellows commented 2 years ago

Thanks. This model was supposed to be fit-able, but close to degenerate. Basically I backed off the gwesp decay until I got something that fit and had decent trace plots at the final iteration. Thus, the MLE is close to a phase transition by design.

Of course, ergm has gone through significant changes and so we should expect that convergence might not be achieved in newer versions due to unstable nature of the MCMCMLE process around phase transition points.

Do you have a suggestion for a best fit achievable by ergm for this data?

krivit commented 2 years ago

Off the top of my head,

fitukErgm <- ergm(ukFaculty ~ edges() + mutual + nodematch("GroupC", diff=TRUE) + nodefactor("GroupC", base=3) + gwesp(.2, fix=TRUE))

seems to be a bit better behaved and runs between 1 and 3 minutes on my system.

More generally, there seems to be quite a bit of unaccounted-for degree heterogeneity in the model.

krivit commented 9 months ago

Following up on this, @ifellows, this has been messing up my reverse-dependency checks for the past few releases. I've suggested a possible improvement above (and I think it works pretty well with the original .25 decay as well on more recent versions of ergm). Can you please incorporate it?

martinamorris commented 9 months ago

watching.

On Sat, Dec 16, 2023 at 10:37 PM Pavel N. Krivitsky < @.***> wrote:

Following up on this, @ifellows https://urldefense.com/v3/__https://github.com/ifellows__;!!K-Hz7m0Vt54!jAo6n217kYJ9f74J7zqQNltt9oo4zcKuyIad9e6yKltkhvvDYGBAC_SImO02hGq_LvP_WBZ_Rx1hdAi8SsjVg2E$, this has been messing up my reverse-dependency checks for the past few releases. I've suggested a possible improvement above (and I think it works pretty well with the original .25 decay as well on more recent versions of ergm). Can you please incorporate it?

— Reply to this email directly, view it on GitHub https://urldefense.com/v3/__https://github.com/statnet/lolog/issues/6*issuecomment-1859053178__;Iw!!K-Hz7m0Vt54!jAo6n217kYJ9f74J7zqQNltt9oo4zcKuyIad9e6yKltkhvvDYGBAC_SImO02hGq_LvP_WBZ_Rx1hdAi8ehes-vs$, or unsubscribe https://urldefense.com/v3/__https://github.com/notifications/unsubscribe-auth/AB6QTYT2JITSKEPCJ3FC4J3YJ2HKRAVCNFSM5VG23CHKU5DIOJSWCZC7NNSXTN2JONZXKZKDN5WW2ZLOOQ5TCOBVHEYDKMZRG44A__;!!K-Hz7m0Vt54!jAo6n217kYJ9f74J7zqQNltt9oo4zcKuyIad9e6yKltkhvvDYGBAC_SImO02hGq_LvP_WBZ_Rx1hdAi8ZvGl5so$ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

ifellows commented 8 months ago

Suggested change has been added. Thanks.