[statnet_help] Decay & Cutoff in gwesp parameters (& convergence issues)

[Steffen Triebel]

Hi Statnet-list,

I am pretty happy with a two-mode ERGM I estimated for a recent project. It converges well, the fit appears to be good, and the results make a lot of sense. However, it came up in discussions with colleagues that it would be great to cater more to the strengths of ERGMs by including more controls for network closure/structure.

 

My perspective on this is a bit more lenient on this and I think that the strength of ERGMs already lies in including dyad-dependent effects. In any case, the amount of structural network effects in my model is on the low end and I would like to at least try to include more (perhaps to improve fit or just guard myself against surprises). Currently, I only control for b1degree(1) because most actors only have affiliations to one group and b1star(2) as a generalized way of controlling for groups that are connected. There are various b1star(2, attr)-effects, but other than that only dyad-independent effects.

 

I am wondering about 3 things currently:

 
Is it a reasonable argument for excluding effects such as gwb1dsp from the analysis that there are no specific theoretical arguments for their inclusion and that their inclusion leads to trouble with model convergence? In the same vein: Would the inclusion of several b1star-effects be considered sufficient as controls for dyad-dependencies? As I understand it, the geometrically weighted structural effects improve fit, but my fit is quite good already.
 
I am unsure about what to do with the decay- and cutoff-parameters. Is decay similar to the alpha-parameter in SAOMs gwesp-effect where a=0 is similar to transitive ties and the higher the parameter, the closer it is to transitive triplets? As for cutoff, I have no idea what that does. I toyed around with various parameters for gwb1dsp, gwb2dsp, gwb1degree and gwb2degree and was only able to get a model to converge that includes gwb1degree with a decay of 0.25. The result is non-significant (by a long shot). I am not sure if I’m just incredibly bad at picking these parameters or if this is a good indication that I should just keep these effects excluded. I tend to think it’s the latter.
 
Short last question: If my model does not convergence and I re-estimate with init=coef(prior_model) and reiterate this 3 times with no notable improvements, am I correct to assume there is not much hope and I should abandon this specification?
 

 

Cheers & Thanks,
Steffen

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman13.u.washington.edu/pipermail/statnet_help/attachments/20221128/50d60c10/attachment.html>