Clarification on the usage of nodemix and nodeofactor/nodeifactor

[gocoyd]

Hello!

I am very new to ERGM and have some questions about model specification. I am interested in ethnic homophily in school classrooms and to get the measures I use nodemix('minor', base = c(2, 3)). As a result it provides me with migrant homophily (mix.minor.1.1) and native homophily (mix.minor.0.0) rates. I do not use just nodemix because I want to get more detailed results.
At the same time I want to account for sender and receiver effects, so I use nodeofactor('minor') and nodeifactor('minor') but the model gives a warning:

Model statistics ‘mix.minor.1.1’ are linear combinations of some set of preceding statistics at the current stage of the estimation. This may indicate that the model is nonidentifiable.

The model works okay when I use either nodeofactor('minor') or nodeifactor('minor') and not them together. Or if I use nodematch instead of nodemix. Can somebody please explain why this is so? I would be very grateful.

[krivit]

Firstly, please take a look at the term help, i.e., ergmTerm?nodemix, ergmTerm?nodeofactor, etc. if you haven't already. In particular, using base= is deprecated.

Secondly, in the future, please provide more complete examples, i.e., the formulas for the models you are trying to fit.

Based on what you've written, if you have an directed network and a nodal attribute with two levels, counting up the ties between and within them gives you four distinct counts: 0-0, 0-1, 1-0, and 1-1. This means that you can have at most a total of 4 statistics involving them, and how you parametrise them depends on your substantive considerations. In this case, I suspect your model also has an edges effect, so you have nodeofactor.minor.1 + nodeifactor.minor.1 - mix.minor.1.1 + mix.minor.0.0 = edges (for every possible network), so one of these is redundant, which is what that message is telling you.

If this doesn't answer your question, please reopen the discussion.