fix issue #567 (ergm), fix issue #572 (svyglm)

[briatte]

Fix issue #567 (warnings with ergm objects)

The code now calls the ergm:::summary.ergm function internally, to avoid a warning about the function not being meant to be called directly.

Note that a problem will subsist with ergm::get.miss.dyads (used in the glance method), which is deprecated. ergm::get.miss.dyads has been replaced with the code from ergm 3.9.4. Note sure it's doing the right thing, though: see my discussion of the issue in #567. Fixed thanks to @krivit.

Fix issue #572 (glance method not working for svyglm objects)

The code adds tidy.svyglm and glance.svyglm. The first one is just a copy of close to tidy.(g)lm. The second one is identical to glance.glm, except:

• The survey:::summary.svyglm function is called instead of summary (which evaluates to stats::summary.glm).

• The BIC is computed differently, using survey:::dBIC(model, maximal = model) (a maximal model can be provided, but the default is to use the same model).

• No logLik is returned, given that svyglm is not fitted via ML.

I have added an example for glance.svyglm, provided by @IndrajeetPatil in #572.

Side comment: BIC in glm and svyglm

The docs for glance.glm and glance.svyglm (in @evalRd return_glance()) say that they return the BIC, but they do not always do so:

library(survey)
data(api)
m <- glm(formula = sch.wide ~ ell + meals + mobility, data = apistrat, 
         family = quasibinomial())
AIC(m)
BIC(m)
logLik(m)
glance(m)

---

(Comments below made obsolete by changes sollicited by @alexpghayes in the comments below.)

The biggest change is that I had to add a dirty hack to finish_glance in utilities.R in order to extract the AIC from a svyglm object, since the result of survey:::AIC.svyglm is not similar to that of stats::AIC (the first one is a 3-element named vector, the second one is a 1-element vector).

A similar hack was already in the code (for the deviance of REML objects created by lmerMod).

[alexpghayes]

Code is very clean and I love the tests! Left some comments -- am trying to make broom more modular and have some suggestions to that extent.

[briatte]

@alexpghayes Hi -- I have implemented your suggestions.

Quick question -- would you be interested in removing some dependencies to the package? I'm pretty sure that I can remove stringr (by using base grepl and the like) and reshape2 (by using only tidyr, although perhaps that's a bit risky since reshape2 is probably more stable than tidyr).

Update -- I've explored both options above, and while removing stringr is doable, it'd make the code much less readable. As for reshape2, as far as I understand, reshape2::melt and tidyr::gather are not identical enough for replacing all calls to the former with the latter. I never really understood that.

[alexpghayes]

Few more comments.

I'm planning on looking at dependencies more systematically in #582 at some point. Personally I find stringr much more readable than base R string manipulation and would actually prefer to move to it, rather than away for it. Same for tidyr vs reshape(), although there's one or two places where we use reshape on 3-arrays that I don't think we can get rid of. So: yes, dependencies are my mind, but cleaning up the code base is a higher priority at the moment.

[briatte]

Hi @alexpghayes

All suggested changes implemented, although let me know if you want glance.svyglm to lose all the comments, which I have left for clarity but which can be removed if need be.

Not sure why the PR creates a conflict.

[briatte]

Removed conflicts with base branch.

[briatte]

Improved code for ergm thanks to @krivit's messages in #567.

[alexpghayes]

Thank you so much!

[briatte]

You're welcome. I like working on broom as a way to get into the details of the models that I use.

I think it'd be a good idea for me to stop adding to this PR for now, and to come back to it if ergm 3.10 introduces other changes.

In the meantime, I'll have a look at how to augment svyglm.

[alexpghayes]

Thank you both for taking care of this and your patience!

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.