anova comparisons for univariate responses
jebyrnes opened this issue · 6 comments
Hello! Quick question. In previous versions of gllvm, one could run something like
anova(mod1, mod2,
which = "uni",
method = "fdr")
This functionality seems to be gone? Is it coming back, or is there another way of doing this?
Hello Jarett. No, I am afraid this functionality is no longer supported; it was an experimental feature that I decided to discontinue for lack of merit and technical backing.
The feature relied on retrieving the (partitioned) negative VA log-likelihood from the C++ side of things, which is not implemented anymore, so there is no other way of doing this than using an old version of gllvm.
Huh....... if folk do want to do posthoc testing for individual species, what would you recommend then?
That depends on what you want to test. If it is predictor effects, there are wald-statistics available in the summary that you can use.
For individual coefficients, yes, but, if I'm trying to assess, say, two models with multiple levels of multiple categorical variables and interactions - something that in a univariate case I'd use a F test or LR Chi Sq - instead I have a lot of tests with treatment contrasts to wade through. I'm trying to see if there is a simpler solution akin to the aforementioned. Does that make sense?
Wald-statistics do not need to be calculated on an individual coefficient basis. Multivariate wald-statistics, though not included in gllvm, could test simultaneously across all predictors for a single species.
I understand what you are after - it is the same I was looking for when I started to think about this - but I do not have a solution. I am happy to think with you about this though, if you do!
The next thing to try might be similar to what classical ordination methods do, or what mvabund does, i.e., re-sampling or permutation testing. However, that gets into all kinds of nasty business with GLLVMs due to sensitivity to starting values and computational intensity of the model fitting.
Huh - yeah, I see your point re: simulation and permutation testing. I admit, my background is not deep enough yet to implement multivariate wald stats - but, if you have some good reading/code examples, I'd be happy to try it out! Just out of my current wheelhouse and would need some start points to run down the road.