gfalbery
commented
3 years ago Hi Jonathan, thanks for the kind message! It's really nice to know ggregplot is useful, although it's also a bit daunting because I really made it just for my own reckless purposes. Happy to help sort out a solution for this if it's helping you. Hit me with some data if you like, or let me know if you're getting errors with the pushes I make now.
The reason INLAModelSel doesn't really have any allowance for complex models is because I initially designed it to perform simple model selection on a pre-spatial model, and then to add more complex (spatial) effects in subsequent models. I later ended up trying to choose between various spatial components, which I have mostly been doing with INLAModelAdd instead because it makes a bit more sense as a forward model selection/model addition process. This is all a long way to say that INLAModelAdd is probably a better (or at least more updated) bet if you're looking to do some model selection with INLA. Among other things, it's got an allowance to fit and pick between multiple random effects, which is obviously missing from INLAModelSel.
What might work for you is if I add a ... formal argument to the
INLAModelSel function and push it (and I'll mirror any changes with
INLAModelAdd too). That way you could put e.g. graph = g, constr = T in
the INLAModelSel call and then graph = graph, constr = constr in the
variable name for the bym2 effect. I'll try pushing that now, but because
INLAModelSel's development has lagged behind my coding ability a bit I'm
not sure how well it'll handle the complexity of the string.
Another possibility is as you suggest I could just add a series of options
for priors, hyperparameters, etc, just like I've done for the random models
-- if the ... option doesn't work I'll play around with some other
approaches. I could also combine all the random effects etc into a named
list where each component of the list has a load of named components that
correspond to the arguments that you stick inside the f(), which might be
a more sustainable option for fitting quite complex stuff.
G
On Sat, Apr 10, 2021 at 6:37 PM Jonathan Simkin @.***> wrote:
Hey! Really love this package and the tutorials you have up on ourcodingclub. I am working through a model selection exercise and I'm specifically using the BYM2 model with R-INLA. I came across your package/tutorial on ourcodingclub and saw you can add the response, covariates and random effects using INLAModelSel function but it doesn't work with the BYM2 model (which has a spatial random effect and an unstructured random effect "iid"). I think this is because when you add the spatial random effect to the INLAModelSel, there isn't an argument to add a graph/spatial weights matrix.
This is what the INLA model would look like:
formula <- as.formula(paste0(resp, " ~ ", paste(covar, collapse = " + "), " + f(idarea, model = 'bym2', graph = g, constr = T, hyper = prior, scale.model = T)"))
inla(formula, family = "poisson", data = chsadf_inla, E = exp, control.compute = list(dic = TRUE))
Then model selection ModelSel <- INLAModelSel(resp, covar, "idarea", "bym2", "poisson", mydata)
Output: Error in INLA::f(idarea, model = "bym2") : The 'graph' has to be provided for model bym2
I think this should be relatively easy to implement? i.e. adding a new argument for the INLA argument 'graph'? Another item would be an offset argument for those using Poisson..
Happy to provide data to work with if necessary!
Thanks!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/gfalbery/ggregplot/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADQWUG2WXEPMO24JYIE5URDTICEDZANCNFSM42WXEVEA .
-- Greg Albery Postdoctoral Fellow, Bansal Lab Georgetown University Washington, DC (Working remotely from London, UK)
@.*** // gregalbery.me google scholar https://scholar.google.co.uk/citations?user=0cLDqU8AAAAJ&hl=en || github https://github.com/gfalbery || twitter https://twitter.com/gfalbery || researchgate https://www.researchgate.net/profile/Gregory_Albery
(I work at all hours but I never expect replies outside 9-5 weekdays)
jdsimkin04
Hey! Really love this package and the tutorials you have up on ourcodingclub. I am working through a model selection exercise and I'm specifically using the BYM2 model with R-INLA. I came across your package/tutorial on ourcodingclub and saw you can add the response, covariates and random effects using INLAModelSel function but it doesn't work with the BYM2 model (which has a spatial random effect and an unstructured random effect "iid"). I think this is because when you add the spatial random effect to the INLAModelSel, there isn't an argument to add a graph/spatial weights matrix.
This is what the INLA model would look like:
formula <- as.formula(paste0(resp, " ~ ", paste(covar, collapse = " + "), " + f(idarea, model = 'bym2', graph = g, constr = T, hyper = prior, scale.model = T)"))inla(formula, family = "poisson", data = chsadf_inla, E = exp, control.compute = list(dic = TRUE))Then model selection
ModelSel <- INLAModelSel(resp, covar, "idarea", "bym2", "poisson", mydata)Output:
Error in INLA::f(idarea, model = "bym2") : The 'graph' has to be provided for model bym2I think this should be relatively easy to implement? i.e. adding a new argument for the INLA argument 'graph'? Another item would be an offset argument for those using Poisson..
Happy to provide data to work with if necessary!
Thanks!