Interactions between covariates

[Raquel-RuizDiaz]

Hello!, I have a fairly easy question:
Is it possible to include interactions between covariates in sdmTMB?

I want to create a delta_gamma model with the following configuration:

fit_dg <- sdmTMB(
biomass ~ 0 + s(temperature) + s(depth),
data = data,
mesh = mesh,
family = delta_gamma(), #< new
spatial = "on",
spatiotemporal = "iid",
time = "year",

silent = FALSE

)

I tried to do it like in a regular GLM changing the following part of the code:
biomass ~ 0 + s(temperature) : s(depth)
and biomass ~ 0 + s(temperature) * s(depth)

none of them worked..

Then, I tried removing the spline and it worked but I got errors:
Warning messages:
1: In stats::nlminb(start = tmb_obj$par, objective = tmb_obj$fn, gradient = tmb_obj$gr, :
NA/NaN function evaluation
2: In stats::nlminb(start = tmb_obj$par, objective = tmb_obj$fn, gradient = tmb_obj$gr, :
NA/NaN function evaluation

So.. I guess it is not possible?

[seananderson]

Interactions should work just fine like that with regular covariates, just not with smooths. This is the same as mgcv, which does the smoother parsing. You can let a smoother interact with another variable in several ways including:

• 2-dimensional smooths: s(x, y) ort2(x, y) (t2 allows different wiggliness in the 2 dimensions)
• by variables: s(x, by = y) where y is continuous (x varies linearly with y) or y is factor (different x smoother for each y)... you likely also want + y in the formula in this case too.

Some references:
https://peerj.com/articles/6876/
https://fromthebottomoftheheap.net/2017/10/10/difference-splines-i/

[Raquel-RuizDiaz]

Thanks a lot for your quick answer!