I have a GAM model (below) where SST_mean and NAO are numerical values, and cycle and region are categorical factors. I checked for concurvity using the concurvity function in mgcv.
m2 <- gam(Strandings ~ s(SST_mean) + s(NAO, bs="re") + Cycle + Region,
family=poisson, data=DAT_ST, method = "REML")
The initial results below had quite high values suggesting there is concurvity in the model.
> concurvity(m2, full = TRUE)
para s(SST_mean) s(NAO)
worst 0.8944583 0.7532177 0.7131497
observed 0.8944583 0.5784295 0.7131497
estimate 0.8944583 0.5309899 0.7131497
I ran pairwise comparisons and it looks like the concurvity issues are actually between the parametric terms not between the smoothed terms or between the smoothed terms and the parametric
> concurvity(m2, full = FALSE)
$worst
para s(SST_mean) s(NAO)
para 1.000000e+00 1.736827e-25 0.001064966
s(SST_mean) 1.746022e-25 1.000000e+00 0.351366528
s(NAO) 1.064966e-03 3.513665e-01 1.000000000
Is there any reason for this, and is this an issue for my models. Or because the parametrics are categorical, is this to be expected for some reason? Someone I read that above 0.8 indicates concurvity that should be addressed but I can't find any specific thresholds in the literature. Are there any references people could recommend please?
