I am fitting a binomial GAM using mgcv in R. My data is presence (1)/ absence(0) of dolphin acoustic detections in 10 minute time windows over ~1 year period. However, I have only ~750 presences to ~55,000 absences.
Var1:Var6 relate to various physical, temporal and environmental variables where Var1 and Var2 are factor variables with 2 levels.
Despite the low number of presences, the model identifies significant relationships with all included variables and confidence intervals on plots are fairly small. However, the deviance explained is <10% and I don't think UBRE should be negative. Could the low deviance explained be driven by the low presence:absence ratio?
ACF plots show no major autocorrelation issues and there doesnt appear to be concurvity either:
Here are example plots of the model output for Var6:
Firstly plotted using
mydf<-ggpredict(fullMod,terms=c("Var6"),full.data = FALSE) plot(mydf) + ylab('Presence') ggplot(mydf, aes(x, predicted)) + geom_line() + geom_ribbon(aes(ymin = conf.low, ymax = conf.high), alpha = .1) + xlab('Var6')
and then plotted using the following code with partial residuals:
visreg(fullMod,"Var6", scale="response", rug=FALSE, partial=TRUE, gg=TRUE,main="Subject Smooths")
I'm not sure what to interpret from the above plot, e.g. what does it mean that most residuals appear close within the confidence intervals and then there are some that are way off? From my very limited statistical knowledge and lots of reading I can't seem to find many appropriate model checking plots for binomial GAMs so advice on this would also be appreciated. For example, I'm not sure how to interpret this plot of fitted vs residuals:
I'm wondering if there is a way to check the robustness of my model? I'm looking to improve my confidence in the model result but I'm not sure where to begin.
Apologies for quite a broad question but my internet searching has not gotten me far. If I can provide any other information that may be helpful please let me know.
Thanks in advance, Laura