jmlondon
commented
1 year ago We decided not to remove the day (and day^2, day^3) interactions with sin1, cos1, etc because we want the ability for haul-out behavior within a day to change over the season (e.g. go from a unimodal distribution to bi-modal or skewed toward afternoon or morning).
After further exploration and testing, we've decided to not add the 4th degree polynomial. It does provide some additional haul-out behavior of females when we would expect them to be nursing. However, adding the day^4 term also means adding additional interactions for e.g. day^4:sin1, day^4cos1, etc. Simple model selection using mgcv::bam() and AIC suggests the benefit of adding day^4 is not outweighed by the cost of adding additional parameters
After some discussion and exploration, we're considering updates to the model specification with a few key differences:
The 4th degree poly allows for some additional flexibility such that the period of time for adult females when they are nursing is represented better. The 3rd degree polynomial seemed to have smoothed over this period too much.
The removal of interactions was done to limit the complexity of the model and allow those features to be captured within the temporal autocorrelation and not as part of the mean. Generally trying to limit terms to those for which we can attribute a biological/ecological meaning.