I am currently modeling the occupation of the Andean tiger cat, (Leopardus tigrinus pardinoides) and its relationship with covariates at the microhabitat (% canopy cover and height, % leaf cover, % litter cover and depth, canopy height and slope), and landscape scale (Forest ammount, Euclidean Nearest Neighbor Distance - forest, Patch Cohesion Index - forest, Edge density - forest, Landscape heterogeneity - Shannon Index, Euclidean Nearest Neighbor Distance - pasture).

Microhabitat covariates were measured at 5m in each cardinal point and then averaged to obtain a mean value of each one for all the stations that detected the species at least once.

Landscape covariates were measured around a 500m buffer obtained for each camera that detected the species at least once.

My problem here is that in both scales, there are highly correlated covariates, but ruling out any of these prevents me from properly exploring whether my developed hypotheses make sense for the species in any of the two scales proposed.

I have modeled the occupancy of a species under the approach of sequential sub-models (see Morin et al. 2020) using the AIC criterion to understand how the occupancy and detection of the species vary as a function of these covariates.

At the microhabitat scale, the models were first run for detection while holding occupancy constant (no effect of covariates). Subsequently, detection was held constant and the individual effect of each covariate on occupancy was evaluated. On a landscape scale, this procedure was repeated only with occupancy.

In most cases, several covariates met my decision criteria in providing information on the trend of my parameters for both landscape and microhabitat (delta AIC < 2).

However, many times, there are several covariates that are rescued and their relative weight is low. Some of these covariates are correlated with other not rescued by the delta AIC scores.

Should I consider performing a PCA of the covariates at each scale to better understand their relationship with these parameters from a comprehensive perspective, or should I eliminate the most correlated covariates at the expense of sacrificing some conceptual elements that will allow me to better understand this system?

I will be very grateful for any advice on this matter.