MaxEnt variable importance flipped after adding a few predictors — multicollinearity issue?

[Emelia Pettit]

Hi all! I just reran my marine Maxent model with some new variables but it seems to just have shifted the importance of pre-existing variables.

In my original model, my top predictors were:

Bathymetry and Minimum SST - Together ~80% of the model

Meanwhile Primary Productivity was only ~4%

Model 2 (revised variables)

Swapped Rugosity for Slope and pH for Nitrate, and added variables pertaining to light and substrate.

My permutation importance was now:

PrimProd_Mean (55%)

Depth (25%)

Temp_Min (5%)

In the first jackknife, Depth and Temp_Min dominate both in isolation and when removed, in my opinion indicating the model relies heavily on a simple depth–temperature gradient.

In the second Jackknife, Primary Productivity becomes the strongest variable while Depth remains important and Temperature drops off, and removing any single variable has less impact, showing that importance is now shared across a correlated gradient rather than concentrated in a few drivers.

Both models have high AUC. Model predictions are somewhat similar but some key hotspots did change.

Suspected issue

While I pruned variables with a correlation of over .7, there does seem to be a wide correlation cluster of depth, temp, primary productivity, ocean chemistry variables and light fallign between .4 and .6 across these variables.

So I’m wondering if:

Primary Productivity is now absorbing shared signal (inflated?)

Temperature + Nitrate are being suppressed (deflated?)

Is this kind of shift in permutation importance normal when adding more variables?

Does this mean Model 2 is actually “better,” or just redistributing the same signal?

Any good way to separate real drivers vs proxy variables here? I've seen PCA's listed as a n option, but I would really love to maintain my environmental variables, since I am most interested in my response curves and working with PCA-derived variables seems much more difficult and less clear.