Why can I have this massive difference in between predicted omission and omission of test samples?
Sara
I'm running a multiescale model in MaxEnt but first I'm selecting the environmental variables at its better scale. When I run the model with all my 100m variables I obtain that the omission of test samples doesn't fit with the predicted omission. However, the AUC of the model is 0.591, not very good but I would have expected a better fit of the omission rate. I've used 5 fold crossvalidation (so 5 replicates). Therefore the graph attached shows the average resutls
Could help me to undersand why I'm obtaining this diference etween predicted omission and omission of test samples? Do you think this graph is normal?
Thank you so much,
Sara
Jamie M. Kass
Not sure what the core of the problem is, but your test omission looks like it's around 1, which means all your test points were outside the predicted area (the binary surface produced by thresholding the continuous surface by some value). This means your model is a poor predictor of the test points, and your AUC is weak because it is quite close to 0.5, which is the value at which the model is performing no better than random. I think we'll need some more details before we can diagnose the problem.
1. What's the background area you're using? Does it include enough of a representation of the surrounding area in addition to the core area your occurrence points fall in? Does it exclude areas your species cannot disperse to, or could not in recent time?
2. Do the predictor variables you're using have enough variability to help the algorithm discern between presence and background?
3. How many occurrence points are you inputting? If you input very few points, random partitioning might perform more poorly than jackknife.
Can you begin by answering these questions? A map of your study area with occurrence points would help too.
Jamie Kass
PhD Candidate
City College, NYC