I'm doing maxent for a couple species. I'm doing 10 replicates with subsampling to hopefully obtain average AUC scores. One only has 15 samples so I did bootstrapping with 20% of the data withheld and that seemed to work fine. The other has 31 samples so I'm trying to do crossvalidation, but I don't know if I'm getting it to work right in Maxent. I have crossvalidation selected under the Replicated runt type, but it seems to just pull 3 or 4 samples for testing. Do I need to change the settings on anything else or supply something like a test sample file?
I am afraid that this is exactly what you should expect. Crossvalidation divides the data into k (more or less) equally sized parts. Each time all but one are use for training and that one for testing. So with 31 samples and k equal to ten you really cannot expect anything else. For such small sample sizes only bootstrapping makes sense. For bootstrapping however you need more replicates (I would say not less than 25 since the variability of subsampling with replacement is high).
Oh yes, I see now, that makes sense. Yeah it's a project for an undergrad. We are getting more data, but it is just coming in slowly and he needs something to present next week. We are just running though things so he can at least show how things work and we will hopefully have better results when all the data comes in. Thanks for helping me and answering all my questions.
- Badges
- Science method
- Similar topics
- Biological Science
- Ecology