Warning: Ignoring non-zero random test percentage because cross-validation in use

[hermani vieira]

Dear Mr.,

Is there a thumb rule to define how to set parameters to "random test
percentage" and "replicates" based on the number of a species
occurrence?

Sorry about my bad English.

Regards,

Hermani Vieira
Geographer

[David Galbraith]

I thought that if a species is not very scarce such that observation data is plentiful then it would be OK to randomly remove as much of the sample as you want (I first tried doing that with 50% subsetted for model testing, and also played around with other percentages), but my supervisor informed me that is unnecessary, since it would be theoretically possible to miss some key areas since computers can not generate random numbers in the base ten system.  So, maybe remove 10 or 20% for model testing, but remember that Maxent is pretty complicated, and depending on the number of observations you have and any bias inherent to your data set might mean that you should adjust the percentage of your sample for testing based on the goals and constraints of your project.

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To post to this group, send email to max...@googlegroups.com.
To unsubscribe from this group, send email to maxent+un...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/maxent?hl=en.

[Fabio Berzaghi]

So generally this is how it works

:

If you use cross validation the number of replicates would be the number of subgroups you are dividing your sample set in

. So if you are using cross validate the random test percentage should be zero.

Otherwise if you want to use a percentage and just run the model a bunch of times then you should use either bootstrapping or subsampling

. In that case you are required to set a random test percentage.

As far as the percentage to choose there is no definite answer

. I have write a paper that was suggesting to use 50% to make robust testing. However I have to investigate what it has been said in this thread that since computers are working in a binary way they might have problems with multiples of 10, but that sounds a bit strange to me.

I hope this helps

On Friday, September 7, 2012 7:58:50 AM UTC+2, margtoo wrote:

David, I am green beyond comprehension on statistics, MaxEnt and modeling in general. I am however doing a uni course that aimed at equipping me in a conceptual way in relation to plant biosecurity. My lecturer has been away for 3 weeks now and I am a full time working external student. I can't get help on my question from him and I'm running out of weekends to do my assignment. I hope you can help. I have 30 samples - just for the purposes of an assignment. We have been told to set the random test percentage to 33% but the screen dump on the assignment shows "cross validate" as being active. From what I can make out, this is negating the random test percentage. Would choosing one of the other options be appropriate?

I am grateful for any plain English answer you can offer me.

Thanks a lot

Margtoo

[Marg]

Fabio

I am very grateful for your help. I won't waste any more of my time on this assignment this weekend, until the lecturer has settled their requirements.

You are very kind.

Margot

To view this discussion on the web visit https://groups.google.com/d/msg/maxent/-/VHXV54LAiaUJ.