My group has just begun to play with it. The case that worked the hardest on it found that where 3.3.3 gave perfectly reasonable models with low and high probabilities in the right places, the new version gave a very small range of suitability values, and really did not make much sense. 

On a more fundamental level, we were disappointed that this version did not include automatic calculation of AIC and AICc statistics that are so crucial for modern model selection exercises. Seems like they have all the information there, so it was a bit odd that they did not include that ...

All the best, ATP

Thank you, Townsend! I guess I will start playing with the new version but proceed to use 3.3.3 for this project as the time is tight (with not much room for experiments).

Yes, it would be nice to have metrics to evaluate model complexity incorporated within Maxent. While those are not incorporated into the program itself, I have found ENMeval really practical for assessing model complexity (among other things)—I am sure you know about it already, but I mention it here for the potential benefit of other readers. 

https://www.researchgate.net/publication/265297099_ENMeval_An_R_package_for_conducting_spatially_independent_evaluations_and_estimating_optimal_model_complexity_for_MAXENT_ecological_niche_models 

Article ENMeval: An R package for conducting spatially independent e...

Dear Eliécer E. Gutiérrez :

Thank you for sharing a new version of Maxent, it's very helpful...

Mariwan, thanks for your message, but thanks go to the developers S. Phillips et al. Best. E

Thanks for the plug, Eliécer! Bob Muscarella and I are now chatting about how best to put maxnet into ENMeval... but not sure when that will be available.

I'd like to point out that from what I understand, the R package maxnet and Maxent v. 3.4 the Java program (maxent.jar) are not the same. I think that maxent.jar v. 3.4 is simply a new iteration of v. 3.3.3k, whereas maxnet is a repurposed call to glmnet (a generalized linear model with regularization built in) with some math in there to make the outputs match with Maxent. This can be done because of equivalancies found between Maxent and GLMs under certain conditions (which are applied in maxnet, see Phillips et al. 2017 in Ecography).

I have done a little testing and find that. for the most part, the predictions between maxent.jar and maxnet are very similar (but not identical) when comparing the same output type, and that the new cloglog output shows the same patterns as logistic, but has higher predictions in intermediate suitability areas, consistent with what Phillips et al. 2017 says. There are still some uncertainties, and please reference these Maxent Google Group posts that present some ongoing work exploring some differences between maxent.jar and maxnet that we've found (and please contribute!): https://groups.google.com/forum/m/#!topic/Maxent/NtKTyF9v8v4

In these posts, there is R code posted that you can use to reproduce the analyses we've run. I'd be interested to hear more about what others have found, and if anyone can explain the differences we've found in the Google Group.

Town, thanks for your input! I was just wondering if you compared cloglog to logistic, or logistic to logisic? Either comparison should not produce the results you got, but I'd be interested to know. Thanks.

Jamie

Thank you, Jamie! Very informative and helpful comment! 

We have had a play with both (in a fairly specific data case), in practise 3.3.4 with the cloglog seems to give more generalised predictions (wider extent across the low/mid values) when compared to logistic scaling. This is penalised by validation metrics so it's fairly unclear if there are performance differences between the two. It is worth noting that our results appear largely similar after application of a thresholding approach. The exclusion of hinge features from 3.3.4 as default was the main driver of observed spatial differences, whereas the difference in predicted values between transforms whilst quite contrasting to look at appears to be largely trivial after the outputs have been evaluated and binned accordingly.

The cloglog approach assumes spatial independence in the presence data (beyond what is accounted for in the environmental data), so presumably there is the potential for an inflationary effect on predictions if this is violated? If you are working with opportunistic data that may exhibit spatial clustering, and aren't looking to benefit specifically from the newfound ability to use Maxent as a regression approach (i.e. training on some continuous response value), I would probably stick with the logistic output (still available in 3.3.4) to avoid this potential issue, and assess predictions on a species by species basis to determine the most appropriate regularisation parameters and feature type combination for each. But there may be some additional advantages to the cloglog that I am missing.

Certainly interested to hear other's experiences with this, thanks for raising the topic.

Alex

Thank you Gutiérrez et. al for this useful discussion, I am also playing around with Maxent 3.4 for the first time and was trying to compare logistic and cloglog outputs. In my case, I am using SWD format of Maxent with linear and quadratic features and as Phillips et. al 2017 have indicated in their software note cloglog transformed output is somewhat greater than the logistic one but species.lambdas file is exactly the same. Therefore, I assume the difference are basically due to different assumptions and transformation parameter used but as Alex indicated I am also knee to understand additional advantages of cloglog transformation, if any.

Has anyone had time to look further into it?

Furthermore, I am preparing for submission a couple of publications where I used maxent v3.3.3. Should I be alright publishing with results obtain from this previous version?

I am still using the old version 3.3.3 as the new one is not supported by Dismo in R, did anyone find solution for this? 

thanks. 

Dear Madan Krishna Suwal, it is possible to run Maxent 3.4.0 using dismo in R, I personally have managed to run samples with data format of maxent using 3.4.0 in R. But as there are lot of online discussions going on regarding the results of version 3.4 (particularly run using R package maxnet) being incomparable to those from version 3.3.3k , I am not sure if I should stick to 3.4.0 or switch back to 3.3.3k. This is becoming a difficult choice when the maxent team has released newer version 3.4.1.

Dear Jamie M Kass, are you sure that maxent.jar 3.4.0 is simply a new iteration of 3.3.3k, if so maxent.jar 3.4.0 logistic output should be same or at least comparable to maxent.jar 3.3.3k logistic output, right? Have you check these two outputs?