Responding to Reviewer's comments on MAXENT manuscripts

[CLEMENT MWEYA]

Dear ALL,

I have attempted to publish my work based on output generated by MAXENT on species distribution.

Has anyone received the same kind of questions? How did you respond?

See comments below;

Reviewer #1: Results of a modeling exercise using Maxent to study the potential distribution of disease vectors in West Africa. Although the goal of this work is laudable, I cannot recommend this manuscript for publication. The primary flaw is the low sample size (<20 observations) and the lack of description of sampling design (which precludes evaluation of the modeling procedure). However, even if the field survey should prove to be adequately performed, the proposed modeling exercise cannot be performed or validated with only 20 observations. Particularly, even using jacknife validation, Maxent does not permit the proper selection of 8 covariates using only 20 observations. Further, jack-knife validation is well known to result in over-fit models. Additionally, the manuscript lacks details important modeling details, including: (1) descriptive information about covariates available for modeling and units of covariates used in final

variable set, (2) details of the modeling workflow (eg pre-processing), (3) rationale for selection and coordinates of the spatial extent used for random background sampling, (4) what optimization procedures were used (tuning of maxent), and (5) what means were used to compute variable importance. Additional aspects the authors should include the following. (1) Even if a withheld test set cannot be constructed (due to scarcity of data), AUC still should be reported to assess performance of model. This is the standard for evaluating performance of a binary model. (2) The current study fails to include any analysis of uncertainty or propagation of uncertainty due to sampling error. (3) No information is presented to communicate the locations of presence and background points selected for modeling. I suggest these be plotted on a map. (4) There is no evaluation of effect sizes. Regrettably, I do not think these obstacles may be overcome with the existing data set

[Florencia Sangermano]

Hi Clement,

 

It seems that the main concern is the sample size, the other comments can easily be addressed by just including the information requested (description of sampling design, information about covariates, details of modeling workflow, how background samples were selected, Maxent tuning, adding the AUC, etc. ).  Below are my comments regarding sample size, it will be nice if others chime in with their experiences working with small sample sizes. 

 

I agree with the reviewer that 20 observations is a small number, however the usefulness of a small sample size strongly depends on the objective of your research… is it an explorative study of the potential distribution of the vector? or, are you trying to  extract the response of the vector  to the different covariates? I think 20 samples will be good for the former … may be not for the later (?)… but you can probably discuss the caveats of a small samples size in the discussion section.   Studies done across different samples sizes show that Maxent is useful to model species with low (5-10) sample sizes (Hernandez et al. 2009, Pearson et al. 2006), although Wisz et al. 2008, suggests a larger number. 

I think the spatial / environmental distribution of those 20 observations is more important than the number of samples per se (you can have 1000 samples distributed in small range of the environmental conditions important for the species and, although is a large sample size, your models will not be good representing the species habitat either), so having the map showing the distribution of the points in the landscape will be really useful.

 

I’m not sure how you partitioned the data for calibration-validation, but with a small sample size you can use the pvalue calculation developed by Pearson et al. (2007), or using Raes and Ter Steege (2007) Null model method.

 

Regarding the analysis of uncertainty, if you do the Pearson et al. (2007) method, you will end up with 20 models- maybe you can use the standard deviation across the models to show model variability depending on the samples used for training… Anyone has other ideas for this? 

 

I’m happy to read the paper and give you comments if you like.

 

Good luck!

 

 

------------

Florencia Sangermano, Ph.D.

Assistant Research Professor, Conservation GIS, Clark Labs

Affiliate Scientist, Graduate School of Geography

Clark University

Worcester, MA 01550

fsang...@clarku.edu

http://wordpress.clarku.edu/fsangermano

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

[CLEMENT MWEYA]

Dear Florencia,

I thank you so much for very helpful comments. Essentially, the objective of my study was just to explore potential distribution of vectors. I will take your comments and update my manuscript for submission to a different journal.

I will also share with you the current manuscript.

Regards,

Clement

--
 
Clement Nyamunura Mweya BSc, MSc (Medical Parasitology & Entomology)
Senior Research Scientist - Medical Entomology
National Institute for Medical Research (NIMR)
Tukuyu Medical Research Center (TMRC)
P.O BOX 538
TUKUYU, TANZANIA
Tel: +255 25 2552214
Fax: +255 25 2552016
Mobile:: +255 788 802003, 0655 802003, 0767 802003
E-Mail: mweyac...@gmail.com, mweyac...@yahoo.com, cmw...@nimr.or.tz