It could be a matter of not using the ideal climatic variables, but not necessarily. The interaction between climate and topography can also give odd results, to the exclusion of one type of topography. Another thought: you do not specify what species this is. If it is a bird for example, not predicting occurrence where the species appears only occasionally may be a good thing.

Apart from MaxEnt you may test other species distribution models using SDM package in R.(https://cran.r-project.org/web/packages/sdm/index.html)

And see which model is performing best for your data sets.

You have not mentioned the predictive variables used and not indicated the characteristics of both areas. My answer is a guess from experience. You may have used BioClim variables only. These may be inaccurate depending an the (unknown) underlying meteorological point data of BioClim (see attached). A second or related issue may be that the range of the elevation (Serban's response may hints on this) or climate variables within your two disjunctive distribution areas differ substantially (see both attachments).

Other SDMs are unlikely to improve the prediction with your current environmental data-set (see second attachment). 

Article Where the bears roam in Majella National Park, Italy

Article Transferability of species distribution models: The case of ...

Thank you both for  the answers! I used bioclim variables and select the best with Jacknife (4 variables).  There is a big difference in elevation between places where the presence is predicted (high elevations)  and Missiones (low elevation).  The animal is a rodent. I'll look at the papers you sent. Thank you very much! 

I agree with Mriganka, despite its usefulness and popularity, MaxEnt is not necessarily the best suited tool for all the modelling works and a carefull thinking of its assumptions and suitability is needed.

Appart from that, a model is only as good as the ideas, science and data behind it. I mean that probably, the problem is not the model, but that the data used are probably not good for what you are trying to do. Hein has already pointed out that BioClim data might be inaccurate. I agree, specially if your study area did not contribute with various stations to BioClim data or if your scale of analysis is much more detailed thatn BioClim data.

The issue of scale is of paramount importance. A really limiting variable can sort out to be unselected by the model if its included in a different scale to that it is "acting". Alttitude, for example, might well be an influencial variable for the distribution of a particular animal, but would fail to enter the model if you're only modelling mountain regions. Additionally, I would suggest to think in other kind of variables, such as of types of woodlands. The distribution of a rodent is probably limited by resource availabitliy, and hence the abundance of nut-producing trees is a fundamental variable to consider. This could be linked to some degree to the type of forest. Of course, it colud be also explained by climatic variables, but I wouldn't expect to be so, at a regional scale.

So, you would need to think carefully on what kind of variables may be limiting the distribution of your species, and at what scale they may be acting.

And last, but not least, extrapolating data from one area to the other is at least difficult, and generally shoud not be done without further cross-validation. It is not clear from your message if model training did include Missiones.If it did not, the problem is probably there. Otherwise, you may have a problem of different weightings of both areas. How many points did you have in Mata Atlántica? If  you have many points there and only 5 in missiones, the model will be really modelling the distribution in Mata Atlántica.

Hope this helps.

Good luck!!

   Asier

Very good points in the above answers.

To that I'll only add this: Let us not forget that what MaxEnt (and indeed many SDM algorithms) do is not exactly "to predict". Often they will simply draw areas where the model variables (bioclimatic or otherwise) are best adjusted to the sampling points, but that may have little predictive value without biogeographical (dispersal, barriers) or ecological (competitors, resources) considerations.

Any missing but important driver (as it seems to be the case here: elevation) may easily tip off the balance. And even if most variables are accounted for, these are often interpolations and (as Asier rightly points out) a more dense sampling for the variables might yield a radically different interpolation at another scale--in turn producing a different SDM.

In sum, it is your field, biological and taxonomical knowledge what will fill the gaps and improve the model's fit.

This seems like an interesting situation.

Why not run the model separately for each region? If the limit to distribution is the realized niche and not the fundamental niche, then different regions may have different realizations, particularly if the animal is more strictly limited by food than climate or some other interaction that differs between regions. 

Roman

Perhaps in one of the places your rodent is actually a cryptic species, e.g. you do in fact have 2 species.

kind regards

Irene

Great discussion here. The first thing I thought about was the threshold rule, if you are using binary outputs, as mentioned by Luis Acosta. A less restrictive threshold rule could include the Misiones records inside the predicted area, but at the same time it can increase type I errors. Try running the models with different threshold rules and check the outputs. I also agree with Mriganka Sarkar, you should try modeling with different methods. Contact me if you need help with the models.

Bruno

Thank you everybody again!

My study is focused in taxonomy, I have genetic and morphologic analysis of the specimens, but I would like to talk about geographic distribution in my paper and that's the reason for the model distribution analysis. It is only one species for sure. Luis E. Acosta, about choosing only 4 variables, I first did run the model with 19 variables and selected those more important with jacknife and after I excluded those hightly colinear (r>0,7 and p

I also encountered the same issue. I was doing SDM of Rhododendron spp in HKH region, but it could not predict in Pakistan, although I supplied with some occurrence from there. 

here, which one is most recommended for SDM?

Hi Caryne

From my limited experience I learnt that variables selection can strongly affect your model and the resulting output. If you haven't read* it yet, Hein's paper is a good example of that.

Thus for your scenario I would consider deeply thinking of the ecology of your species, and what you think are the main limitations to its dispersal-> distribution. Then try find good quality data layers for such ecologically relevant variable and run the model with those.

You should be able to find people from your area in MAXENT related posts and they could direct you to good sources for obtaining good quality environmental variables

* and yeah thresholds and maxent are a pain. Be sure to read the literature available (below some examples) and experiment a lot with different thresholds

- Escalante, T., Rodríguez-Tapia, G., Linaje, M., Illoldi-Rangel, P., & González-López, R. (2013). Identification of areas of endemism from species distribution models: threshold selection and Nearctic mammals. TIP, 16(1), 5-17.

- Liu, C., White, M., & Newell, G. (2013). Selecting thresholds for the prediction of species occurrence with presence‐only data. Journal of Biogeography, 40(4), 778-789.

- Norris, D. (2014). Model thresholds are more important than presence location type: Understanding the distribution of lowland tapir (Tapirus terrestris) in a continuous Atlantic forest of southeast Brazil. Tropical Conservation Science, 7(3), 529-547.

- Rebelo, H., & Jones, G. (2010). Ground validation of presence‐only modelling with rare species: a case study on barbastelles Barbastella barbastellus (Chiroptera: Vespertilionidae). Journal of Applied Ecology, 47(2), 410-420.

regards