Statistical analysis for Maxent
Patrick Beignet
Dear Maxent-group,
I am a newbie in using Maxent and I am struggling with some questions. For my
PhD thesis, I have to search for new sample locations for an ant species and I
wanted to use ENM to establish a species distribution map, which I would use as
a template map.
I have 42 known locations and I used the 19 bioclim variables from WorldClim,
as well as the altitude. Furthermore, I calculated the curvature and roughness
in ArcGis.
In a first step, I tried to reduce the environmental variables using a
correlation matrix in Statistica. I omitted variables with a threshold >0.8
and created first ENM models with 11 variables, which seemed good (ROC test and
training values above 0.9).
In a second step, I wanted to re-check the correlation using ENMtools by Dan
Warren. By calculating the correlations I got different correlations. However,
I again omitted correlated variables with a threshold >0.8 and created
models with 9 variables (ROC test and training values were a little bit lower,
but still approx. 0.9).
I used the default settings, except for a test percentage of 25%, write
background predictions, create response curves, do jackknife to measure
variable importance and I sometimes used the raw instead of logistic output
(for further statistical analysis, which still should not affect model
calibration). Do I have to change the regularization multiplier and do i
have to use only linear and quadratic features, instead of all?
I'm also thinking of re-checking the correlations again by using all
variables in Maxent and doing the jackknife to measure variable importance. So,
I would gradually reduce variables, reduce overfitting and increase the model's
reliability. Any suggestions here?
I also used statistical analysis to compare the models, but I am not sure,
which statistical anlysis to use.
The AUC of the ROC values seems to systematically undervalue models that do not
provide predictions across the entire spectrum of proportional areas in the
study area (Peterson et al. 2008), so I will use other statistcal analysis too.
Still, the calibrated models showed values above 0.9.
The p-value for jackknife validation cannot be used, since the pvalue compute
programme is designed for smaller sample sizes <25 (Pearson et al. 2007)
The partial ROC might be a possibility to check for a firmer foundation for
evaluation of predictions from ecological niche models (Peterson et al. 2008),
but I haven't calculated it yet
The TSS (True Skill Statistics) might also be an option, too and when
calculated, I got values from 0.6 to 0.7, which seem sensible to me.
The AIC, AICc and BIC are also important for deciding which model seems to be
the most appropriate. Here, the values calculated were a little bit higher than
1000 and as far as I understood, the lower the values, the better the model? Does
anybody know a paper, where the AIC, AUCc and BIC are described for Maxent?
Which statistical analysis or analyses should I use for model selection?
Should I use a combination of different analyses?
If I finally established a sensible model, do I have to use the replication
possibility to improve my models reliability? Or how to proceed? Or am I good
to go and use the ENM for creating a location map?
Hope anyone can help me?
Patrick