[GRASS-user] Logistic Regression

[Paola Salmona]

Good afternoon everybody!
I'm trying to make a landslide susceptibility map based on the statistical analysis of the occurrence of landslides in correlation of certain predisposing factors (i.e. slope, aspect, etc.) reclassified into discrete classes (i.e. every 10 degrees, North, West, South, East, etc.)
Since the occurrence /non occurrence of a landslide is a dichotomic variable,  Logistic Regression is generally considered a most suitable model.
To manage the several predisposing factors, I have tried with the command r.regression.multi, but I am not sure about a few issues.
I have assumed that the "map for y coefficient" is the map with observed data, but do I have to use a 0 = no landslide / 1= landslide map or a logit map expressed as log (P + e1-8) - log(1- P + e1-8)?
In both cases, when comparing the estimates map with observed landslides, most landslides are located within the areas with the highest values of susceptibility , but it is difficult to quantify the reliability of the model because the AIC index assumes scarcely comparable values (i.e. 6258510.340522, 6258514.420749.... ) and the other indexes reported in the output txt file, according to literature, are not suitable to evaluate a Logistic Regression model. I have then used the addon r.edm.eval to calculate the Area Under the Curve and the results appear to be coherent with the comparison (around 88%), but this command only works if as "layer containing references classes" I use the  0/1 landslide map.
Have I messed up in the procedure or are the commands I 've used not correct for this kind of analysis?
Thank you and best regards
Paola
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/grass-user/attachments/20240528/94e8bb99/attachment.htm>