Sorriso-Valvo M., Greco R., and Catalano E., 2009, Spatial prediction of regional-scale mass movement using Logistic Regression analysis and GIS., 57-08 (2009): 263–280.,
URL: http://www.cnr.it/prodotto/i/67056

The spatial prediction of mass-movement phenomena (y) can be achieved by assessing the probability (Py) that in any given point of the area under study, such a phenomenon exists (y = 1) or not (y = 0). The problem can be faced using a multivariate statistical procedure named Logistic Regression (LR), by means of which the probability that the value of a dichotomous variable (event exists/does not exists) is 1 (= event exists). The LR procedure is robust_ the independent variables can either be categorical or parametric and no fitting of their frequency distribution with normal distribution is required. The procedure is based on the fitting of a probabilistic regression model on a sample made of affected and non affected sites, and then applying the optimised model in the remaining study area. For best results, affected and non affected sampling sites should be of comparable extent. The procedure has been performed in the whole Calabria Region, South Italy, ca. 15,000 km2. Transforming the categorical independent variables into ordinal variables based on observed landslide frequency for each class, allowed improving the performance of RL. Validating the performance by means of a regression of expected values on observed values for different types of mass-movement yields R values in the range 0.97 - 1.00, and the regression line slope ranging from 0.95 to 1.33, if the threshold for existent event (y = 1), based on prevalence criterion, is set at Py > 0.5. ROC curve show a negative flex when the threshold is close to 0.75, in correspondence of which ca. 80% of true unstable and ca. 75 of true stable cells are correctly predicted. Thus, P(y) = 75% can be adopted a threshold value for maximum probability class for mass-movement presence.

Data from https://intranet.cnr.it/people/