Most studies focusing on landslide spatial analysis have considered the relationships between predictors and landslide occurrence as fixed effects. Yet spatially varying relationships, i.e. non-stationarity, often occur in any spatial data set and should be theoretically considered in statistical models for a better fit. In Skagafjörður, a landslide-rich north–south oriented area located in northern Iceland, we investigated whether spatial non-stationarity in the relationships between paraglacial variables (glacio-isostatic rebound and post-glacial debuttressing, both captured in this area by latitude) and landslide locations is detectable. To explore the non-stationarity of factors that predispose landslide occurrence, we performed two logistic regression models, one global (GLR) and the other enabling the regression parameters to vary locally (geographically weighted logistic regression, GWLR). Each model was computed with two types of outcome, one based on the entire masses of landslides and the other only on the scarps of landslides. GLR results reveal that increasing latitude is associated with increasing probability of landslide occurrence, confirming that post-glacial rebound is of prime importance at the regional scale. Nevertheless, GWLR indicates that this relationship is absent or reversed at some locations, meaning that the influence of paraglacial and other predisposing factors of landsliding (slope, valley depth and curvature) vary at the local scale. This result sheds light on the spatial clustering of three subzones where landsliding drivers are homogeneous. We conclude that a GWR-based approach provides some significant inputs for spatial analysis of mass movement processes, by identifying multi-scale process control zones and by highlighting local drivers, indecipherable in global models.
On fixed effects estimation in spline-based semiparametric regression for spatial data
