A0429
Title: Spatial confounding in gradient boosting
Authors: Lars Knieper - Georg-August-University of Goettingen (Germany) [presenting]
Thomas Kneib - University of Goettingen (Germany)
Elisabeth Bergherr - Georg-August-Univerität Göttingen (Germany)
Abstract: 'Inside Airbnb' offers high-dimensional data sets of Airbnb accommodations, which include variables with information about the accommodations as well as their coordinates. There is a clear trend to higher prices in city centers, even though it might not be completely the center itself causing higher prices but the proximity to sights and alike. Hence, when complementing 'Inside Airbnb' data with covariates that describe the location, these covariates are correlated with the spatial trend of prices. This collinearity between covariates and spatial effects can lead to a bias in the corresponding fixed effects estimates, known as spatial confounding. Recently, the Spatial+ approach suggests regressing the spatial effect in the covariate first before estimating the model of interest. Drastic spatial confounding is observed in gradient boosting due to its step-wise procedure. The suggested two-step approach is applied, and its ability to correct spatial confounding for gradient boosting is also confirmed. A major advantage of this correction approach is that the gradient boosting algorithm itself does not need an alteration as the correction happens beforehand. Additionally, correcting also non-confounded covariates does not decrease estimation performance.
