Using geolocation data in spatial-econometric construction of multiregion input-output tables: a Bayesian approach

Abstract

Interregional input-output tables for Poland at NUTS-3 level are built by using the Bayesian approach to spatial econometric analysis. I apply the multi-equation Durbin specification proposed by Torój (2021) to derive the sample density and Statistics Finland (2006) regional I–O tables to derive the prior hyperparameters. This prior aims to introduce additional information in the presence of noisy spatial data, but also to avoid the areas where the spatial decay profiles representing the supply geography become insensitive to the parameter values of the selected functional form. To measure the distance, the real-world driving distance between the most populated cities of the regions from Google Maps is used. Posterior distrubutions indicate that the agricultural commodities and advanced services are supplied to the most distant locations, whereas the simple services – to the least distant ones; the result for the former group of sectors is characterized with the highest uncertainty. The illustrative simulation indicates that 82.2% of the indirect effects occur in the home region, with a posterior-based confidence interval from 71.5% to 92.4%. The results do not change qualitatively when I use the driving time (averaged over 42 equidistant moments in a 7-day week) as the alternative measure of distance, but the hybrid time- and distance-based model is strongly preferred in the Bayes factor comparison, since for all sectors except industry (NACE sections B-E), the time-based metric turned out to be dominant. When commuting is taken into account in the induced effect calculation (measured with mobile geolocation data), 4.9% of the induced effects are relocated from the home region (central point in a big agglomeration) to the other regions, especially the surrounding ''ring''.