Predictions of future land use areas are an important issue as land use patterns significantly impact environmental conditions (biodiversity, water pollution, soil erosion, and climate change) as well as economic and social welfare. In order to improve the prediction accuracy of aggregated land use share models, we propose in this paper a methodological contribution by controlling for both unobserved individual heterogeneity and spatial autocorrelation. Our model is a land use shares model applied to aggregated data in France. Our data-set is a panel which covers both time series observations from 1992 to 2003 and cross-sectional observations by Departement (equivalent to NUTS3 regions). We consider four land use classes: (1) agriculture, (2) forest, (3) urban and (4) other use. We investigate the relation between the areas in land in different alternative uses and economic and demographic factors influencing land use decisions. Based on the comparison of prediction accuracy of different model specifications, our findings are threefold: First, controlling for both unobserved individual heterogeneity and spatial autocorrelation outperforms any other specification in which spatial autocorrelation and/or individual heterogeneity are ignored. Second, accounting for cross-equation correlations does not seem to improve the prediction performances and finally, ignoring individual heterogeneity introduces substantial loss in prediction accuracy. (C) 2012 Elsevier B.V. All rights reserved.
