Anticipation of land use change through use of geographically weighted regression models for discrete response

Abstract

Geographically weighted regression (GWR) enjoys wide application in regional science, thanks to its relatively straightforward formulation and explicit treatment of spatial effects. The application of GWR to discrete-response data sets and land use change at the level of urban parcels has remained a novelty, however. This paper describes work that combined logit specifications with GWR techniques to anticipate five categories of land use change in Austin, Texas, and controlled for parcel geometry, slope, regional accessibility, local population density, and distances to Austin's downtown and various roadway types. Results of this multinomial logit GWR model suggested spatial variations in - and significant influence of - these covariates, especially roadway vicinity and regional access. A 1% increase in the distance of an undeveloped parcel to the nearest freeway, for example, was estimated, on average, to increase the probability of residential development by 1.2%, while the same increase in distance to a major arterial was estimated to increase the probability by 1.8%. Conversely, proximity to roads (through reductions in such distances) was estimated to boost the likelihood of nonresidential development (e.g., 9.0% in the case of commercial development in response to a 1% decrease in distance to such arterials). The logsum accessibility index was estimated to exert an average positive influence on commercial, office, and industrial development tendencies, while it dampened land use transitions from an undeveloped state to residential uses. Comparisons of results with a spatial autoregressive binary probit (with the use of all developed land use categories as a single response) and GWR binary probit also provided some insights. The latter seemed to surpass the former in its account for spatial effects, as reflected by a lower Akaike information criterion value.