Determining transit impact on Seoul office rent and land value: an application of spatial econometrics

Abstract

This study posits that there may be a systematic bias in measuring the transit’s endogenous impact on land values in a built up area due to discrimination by location in the city. Studies of transit value-added effect report mixed results about the capitalization of station proximity. The question is not whether a transit station influences nearby land values, but how and where location determines the impacts. Examining 731 office rentals and land values in Seoul, this study finds that value premium over better accessibility to a station decays with increasing distance from the central business district (CBD) and significantly depends on the development density of the station area. Overall, station benefits seem to exist in Seoul, but they look more notable in centers with higher centrality. This makes a hierarchy of regression coefficients for station proximity by location, i.e. the beta in the CBD is the highest and those in the subcenters are next, while that in other areas is the lowest. Study findings imply that the potential of more compact and denser developments within station areas seems higher in a dense inner city, providing evidence for the concept of ‘compact city.’ Questions concerning model specification in the hedonic approach are raised: in research sampled heavily from the suburbs, the coefficient may be underestimated where this benefit actually exists. Also, due to the incongruence of station area with station value-added area, using a dummy variable seems intrinsically risky. This study shows that estimation with spatial models outperforms OLS estimation in the presence of spatial autocorrelation. Also, there is a strong spatial autocorrelation even in the SAR residuals where the omission of key variables still influences the estimation. Overall, spatial lag and error term variables greatly improve the fitness of regression equations; however, the latter seemed more useful than the former in this study. One thing to note is that the latter seems more sensitive to the choice of weight matrix than the lag variable. There may exist a unique weight scheme proper for the data structure which cannot be known in advance.