Abstract
Modern cross-sectional land price models are “functional” models in which relations between land price and only some attributes (e. g. accessibility to CBD) are considered. However, the expectation factor exists in the determination of land prices, but is not captured by these attributes independently. In fact, the spatial spillover of land prices in metropolitan areas under the Japanese bubble economy shows that the expectation factor is based on the spatial relationship “structure”.
This study, therefore, aims to econometrically identify the functional and structural factor respectively, and estimate “structural-functional” models for land prices with the Spatial Autocorrelation Analysis method. Specifically, we analyze the residential and commercial land in Tokyo's 23 wards during a 10-year period (1984-1993) which includes the bubble and post-bubble period.
The estimation is developed by maximum likelihood method based on asymptotic theory, thus it is necessary to consider smal sample properties of estimators. We investigate this problem using both artificially generating data and raw data which is used in the estimation of “structural-functional” models of land prices.
The results of this study are concluded as follows:
(1) “Structural-functional” models for land prices have consistently high fit for the period. The strength of spatial relationships measured as ρ is strongest in 1987 and is more considerable in the residential land than in the commercial land. In fact, the spatial structural factor is dominant in the residential land prices, but the functional factors is relatively dominant in the commercial ones.
(2) The estimates for spatial autoregressive paramerers (ρ) and their variances tend to be underestimated and the extent of underestimations mainly depends on the size of samples and the true values of ρ. Also, the reproduction of the estimated coefficients (β and ρ) in the “structural-functional” models of land prices can be almost accepted only if the coefficients are significant. Specifically, the error ratios are under 10% in 500 repetitions.
