Application of pooled data techniques in the calibration of spatial interaction models
One of the problems with current calibration techniques for spatial interaction models is that they have been limited to cross-sectional data for a single point in time. Yet we know that the phenomena described by these single cross-sectional calibration techniques vary over time. One way to improve model calibrations with respect to this problem would be to pool the data used for calibration in the temporal dimension, i.e., to calibrate using multiple cross-sections of data.
Two pooled data techniques for calibration (the SDCP and TSCP methods) are developed in an effort to explain not only the temporal but also the spatial dimension of geographical regions in a spatial interaction manner. Numerical experiments using artificial data and real city data (the Sacramento, California, metropolitan region) are applied to these methods. The techniques in this study use primarily the multivariate logit method, the maximum likelihood criterion, and Kullback's divergence (or DRAM criterion for likelihood function) criterion for pooled calibration: MATLAB and GIS serve as tools for data preparation, aggregation, manipulation, processing, optimization, and presentation; Goodness-of-Fit, Reliability in forecasts of zonal changes in activity levels and Absolute Percentage Error serve as tools for comparison with the calibration technique for single cross-sectional data.
The pooled data techniques with highly acceptable Goodness-of-Fit can (1) smooth absolute percentage error values in the calibration and reduce the number of large absolute percentage error values in the calibration, thus forecasting the number of households with smaller variation; (2) capture the dynamics of residents over time (due to the use of more information), which the single cross-sectional technique alone cannot; and (3) provide an interpretation in which the pooled values of parameters are a closer approximation of the underlying behaviors of people over time.
Consequently, updating the pooled parameter values from the past with available data sets for a recent year (e.g., 2000) would improve the quality of residential location forecasting and the clarity of long-term trends due to its stability of model forecasts.
Area planning & development;
0999: Area planning & development