Abstract
Separate space- or time-lags have been considered regularly in data analyses; as space–time models are more recently being studied extensively in data analytic fashion, joint estimation of both lags has to be considered explicitly. This paper addresses this issue, taking into special consideration parametric parsimony together with specification richness; use of the bivariate Poisson frequency distribution is advocated and applied to an empirical case. The relation of this approach to random effects specifications is investigated. Data for Belgian regional products constitute the empirical case study.
Similar content being viewed by others
Notes
This decomposition is similar to that for principal components analysis (PCA) in multivariate statistics. A key difference is that PCA eigenvectors are used to construct synthetic variates that are linear combinations of attribute variables, whereas here the eigenvectors themselves are used as synthetic variates. De Jong et al. (1984), and Tiefelsdorf and Boots (1995), show that the eigenvalues associated with these eigenvectors cover the complete range of spatial autocorrelation possibilities for a given connectivity matrix. As such, the eigenvectors identify global, regional, or more local components of spatial autocorrelation.
References
Bresson G, Pirotte A (1995) Econométrie des séries temporelles, Théorie et applications. Presses Universitaires de France, Paris
de Jong P, Sprenger C, van Veen F (1984) On extreme values of Moran’s I and Geary’s c. Geogr Anal 16:17–24
Fylstrom D, Lasdon L, Watson J, Waren A (1998) Design and use of the Microsoft Excel solver. Interfaces 28(5):29–55. doi:10.1287/inte.28.5.29
Gi L, Zhang T (2007) Loglinear residual tests of Moran’s I autocorrelation and their applications to Kentucky breast cancer data. Geogr Anal 39(3):293–310. doi:10.1111/j.1538-4632.2007.00705.x
Griffith D (2000) A linear regression solution to the spatial autocorrelation problem. J Geogr Syst 2(2):141–156. doi:10.1007/PL00011451
Griffith D (2003) Spatial autocorrelation and spatial filtering: gaining understanding through theory and scientific visualization. Springer, Berlin
Jones JIII, Casetti E (1992) Applications of the expansion method. Routledge, London
Kaashoek JF, Paelinck JHP, Zoller HG (2004) On connectropy. In: Getis A, Mur J, Zoller HG (eds) Spatial econometrics and spatial statistics. Palgrave, London, pp 217–231
Ord JK (1972) Families of distribution functions. Griffin, London
Ozaki T (1985) Non-linear time series models and dynamical systems. In: Hannon EJ, Krishnaiah PR, Rao MM (eds) Handbook of statistics, vol 5: time series in the time domain. North Holland Publishing Company, Amsterdam, pp 25–84
Paelinck JHP (2005) Lotka-Volterra models revisited: a mixed specification with endogenous generated sdls-variables, and a finite automaton version, paper presented at the Las Vegas meetings of the North American Regional Science Association, November 2005
Paelinck JHP, Klaassen LH (1979) Spatial econometrics. Saxon House, Farnborough
Taylor L (1974) Estimation by minimizing the sum of absolute errors. In: Zaremba P (ed) Frontiers in Econometrics. Academic Press, New York, pp 169–190
Tiefelsdorf M, Boots B (1995) The exact distribution of Moran’s I. Environ Plan A 27(6):985–999. doi:10.1068/a270985
Author information
Authors and Affiliations
Corresponding author
Appendix A
Appendix A
1.1 Significance of least absolute difference (LAD) estimators
LAD estimators are derived from a square sub-matrix of the usual X-matrix in the linear model (see Taylor 1974, pp. 169–190):
Call that sub-matrix X *; then (the capped \( {\hat{\varvec{\upbeta }}}\) being the vector of estimators):
where y * is a sub-vector of y corresponding to X *; if \( \varvec{\upbeta} \) is k × 1, then X * is k × k and y * is k × 1.
and if ε is symmetrically distributed around zero, even if X * is a random matrix, then \( \varvec{\upbeta} \) is unbiased (Taylor, Theorem 5); under the usual assumptions, its variance is given by
To derive X *, one computes, for all observations,
where \( \mathop = \limits^{\Updelta } \) denotes “per definition,” and for the relevant range of y, e* = 0 (from the properties of the LAD-estimators; as in practice, here a reduced gradient method—Fylstrom a.o., 1998—is used instead of linear programming; e* should correspond to the k smallest errors).
Then \( \hat{\sigma }^{2} \) can be computed from RSS/(n – k − 1), as the average of e is not necessarily equal to zero. The square root of the diagonal of the expression in Eq. 10 (with estimated σ2) gives the desired standard deviations, which then can be used to conduct classical t tests.
From RSS and the total sum of squares, \( {\text{TSS}} = \left[ {{\mathbf{y}} - {\varvec{\upmu }}({\mathbf{y}})} \right]^{\prime } \left[ {{\mathbf{y}} - {\varvec{\upmu }}({\mathbf{y}})} \right], \) and the pseudo- R 2 can be computed as
and submitted for a classical F test.
Rights and permissions
About this article
Cite this article
Griffith, D.A., Paelinck, J.H.P. Specifying a joint space- and time-lag using a bivariate Poisson distribution. J Geogr Syst 11, 23–36 (2009). https://doi.org/10.1007/s10109-008-0075-3
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10109-008-0075-3