kets do not diminish immediately. Instead of estimating unrelated equations in the style
of (3.3) for different periods of time, a spatial SUR analysis is preferred. Contrary to the
familiar SUR model, the system
yii = β0 +β1 ∙ xli + ∑Y1 j ∙ω j + ν1i
S1
(3.4) y2i = β0 +β1 ∙ xl2i +∑γ2j ∙ωj +ν2i
S2
yTi = β0 +β1 ∙ xTli + ∑γTj ∙ωj +νTi
ST
is formed by the time periods for all n spatial units ( i=1,2,...,N), see Anselin (1988, pp.
137). The sets S1, S2, ..., ST of eigenvectors may be composed of different elements. If
the choice of spatial components for the sets St, t=1,2, ...,T, proves to be successful, the
disturbances νti will be contemporaneous uncorrelated:
(3.5) Cov(νti, Vti) = E(Vt1-Vtl) = 0.
Dependencies over time can be modelled by defining the NxN covariance matrix Σ,
(3.6) Σ = E(Vt -V’s) = σts-IN,
in which σts denotes the covariance between the periods t and s, t≠s, and νt an Nx1 vec-
tor of the disturbances in period t. With this the time-dependent autocorrelation struc-
ture of the spatial SUR system (3.4) is given by the NTxNT covariance matrix
(3.7) Σ* = Σ 0In.
Provided that T < N the covariance matrix Σl can be estimated directly from the data.
As dependencies over time are taken into account, parameters are estimated more effi-
ciently. In previous studies space-time models have been employed where spatial effects
were captured in a restrictive way by first order autoregressive lags, see e.g. Beck and
Katz (2001), Elhorst (2001) and Beck and Gleditsch (2003). Probably due to high com-
putational demands empirical applications of the spatial SUR model are rarely to find.
Exceptions are the works of Florax (1992) and Fingleton (2001). The advantage of Grif-
fith’s approach consists in allowing to set up the model straightforwardly with ordinary
devices taken from regression analysis.