scientific article; zbMATH DE number 3945186
The model considered is of the state space form \(X(t)=\Phi X(t-1)+\psi z+W(t),\quad Yt=M(t)X(t)+v_ t\). The special feature is the use of this representation to model a space-time situation so that X(t) may be decomposed into L parts, according to space locations, with m observations made at each location for each t. The term Starmax refers to such a space-time ARMAX model. Thus X(t) has components \(x_{t-j},\quad j=0,1,...,k-1\) and the state equation is \(\Sigma \Lambda_ jD_ jx_{t-j}=\Psi z_ t+w_ t,\quad \Lambda_ 0=D_ 0=I\), with the \(\Lambda_ j\) diagonal and representing special time lag effects whereas D describes interrelations between spatial components and may be completely prescribed. The method of fitting of the model is based on taking \(w_ t\), \(v_ t\) as Gaussian and serially independent and using the EM algorithm. The E step is effected by using the Kalman smoother to calculate the conditional expectations of X(t) given the data \(Y_ t\). The M step then becomes simple. The paper concludes with a discussion of modifications needed when there are observations missing at some locations.
This page was built for publication:
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3716145)