Spatial Statistics Manuscripts
This page contains links to various spatial statistical working papers and spatial papers under review.
A conditional spatial autoregression (CAR) specifies dependence via a weight
matrix. Using a doubly stochastic weight matrix allows users to interpret the CAR
prediction rule as a semiparametric prediction rule and as BLUP with smoothing in addition
to other interpretive benefits. We examine standard and doubly stochastic weight matrices
in the context of two illustrative data sets to demonstrate feasibility and statistical
(1.7 megabytes in size)
This manuscript uses matrix exponentials to hold the log-determinant term constant in the likelihood. Consequently, the maximum likelihood estimates have a closed-form in terms of the eigenvalues of a small matrix. The spatial autoregression involving 57,647 observations takes under 1 second to compute using Fortran 90 on a 600 Mhz PC (conditional upon knowing the neighbors).
The Fortran 90 code (source and PC executable) and data used in the paper reside in a 3.7 megabyte zip file.
While we strongly urge using the pdf files, you can also view the HTML Version .
This manuscript uses matrix exponentials to hold the log-determinant term constant in the likelihood. This greatly simplifies the computation of Bayesian estimates via MCMC methods when the spatial structure can change across iterations. The paper demonstrates the computational advantages on data of various sizes (largest size was over 35,000 observations), shows the Bayesian estimator performs well with generated data, and examines the spatial structure of spillovers and externalities in an urban setting.