Regression is a widely used statistical technique in the social sciences, but most regression analyses applied to spatially referenced data that take account of latent spatial autocorrelation assume normally distributed error terms and involve maximum likelihood estimation. In traditional statistical analysis this situation results in a nonlinear structure, leading to estimation procedures that require slow iterative solutions. This makes spatial regression computationally intensive, thus slowing the dissemination of spatial statistics and spatial econometrics techniques. This project will explore useful approximations to this normalizing factor in spatial regression to simplify the analysis of intermediate and large size data sets and to demonstrate the utility of these simplifications. The approach will involve simulation experiments with timing functions to evaluate the costs and benefits of reductions in numerical intensity. Experiments will be conducted in a supercomputer environment. This project will assist in the removal of barriers to the dissemination of spatial statistical procedures, and it should reduce the numerical intensity normally associated with the use of spatial statistics in social science research.
