ABSTRACT
This chapter discusses regression models specifically developed for processes described by spatially autocorrelated random variables. The effect of spatial autocorrelation, or apparent spatial autocorrelation, on regression models depends on how these spatial effects influence the data. In particular, specification of homoscedastic errors when in fact the errors are heteroscedastic can lead to both biased estimates of the regression coefficients and indication of spatial autocorrelation when none really exists. The conditional autoregressive, or conditional autoregressive (CAR), model is based on an alternative description of the probabilistic structure of the data. J. W. Lichstein et al. present a very thorough and extensive comparison of ordinary least squares and CAR models, with and without trend terms, for habitat suitability models for southern Appalachian songbirds. L. A. Waller and C. A. Gotway provide a thorough discussion of the issue of heteroscedastic residuals in spatial regression models.
