Fast Weighted Regression for Non-Gaussian Time and Space Data

Keywords: GEE, conjugate gradient, generalised linear model

Abstract: Marginal generalised linear models offer a computationally efficient way to do regression with non-Normal time and space data. The conjugate gradient algorithm allows sparse working correlation structures to be used with only a constant factor extra computational complexity over unweighted estimation. With the use of robust weighted empirical variance estimators for valid inference this gives an increase in statistical efficiency without sacrificing the ability to perform data analysis interactively in moderate sized data sets.