Rework glm.q to use new covariance routines