I have just noticed a difference in behavior between coxph and lm/glm: if one or more of the coefficients from the fit in NA, then lm and glm omit that row/column from the variance matrix; while coxph retains it but sets the values to zero. Is this something that should be "fixed", i.e., made to agree? I suspect that doing so will break other packages, but then NA coefs are rather rare so perhaps not. Terry Therneau