For both linear mixed models and marginal models, the generalized least squares estimator ????̂ = (

XTV−1X

)−1 XTV−1y is useful. Suppose V is not the true var(y) but we can consistently estimate var(y) empirically by S. Describe a way to do this. Explain why it is then sensible to estimate var(????̂) by

(

XTV̂ −1 X

)−1 XTV̂ −1 SV̂ −1 X

(

XTV̂ −1 X

)−1

.