Estimates with lme(...varPower())
Dear Douglas Bates,
It would be more common to use the predicted response rather than the observed response for the varPower weights, as in varPower(form = ~fitted(.))
Following the example on page 223 of PB, I had used that first, but did not achieve convergence. Will try with more iterations.
If you use the observed responses then abnormally low observed responses will have unusually high weights in the fit, leading to low predictions, as you observed.
Dieter Menne