When we wrote the nlme package we tried to be very careful about
starting estimates and created the lmList function to allow evaluation
of coefficients from within-group linear fits. It turns out that it
is most useful for longitudinal data where there are enough
observations on each subject to do within-subject fits; that is, for
data like the sleepstudy data. Subsequent developments have made it
easy enough to fit the mixed model reliably without needing a lot of
preliminary work to get starting estimates. (Similarly, the
groupedData class was a good idea at the time but no longer
necessary.)
Would anyone miss the lmList function if it was removed from future
versions of lme4?