Dear S-Plusers,
I am bootstrapping a mixed model analysis of family data. (I am resampling
families.) There are about 300 families containing about 670 individuals
represented in the data. Each individual is assessed up to 7 times. Each
assessment produces values of up to 4 different measures. Because of the
family, temporal, and within assessment dependence I am using 'lme' to fit the
model. I am interested in the various parameters of the random part of the
model. The calls to 'lme' look like this:
lme(fixed = formula.fixed, data = resampled.data, random = list(family = ~-1 +
measure.type, subject = ~-1 + measure.type), correlation = corSymm(form = ~1 |
family/subject/time.point), weights = varIdent(form = ~1 | measure.type),
method = "REML")
The fixed part of the model (formula.fixed) is large. It contains 28 terms.
Some of the "families" have rather spotty data, for example, one person
assessed at one time point. Perhaps for these reasons, for most resamples
'lme' issues a warning like this: "Singular precision matrix in level -1, block
16". But this is just a warning. 'lme' still generates plausible looking
parameter estimates. I still get frequent warnings even when I eliminate the
families with the scantiest data. I get fewer warnings if I use the jackknife
instead of the bootstrap. My question is, can I just ignore the warnings and
use the estimates or does the warning indicate that the corresponding estimates
are garbage? Thank you.
-- Steve Ellis
----------------------------------------------------------------
|