On Thu, 19 Aug 2004, Duncan Lee wrote:
> I want to run a GEE for a time series of counts. The data are daily
> respiratory mortality counts and so there aren't any
'clusters' in the
> longitudinal sense. Neither the gee or geese packages work. The gee one
> wont run at all and the geese one produces NaNs or just runs
> indefinitely depending on how long the time series is. Any ideas how to
> make these work of any other packages that might do the trick?
>
There are sandwich estimators for time series, but they are slightly more
complicated than for GEE, because there are no simple replicates. The
"sandwich" package provides some of these.
One of these, the Newey-West estimator, is closely related to an ad hoc
estimator that lots of people have invented. If you break your time
series into chunks that are long enough to be roughly independent you
could just use GEE. The Newey-West estimator does this, but averages over
different places you could have put the break points (eg with month-long
chunks should they start at the 1st of the month, the 2nd, the 15th,...).
These estimators are also closely related to the block bootstrap and the
subsampling estimators for time series. One place to find descriptions of
this is
Lumley T, Heagerty PJ, (1999) "Weighted Empirical Adaptive Variance
Estimators for Correlated Data Regression" Journal of the Royal
Statistical Society, Series B.61:459-477
Econometricians will point out (with some justification) that they
invented all this stuff decades ago. On the other hand, AFAICS they
didn't point out to anyone that you could do this for generalized linear
models as well.
-thomas