We are currently analyzing data on children clustered in day care-centers (DCC).
We have tried to use geepack and gee libraries to estimate an overall incidence
rate for absences (=number of absences/risk time) by specifying
geese(number.absences ~ offset(log(risktime)), id=day.care.id,
family=poisson("log"), data=dcc,
corstr="exch",
sca.link="log", cor.link="fisherz")
gee(number.absences ~ offset(log(risktime)), id=day.care.id,
family=poisson, data=dcc, corstr="exchangeable")
However it returns a value error of 1 ,in some cases it returnes NaN estimates,
andin the case or gee, it hangs. We intend eventually to add other covariates we
are interested in.
Our clusters (day-care centers) include about 50 children each, and in one case
over 100. By taking a smaller number of children in each day care center, we
managed to obtain convergence, but as long as the cluster size was under 25
(i.e. no day care center larger than 25 children).
Is the geese and gee functions limited by the size of the cluster? And if so,
are there any suggestion how to go around the problem?
Thank you for your help.
Sincerely,
Sharon
===========================================Sharon K?hlmann Berenzon, Ph.D.
Statistician
Dept. Epidemiology
Swedish Institute for Infectious Disease Control (SMI)
Sharon.Kuhlmann at smi.ki.se
tel. +46-8-457 2376; fax. +46-8-30 06 26
