[R] Modelling an "incomplete Poisson" distribution ?

Emmanuel Charpentier charpent at bacbuc.dyndns.org
Sat Apr 18 19:28:56 CEST 2009


Dear list,

I have the following problem : I want to model a series of observations
of a given hospital activity on various days under various conditions.
among my "outcomes" (dependent variables) is the number of patients for
which a certain procedure is done. The problem is that, when no relevant
patient is hospitalized on said day, there is no observation (for which
the "number of patients" item would be 0). 

My goal is to model this number of patients as a function of the
"various conditions" described by my independant variables, mosty of
them observed but uncontrolled, some of them unobservable (random
effects). I am tempted to model them along the lines of :

glm(NoP~X+Y+..., data=MyData, family=poisson(link=log))

or (accounting for some random effects) :

lmer(NoP~X+Y....+(X|Center)), data=Mydata, family=poisson(link=log))

While the preliminary analysis suggest that (the right part of) a
Poisson distribution might be reasonable for all real observations, the
lack of observations with count==0 bothers me.

Is there a way to cajole glm (and lmer, by the way) into modelling these
data to an "incomplete Poisson" model, i. e. with unobserved "0"
values ?

Sincerely,

						Emmanuel Charpentier




More information about the R-help mailing list