[R] Poisson Regression

[Martyn Plummer]

On Fri, 2003-03-28 at 12:49, Meles MELES wrote:
> Hi,
>         somebody asked me to do a Poisson regression on cancer incidence over
> years to see wether if there is a descending or an ascending tendancy.
> I tried with R, but it complains that the data are not integer but
> floating poing data. So, as I never did that sort of thing before, I'm a 
> bit lost.
> 
> the data serie looks like this
> SEX  PERIOD     INCIDENCE
> 1    [1978;1982]  57,3
> 1    [1983;1987]  58,3
> 1    [1988;1992]  65
> 1    [1993;1997]  56,8
> 2    [1978;1982]  4,6
> 2    [1983;1987]  5
> 2    [1988;1992]  5,1
> 2    [1993;1997]  6,2
> 
> Any suggestions?

You don't have enough information. You need the number of events (y) and
the number of person-years at risk (pyar) separately. What you have is
the incidence rate (y/pyar) * (some scale factor).

With the correct data, your model will look something like this

glm(y ~ sex + period + offset(log(pyar)), family=poisson)

Judging by the PERIOD variable, I guess that you are looking at cancer
incidence data from volumes V to VIII of "Cancer Incidence in Five
Continents". In this case it is possible to reconstruce the original
age-specific incidence data from the published incidence rate and the
population pyramid data.

Martyn