Date: Sun, 7 Feb 1999 08:20:05 +1100
Reply-To: Bigpond <tchur@BIGPOND.COM>
Sender: "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From: Bigpond <tchur@BIGPOND.COM>
Subject: Re: incidence rates & Poisson C.I.
Content-Type: text/plain; charset="iso-8859-1"
Panos Papanikolaou writes:
>I have used SAS extensively for data manipulation with big
>cross-section data and relatively a ittle bit for statistical
>analysis . So, I would appreciate very much if I would have your
>understanding and assistance to comprehend a few things about the
>statistical capabilities offered with SAS. For this reason, I am
>explaining my situation below.
>
>I have very small data set (N=7) which consists of Diabetes incidence
>rates observed over a seven years. I have been told to determine
>computationally the confidence intervals of the incidence rates
>assuming the Poisson distribution. Notice that in this area there is
>a piece of work that reports the incidence rates together with 95%
>Confidence Intervals computed (or estimated) separately for every
>single year (or observation if you like) and for the average value of
>incidence rate as well. The authors of the papers fitted Poisson
>regression models to the number of cases.
>
>With this background information, I wonder how i would possibly use
>SAS to compute the 95% Confidence Intervals determined seperately for
>each single case and for the overall mean value. I tried to find my
>way around but without any success. So, I would be very grateful to
>you all if you would offer me your advice and comments on the above
>problem; especially if this help would be accompanied by the
>provision of any specific references to SAS codes and/or statistical
>analysis in regard to the above.
>
>Thank you very much indeed for taking the time to consider my
>request. I look forward to hearing from you.
>
>Yours sincerely
>Panos Papanikolaou
>Research Fellow
>SONS
>UWCM
>Great Britain
Panos,
You don't mention whether the rates are to be age and sex adjusted as well,
but unless the rates are for a populations which have identical or nearly
identical age and sex distributions, they probably should be, particularly
since the incidence of diabetes (I am assuming that it is Type II or NIDDM
that you are talking about) increases sharply with age after about 45 years
of age.
There are two approaches: a) you can calculate traditional direct age (and
sex) standardisation and calculate Poisson CI's using a method described in
the early 1990's by Prof Annette Dobson et al - do a Medline search for
Dobson A in the author field and look for a paper in Statistics in Medicine
by Dobson and ??Kuulaamsa?? (sorry, a Finnish name with repeated vowels, not
sure of the spelling), titled something like: "Confidence intervals for
weighted sums of Poisson parameters". From this, you should be able to write
SAS code to implement the method they describe. If you have difficulties, I
would be happy to assist given some evidence that you have made an honest
attempt.
Alternatively, you can fit a log-linear model with a Poisson link function,
with age, sex and year as covariates. Use PROC GENMOD for this - there is a
example in the SAS documentation of how to model rates, complete with CI's.
Hope this helps,
Tim Churches
| 
