On Fri, 29 Nov 2002 11:21:28 -0000
Andrew Corroll <acorroll@caci.co.uk> wrote:
> Hello
>
> I am trying to adjust a logit model that was built on a 50:50 case control
> sample so that it reflects the true proportion of cases this number is <
> .001 of the population. I know why and how I can adjust for the intercept,
> but I would like some help in locating information about adjusting the
> coefficients for the covariates, or if infact if this is a good idea.
>
>
> Regards
>
>
> Andrew Corroll
>
>
Andrew - this is a good thing to worry about, because most people do a crude
unconditional prevalence adjustment, ignoring the fact that intercepts only
have meaning when adjusting for covariates, which means conditioning on all of
them. In other words, you can't just use the ratio of odds of target
prevalence to original prevalence (odds=1) as an intercept adjustment as is
commonly done. The only really safe way to adjust the intercept is to have
covariate data in both an oversampled (50:50) and randomly sampled population.
Compute Xnew*BetaOld and use it as an offset term in the random sample, in a
model with no other covariates, to get a new intercept.
--
Frank E Harrell Jr Prof. of Biostatistics & Statistics
Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences
U. Virginia School of Medicine http://hesweb1.med.virginia.edu/biostat
|