search

alexpkeil1 / Case-control-causal-review

GNU General Public License v3.0
0 stars 0 forks source link

Comment on ccc_interceptcorrection #2

Closed MiceMesidor closed 1 week ago

MiceMesidor commented 1 week ago

Hi Alex, Thanks for all the work on the R code. I don't have much more to add to it😊. I was wondering if it wouldn't be easier for people to calculate the offset using the same formula as in the document. https://github.com/alexpkeil1/Case-control-causal-review/blame/73dcc49e60690137a7a85bf5aa94a5bae1b89d89/code/ccc_interceptcorrection.R#L26 Let's say something like this:

n cases

n1 = sum(ccdat$y==1)

n controls

n0 = sum(ccdat$y==0)

intercept correction: population log-odds

ccdat$c0 = log((q0/(1-q0))/(n1/n0)) On the other hand, I suppose you add the minus sign to avoid convergence issues, or is there some other reason? Thanks

alexpkeil1 commented 1 week ago

I think either way is fine, and I agree it should match the paper, so I've updated it in commit ddb004c883c0bd71206fad9942101a422f58de0f. I was using the basic formula I've seen from Greenland (among others), but they give the same result since the sample size cancels out in the ratio. The negative sign is necessary because the offset term goes on the right side of the model, so it takes the overestimate of the baseline odds (due to undersampling non-cases from the population), and reduces it. Otherwise it doesn't work correctly in R (you can confirm by looking at the estimated intercept in the "m1" model fit to the cohort and the intercept in the "m_interceptcorrected" model, which should be approximately equal).