The setup for this question is as follows. Lets say we have count and denominator data for six different study sites, by month for 12 months. They're all using a uniform diagnostic standard at the time, so theoretically their counts are all roughly the same. The goal of this is to get an estimate of the typical monthly case count per person-time each study site gets, though we allow it to vary by month in the event that it changes dramatically.
So essentially, a pretty standard Poisson regression estimation of an incidence-density.
Except two of the sites change their diagnostic procedure to a more accurate one somewhere in the middle of the study period. Lets say Site 1 changes 2 months in, and Site 2 changes 8 months in, for the sake of the example.
The new procedure is believed to be of improved accuracy, but there's no direct data from the sites to produce our own accuracy measures, we'd have to go off published sensitivity/specificity figures from the two tests in the literature.
I'm trying to figure out a way to get around just out and out discarding the data from after the protocol change. Any ideas?
An update: One of the sites in question is about 3% of the total study person-time but about 10% of the cases, both because of the better test, and by the looks of it, an outbreak at the site at the time. So while "toss it" is still an option, it's not my favorite option.