We are estimating the impact of an intervention on the practice of
health providers. One outcome measure is the ratio of metformin
prescribed to diabetic patients. This ratio is defined as the ratio
between the number of diabetic patients prescribed with metformin and
the total number of diabetic patients treated by the provider. We try
to examine whether metformin ratios are different for treatment group
and unexposed group before and after the intervention date. Xtgee is
used to estimate the impact of the intervention.
But there is a problem: the number of diabetic patients varies a lot
from one provider to another. Some had only one or two diabetic
patients in a quarter, while others had more than 50. If a provider
has only 2 diabetic patients, one of whom was prescribed with
Metformin, then one patient's drop from being prescribed with
metformin would lead to a change of 50% in the metformin ratio. In
contrast, a health provider with 100 diabetic patients will have very
stable metformin ratio. Thus, it is inappropriate to treat all
providers the same in the regression.
Should we use the pweight option in the regression? Or is there any
other better approach that works for this study? Thank you
