All, would welcome some advice on how to model this!
I've got a natural experiment whereby staff at 2 branches (of around 30) were given a new process around mid-year. So I have x thousand staff, assigned to either 1 of 2 treatment branches, or one of 30 ish control branches. I have measures of my outcome variable per month or during the entire period (eg, Jan to June then July to December). I also think there is an interaction between the outcome variable and customer type.
What is a good way to measure the impact of the change in policy, controlling for the baseline effect in the branch before?
Initial plan was to have simple OLS with dummy effect per branch:
''' "Outcome ~ C(Branch) + C(CustomerCategory)*C(TreatmentPeriod)*C(TreatmentBranch)" '''
This suggest it works, but obviously due to autocolinearity, the C(branch) for treated branches is replaced by the TreatmentBranch effect. It's also messy because of all my dummies.
I also looked at panel data or mixed models, but I do wonder if there is a better way of modelling this - some sort of diff-in-diff or synthetic control, but I've never seen that with multiple treatment and control groups and trying to identify one overall effect. My main challenge is that my treatment was also distributed by group, so not sure how to both baseline by group and isolate my treamtment
