lloydheng
commented
1 year ago Moving forward, we're no longer imputing missing months with the mean of the observed months in the baseline year. Restaurants are penalized in their weights if they're missing months (and consequently transaction count).
This means the numerator is a simple sum of restaurant-treatment-specific transaction count in the baseline year, and the denominator is a simple sum of the treatment-specific transaction count of all restaurants in that treatment during the baseline year. We continue multiplying it by treatment N to center weights around 1.
Yes, again 🙀
This affects a minority of restaurant-pairs, so I'm moving ahead while we decide on this. Original set-up:
$$weight{ij} = \frac{\frac{12}{length(t{ij})} \times \sum count{ijt}}{\sum count{jt}} \times length(i_j) $$
Where:
If a restaurant is missing data in the baseline-year, we are essentially imputing the missing months with the mean of the remaining observed months during that period. Conceptually, I think this implies that we are up-weighting restaurants with missing baseline observations, which sounds counterintuitive to me.
Even if we think that's good, we should include the imputed months' counts in the denominator so the numerators sum to the denominators, yes?
Discuss.