results with xthdidregress change with order of control variables

[Olov Isaksson]

Hi,

I am trying to estimate the ATET (e.g. on sales) of a staggered implementation of a new technology at a retailer. I have a panel with different stores over time. The stores had the technology implemented at different points in time and we have reasons to believe that the treatment effect is both dynamic and heterogeneous. I am using xthdidregress and have run into some strange results. The ATET changes depending on the order in which I specify my control variables. I get different outcomes for the two specifications below:

xthdidregress aipw (sales i.store_id store_traffic) (treated) , group(store_id) controlgroup(notyet)

xthdidregress aipw (sales store_traffic i.store_id) (treated) , group(store_id) controlgroup(notyet)

where:
sales is the sales in store i at time t
store_id is a unique identifier for each store
store_traffic is the number of visitors in store i at time t
treated is an indicator variable that takes the value 1 if a store has received treatment (after technology has been implemented). Otherwise 0.

The only difference between the two specifications is the order in which I write the controls store_traffic and i.store_id.

Does anyone have a logical explanation for this? Thanks in advance!

[FernandoRios]

Hi Olov
The problem you are facing is also present if you use csdid (my implementation of the method).
The reason of the problem comes from the i.store_id. Specifically, you cant add Store fixed effects, because it violates the overlapping assumption (only some "stores" are treated).
If you drop that, your results should be consistent.
F