Event-study results change when extending time window (reghdfe)

[marie carmel]

Hello everyone,

I’m running a stacked event-study using reghdfe in Stata, and I’ve encountered an unexpected issue that I’d like to understand better.
First, this is how I generate my stack for each treatment year :

* Save dataset in memory, so we can call this function multiple times.

preserve


* Determine earliest and latest time in the data.

* Used for feasibility check later

sum `timeID'

local minTime = r(min)

local maxTime = r(max)




*variable to label sub-experiment if treated in focalAdoptionTime,

gen sub_exp = `focalAdoptionTime' if `adoptionTime' == `focalAdoptionTime'



*Now fill in this variable for states with adoptionTime > focalAdoptionTime + kappa_post



*Include only never treated

replace sub_exp = `focalAdoptionTime' if never_treat == 1



*note, this will include never treated, because adopt_year is ., which stata counts as infinity

*replace sub_exp = `focalAdoptionTime' if `adoptionTime' > `focalAdoptionTime' + `kappa_post'



*Keep only treated and clean controls

keep if sub_exp != .



*gen treat variable in subexperiment

gen treat = `adoptionTime' == `focalAdoptionTime'



*gen event_time and

gen event_time = year - sub_exp



*gen post variable

gen post = event_time >= 0



*trim based on kappa's: -kappa_pre < event_time < kappa_post

keep if inrange(event_time, -`kappa_pre', `kappa_post')



*keep if event_time >= -`kappa_pre' & event_time <= `kappa_post'

gen feasible = 0

replace feasible = 1 if !missing(`adoptionTime')

replace feasible = 0 if `adoptionTime' < `minTime' + `kappa_pre'

replace feasible = 0 if `adoptionTime' > `maxTime' - `kappa_post'

drop if `adoptionTime' < `minTime' + `kappa_pre'




* Save dataset

compress

save "subexp`focalAdoptionTime'", replace



restore



When I estimate my model over the window [–3,+2], I get positive post-treatment coefficients.
However, when I extend the window to [–3,+3], the coefficients for +1 and +2 become much smaller and negative — even though the treated and control groups are identical across both samples.
To make sure the change wasn’t due to sample composition, I:

1. Fixed the treated group (same departments in both windows).
2. Used only never-treated units as controls.
3. I also verified that the number of observations is identical across windows for each window. The only two exception are two stacks (2016 and 2019) gain extra observations (only for that time) at event_time == 3.

When I restrict the [–3,+3] sample to if inrange(event_time,-3,2), the results match perfectly with the [–3,+2] run — so the change clearly comes from the new event_time == 3 observations in those two stacks.

My interpretation is that adding those +3 observations shifts the within-stack means absorbed by sub_exp#event_time, which in turn re-centers all the other coefficients. Even though I extend the time window, I should not except my coefficient for time 1 and 2 to be so different.

Has anyone experienced something similar?