FEs with no variation using reghdfe

[Adrienne Wold]

Hi,

I am trying to run a regression that uses the following design: measuring the effect of a treatment whether treatment is defined as being surveyed X days after an event and control is X days before the event, and my data groups together several different events. I use the following specification using the reghdfe command accessed via ssc install:

reghdfe  y i.treat $controls , a(event_id codenum) vce(robust)

where event_id are event FEs and codenum are locality FEs. Some events only have respondents in one of the groups (only treatment or only control respondents). But looking at bys event_id: tab treat if e(sample) I see that these are not being dropped from the estimation sample. Including vs manually excluding these events changes the output significantly, and I'm not sure how to proceed (exclude or include these events).

[Andrew Musau]

In the absence of controls, the coefficient on the treatment indicator will be the same because the model is effectively comparing means between groups. However, the standard errors would differ because the degrees of freedom change with the inclusion or exclusion of events, impacting the estimation of variability.

In the presence of controls, including the extra observations influences the coefficients on the control variables, which in turn affects the treatment effect estimate because these coefficients reflect the impact of treatment while holding other factors constant.

Excluding events with respondents in only one group (treatment or control) could bias the results if the occurrence of such events is not random but instead related to unobserved factors correlated with the outcome. However, if you have a theoretical reason to narrowly define your sample, exclusion might be justified.

[Adrienne Wold]

Thanks for this informative response. I'm not sure I understand why these events are not being dropped though - without variation in treatment and event fixed effects, what variation is being used for estimation?

[Carlo Lazzaro]

Adrienne:
as per FAQ, sharing via -dataex- an excerpt/example of your dataset that allows interested readers to replicate your research question, could help.