Keeping singletons in reghdfe

[William Rossi]

Dear community,
I have a curiosity about the reghdfe command. It suggests to drop singletons because inference could be biased if I use cluster robust standard errors.
I read the paper of Sergio Correia where he explains why this could be an issue and it seems to suggest that this would be a problem with many FE.

My question is: Should I be worried of the same problem in a simple TWFE regression (with only time and individual FE) or does this become a problem with multilevel FE.

Kind regards,
William

[Andrew Musau]

reghdfe is from https://github.com/sergiocorreia/reghdfe/ (FAQ Advice #12).

I have a curiosity about the reghdfe command. It suggests to drop singletons because inference could be biased if I use cluster robust standard errors.

The paper may argue in favor of this, but the command does this by default. So, if you use the command with its defaults, it is not a suggestion. The inclusion of singletons affects the estimation's degrees of freedom, which, in turn, impacts the calculation of the standard errors. Whether you have one fixed effects variable or multiple variables does not change this, although the issue may be exacerbated with many fixed effects variables due to the lower initial degrees of freedom.