clustered standard errors one per observation?

[Max Piper]

Hello,

in an investigation of a panel of firms across several years, I want to control for industry and year fixed effects and cluster standard errors by firm. I read multiple times in works I cited that they clustered their standard errors by firm and year. However, there are not multiple observations per firm and year in their or in my data. That is, every observation is uniquely identified by one firm-year combination - one observation is one firm-year.

I understand that clustered standard errors control for heterogeneity within each group of the variables that one clusters by, and that this adjusts standard errors accordingly.
I would create the groups as:

egen c=group(firm_id year_id)

and cluster on c, i.e.

reg firm_value ROA ln(TA), vce(cluster c)

(following: https://www.stata.com/statalist/arch.../msg00424.html).

Perhaps I have a wrong understanding of this, but how can one control for heterogeneity within groups (i.e. firm-year-combinations), if one only has one observation per group? What does heterogeneity within groups then even mean? Is there an intuitive explanation for this? How could the results be interpreted - if at all?

[Phil Bromiley]

I suspect that what they've done is use gls to allow within-firm over years, and across-firms within years correlations. With panel data, this is xtgls (although there are other procedures in Stata that may give equivalent estimates). I don't see how it makes any sense to have a cluster per observations.