LSDV vs. FE postestimation predict fe_effects, u

[Jonas Lee]

Dear everyone,
I regressed this LSDV model in Stata
regress delta_yield delta_BA i.BondID, vce(cluster BondID), usually one is not interested in the individual dummy estimates but those are important for my further research.
I also conducted the FE regression with xtreg delta_yield delta_BA, fe vce(cluster BondID) and performed afterwards predict fe_effects, u to get postestimation of the individual fixed effects estimates per individual. Although both are giving me different estimates and I am unsure what approach to use and what exactly the predict command in Stata performs. Any help is highly appreciated!

[Sebastian Kripfganz]

They different estimates are not fundamentally different. They just have a different reference point.

When you run the regress command, the dummy for the first BondID is omitted; this becomes your reference group. All other dummy coefficients are then in relation to this omitted group.

When you run xtreg, fe, the predicted fixed effects are in relation to their overall mean; thus, it is guaranteed that the average over all fixed effects has mean zero.

You can verify that the difference between any two dummy coefficients from regress equals the difference between the corresponding fixed effects predicted by xtreg.

In any case, when the number of BondIDs is large relative to the number of time periods, none of these effects are consistently estimated. Their usefulness for further analysis is thus generally very limited.

[Jonas Lee]

They different estimates are not fundamentally different. They just have a different reference point.

When you run the regress command, the dummy for the first BondID is omitted; this becomes your reference group. All other dummy coefficients are then in relation to this omitted group.

When you run xtreg, fe, the predicted fixed effects are in relation to their overall mean; thus, it is guaranteed that the average over all fixed effects has mean zero.

You can verify that the difference between any two dummy coefficients from regress equals the difference between the corresponding fixed effects predicted by xtreg.

In any case, when the number of BondIDs is large relative to the number of time periods, none of these effects are consistently estimated. Their usefulness for further analysis is thus generally very limited.

Dear Sebastian, many thanks for your comments, they are highly appreciated! Relating to your last statement, the number of BondIDs is 120 where as the number of observations are in total 35,000+. My goal is primary to capture the unobserved heterogeneity in my data. Would you recommend in my case to use the LSDV Dummies is relation to their reference group or the predicted fixed effects from Stata to capture those? I calculated both scenarios on Stata, but the mean form my dummies is slightly negativ whereas the mean of the predicted fixed effects is slightly positiv. Unfortunately, the sign is necessary for the interpretation and I am unsure what approach is more unbiased. Your comment on this would be very useful! Cheers!