Generate residuals for all observations (outside e(sample)) using reghdfe

[Meng Zhang]

I’m seeing that when I run:
reghdfe dep_var if treatID == 0, absorb(id year) resid(resid_all)
the stored residuals (resid_all) are only non-missing for the treatID == 0 observations (i.e., those actually used in the regression). All treatID == 1 rows show missing.

Because I restrict the estimation to the control subsample, reghdfe doesn’t compute or cache FE-adjusted predictions or residuals for the excluded (treated) group. However, for my two-stage DID I need residuals (or fitted values) across the entire panel, not just the sample that was regressed on.
I cannot simply switch to regress + predict because my dataset is enormous (millions of obs) and I have dozens of high-dimensional fixed effects—regress cannot handle that many FEs efficiently.
Has anyone discovered a workaround or option—within reghdfe or via a small manual step—that allows one to generate residuals for every observation, including those outside of e(sample)? Thanks!

[Clyde Schechter]

regress cannot handle that many FEs efficiently

Yes it can, if you are using the current version of Stata. -regress- now has an -absorb()- option.

Of course, if you just do a regression on the entire sample, you will get different results from what you got on the treatID == 0 sample. Now, I suspect that the motivation for doing that subset analysis in the first place is because you believe there is, or may well be, a difference in what you are estimating between the treatID == 0 and the complementary group. So why not do

reghdfe dep_var i.treatID, absorb(id year) resid(resid_all)

That will reflect any difference across the treatID groups. Or, if what you are really after is a model in which the difference between the treatID groups is deliberately ignored (constrained to be 0), -reghdfe dep_var, absorb(id year) resid(resid_all)- will do that.

[Meng Zhang]

Hi Clyde, thanks for your reply.
I’m using Kyle Butts’s heterogeneous‐robust two‐stage DID estimator (did2s) on a massive panel with dozens of high-dimensional fixed effects. I’m following his “Large Datasets or Many Fixed Effects” workflow (https://github.com/kylebutts/did2s_stata):

1. Stage 1: Regress y on all fixed effects using only the control group (treat== 0), then predict residuals for every observation in the full sample.
2. Stage 2: Regress those full-sample residuals on the treatment indicator (treat).

It’s critical that I residualize exclusively on the control subsample but still obtain residuals for all units before moving to Stage 2. Since I’m on Stata 16 MP and don’t have the built-in regress, absorb() option, I’ll try upgrading to Stata 18 and see if that resolves the issue—any other tips or workarounds would be greatly appreciated!

[Joao Santos Silva]

Dear Meng Zhang,

See if the following trick works.

clear all
sysuse auto
reghdfe price mpg if rep78!=., a(fe=foreign)
predict xb
qui egen double pair_FE=max(fe), by(foreign)
g yhat=xb+pair_FE
reg price mpg i.foreign if rep78!=.
predict yhat1
su y*

Best wishes,

Joao

[Meng Zhang]

Dear Meng Zhang,

See if the following trick works.

clear all
sysuse auto
reghdfe price mpg if rep78!=., a(fe=foreign)
predict xb
qui egen double pair_FE=max(fe), by(foreign)
g yhat=xb+pair_FE
reg price mpg i.foreign if rep78!=.
predict yhat1
su y*

Best wishes,

Joao

Dear Joao,

Thank you for your rapid and insightful reply—it works beautifully. I also extended your approach to cases with multiple fixed effects, and it functions as expected. Below is the code I used:

 use https://github.com/kylebutts/did2s_stata/raw/main/data/df_het.dta, clear
egen unique_id = group(state unit)
 
capture program drop did2s_est

program did2s_est, rclass
    version 13.0
    reghdfe dep_var  if treat == 0, absorb(new_id year, savefe)
    cap drop xb pair_FE1 pair_FE2
    predict xb
    qui egen double pair_FE1=max(__hdfe1__), by(new_id)
    qui egen double pair_FE2=max(__hdfe2__), by(year)
    tempvar dep_var_resid
    gen `dep_var_resid'= dep_var- xb-pair_FE1-pair_FE2
    regress `dep_var_resid' ib0.treat, nocons
end

xtset unique_id year
sort unique_id year
bootstrap, cluster(state) idcluster(new_id) group(unique_id) reps(100): did2s_est

I greatly appreciate your time and guidance on this. Any further suggestions are most welcome.

Best regards,
Meng

[Joao Santos Silva]

Dear Meng Zhang,

Glad it worked, credit should go to Tom Zylkin who showed me the trick (for ppmlhdfe).

Best wishes,

Joao

[Koko DIBLONI]

Dear Joao Santos Silva,

Could you give me a hand, please?

I'm analysing the effects of Fiscal consolidation episodes on FDI inflows. But I face a technical problem when I use lagged variable. Following my code:

ppmlhdfe in_Flow_per_r L2.Fisc_r log_GDP_r log_POP_r res_rents_r remit_gdp_r access_elec_r gov_eff_r gdp_growth_r fin_dev_r elec_consump_r , vce(cl iso_r iso_p) absorb(iso_p#iso_r iso_p#year) nolog d keepsingletons separation(none)


return: ppmlhdfe in_Flow_per_r L2.Fisc_r log_GDP_r log_POP_r res_rents_r remit_gdp_r access_elec_r gov_eff_r gdp_growth_r fin_dev_r elec_consump_r , vce(cl iso_r iso_p) absorb(iso_p#iso_r iso_p#year) nolog d keepsingletons separation(none)
not sorted

But I already did this:
egen ID=group(iso_r iso_p)
xtset ID year

sort iso_p iso_r year

[Joao Santos Silva]

Dear Koko DIBLONI,

I believe this question was answered elsewhere; please do not post questions multiple times. Anyway, are you sure you want to use the option "separation(none)"?

Best wishes,

Joao

[Koko DIBLONI]

Sorry for my mistake with the post.

Regarding the use of "separation (none)", I may have misunderstood its purpose. I'm still new to working with "ppmlhdfe". If you could kindly explain the issue with this option, that would be really helpful.

Best regards,
Koko

[Joao Santos Silva]

Dear Koko DIBLONI,

That option negated the main advantage of ppmlhdfe, which is to ensure the estimates exist. Please check the help file.

Best wishes,

Joao