rifhdreg works on only 32 observations

[Nancy Barakat]

Hi all
I am running an unconditional quantile model on panel data from WBES to tackle the effect of firm-level mobile money adoption on energy intensity. From what I have read, I can use either xtrifreg or rifhdreg. I have used rifhdreg because it allows me to consider more than one fixed effect (firm, year, industry). Actually, it produces a perfect outcome for me; however, it runs on only 32 observations out of 700.
I would be grateful for any recommendations on how to remedy this .
Thanks a lot in advance

This is the code that I used:

rifhdreg EE20152 MM finance_measure R_D  human_capital labprod2015 exporter foreign_ownership ln_age i.firm_sizeWBES [pweight=average_wweak ] , rif(q(25))  vce(cluster combined_numregion ) abs(unique_id_01 year industry_type )

[FernandoRios]

Hi Nancy
What you describe is a problem that may be related to your sample
Specifically, what would happen if you run the same regression with reghdfe and no rif() option. Do you have the same results?
What about running a simple Regression where the "fixed effects" are added (without i.) to the model

Better yet, Consider all the variables in your model, including those used for cluster, fixed effects and weights. And provide summary statistics of them
F

[Nancy Barakat]

Hi Fernando
Thanks a lot for your fast reply. yes it doesnt differ with reghdfe and without i. option for fixed effects.

[FernandoRios]

that means that you have other problems in your data. its not because of rifhdreg, but because of missing data in your model
If you provide summary statistics for your data and description of it, it may provide some clues

[Nancy Barakat]

here is the summarised characteristics of my main variables:
Variable | Obs Mean Std. dev. Min Max
-------------+---------------------------------------------------------
EE20152 | 1,583 .0610209 .1240138 0 3.642857
MM | 3,343 .5145079 .4998642 0 1
finance_me~e | 4,668 2.16838 .6045419 1 3
R_D | 3,320 .2331325 .4228896 0 1
human_capi~l | 2,921 68.06607 35.44003 0 100
-------------+---------------------------------------------------------
labprod2015 | 4,146 4.673682 2.239081 1.163836 32.01134
exporter | 4,996 .1661329 .3722372 0 1
foreign_ow~p | 4,911 9.515598 27.31927 0 100
ln_age | 4,840 2.533776 .8534495 0 4.820282
|
firm_sizeW~S |
SMALL | 4,996 .5456365 .4979628 0 1
-------------+---------------------------------------------------------
MEDIUM | 4,996 .3040432 .4600471 0 1
LARGE | 4,996 .1503203 .3574208 0 1
|
combined_n~n | 4,996 11.52142 7.354198 1 25
|
industry_t~e |
OTHERSERV~S | 4,014 .2805182 .4493084 0 1
Manfactur~g | 4,014 .5201794 .4996549 0 1
-------------+---------------------------------------------------------
RETAIL | 4,014 .1993024 .3995256 0 1

[FernandoRios]

You see how there are so many missing values
it may be that only those 32 are available for all variables

[Nancy Barakat]

Actually I tried to run conditional quantile model before trying with unconditional quantile modelling and it worked on the whole sample. I have resorted to unconditional given its unrestricted assumptions to particular distribution type, that's why I am a little confused.

[FernandoRios]

Well, if you could share the data with me (find my email on the helpfile for rifhdreg), i can try to see if the error is due to your data or the command.

[Nancy Barakat]

Thanks a lot for your help, I sent an email. Highly appreciated

[FernandoRios]

Ok I just figure out your problem
You are absorbing unique_id_01, which identifies almost all observations uniquely. This means that after including that fixed effect, there is no variation left in the data to do any analysis.
Now, reghdfe (and by default rifhdreg) drops data where Fixed effects are singletons (no variation), which leaves you with only 32 observations.
If you drop this, you only have about 600obs left to do any analysis. So you still have to address that problem
HTH
F

[Nancy Barakat]

Thanks Fernando for your help. I really appreciate it. So, what should I do in that case, if I dont include it then I wont take into account firm time invariants? Is there any recommended remedies to such an issue?

[FernandoRios]

As i mentioned earlier, you basically have a pool crossection. In that context thinking about Firm time invariant factors makes no sense.