Different results when using reg/reghdfe compared to xtreg

[Cedric Lin Huang]

Dear Statalisters,
I'm running a model with fixed effects using reg, reghdfe and xtreg. Basically reg and reghdfe give me almost identical results except for the fact that results are less significant in reghdfe. But when I use xtreg, the results are dramatically different. One of the coefficient even has a different sign and it looks problematic. The codes I use to produce the results are presented below.
For reg I have:

reg hours_worked $control i.year i.industry1 i.level1, vce(robust)

Similar to this I have the following for reghdfe

reghdfe hours_worked $control , vce(cluster statefips) abs(year industry1 level1)

For xtreg I have:

xtreg hours_worked $Control i.year i.industry1 i.level1, fe vce(robust)

For reg the results arecoef given by reghdfe are identical)

	Coef.	Std. Err
Local wage	0.354	0.008
Wage difference ratio	0.547	0.003
Unemployment	0.029	0.002
CPI	0.589	0.026

For xtreg the results are

	Coef.	Std. Err
Local wage	-0.268	0.065
Wage difference ratio	0.046	0.015
Unemployment	-0.072	0.005
CPI	3.99	0.208

Could someone give me a hint as to why the two different methods give completely different results?
Thank you all so much

[Andrew Musau]

xtreg requires you to declare a panel identifier, and estimation after specifying the -fe- option includes those panel fixed effects.

[Cedric Lin Huang]

xtreg requires you to declare a panel identifier, and estimation after specifying the -fe- option includes those panel fixed effects.

Hi Andrew, I didn't include that part but the panel identification is identical throughout the different regression approaches that I've used.

[Andrew Musau]

You declare data to be panel using xtset

[XT] xtset -- Declare data to be panel data
(View complete PDF manual entry)


Syntax

Declare data to be panel

xtset panelvar

xtset panelvar timevar [, tsoptions]

Therefore, what I mean in #2 is

webuse grunfeld
xtset company
xtreg invest kstock, fe

already includes the company fixed effects as I have declared "company" as the panel identifier. With regress (including indicators), I would estimate the model as

regress invest kstock i.company

[Cedric Lin Huang]

@Andrew Musau
Thanks for the comment. As I'm working with individual data, User ID is my panel identifier and does not appear in any of my fe specifications. Do you think this could be the problem? (not including User ID in the specifications) My study is about min wage and how it affects the number of hours that individuals work.

[Andrew Musau]

My study is about min wage and how it affects the number of hours that individuals work.

Yes, as your unit of analysis is the individual and you have panel data, you should take into account individual fixed effects.

User ID is my panel identifier and does not appear in any of my fe specifications.

It does in your xtreg estimation as you have xtset your data. The -vce(robust)- in xtreg is equivalent to -cluster(User_ID)-, so the estimation in reghdfe (SSC) should be

reghdfe hours_worked $control , vce(cluster User_ID) abs(year industry1 level1 User_ID)

Some of the fixed effects may be redundant, e.g., if individuals are nested within industries - but it does not matter that you include both variables within -absorb()-.

[Cedric Lin Huang]

@Andrew Musau
Again thank you so much for all the comments. They have been extremely helpful.
Now that I've updated my codes for reghdfe, they are giving me identical results compared to xtset.
Like you said in my previous reg regression I didn't include individual fixed effects. I've tried to include i.user_id (code is given below) but now stata is saying that maxvar is too small. I've tried to set the maxvar to 120,000 but the same problem still persists.

reg hours_worked $control i.year i.industry1 i.user_id, vce(robust)

Given the nature of my study do you think it's safe to disregard reg and just focus on xtset and reghdfe? Is there a way to bypass the maxvar issue?

[Andrew Musau]

With all those indicators, you are asking regress to invert a very large matrix and it is telling you that it has no scope to do that. However, regress does allow areg syntax, so you can do

reg hours_worked $control i.year i.industry1 , absorb(user_id) cluster(user_id)

Two points:

1. The estimator does not really matter. You have panel data, so you should prefer xtreg and reghdfe over regress.
2. If using regress or reghdfe, the option -robust- computes Huber-White standard errors. These are not appropriate with panel data, so you need to specifically cluster on the panel variable. xtreg gets the clustering correct with the option -robust-.

[Cedric Lin Huang]

@Andrew Musau Thank you for the comments. I think all my problems are solved.