Fixed effects along with robust standard errors at firm level

[Farah Ali]

Hi,

I am working on a model following the approach used by Benlemlih and Bitar (2018), where they have used industry fixed effects and also stated to use OLS with robust standard errors clustered at firm level. Whilst I was trying to apply the approach, I am not sure as how both of these will be incorporated into one model. I am thinking on using the following model, but not sure if its sufficient.

xtreg DV IV controlvar i.year i.SIC fe cluster(SIC)

Does this model suffice all of above mentioned methods (ndustry fixed effects and also stated to use OLS with robust standard errors clustered at firm level) or do i need to use any other additional model?

[Clyde Schechter]

Well, it is bad enough when people post questions that require the responder to read their mind. You have posted a question that requires the responder to read the mind of an author of some article that you give only scanty information about, not enough for somebody not already familiar with it to find the article. (Remember that this is a multi-disciplinary, international forum and things that may be folklore in your circle may be mystifying to others.)

There is a sense in which -xtreg- uses OLS: the estimates of the coefficients and standard errors are actually calculated by applying OLS to the demeaned data. But that may or may not be what B & B had in mind. Maybe the ran two models, one using -xtreg, fe- (fixed effects) and a separate one using OLS on the pooled data (-regress-). To resolve that would require, at the least, carefully reading the article. For that matter, it might not even be clear after carefully reading the article and you might have to contact one of the authors.

If you want more concrete advice here, please post back with either a complete reference, or, better (because many responders will lack the time or the access to the journal to retrieve the article) the most relevant quote from the article describing what they did. Perhaps we can help you interpret their words.

By the way, the advice not to put up abbreviated references is in the Forum FAQ, as well as lots of other good advice that, if followed, improves your chances of getting a timely and helpful response. It is well worth your time to read them and take them to heart.

[Farah Ali]

Hi Clyde

My apologies for creating an awkward post......I have carefully noted your advice and will follow that now as well as in future.....

I have attached the reference for the paper i was discussing about... "Benlemlih, M., & Bitar, M. (2018). Corporate social responsibility and investment efficiency. Journal of Business Ethics, 148(3), 647-671." I will also quote their words as described.

"We include industry dummy variables to control for industry fixed effects, which may affect the relationship between firms’ investment efficiency and social performance scores. Industry dummy variables are based on the first two digits of the SIC code. We also include dummy variables for each year in our sample period (i.e. year fixed effects) to control for changing economic conditions.
We use ordinary least squares (OLS) specifications with robust standard errors adjusted for both heteroscedasticity and clustering of observations. More precisely, we use Petersen’s (2009) one-way cluster-robust standard errors approach at the firm level. This technique is shown by Petersen (2009) to be the preferred method for estimating standard errors in corporate finance applications using panel data".

[Clyde Schechter]

This sounds to me like they did

regress DV IV controlvar i.year i.SIC,  vce(cluster SIC)

which would be completely equivalent to:

xtset SIC
xtreg DV IV controlvar i.year, fe vce(cluster SIC)

[Joro Kolev]

From the description in #3 seems that they have

 
 regress DV IV controlvar i.year i.SIC,  vce(cluster FirmID)

where FirmID is the firm identifier variable. They say they cluster at the firm level (not at the industry level which is the SIC code).

[David Benson]

EDIT: This crossed with Joro's post.

It sounds to me like they included dummy (indicator) variables for the SIC code, and clustered the standard errors at the firm level. I am not familiar with "Petersen’s (2009) one-way cluster-robust standard errors."

Also, you could email the authors and they could tell you.

The tables say:

All the models include industry and year fixed effects...Robust t-statistics corrected for clustering at the firm level are presented in parentheses.

* No firm-fixed effects
regress DV IV controlvar i.SIC i.year,  vce(cluster firm_id)

* Both of these add firm fixed effects (one xtsets first, the other doesn't)
xtset firm_id year
xtreg DV IV controlvar i.SIC i.year, fe vce(cluster firm_id)

* If you don't want to xtset
xtreg DV IV controlvar i.SIC i.year, fe i(firm_id) vce(cluster firm_id)

[Joro Kolev]

. I am not familiar with "Petersen’s (2009) one-way cluster-robust standard errors."

I can explain in great detail and with plenty of irritation what are the so called "Petersen’s (2009) one-way cluster-robust standard errors."

These are the Stata clustered standard errors of Rogers, W. H. 1993. Regression standard errors in clustered samples. Stata Technical Bulletin 13: 19–23. Reprinted in Stata Technical Bulletin Reprints, vol. 3, 88–94.

Which a little bit later were shown by Arellano to be perfectly fine in fixed effects models, Arellano, M., 1987. PRACTITIONERS’CORNER: Computing Robust Standard Errors for Within‐groups Estimators. Oxford bulletin of Economics and Statistics, 49(4), pp.431-434..

Another very relevant reference is Froot, K. A. 1989. Consistent covariance matrix estimation with cross-sectional dependence and heteroskedasticity in financial data. Journal of Financial and Quantitative Analysis 24: 333–355.

The point here is that Petersen (2009) is a survey article summarising methods that were invented 22 years before the survey, and programmed in major software 16 years before the survey he wrote.

Why would anybody call those "Petersen’s (2009) one-way cluster-robust standard errors," is beyond me.

[Farah Ali]

Thankyou all. I will try these codes and hope they work....