Adding industry fixed effects or clustering

[Sebastiaan Fromme]

Dear all,

I am doing an event study concerning abnormal stock returns around M&A announcements. I have a dataset in which I have included over 4000 deals done by almost 2000 different companies from 1996-2016. I have corporate governance and other financial data of the acquirer for each deal. I am doing a standard OLS regression in which the abnormal returns are my dependent variables, the corporate governance data is my independent variable and I have included some financials as controls. I have included time fixed effects.

I am a bit in doubt how I should proceed with the adding of industry fixed effects or clustering. Some papers that I have read have included industry fixed effects. They usually do this with the use of the Fama & French industry division. I have this information also in my dataset. However, other papers (also in this field of research) state that they adjust for acquirer clustering. I also have the CUSIP number of everyfirm in my dataset, should I use this number to cluster a the acquirer firm's level? I looked a bit online but I could not find a simple answer to what the difference is between the two and what I should do.

Can someone tell what the precise difference between industry fixed effects and clustering is and which one would be more suitable for my research?

Thanks in advance.

Sebastiaan

[Carlo Lazzaro]

Sebastian:
welcome to this forum.
Some comments about your post:
- you are dealing with a N>T panel dataset with a continus dependent variable. Please note that rarely (pooled) OLS outperforms -xtreg- when it comes to panel data regression;
- if you stick with pooled OLS, you must cluster your standard errors on -panelid-, as, due to the panel data structure of your dataset, observations in each panel are not independent;
- you can also add -i.industry- as a predictor and -cluster- your standard errors as well.

[Sebastiaan Fromme]

Dear Carlo,

Thanks for your reply. For my regression is use now: reg y x i.year i.industry, robust in Stata. But since my dataset is an unbalenced panel dataset I should not use it if I understand you correctly. If I use xtreg I should speficy my panel with -xtset-, in my case would that be the CUSIP for the x and the announcement date as t, so xtset CUSIP DateAnnounced, right?

Then my regression would like a bit like this: xtreg y x i.industry i.year, robust

In that regression I have accounted for my unbalanced panel data, I have clustered my standard errors and I have accounted for hetereoskedasticity by using robust standard errors. Is this correct and is there anything else I should add to this regression to make it more 'correct'?

Thanks in advance.

Sebastiaan

[Carlo Lazzaro]

Sebastiaan (I'm noticing that I've mispelled your name in my previous reply. Sorry for this):
1) it's rare the case when pooled OLS outperforms -xtreg- in dealing with panel data, regardless they're balanced or not (Stata can handle both without any problem; so you could not have accounted for it, simply because Stata did it for you);
2) you -xtset-ting the data is correct;
3) under -xtreg- (by the way: what follows does not hold for -regress-), robust/cluster standard errors do the same job: accounting for both autocorrelation and/or heteroskedasticity.
4) your regression code looks correct for your research goal. However, you should take the matter furtrher and check whether the variance of the random effect differs from 0 (see -help xttest); if that were the case, you should check whether -re- specification fits your data better than -fe- specification. As you have invoked non-default standard errors, -hausman- is not a choice. You should switch to the user-writte programme -xtoverid- (type from within Stata -search xtoverid- to install and use it). Unfortunately, being still very useful but a bit old-fashioned, -xtoverid- does not support -fvvarlist- notation (that is: -i-industry-; -i.year-): see -help xi.- for a fix.

[Sebastiaan Fromme]

Dear Carlo,

Thank you for advice. The difference between the -re-specification and the -fe- specification and how to test it is clear. I'm just wondering how I should choose between the OLS pooled regression with the cluster option or the -xtreg- option. Should I do this based on the Breusch-Pagan Lagrange multiplier test? If I can't reject the null hypothesis for this test does that mean that I should use pooled OLS and not use the -xtreg- RE model?

Regards,

Sebastiaan

[Carlo Lazzaro]

Sebastian:
you're correct.
If -xttest0- cannot reject the null, go pooled OLS with clustered standard errors

[Sebastiaan Fromme]

Hi Carlo,

-xttest- gives me results way above <0.05 which means that I should go for the pooled OLS model with clustered standard errors. Therefore I should not go for the -xtreg- model neither with the fe specification nor with the re specification. For my research I'm trying to use the same methodology of Masulis, Wang and Xie (2007) "Corporate Governance and Acquirer Returns". They state before their regression results that "In parentheses are t-statistics based on standard errors adjusted for heteroskedasticity (White (1980)) and acquirer clustering". Do you think that with the pooled OLS with clustered standard errors I have matched this?

Thank you very much for your help!

Sebastiaan

[Carlo Lazzaro]

Sebastian:
under -regression-, -cluster()- option allows for intragroup correlation, whereas -robust- accounts for heteroskedasticity.
I do not know the literature you mention; that said, I would go pooled OLS with clustered standard errors anyway.