Pooled OLS or xtreg

[Jos Heuvel]

Dear all,

I am currently working on my master thesis and I had some questions on what method to use. The problem is that don't know if I should use a panel data with random or fixed effects or just a pooled OLS.

For my thesis I investigate the relation between the money that companies spend on Lobbying and the abnormal returns around the signing of the CETA (the trade agreement between the European union and Canada). In order to do so I conducted an event study with a data tool that our university offers and generated abnormal returns for several event periods.

Data
My data consist of a data set with around the 1300 companies who are listed in a CETA (EU or Canadian companies) country or in a Non-Ceta country (Companies from the us etc.). From the1300 companies 300 are listed by the European union as companies that lobby. As control variables I have growth, total assets , Tobins Q and leverage.

My idea was to run the following regression in stata:

reg Abnormalreturns TobinsQ LN(TOTALASSETS) Leverage DotheyLobbydummy CETACOUNTRYDUM DotheyLobbydummy*CETACOUNTRYDUM

The idea is that this easy OLS regresion would show me the effects of lobbying and the effect of being a Ceta country on the abnormal returns. After surfing the web I Also saw people who use a paneldata method for this. After reading the panel data explanation this also seemed a reasonable method to conduct.

The problem i have now is that I don't know which method is preferred and if the panel data is even possible with my data since most of the panel data regressions have several time periods to compare and mine has only one, being the abnormal return at the chosen event period.

if someone could please explain to me what the best methods is to use and how to do so that would be fantastic cause i really don't know what to do ?

a sneak peak of my data




Kind regards,

Jos Heuvel

[Carlo Lazzaro]

Jos:
if you have one-wave data only, go -regress..
If you have both cross-sectional (N) and time-series (T) dimensions in your data (that is, the same company is measured >1 time on the same variable), consider -xtreg- if N>T or -xtgls- if T>N.
As an aside, please post waht you want to share with interested listers via CODE delimiters and avoid posting screenshots, as they cannot be elaborated on (see the FAQ on this. Thanks).

[Jos Heuvel]

Hey Carlo,

First of all thanks for your response! unfortunately my question isn't fully answered yet. Yhe problem is that I use the outcome of an event study as my data. So basically my depend variabel are the abnormal returns generated by my event study. Because I use this abnormal returns I don't understand if I should use xtreg or an OLS.

Cause if I use the abnormal returns as dependent variable then it looks like one wave data, but if you think about how abnormal returns are generated then I am not sure anymore

Kind regards,

Job Blankenspoor

[David Benson]

Jos:

In your specific example, you will probably use pooled OLS because you will only have 1 observation per firm (the day CETA was signed). You may need to do something in EVENTUS (I assume that is the tool you are using) because all of the events take place on the same day and so there is going to be correlation across the standard errors across the firms (because they are all happening on the same day you can't assume cross-sectional independence, but EVENTUS has a built-in command to handle this, just search for the EVENTUS user manual). However, studies that have examined how the stock market reacts to US Supreme Court decisions have all had to do the same thing.

1. Take a look at McWilliams, Turk, and Zardkoohi (1993) https://onlinelibrary.wiley.com/doi/...1993.tb00888.x, They examine the effects of Supreme Court decisions on pending mergers. Note that they found that most of the market reaction occurred when the case was argued, rather than when the Supreme Court's decision was announced.

2. Bhagat and Romano (2002) "Event studies and the law" and (2007) "Empirical studies of corporate law" both discuss the issue that you are dealing with (how to do event studies about effects of a particular law passing.

3. Also take a look at Fisman (2001) "Estimating the Value of Political Connections." American Economic Review, 91 (4): 1095-1102. DOI: 10.1257/aer.91.4.1095

4. Corrado (2011). "Event studies: A methodology review" also has some useful suggestions for dealing with your issue (see https://onlinelibrary.wiley.com/doi/...X.2010.00375.x )

5. You might consider using the log of the amount companies spend on lobbying rather than just a yes/no variable for whether they lobby.

Hope that helps!
--David

[Jos Heuvel]

Hey David,

First of all thank you for your response. The papers you posted are very useful and they real add value to my thesis!
However, I still had a question about what you meant with the cross-sectional dependence. What is the reason that you think there is correlation across the standard errors across the firms ?

Is it because of the fact that tools like eventus work with a market model and the fact that using the market as a benchmark creates correlation across the standard errors, due to the event influencing the entire market? Because in order to prevent this problem I have determined my abnormal returns with a Mean adjusted return model. Therefore my abnormal returns only depend on the previous returns of the company and I think that therefore the correlation is solved.

Could you let me know if this is what you meant by the correlation across the standard errors?

And as last note, I unfortunately do not use eventus but a program that is made by my university.

I hope to hear from you soon!

Kind regards,

Jos Heuvel

[David Benson]

The lack of cross-sectional independence would occur if the event for all 1300 companies in your sample occurred on the same date (for example, the law went into effect on the same day for all of the different companies). This is different than when you are looking at the stock market reaction when the events occurred on different days (i.e. the law passed on different dates in different states/provinces, or companies announced acquisitions on different dates).

If the events happen on the same day, it is often called "event clustering" or "event-date clustering". You can go look at the papers by James Kolari and Seppo Pynnönen. The working paper version was called "Event-Study Methodology: Correction for Cross-Sectional Correlation in Standardized Abnormal Return Tests." (which I think you can still find online). The published version is https://doi.org/10.1093/rfs/hhq072

Anyway, I am not sure off the top of my head if using the Mean Adjusted Returns Model (MAR) solves the problem.

Hope this helps!