Fixed effect model doesn't work...Where is the mistake?

[John Snow]

Hey everybody,

you probably heard it a thousand times but I have serious troubles with my fixed effect model. But lets start from the beginning:

I'm currently doing a panel regression with 23 different european countries for the time period 2001-2012. In my research I want to explore the impact of sovereign credit ratings on the M&A activity. I there fore gathered all the data including GDP, Inflation, Interest rates, Exchange rates, M&A volume and some other control variables. However, from a logical and theoretical perspective I would say that one would apply a fixed effect model for this kind of thematic (thats actually also want all other researches did so far). However, when using a fixed effect model I only get very small t-statistics (around 0) for all my variables, what is pretty surprising. At least for GDP I should find something significant... The Hausman test also suggests to take a random effects model. When using an re I indeed get better results but the standard error and coefficient of the rating is extremely high and all other coefficients are extremely low. So currently I'm not really sure how to go one. One the one hand I have more or less good results when using a random effects model but in theory a fixed effect model should be appropriate. I don't know where the mistake could hide as I think my data is pretty good and also the correlation table totally makes sense.

My code which I used for the fixed and random effects model:

egen country1 = group(Country)

xtset country1 Year, yearly

xtreg MAvolume LnGDP LnRating Inflation InterestRates ExchangeRates TradeOpeness, fe


I also posted my results and I hope that you understand my issue

Thank you very much for your help!

[Anat Tchetchik]

John, from my experience random effects model are rarely consistent, and understand your concern.
Have you tried adding to the fixed effect the option vce(clustered cuntrynum)? your "within" R2 is very low in both the fe and the re models.

[John Snow]

Hi Anat,

thank you for your quick response. I now tried it and the R square is still very low as well as the t statistics..

[Carlo Lazzaro]

John:
I'm under the impression that -re- specification gives you "better" results (whatever that means) because you do not have time-invariant predictors.
Whreas the within-Rsq is very low in both -xtreg- specifications, the between-Rsq in -re- sounds good.

[Joao Santos Silva]

Dear John,

Let me add a few notes to the earlier comments:

1 - The Hausman test you used is unlikely to be invalid in this context because it assumes homoskedasticity and no serial correlation.
2 - In any case, the standard errors of the FE estimator are so large that it would be difficult to reject the null that the RE estimator is valid; that is, the test is unlikely to have reasonable power.
3 - It looks as if most of your regressors will have a lot of variation across counties, but little variation over time. Therefore, I am not surprised that when you include fixed effects you find that few variables are significant.
4 - Maybe you can try to address the problem in 3 by using growth rates; for GDP that would make sense, but maybe not for other variables.
5 - I assume your dependent variable is a count and therefore you should consider an approach that takes that into account; have a look at -xtpoisson- and consider using FE.
6 - Make sure you use appropriate standard errors.

Best wishes,

Joao

[John Snow]

Hi Jaa,

thank you for your comprehensive response.

So you mean the suggestion of the Hausman test to use a random effects model is valid and therefore a random effects model is appropriate? I already tried to include GDP growth but it was even more insignificant than GDP. The M&A activity is not a count but a sum of investment inflows. I now have some valid results for a random effects model, where the coefficients and the standard errors make more sense. I'm currently thinking to stick with the random effects model as might be appropriate in Europe because the countries do not differ that much from another. Especially when considering that most of the countries in the sample are in the EU. What do you think?

[Joao Santos Silva]

Sorry, John, I meant that the test in unlikely to be valid! So, I would not trust it at all.

Thanks for clarifying the nature of the dependent variable; I think it is worth looking at Poisson regression.

I assume you included GDP growth in place of GDP, not in addition to GDP, right?

About whether or not to include the FE, that is a modeling decision and I do not think that there is a right or wrong answer. All depends on what you want to do.

Best wishes,

Joao