How to interpret country dummy results?

[thomas bertje]

Hello everyone,

I recently performed panel data analysis on a panel of about 1100 firms from different European countries. My supervisor advised me to add country dummies which I did, but when I look at the regression results I don't know how to interpret the results of these dummies. I will add the picture of my results to this message and if there is any additional information I have to provide, please let me know.

Thanks in advance,

Thomas

[Carlo Lazzaro]

Thomas:
unfortunately, you provided what Stata gave you back (althoug you would be better off with using -CODE delimiters- for that purpose; see the FAQ on this topic and, in more general terms, on how to post more effectively. Thanks.), but what you typed is missing.
What is your dependent variable? Is it logged? And what about the predictors?
Did you use -fvvarlist- for creating country dummies or did you plugg them in by hand?
Are you dealing with a large N, small T panel dataset (or the other way round)?
I assume you went -xtreg-; but -fe- or -re- (with a bit of guess-work, I rule out -be-)?
Please, provide some more details. Thanks.

[thomas bertje]

Dear Carlo,

Thank you for your quick response. My dependent variables are dummy variables EARNS (Earnings restatement) MANAG (Management cost/risk techniques) PROF (Profit warnings) SHARE (Policy on insider trading) where for every variable 1 means a firm has applied/used it and 0 means it isn't. I am not sure what logged means, but it is panel data over 6 years and the DV, IVs, controls and moderator variable are arranged accordingly. I plugged in the dummies by hand and then used encode. my N is 1072 firms over 6 years, so 6432 observations. I don't seem to have any T-values, but I have z values. I went xtset and xtreg, re robust yes. Thanks a lot for the effort, my apologies if I am not following the rules well.

Regards,

Thomas

[Carlo Lazzaro]

Thomas:
I suspect you report your independent variables as dependent ones.
I assume that your dependent variable (-FERROR-) is continuous.
By logged dependent variable I meant -ln(FERROR)-.
By large N, small T panel dataset I meant a panel dataset with more observations (N) than time periods (T).
How did you select -re- specification? On the gounds of the -hausman- test outcome (by the way, please note that -hausman- test supports default standard errrors only).
To create categorical variables, please take a look at -fvvarlist-.
Just out of curiosity: which countryid is abbreviated as IT?
As far as the country dummies are concerned, you may want to test their jointly statistical significance via -testparm-.
Their interpretation sounds like this: when adjusted for the other predictors being located (say) in BE vs reference country inccrease the between panel difference of FERROR by about 21%.

[thomas bertje]

Dear Carlo, thanks for your swift response again. If I did report my independent variables as dependent ones, how did I make this mistake and how can I undo it? As far as I know I ran the regression like this: xtreg IV DV1234 CONTROL1234. Or do you mean I reported them as continuous in stead of categorical?

my dependent variable FERROR is continuous yes. Is there a reason to use the natural logarith of my DV? It ranges from -6 to 3.33.

I have more observations than time periods yes.

I selected -re- on the grounds of the hasuman test, correct. What does that mean, that it only supports default standard errors? That there is no sense in adding -robust- or?

Good question, IT was a mistake I made, it should be ITA for Italy. Thanks for spotting that, you have a keen eye.

[Carlo Lazzaro]

Thomas:
I meant that in your previous post you miswritten dependent for independent. that's all.
As reported in any decent econometrics textbook, log-linear model are often quaite easy to explain. However, ifyou have many 0s and/or negative values in -FERROR-, logging it will produce a lot of missing values (and the related observations will be listwise deleted from the following -xtreg,re-).
It is impossible that you've run -hausman- with robustified standard errors, nor is advisable to add -robust()- after that -hausman- test with defaul standard errors (that is, no robust, cluster, bootstrap or jackknife options) indicates that -re- is the way to go. If you suspect heteroskedasticity and/or autocorrelation and you wisely robustify/cluster your standard errors, then you have to compare -fe- and -re- specifications via an user-written programme named -xtoverid- (type -search xtoverid- from within Stata to install it).
I'm Italian:so ITA vs IT hit my eyes!

[thomas bertje]

Thank you for everything Carlo and pardon my mistake. I love Italy and the Italians!