Panel Data Techniques

[Joyce Parreira]

Hello,
I'm running a fixed effect model (after proven by Hausman Test) and I'm not sure if the code is right. I'm using "xi: reg" and ",rob" and I saw in a forum here that "xtreg" and ",fe" should give me the same coefficients but it's not happening.

These are the codes:
xi: reg reconsumption lngdppctogether gdpgrowth govfinalconsumpexpendgdp gdpperunitofenergy trade energyimports energyuse oilprice lnpoptotal popdensity urbanpop accesstoelectricity lnrailways lnairtransport i.country, rob

xtreg reconsumption lngdppctogether gdpgrowth govfinalconsumpexpendgdp gdpperunitofenergy trade energyimports energyuse oilprice lnpoptotal popdensity urbanpop accesstoelectricity lnrailways lnairtransport i.country, fe

I'm sorry if it is a silly question. I'm new at Stata.
Thanks!

[Joro Kolev]

Somewhere before -xtreg- you should have

xtset panelvar timevar

what are the panelvar and the timevar that you have set?

[Carlo Lazzaro]

Joyce:
welcome to this forum.
Some comments about your query:
- it's rare (although possible) that (pooled) -regress- outperforms -xtreg- with panel data. This can happen if, after running -xtreg,fe- with default standard errors (SEs), the F-test appearing below the outcome table lacks statistical significance; if that were the case, switch to pooled OLS;
- if, as warmly recommended, you use -fvvarlist- for creating categorical variables and interactions -xi:- is redundant;
- the -robust- and -cluster- options for SEs do the same job under -xttreg-, but act differently under -regress-. In your -regress- code, you have taken heteroskedasticity into account, but not autocorrelation (that requires -cluster-);
- last but not least, as per FAQ please post also what Stata gave you back, so that interested listers can see what's the matter with your "weird" results. Please use CODE delimiters to share what you typed and what Stata gave you back. If you're not familiar with code tags, take a comprehensive look at the FAQ, that highlight relevant posting related topics in addition to CODE delimiters. Thanks.

PS: crossed in the cyberspace with Joro's helpful advice.

[Joyce Parreira]

Joro:

Thank you!

I wrote

xtset country

xtsetyear

[Joyce Parreira]

Carlo, Thank you!

I need to admit that I got confused with your answer.

-After runnig - xtreg,fe-, the F test result was Prob > F = 0.0000. I think that is good, right?
-How do I account for autocorrelation and heteroskedasticity at the same time? Using the code

cluster(countey)

?
-I'm reading the FAQ now, thanks!

[Joro Kolev]

I think Carlo is answering you at more advanced level relative to the errors that you are currently committing.

It you had properly xtset your data by

xtset country year

and then you tried

xtreg y x i.country, fe

Stata would have reported to you that all the i.country dummies are being dropped.

I think it is best at this point if you just do:

xtset country year

xtreg y x, fe cluster(country)

Note that you must not include the i.country term.

[Joro Kolev]

xtset country year

xtreg reconsumption lngdppctogether gdpgrowth govfinalconsumpexpendgdp gdpperunitofenergy trade energyimports energyuse oilprice lnpoptotal popdensity urbanpop accesstoelectricity lnrailways lnairtransport i.year, fe robust cluster(country)


Note the difference between this, and your code, the -xtset- is different, and then instead of including i.country (which you cannot include) you include i.year.

[Joyce Parreira]

Joro,
In all of my regressions I have being using

 xi: reg reconsumption lngdppctogether gdpgrowth govfinalconsumpexpendgdp gdpperunitofenergy trade energyimports energyuse oilprice lnpoptotal popdensity  urbanpop accesstoelectricity lnrailways lnairtransport  i.country, rob

When I use

xtreg reconsumption lngdppctogether gdpgrowth govfinalconsumpexpendgdp gdpperunitofenergy trade energyimports energyuse oilprice lnpoptotal popdensity  urbanpop accesstoelectricity lnrailways lnairtransport  i.country, fe

Only some of the countries are omitted.

If I want to see the country specific effect for the country, why not use i.country?

I tried the code

xtreg y x, fe cluster(country)

and stata gave me that "panels are not nested within clusters r(498)"

[Joro Kolev]

The -xtreg, fe- applies the within transformation to the data, and after that everything that is constant within the panel variable (country in your case) drops out/becomes collinear with the constant.

The within transformation is equivalent to the dummy variable regression, that is including i.country. Therefore

reg y x i.country

is equivalent to

xtset country year

xtreg y x, fe

and in the second you do not need to include, and in fact cannot include the i.country terms.

The only reason that I can think of Stata not dropping all country dummies, but only some is that you might have unbalanced panel.

As for the rest that you say, I do not see your data, so I do not know why "panels are not nested within clusters r(498)", your panels are country, so I cannot guess from where this error message comes from.

[David Benson]

Joyce,

I would try tabulate country to see if there is anything funny with that variable (for example, some observations "change" countries over time, or one country has multiple obs for the same year).

See this Statalist topic, "panels are not nested within clusters" and this one (on StackExchange) Stata error: “panels are not nested within clusters”

[Carlo Lazzaro]

Joyce:
if, after runnig - xtreg,fe-, the F test result was Prob > F = 0.0000, pooled OLS is not the way to go and -xtreg,fe- should be considered.