Procedure to use XTPCSE correctly

[Francisco Castellanos]

I am working on extending a paper whose regressions were obtained by OLS on a panel data (aprox. 20 years of data for 50 countries) using year and region (not country) dummies. I may think they pooled the data since the next criterion: R2 and F-statistic are the only test values presented for each regression.


Facts:

1) By now, I have the data for the next 20 years for the same countries (not compeltely balanced).

2) The model presents serial correlation (according to the xtserial command, and obviously for the type of the info).

3) It is implicitly obvious the panel-specific AR1 autocorrelation structure since countries follows their own cycles.

4) Due to the fact that the info is for countries the estimation option "hetonly" is discard.

5) Variables are monetary measures for countries.


Questions:

1) Does running

xtpcse depvar indepvar, corr(psar1) pairwise

is the best way to estimate the model?

2) If the answer to the previous question is "Of course not, please don't be a fullish" What do I have to do? Do I have to add rhotype(tscorr)? What do you recommend about the estimation?

I just want to be sure if I am running correctly the model and avoid being a rocky disappointment.

I will appreciate your answer and advice.

[Carlo Lazzaro]

Francisco:
the issue rests on the fact that -xtpcse- is conceived for dealing with small N, large T panel data (something that is opposite to your dataset).
Hence, provided your depvar is continuous, I would go:

xtreg depvar indepvar, vce(robust)

By the way, I would find weird that panel data do not include serial correlation.

[Francisco Castellanos]

I am so thankful for your support.

In order to follow your advice and considering that now, with xtreg, I have the chance of controlling for trends across the globe and macroeconomic trends, I will create a sensitivity analysis by adding time and region dummies to the basis model you proposed to me. It will look as:

xtreg depvar indepvar, vce(robust)

xtreg depvar indepvar i.time, vce(robust)

xtreg depvar indepvar i.time i.regions, vce(robust)

Have a great weekend!

My best regards, Francisco.

[Carlo Lazzaro]

Francisco:
what you propose makes sense, provided that your goal is performing a sensitivity analysis (as you clearly state) instead of hunting for "the best" (whatever that means) model.
As a a closing-out remarks, you codes imply a random effect specification. What does -hausman- specification test (taking its limitations into account) tell you?

[Francisco Castellanos]

Exactly, I am not looking for "the best" model (because that is not even ethical) but for the appropriate specification to obtain the more accurate coefficients.

Regarding the hausman specification test, it cannot be implemented for vce(robust) models. Then, when I run the complete model withouth the vce option, FE omit two variables because of collinearity. In this case it is not correct to running a test for dispair models.

Therefore, I tried both models with the same available variables and the hausman-test Prob>chi2 is 0.000 rejecting the null hypothesis. Implying FE is more appropriate than RE.

In this case, FE omits important variables. So, what is the specification you recommend me to use? I think it is kind of rude just to use RE because it is the model that fit better the panel data I have.

I will really appeciate your support.

My best regards, Francisco.

[Carlo Lazzaro]

Francisco:
thanks for providing further details.
I would disagree about your statement

In this case it is not correct to running a test for dispair models.

, because -hausman- works considering the coefficients which are present in both FE and RE models.
As far as -hausman- test with -vce(robust)- is concerned, you may want to take a look at http://www.stata.com/statalist/archi.../msg00470.html.
As per FAQ, more helpful replies to your query may come alive, conditional on your posting what you typed and what Stata gave you back (using code delimiters, please). Thanks.

[Francisco Castellanos]

Dear Dr. Lazzaro,

Thanks for your support, by now, I will reconstruct the data sample to avoid any concern on it and being 100% it is correctly built (just to be more accurate, one never knows if the problem is how we build the dataset). Then, to extend the original paper that I aiming replicate, I will start with robust pooled-OLS (with year and region dummies); then, as second step, I will propose the robust GLS-RE sensitivity analysis (with and withouth dummies), as you recommended it to me (obviously with the hausman test verification). And, finally, I will try a robustness check with any possible methodology on panel data.

I have really learned much with your guidance.

Sincerely, Francisco.

[Carlo Lazzaro]

Francisco:
thanks for providing further details.
Please, as anybody else on this list, call me Carlo.

[Nurshamimitul Ezza]

Dear all, sorry for interrupt here but I have further question regarding xtpcse command. My data has 3000 observations covers for 15 years period. I am referring to Hoechle 2014 to conduct robust testing for my data as it suffers heterogeneity and autocorrelation. I am planning to use xtpcse command but noted from the above, Carlo mentioned that xtpcse- is conceived for dealing with small N, large T panel data.

So,perhaps what is the correct robust method for my data? I am kind of a novice in the research area and as such your help is really appreciated.

Regards, Ezza

[Carole J. Wilson]

I would start with Carlo's recommendation in #2.

[Nurshamimitul Ezza]

Dear Carole, tq for your response but I am not familiar with vce robust command. Can you elaborate further when to use vce robust?

Like my case, I employed static panel data and I design my analysis as follow;-s

1. conduct pool reg, FE and RE;end result FE is the best model
2. conduct diagnostic check; end result confirms there is hetero and auto prob
3. to ractify the above problem, conduct robust cluster regression (as Hoechle 2014 suggestion) command xtreg (cluster)
4. to ensure the model is robust enough, I intend to do further robust testing which is either using xtpcse/xtgls/xtscc also as Hoechle 2014 suggestion

Kindly advice. TQ

Regards, Ezza

[Carlo Lazzaro]

Ezza:
some remarks about your previous post:
- instead of describing what you're intended to do, it's better to report what you typed and what Stata gave you back via CODE delimiters (as per FAQ), even if what you post is your first attempt in performing the regressions you'after;
- please provide full reference of the quoted article (as per FAQ);
- it is not clear which kind of rubustness you are interested in. In all likelihood it refers to standard errors, but this is only my wild guess.

[Aneta Blazejowska]

Which method should I use for T=11 and N=50 if I have cross-sectional dependence, heteroskedasticity and serial correlation? xtgls, xtpcse or xtscc? I examined US counties so I expexted cross-sectional dependence.

[zana pekmez]

The choice of the autocorrelation option (ar1 vs psar1) when implementing the linear regression with panel-corrected standard errors significantly impacts my estimation. The statistical significance is lost for my most important explanatory variable once I use psar1 option.
I could not find in the literature that there exist a specific test for the selection of ar1 vs psar1.
How do I decide?
I know that my panels have cross-sectional dependence (tested & confirmed). Would this be a sufficient rational to conclude that autocorrelation structure could be the same across all panels.
I also know that in their article Beck & Katz (1995) make a case against using psar1 option with xtpcse estimation.
Thank you for your feedback in advance!
Zana

[Panika Jain]

Lazzaro, can you provide me a reference for this fact that pcse technique can be applicable when T > N. I have come across many research papers that have used pcse when N >T citing ( Reed and ye, 2011). I have 119 countries with 16 years. I have conducted full sample and sub-sample analyses (in all cases N is greater than T) with both pcse and xtreg, vce (robust). Is it appropriate to apply pcse technique when N is greater than T. If not than what is the best method with such a panel structure to deal with hetroscedatsricty, autocorrelation and cross-sectional dependency,