xtregar or xtreg vce(robust)

[Harsha Konara]

Dear friends

I am looking at a panel data set (from 1960 to 2015) where the dependent variable is military burden of Asian countries and independent variables consist of several determinants such as income, number of borders etc. I want to look at how change of military burden in China or India has impacted on military burden of other Asian countries. With my data set I have the problem of autocorrelation (according to the xtserial). While reading on this, I found that I can use xtregar command or xtreg vce(robust). Can anyone suggest me what would be the best method to use to correct autocorrelation in my case and I want to control for country fixed effects also.

Thanks in advance

Harsha

[Carlo Lazzaro]

Harsha:
if you have an N>T panel dataset (where N is the number of panels and T is the time variable), go -xtreg-, robust (or -vce(robust) or, again -vce(cluster panelid)-);
if you have a T>N panel dataset, go -xtregar-.

[Harsha Konara]

Hi Carlo

Thanks for your reply. My T>N since I am covering a 55 years and number of countries only 20. When using xtregar is there any difference when explaining coefficients compared to xtreg. According to my understanding in xtregar command we use AR(1) model where lag of dependent variable is considered as an independent variable. Is that the only change ?

Thanks agiain

Kind Regards
Harsha

[Carlo Lazzaro]

Harsha:
-xtregar- uses AR(1) as a distubance, not as a predictor; AR(1) is the substantive feature of -xtregar-.

[Harsha Konara]

Hi Carlo

Thanks again for the reply. In that case is there any difference in interpreting coefficients

Thanks
Kind Regards
Harsha

[Carlo Lazzaro]

Harsha:
no difference.
For more details, see -xtregar- entry in Stata .pdf manual.

[Phil Bromiley]

Carlo,

Would you have a citation I can use when I go by your guideline? (if you have an N>T panel dataset (where N is the number of panels and T is the time variable), go -xtreg-, robust (or -vce(robust) or, again -vce(cluster panelid)-);if you have a T>N panel dataset, go -xtregar-.)

Phil

[Carlo Lazzaro]

Phil:
the recommendation to switch from -xtreg- for N>T panel dataset to -xtregar- for T>N panel dataset is reported in https://www.stata.com/bookstore/micr...metrics-stata/, Chapter 8: 277-278.
My advice to invoke -robust- or -cluster- option if heteroskedasticity and/or autocorrelation is detected in N>T panel dataset came from the original poster description; conversely, these options are not available for -xtregar-.

[Phil Bromiley]

Thank!
Phil

[Harsha Konara]

Hi Carlo

Thanks for your above replies. Now I am looking at a much larger sample where N>T. In that case I am thinking of using -xtreg-, robust (or -vce(robust) or, -vce(cluster panelid). But my concern is one of my independent variable (most important one actually) is military burden of China. But this variable is not independent across panels since it is the same values for all the countries. Will that be a problem in using xtreg vce(panelID)

Thanks
Best Regards
Harsha

[Carlo Lazzaro]

Harsha:
coefficients of time-invariant predictors are not estimated under the -fe- specification-. This is the only problem.

[Khati Zolfaghari]

.

[Jonathan Anderson]

Further to the above discussion. Does xtregar have an equivalent option for robust standard errors or are they not relevant in this case? If they are not relevant then why is this the case?

[Eric de Souza]

The answer is given in the help file: -help xtregar-
xtregar fits cross-sectional time-series regression models when the
disturbance term is first-order autoregressive. xtregar offers a within
estimator for fixed-effects models and a GLS estimator for random-effects
models.