Serial Correlation in Cross Sectional Data

[Zygimantas Svitojus]

Hi,

I have a cross sectional data and I want to test for serial autocorrelation. Does it make sense to say that this problem is automatically solved since the regression is cross-sectional? If not, what STATA command I should test for serial correlation?

Thanks

[Carlo Lazzaro]

Zygimantas:
no, the problem (if any) is not automatically fixed thanks to the cross-sectional design of your dataset.
You may want to consider -estat vce,corr- after -regress-.

[Zygimantas Svitojus]

Thanks for your answer.

I have one more question. If I use clustered standard errors in my analysis, can I say that it solves serial correlation issue (if it exists) ?

[Carlo Lazzaro]

Zygimantas:
correct, provide that your sample size is large enough (say =>30 observations, as a rule of thumb; otherwise, clustered standard errors can be more biased than their default counterparts).
See also

http://cameron.econ.ucdavis.edu/research/Cameron_Miller_JHR_2015.pdf

[Laiy Kho]

Carlo Lazzaro :

You may want to consider -estat vce,corr- after -regress-.

What is the threshold to estimate if there is serial correlation?

[Carlo Lazzaro]

Laiy:
take a look at the article that describes the community- contributed module -xtserial-. Its link is available from -xtserial- helpfile.

[Laiy Kho]

Hi Carlo,

I am assuming this is the article you are referring to : https://journals.sagepub.com/doi/pdf/10.1177/1536867X0300300206


However, I do not find any information on a threshold that could be used to determine if there is serial correlation. It discusses xtserial test with null hypothesis of no serial correlation as opposed to a correlation matrix that is provided by estat vce,corr. My sample size is 1 million. Would a correlation coefficient below 0.8 be considered to indicate serial correlation?

[Carlo Lazzaro]

Laiy:
from the helpfile of the community-contributed module -xtserial-:

Under the null of no serial the residuals from the regression of the first-differenced variables should have an autocorrelation of -.5.

That said, does the p-value appearing as a footnote of the -xtserial- outcome table reach statistical significance?

[Laiy Kho]

Carlo:

Thank you so much! Could you provide the link to xtserial helpfile? I am unable to find it. There is only a link for installation.

That said, does the p-value appearing as a footnote of the -xtserial- outcome table reach statistical significance?

Isnt xtserial command used for panel data only? I am dealing with cross sectional data with groups.

[Carlo Lazzaro]

Laiy:
you're correct.
If you are actually dealing with cross-sectional data, you can calculate the correlation of the residuals with -pwcorr- once you have obtained -residual- from -predict residual, res- (and can safely forget -xtserial-. I often work with panale datasets and I'm a bit biased in this respect).
That said, the risk of serial correlation creeps up when your observations belong to higher level groups (e.g., other things being equal, two twins are in all likelihood more similar than two individuals that do not share the same genes).
Eventualy, -estat vce,corr- can give you an idea of the coefficient correlation (that is another topic).
Sorry for the confusion.