Fixed effect and AR(1)

[Alexandre Gassmann]

Dear all,

I am dealing with time-series cross-section for an analysis.
For my analysis, I want to estimate a fixed model but I have to deal with auto-correlation. From what I read there are two possibility to solve this:
1. introducing a lagged dependent variable
2. AR(1) model
In stata I would have:
Solution 1: xtreg dependent_var independent_var L. dependent_var, fe
But also from what I read this is method is biased with fixed effect one solution would be to use xtabond but then with xtabond I cannot use no longer fixed effect. I also see that including a lagged variable can biased the result because it captures a lot of the variation.

Therefore, another solution is to use the AR(1) model with fixed effect. But how do I implement such command in stata?
I saw different command such as
xtgls (but I have N>T)
xtpsce (but it gave me an error message "no time periods are common to all panels, cannot estimate disturbance covariance matrix using case wise inclusion"

I would be very grateful, if someone could help me with this issue.
Thank you,
Alexandre

[Carlo Lazzaro]

Alexandre:
welcome to the list.
If, as I read from your post, you're dealing with a large N, small T panel dataset, you can go -xtreg, fe- with robustified or clustered standard errors (they do the same job under -xtreg-) as they take care of both heteroskedasticity and/or autocorrelation (hence, no lag needed which, as you write, would bias the inference under -xtreg- and will impose switching to dynamic panel data models, such as -xtabond-):

[Alexandre Gassmann]

Dear Carlo,

Thank you for you quick answer.

Therefore if I understood well by doing:
xtreg dep_var indep_var, cluster(id) it will solve my problem.

But if I may ask one more thing: what would be the difference between the solution above and using a dynamic data model in terms of interpretation?

best regards,
Alexandre

[Carlo Lazzaro]

Alexandre:
your interpretation is correct.
See -xtreg- and -xtabond- entries in Stata .pdf manual to get a preliminary view about the difference between those two approaches (which are not interchangeable).

[Eric de Souza]

An AR(1) model requires strict exogeneity of the regressors. With a lagged dependent variable as regressor this assumption is violated.