Running Stata 14 on Windows 10. I'm trying to make sure I'm implementing
actest correctly to test for autocorrelation in residuals in a panel data context. In my dataset, two different approaches --- minimizing the AIC and SBIC (over several specifications that use an increasing number of lagged DV terms), and testing the significance of subsequent lag terms (over that same series of specifications) --- support the use of two lags on the dependent variable (i.e., an AR(2) process). So, after I
xtset country year (because most of my specs use
xtreg, but that doesn't support
actest in postestimation) I run my preferred / main specification, which is of the form
Code:
reg y x lag1y lag2y i.country, cluster(country)
Where y and x refer to y
it and x
it, lag1y is y
i,t-1, and lag2y is y
i,t-2. I'm interested in knowing if the residuals from this equation have any serial autocorrelation, so as a post-estimation command I run
Code:
actest, lags(3) cluster(country)
And, happily, find that the test rejects the null hypothesis for the first and second lags, and fails to reject the null for the third lag. But I am not sure if I'm running this test correctly. Should I instead run this set of commands?
Code:
reg y x i.country, cluster(country)
actest, lags(3) cluster(country)
The example provided in the
actest help file suggests that we might want to run the original regression without lags, but it is not very clear (at least, not to me). At least one published (and seemingly StataCorp-sanctioned) implementation of
actest, provided in
Ariel Linden's Stata Journal article introducing the
itsa command, suggests that the postestimation application of
actest should be run after a regression that includes desired lag terms (see Section 4.1 of that article). I was unable to find other resources online, including in this forum, to indicate clearly either way.
Many thanks.
0 Response to actest for postestimation: run the original regression with lags, or without?
Post a Comment