URADF vs. ADFAUTOSELECT

[raphael]

Hi, I was wondering when and why one should use one of these procedures instead of the other.
Thanks a lot,
Raphael

[raphael]

OK so it seems ADFAUTOSELECT is the more recent one. My next question then is on the choice of information criterion. I guess it depends on the type the variable we are looking at. With the variables I am using, the different criteria yield very different lag lengths, eg. 2 (BIC) vs 12 (AIC). Any book / reference paper on this ?

[TomDoan]

If you're interested in a relatively small number of tests (that is, you're just doing a series or two), the main question is whether it matters. Do you reach the same conclusion regardless of method used for selecting lags? When you can do it, a lag sensitivity table may be even a better choice than any automatic procedure. If the conclusion depends upon the number of lags (once you've added enough to roughly whiten the residuals), then the evidence isn't particularly clear.

[raphael]

Unfortunately I have a long list of variables. Can the lag lengths selected by ADFAUTOSELECT be used in other tests, notably KPSS or PP ?

[TomDoan]

No. In the KPSS and PP tests, the lags are widths of Bartlett windows. In effect, they are lags in a moving average, not in an autoregression like ADF.

[raphael]

Thanks. By the way, I was reading ADFAUTOSELECT, and I am probably wrong but there seems to be a mistake in the IC formulas. For example, the code is:

compute aic = log(sigsq)+2.0*%nreg/%nobs

but shouldn't it be

compute aic = %nobs*log(sigsq) etc.

cheers

[TomDoan]

The first formula is the second divided through by %NOBS. Since %NOBS should be the same when comparing two models, the two formulas will give the same rankings. The formula with /T has the (minor) advantage that the values it produces tend to be easier to read, being within a order of magnitude or so of 1.0.

[raphael]

Cheers. About that sensitivity table, do you mean testing the residuals for serial correlation ? I have 11 variable so this sounds fastidious (I can't even imagine how I could present this in my paper). Isn't there a procedure for KPSS/PP lag length selection ?

[TomDoan]

Again, it only matters if you get qualitatively different results from two lag lengths that pass some form of "whiteness" test (for the ADF). There are some formulas for picking "optimal" bandwidths for Bartlett windows. See Andrews(1991), "Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation," Econometrica, vol. 59(3), 817-58. They're not quite as obvious as the AIC and BIC, since they aren't based upon a regression likelihood function.

[raphael]

OK, thanks, it looks much too complicated for me. I have one last question on the subject, which concerns DFUNIT.SRC: what is the difference between the options [INTERCEPT]/NOINTERCEPT and TREND/[NOTREND] on one side, vs. DET=NONE/[CONSTANT]/TREND on the other side ?

[TomDoan]

None. The INTERCEPT and TREND options date back to the original DFUNIT procedure. The DET option was added more recently to make it more similar to other similar procedures.