Skip to content
Prev 4379 / 15274 Next

standard error and p-value for the estimated parameter in AR model

Hi

Yes dependance of regressor and errors has the effect that your 
estimator is biased. Hamilton (p 215) discusses the case of AR() with 
iid errors:

"the OLS coefficient gives a biased estimate in case of an autoregression 
and the standard t and F statistic can only be justified asymptotically. "


So as you point right out, normal distribution instead of student should 
be used for the p-values! (I'm not sure whether student distribution 
can't be used if you make the assumption that the errors are Gaussian. )

Note however that those results are derived for the OLS estimator, which 
is not the estimator by default in ar().

For small sample p-values, bootstrap methods could be used. Introductory 
discussion can be found in Maddala p 323 (available on google books, 
type: "the procedure for the generation of the bootstrap samples").

Matthieu

markleeds at verizon.net a ?crit :