On Thu, 27 May 2010, Ignacio Diaz-Emparanza wrote:
I am a little confused about the behaviour of AIC, HQC and BIC in
the standard ols command. I estimated some AR(p) models with different p
values with the 'ols' and the 'var' commands. The last with only one
variable which I know it is not designed for, but ... it works.
I see that the BIC obtained through both commands:
1- are not the same. I suspect may be some different scale factor in the
formulas but this should not be a problem.
Well spotted: in fact, when we compute the information criteria for VARs,
we use as the sample size the total data span (T), whereas for OLS we use
the number of datapoints actually used (T-number of lags). To see what
I'm talking about, try this little modification of your script:
function void ARF(series y, scalar maxorder)
var i y --quiet
genr C[i,2] = $bic
list xlist = lags(i,y)
ols y 0 xlist --quiet
genr C[i,3]= $bic / $T
2- leads to different specification, and yes, this is a problem. In
concrete, the BIC in ols would suggest the maximum number of lags, and
this does not seem to be correct if one have a look at the correlograms.
I agree it is a problem, but only to an extent: information criteria are
likelihood-based, and as such have asymptotic validity only. Allin thinks
that this could be the outcome of the policy of "trying to agree as much as
possible with other packages". If package A uses T and package B uses T-k,
well... the issue "internal consistency vs agreement with others" has been
debated to death and I really wouldn't know what to say.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche