[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: Re: Akaike information criterion
----- Original Message -----
From: "Jennifer Ng" <firstname.lastname@example.org>
Sent: Tuesday, September 16, 2003 9:17 PM
Subject: st: Akaike information criterion
> Hi all,
> I would like to get the Akaike information criterion (AIC) to figure out how
> many lags in my model. I checked the reference guide and figured out that
> estimate can help.
> So, my program is as followed:
> Regress Y X Z
> est store AIC
> estimates stats *
> Then I saw that in the reference page 353, it shows the equation of AIC = -2
> log (likelihood) + 2df.
> I remembered that AIC = log (sum(e-sq)/n) + 2k/n and I checked a few
> textbooks and the AIC is the latter one. Are they equivalent? Did I use
> the wrong command/equation?
Yes they are equivalent. On page [R] 353 the "df" term refers to the number of
parameters estimated, including the constant.
The most commonly used definition and the one given by Akaike is AIC = -2
ln(likelihood) + 2#parameters.
For the case of linear regression models it can be written as AIC =
n*ln(sigma^2) +2*parameters, where sigma^2 = RSS/n.
In addition to -est stats *- , you might find J. Scott Long and Jeremy
Freese's -fitstat- helpful.
Hope this helps,
* For searches and help try: