[R] Re: R-help digest, Vol 1 #80 - 14 msgs

[Adrian Trapletti]

> Subject: [R] LRT in arima models
> Date: Mon, 17 Feb 2003 11:53:04 +0100
> From: "vito muggeo" <vito.muggeo at giustizia.it>
> To: <r-help at stat.math.ethz.ch>
>
> Dear all,
>
> For some reason I'm evaluating the size of the LRT testing for the effect of
> some explanatory variable in arima models.
> I performed three different simulations
> y<-2+arima.sim(n, model=list(ar=ar.p,ma=0),sd=1)
> with different ar parameter, namely 0.7, 0.5, and 0.2. n=100.
>
> Out of 1000 replications performed for each ar, the H_0 (no effect of
> x<-1:n) was rejected 77, 67 and 53 times with ar=0.7, 0.5 and 0.2
> respectively. (sigma was assumed known in calculating the statistic test and
> the nominal significance level is 0.05)
>
> That is, the LRT seems to become somewhat anti-conservative when the ar
> parameter increases.
> I am aware the 1000 replictions are few to assess the size, but the "trend"
> in the sizes seems noticeable.  Am I wrong?
> Am I missing some theoretical issues or could it depend on any computational
> apect related to the the arima() function?
>
> Any comment is coming?
> Many thanks
> best,
> vito

Just a general comment:

I am more surprised that the test is still doing that well: An AR coefficient of 0.7
introduces quite a bit of dependence in your DGP. Hence the number of "independent
observations" is considerably smaller than 100. Therefore asymptotic theory might be a
bad approximation and small sample effects might lead to severe size distortions.

Maybe someone can explain more formally whats going on here?

best
Adrian

--
Dr. Adrian Trapletti             Phone :             +41 (0) 1 994 5631
Trapletti Statistical Computing  Mobile:             +41 (0)76 370 5631
Wildsbergstrasse 31              Fax   :             +41 (0) 1 994 5631
CH-8610 Uster                    Email :  mailto:a.trapletti at bluewin.ch
Switzerland                      WWW   : http://trapletti.homelinux.com