Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: RE: Number of observations with ARMA models
From
Nick Cox <[email protected]>
To
[email protected]
Subject
Re: st: RE: Number of observations with ARMA models
Date
Tue, 17 Jan 2012 14:52:54 +0000
We agree for least one good reason: that is never my problem. My
problem is usually wanting to see overall smooth patterns in time
series, which makes me reluctant to throw anything away in most cases.
On Tue, Jan 17, 2012 at 2:04 PM, Christopher Baum <[email protected]> wrote:
> <>
> Nick wrote
>
>> Your second question has no easy answer that I can see, because of the two different perspectives on the first. If this were my problem, I would fit each model to the largest amount of data possible.
>>
>> Nick
>> [email protected]
>>
>> From: [email protected] [mailto:[email protected]] On Behalf Of Steven Trypsteen
>>
>> ...
>
>> I would like to understand this because I would like to use the Log likelihood test (command: lrtest) to determine if two models are equivalent. For example I would like to check if AR(8) is equivalent to AR(2). How should I restrict my sample then to make sure the estimation procedure uses the same number of observations?
>
>
> I beg to differ. If you are using a test to determine optimal lag length, it is imperative that all models be fit over the same observations. (Think about R^2: the total sum of squares will be changing otherwise). The easiest way to implement this is to do something like
>
> reg y L(1/8).y
> g byte es = e(sample)
> reg y L(1/4).y is es
> reg y L(1/2).y if es
>
> etc. Then all regressions will be run over the common (shortest) sample.
>
> Such logic is implemented in many time-series routines in Stata (e.g., unit root tests).
>
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/