|Date: ||Sun, 24 Sep 2006 21:30:31 -0700|
|Reply-To: ||David L Cassell <davidlcassell@MSN.COM>|
|Sender: ||"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>|
|From: ||David L Cassell <davidlcassell@MSN.COM>|
|Subject: ||Re: Modelling several time series together using Proc Mixed?|
|Content-Type: ||text/plain; format=flowed|
djrk0003@COMCAST.NET wrote back:
>----- Original Message -----
>From: "David L Cassell" <davidlcassell@MSN.COM>
>Sent: Friday, September 22, 2006 12:37 AM
>Subject: Re: Modelling several time series together using Proc Mixed?
> > djrk0003@COMCAST.NET wrote:
> > >
> > >Chen presented the problem below. I would like to follow up on this
> > >but in a slightly different direction. Suppose, I have a new data set
> > >I
> > >need to score used the model developed below. For the new data set, I
> > >like to forest for each ID over a 72 month period. Is this feasible? If
> > >how?
> > >
> > >Thanks,
> > >
> > >Doyle.
> > Are you saying that you want to create an entire new data set based
> > on *no* values from the given ID? That does not sound reasonable.
> > In a time series, you need at least *some* starting values for
> > autoregressive
> > models, and at least *some* starting error terms for moving average
> > models, and all of the above in ARIMA(p,d,q) models where p and q are
> > greater than 0.
> > Are you going to have some starting values? Are you going to have
> > a model from PROC AUTOREG to use to predict some expected values
> > based on a set of regressors?
> > Without some more detail, I can't help you much.
> > David
>David and all, here is a clarification of my problem:
>Chen presented the problem below. I would like to follow up on this issue:
>First, can this problem be modeled a longitudinal/panel/repeated measures
>Second, suppose I have developed this model and have a new data set say
>observed only at time 1 for each ID and would like to forecast the price
>each id say over 10 time periods, Is this feasible? If so how?
>I have a date set which looks like this( I made up the data)
>ID Time Price
>1 1 10
>1 2 11
>1 100 150
>2 1 9
>2 2 10
>2 100 189
>50 1 11
>50 100 190
>I want to predict the value at time 101 for each different ID.
>Obviously I can model 50 independent time series.
>My questions are that
>1. Is it possible that I can treat this data as repeated measurement and
>Proc Mixed to build
> one whole model.
>2. If so, how can I write the formula?
> Proc Mixed data=test;
> class ID;
> model Price = Time ID;
> repeated time / subject=ID;
>I just want to catch the interaction between different IDs. But I am not
>sure that Mixed model
>can do that. I just began to learn the mixed model stuff. So I can easily
>make mistakes on basic concepts.
 Is it possible? Yes.
Is it appropriate? Maybe yes, maybe no.
PROC MIXED does have a couple time series covariance structures,
but it is not a time series modeling system. You should really be
using SAS/ETS. I would recommend PROC AUTOREG here, although
you may need PROC TSCSREG or PROC PANEL given the panel
nature of your data.
If it turns out that your time series model is so simple that an AR(1)
or ARMA(1,1) or ARH(1) covariance structure can handle all the
temporal structure, then yes, PROC MIXED can be used. Until you
assess that, I would have to recommend against using it.
 Let's avoid this until we're fairly sure that this is a reasonable
David L. Cassell
3115 NW Norwood Pl.
Corvallis OR 97330
Try the new Live Search today!