Date:         Sat, 2 Jul 2005 23:12:52 -0700
Reply-To:     DavidL Cassell <davidlcassell@MSN.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         DavidL Cassell <davidlcassell@MSN.COM>
Subject:      Re: Proc MIXED
In-Reply-To:  <MC2-F16bB1kOXXGQ3Wr00018b1f@mc2-f16.hotmail.com>
Content-Type: text/plain; format=flowed

phlarsen@YAHOO.COM wrote:
>I need some guidance on how to setup and interpret results from the PROC
>MIXED statement.
>
>I have the following model of sector-level production: Q (i.e. sector
>output) is a fcn of (capital, labor, energy).  The input data is a panel
>data set with cross sections of year and U.S. state.  The error structure
>of
>the data is best modeled with an AR(1) process AND heterscedasticity. So
>far, I have the following code to make this data stationary (and non-biased
>inference possible):
>
>proc mixed data=modeldata;
>     by sector;
>     class state year;
>     model LN_Q = LN_KAP LN_E LN_L /s;
>     repeated /type=ar(1) subject=fips group=fips;
>run;
>
>Two questions:
>1) I still don't understand what the group, subject, and class statements
>are really doing in PROC MIXED.  Considering the structure of my panel, is
>the above code correct?  All I'm trying to do is to account for state
>effects and year effects in the error term.

The CLASS statement spells out for the proc which variables are discrete.
That way, the proc knows how ot handle them, rather than treating them
as continuous variables and giving them the 'regression' treatment.  Also,
the proc needs the SUBJECT= varaible to be discrete, so it can work with
the different levels properly.

The GROUP option in the REPEATED statement gives you the ability to
tell the proc to treat each level of GROUP as having (potentially) different
covariance parameters.  As I recall your previous post, this is something
you
wanted when you were trying to use Parks' method.  It doesn't give you
the 'contemporaneous correlation' across FIPS codes that Parks' method
gives,
but I don't know if that is needed.  (It's pretty hard to tell from way over
here. :-)

The SUBJECT option defines the block-diagonal nature of your R matrix.
It means that years will be correlated within FIPS code, but independent
of records for years in other FIPS codes.

Given your TYPE= option and your SUBJECT= and GROUP= options,
I'd say that you're going to end up estimating 2*K variance parameters
in your R matrix alone, where K is the number of distinct FIPS codes you
have.
That could be messy.

So...  Why did you give up on PROC TSCSREG?  You're still not duplicating
Parks' method.

>2) How do I generate estimated coefficients for KAP, E, and L as well as
>fit
>statistics.  All I can see for output choices are covariance parameters,
>etc.  All I really want is the beta hats...

Well, you'll need to take the estimates for your logged variables and
back-transform
by taking exp(variable).

Or you could do this in something like PROC NLMIXED, and incorporate the
logs
in your model.

>Any help is greatly appreciated.
>
>Have a great weekend.  I'll be working on this.

Take sometime off and have a good weekend yourself.  Remember, all work
and no play makes Jack take a fireaxe to his family in a snowbound hotel in
Colorado and...  Oh never mind.  :-)  :-)

David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today - it's FREE!
http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/