Colleagues
I hope I can get some help from the list. I am getting conflicting
advice from two colleagues. In essence, I am trying to figure out when
I need to divide a coefficient by (1-rho) so that it can be interpreted
parallel to the regression weight from a standard OLS.
See more below.
===
In thinking more about your problem, I gave you incorrect advise. The
beta/(1-rho) transformation is appropriate if you have a lagged
dependent variable with coefficient rho, but is not needed if it is just
the error term that has the lagged impact. The estimated equation is
y(t) - rho * y(t-1) = beta * (x(t) - rho*x(t-1)) and the estimated beta
is the natural beta because both sides of the equation have been
transformed by the lag polynomial 1 - rho*L.
Sorry for the confusion. This should make your task easier.
JP
Paul Gronke wrote:
>
> We are working on a paper where we replicate some work done about a
> decade ago, examining the impact of voting by mail on turnout in Oregon.
>
> The scholar analyzes statewide turnout data, including primary and
> general elections. The turnout model includes a relatively standard set
> of indicators, such as type of election (on/off year, senator, governor,
> etc), competitiveness, etc.
>
> They employed feasible generalized least squared model (Beach/McKinnon
> estimator) because they said the data showed autoregressive properties
> (AR 1).
>
> We have replicated their work using OLS.
>
> Then we replicated the AR(1) model two ways: 1) using the Prais-Winsten
> estimator, and 2) using full information maximum likelihood.
>
> Our question regards reporting the coefficients for the AR(1) model. We
> want to compare the marginal effects for the OLS and the FGLS and GLS
> models. A colleague at a conference told us that the coefficients for
> the FGLS and GLS models need to be transformed to make them comparable
> to regression weights, in this way:
>
> b / (1-rho).
>
> Ok, we have that. We we wondering a) do you agree with this
> recommendation, and b) is there an easy way to do this in Stata. I
> notice that Stata saves a vector of coefficients in e(b) and saves the
> value of rho in e(rho).
>
> We want something like REPORT THE VECTOR THAT RESULTS FROM e(b) /
> (1-e(rho))
>
> How do I do this? Is there something likea "calculate" command in Stata?
>
--
Paul Gronke Direct phone: 503-517-7393
Dept. of Political Science E-Fax (preferred): 440-274-8159 **NEW**
Reed College Local fax: 503-777-7776
3203 SE Woodstock Blvd
Portland, OR 97202 http://www.reed.edu/~gronkep
**********************************************************
Political Methodology E-Mail List
Editors: Melanie Goodrich, <[log in to unmask]>
Delia Bailey, <[log in to unmask]>
**********************************************************
Send messages to [log in to unmask]
To join the list, cancel your subscription, or modify
your subscription settings visit:
http://polmeth.wustl.edu/polmeth.php
**********************************************************
|