Well, I can't really comment much on the Fama MacBeth approach since I never
heard of it before today and work in a completely different field. But, I
believe that slopes-as-outcomes or stagewise regression can work quite well
in certain contexts -- particularly when the first stage regressions have
little uncertainty and serve to get at the quantity of interest. In my work
analyzing energy usage of buildings over time, the first stage regression
(for each building for each year) is merely to adjust for weather
differences (heating degree days) between years and the individual OLS
models typically have an R^2 of 0.95. The results from those regressions
are then used (indirectly through predicted outcomes) as the dependent
variable to analyze differences between buildings. I realize that this
analysis may be more efficiently (in a statistical sense) done within a
multilevel modeling framework, but I think little is lost by simplifying it
into stages since the first level models have such good fits while the gains
from the stage-wise approach are substantial in terms of computing time and,
in some ways, interpretability. These advantages are particularly
noticeable when dealing with tens of thousands of buildings at once. I'm
not sure how long -gllamm- might take to fit a multi-level model with 20,000
cross sections and 500,000 observations...do you think it's worth trying or
will it take several weeks/months/years?
Michael Blasnik
[email protected]
----- Original Message -----
From: "Stas Kolenikov" <[email protected]>
To: <[email protected]>
Sent: Tuesday, October 12, 2004 12:37 PM
Subject: Re: st: Fama-MacBeth regressions
> On Tue, 12 Oct 2004 11:32:48 -0400, Subhankar Nayak <[email protected]>
wrote:
> > Fama-Macbeth approach is an innovative two-stage approach meant to
minimize
> > within-portfolio variance while capturing the across-portfolio
> > characteristics...
> > Their 1974 paper is not a landmark in terms of econometric modelling,
but
> > the approach is nice.
> > Their approach is meant to test Capital Asset Pricing Model (CAPM).
>
> [explanations absorbed; a couple of questions remained]
>
> 1. Looks like there is quite a bit of arbitrary tune up: why 24
> months? why 10 portfolios?
>
> 2. Are the standard errors corrected for the multi-stage estimation?
> You can still cast this problem in terms of linear filtering of the
> original data, as all of the testable coefficients should be linear
> functions of the original prices. Hence you can come up with correct
> standard errors just by the virtue of this linearity. (Am I missing
> something?) The standard errors that come out of the last regression
> may be wrong by a factor of 5 or so.
>
> You or Michael may comfort me on this though, as I am not familiar
> with the area (but curious to learn something new if this does not
> take too much of your time :))
>
> --
> Stas Kolenikov
> http://stas.kolenikov.name
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
