home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!paladin.american.edu!gatech!emory!swrinde!network.ucsd.edu!flim
- From: flim@weber.ucsd.edu (Francis Lim)
- Newsgroups: sci.math.stat
- Subject: Follow-up: ACE and Splus, why transformations don't sum-up!
- Date: 6 Jan 1993 18:47:27 GMT
- Organization: University of California, San Diego
- Lines: 63
- Message-ID: <1if9fvINNah9@network.ucsd.edu>
- NNTP-Posting-Host: weber.ucsd.edu
-
- From charlie@umnstat.stat.umn.edu Wed Jan 6 08:18 PST 1993
- From: "Charles Geyer" <charlie@umnstat.stat.umn.edu>
- To: flim@weber.ucsd.edu
- Subject: Re: ACE (Alternating Conditional Expectations)
- Newsgroups: sci.math.stat
- Organization: School of Statistics, University of Minnesota
-
- In article <1ib1luINN84d@network.ucsd.edu> you write:
- >Hello netters. I have a simple ACE (Breiman and Friedman 1985) question.
- >
- >ACE stands for Alternating Conditional Expectations, and is a method
- >of estimating the relationship between the explanatory and response
- >variables, transforming the variables to maximize their correlation.
- >
- >The presumed relationship between Y and X, for example, is
- >
- > \theta(Y) = \phi(X) + e
- >
- >where e is random error.
- >
- >One can use either BLSS or S to invoke ACE.
- >For both statistical packages, I found that \hat{\theta}(Y) did NOT
- >equal \hat{\phi}(X), where \hat symbolizes the estimated
- >transformations.
- >
- >In fact, \hat{\theta}(Y) was not even proportional to \hat{\phi}(X).
-
-
-
- You said it yourself
-
- \theta(Y) - \phi(X)
-
- is the residuals -- random error.
-
- You might find the output simpler to understand if you use the option
- to disallow nonlinear transformations of the response. Then you get
-
- Y = phi(X) + error
-
- and it looks much more like the usual regression situation. You don't
- then expect Y to be proportional to phi(X). [Actually it won't do this
- it insists on at least a standardizing linear transformation of Y, but
- the points the same.]
-
- --
- Charles Geyer
- School of Statistics
- University of Minnesota
- charlie@umnstat.stat.umn.edu
-
- ================================================
-
- Thanks Charles, yes it all makes sense. ACE attempts to maximize
- the correlation between \theta(Y) and \phi(X).
- Thus if \hat{\theta}(Y) were proportional to \hat{\phi}(X),
- the sample correlation would be 1. This would happen only in
- extreme cases! Thanks again.
- --
- -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
- FRANCIS W. LIM | email: flim@ucsd.edu | dept phone: (619) 534-3383
- Dept of Economics, Univ of Cal at San Diego, La Jolla, CA 92093-0508
- -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
-