NetNews Usenet Archive 1993 #3/NN_1993_3.iso/spool/bit/listserv/statl/2421

Text File | 1993-01-21 | 1.8 KB | 45 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!paladin.american.edu!auvm!STAT.UBC.CA!PING
X-Mailer: ELM [version 2.3 PL6]
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.stat-l
Date:         Thu, 21 Jan 1993 15:52:46 PST
Sender:       STATISTICAL CONSULTING <[e-mail]>
From:         Ping Ma <[e-mail]>
Subject:      Re: r for nonlinear regression
In-Reply-To:  <[e-mail]>; from "BILL SIMPSON" at Jan 21,
              93 10:50 am
Lines: 31

> the sd of the observation.  If unweighted, this is the same formula one
> might use for linear regression: sqrt(1-(sse/ssto)).  I don't see how it makes

Note that 1-sse/ssto = (ssto-sse)/ssto = propotion of total variation
in the response variable explained by the model (a nonlinear model
in this context). This proportion is also called the coeff of determination,
which measures how well a model fits the data.

> any sense for nonlinear regression.
>
> This can't make sense because it measures degree of fit TO A LINE, not to the
> nonlinear regression function.  (It seems to me.) The only r that would

The above is only a special case of linear regression --- one explantroy
variable. In this case, sqrt(1-(sse/ssto)) gives the correlation coeff
of the pair of variables (the response and explanatory) measuring the
strength of linear association of the pair of variables. Since the model
used is a linear model, the "fit" amounts "to a line". The fit could be
a hyper-plane (for linear mutliple regression) or a curve (for
non-regression), etc.

Hope this help.

-------------------------------------------------
Ping Hang Ma,
Statistics Department, UBC
Vancouver, B.C., Canada
E-Mail: [e-mail]
Tel:  [phone]
Fasx: [phone]
-------------------------------------------------