Dianne: the advice of your teachers was a really reasonable one.But what you should do is checking how much your results using linear regression can be trusted. Most programs give you regression diagnostics and make sure that they had been activated (The new Cohen& Cohen 3rd ed. which was written by Steve West and Leona Aiken has a lot of helpful infos and strategies about these problems).You need no new nonparametric linear regression program but you can transform your data to ranks, which is today a few mouseclicks away only and then do the linear regression with the Spearman rank correlation matrix (which results when you use your usual correlation program). This is an analogue to what they meant as a nonparametric version. Compare the R-squares and the regression weights you get with the two approaches and check how stable and comparable they are between the two approaches. If all assumptions hold, the rank based coefficients should be only slightly lower due to the inevitable loss of information and power. If they disagree to such an extant that you have to substantially change your theoretical and substantial conclusions, than you know that outliers probably have fooled you, but good regression diagnostics would have told you that already in the first place. HTH Werner W. Wittmann,University of Mannheim; Germany ----- Original Message ----- From: "Dianne Worth" <[EMAIL PROTECTED]> To: <[EMAIL PROTECTED]> Sent: Thursday, June 26, 2003 4:46 PM Subject: Re: Nonparametric test for regression?
> Okay, mea culpa. > > I think linear regression IS the proper method of > analyzing the data. Back in my NP class (of course), > we were encouraged to use both para- and nonpara- > methods to see if we arrived at similar results. > > Perhaps it was simply an exercise, but I would think > it could come in handy at times. > DW > > --- Donald Burrill <[EMAIL PROTECTED]> wrote: > > Details, please. Linear regression is a method of > > analysis, not a test. > > Having performed a regression, one might then be > > interested in testing > > whether the slope coefficient was different from > > some specified value, > > or in constructing a confidence interval around the > > slope, for one (or > > more) predictor(s); and similarly for the > > intercept. > > > > If you are asking whether there is a > > "non-parametric" analogue of linear > > regression that would provide an estimate of the > > slope of a line > > relating the response variable to the predictor, yes > > there is at least > > one. Testing hypotheses about (e.g.) the _value_ of > > that slope would be > > more troublesome, though. > > > > What did you actually have in mind that you wanted > > to do? > > And why do you [apparently] assume that linear > > regression is not a way > > of doing it? > > > > On Wed, 25 Jun 2003, Dianne Worth wrote: > > > > > Does anyone know if there is a NP test to compare > > with > > > linear regression? > > > > > > > ----------------------------------------------------------------------- > > Donald F. Burrill > > [EMAIL PROTECTED] > > 56 Sebbins Pond Drive, Bedford, NH 03110 > > (603) 626-0816 > > > > . > > . > > > ================================================================= > > Instructions for joining and leaving this list, > > remarks about the > > problem of INAPPROPRIATE MESSAGES, and archives are > > available at: > > . http://jse.stat.ncsu.edu/ > > . > > > ================================================================= > > > __________________________________ > Do you Yahoo!? > SBC Yahoo! DSL - Now only $29.95 per month! > http://sbc.yahoo.com > . > . > ================================================================= > Instructions for joining and leaving this list, remarks about the > problem of INAPPROPRIATE MESSAGES, and archives are available at: > . http://jse.stat.ncsu.edu/ . > ================================================================= . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
