On 03-Jun-05 Michael Grant wrote:
> > I presume the reference is to the 'geometric mean > functional regression' or the 'line of organic > correlation' or 'reduced major axis regression'. If > so, this is relatively easy alsmost trivial to > implement in R.
This somewhat contentious method is indeed trivial to implement in R. The idea is that if you plot the two regression lines (y on x, x on y) on the same axes (y vertical, x horizontal), the slope of the GMR is the geometric mean of the slopes of these two lines.
Since the slope of the y-on-x line is Sxy/Sxx, and the slope of the x-on-y line is Syy/Sxy, the GMR slope is therefore sqrt(Syy/Sxx) = sd(y)/sd(x).
All three lines go through the same point, (mean(x),mean(y)).
> Maybe it's in a package, but I never looked.
> I worked from Helsel's description in his classic water > resources statistics book. See Chapter 10 here: > > http://water.usgs.gov/pubs/twri/twri4a3/
The method goes back a lot further than suggested here. It seems it was proposed in oceanography by H. Sverdrup in 1916, and very influentially promoted by W.E. Ricker (e.g. Jnl Fisheries Research Board of Canada, 1973, vol. 30, 409-434).
> Now, if you are after confidence intervals or > prediction intervals, I haven't found anything on that > yet. Seems that I did something a couple of year ago > by hacking some approximate residuals using the LOC > line and the data, and then feeding that into the CL > and PL equations for OLS. (Be advised that I'm not a > statistician and did that in the spirit of > approximation--who knows? :O) )
The uncertainty properties, and indeed the interpretation, of this method are elusive. You can, of course, resort to whatever stochastic modelling you choose (including simulation and bootstrap) to estimate the variability of the slope sd(y)/sd(x) and of any predictions you may want to make.
However, the method shows its indeterminate side to the extent that the relationship between y and x is loose rather than tight.
At one extreme, where the correlation between x and y = 1, the two regression lines (y on x and z on y) and the GMR all coincide. No problem here.
At the other extreme, where there is no correlation, the GMR method still gives you a definite answer (sd(y)/sd(x)) even though by normal standards there is no relationhip between y and x. In the latter case, the slope of the GMR depends solely on the two SDs, and we may well ask what is being estimated here (apart from the ratio of the SDs).
(Of course, if you go back to the "primitive" definition, you find yourself evaluating sqrt(0 * inf), which is indeterminate; and this is a better outcome than sd(y)/sd(x), but still falls short of telling you directly that y is independent of x).
As you approach the r=0 situation, you therefore have to be mindful that the GMR method will appear to provide a definite answer to a question which in reality has at best a vague answer, i.e. there is a major problem of interpretation.
Therefore I would be suspicious of results obtained by "blind" application of the GMR method which were not accompanied by a good discussion of grounds why the results can be expectd to be meaningful in the particular case where it has been applied.
> By coincidence I've been looking at this again > recently. Maybe bootstrapping.... > > Regards, > Michael Grant > > --- Kjetil Brinchmann Halvorsen <kjetil@acelerate.com> > wrote: >
> > ______________________________________________ > R-help@stat.math.ethz.ch mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide! > http://www.R-project.org/posting-guide.html --------------------------------------------------------------------E-Mail: (Ted Harding) <Ted.Harding@nessie.mcc.ac.uk> Fax-to-email: +44 (0)870 094 0861
Date: 06-Jun-05 Time: 10:20:01 ------------------------------ XFMail ------------------------------ ______________________________________________R-help@stat.math.ethz.ch mailing list
This archive was generated by hypermail 2.1.8 : Fri 03 Mar 2006 - 03:32:23 EST