The pursuit projection packages ppr is an excellent contribution to R. It is great for one-to-three ridge fits, often somewhat intuitive, and for multi-ridge fits, where it at least describes a lot of variance. Like many folk, I need to report the fits obtained from ppr to the greater, outside, non-R world. It is fairly obvious how to use the terms alpha and beta to report on directionality and importance. It has proven difficult to report on the spline fits generated. We are moving into some "cryptanalysis" of the uncommented "predict" code with the "ppr" method in order to locate the information, and can report, if warranted. The question: How can one simply recover the spline knots and the spline parameters associated with a particular fit? Are we missing something obvious, or has there been contributed code that we could make use of? We have considered making spline fits of the spline fit variables, but this seems a bit obtuse. In our case, there are usually several thousand rows of the predictor variables, so the exact description of the knots is necessary but not very problem-dependent. A second (rhetorical) question: Can more information be associated with the differing projection directions chosen for the fit? We have made a second analysis with sphered data, and run arbitrary subsets to assess contributions of spline fits along the various directions, and computed correlations with fitted and (fitted+residual) data. Maybe there's a more standard approach we're missing. Thanks for your advice! Bob Chatfield / NASA Ames Research Center
It is normal to report smooth curves via plots of smooth curves. There are examples for ppr in MASS4 p.241 (referenced on the help page). This is done by the plot() method. Please do consult the references in the documentation: those who very carefully documented these tools put them there because they do give additional information. On Tue, 27 Jun 2006, Robert Chatfield wrote:> The pursuit projection packages ppr is an excellent contribution to R. > It is great for one-to-three ridge fits, often somewhat intuitive, and > for multi-ridge fits, where it at least describes a lot of variance. > > Like many folk, I need to report the fits obtained from ppr to the > greater, outside, non-R world. It is fairly obvious how to use the > terms alpha and beta to report on directionality and importance. > > It has proven difficult to report on the spline fits generated. We are > moving into some "cryptanalysis" of the uncommented "predict" code with > the "ppr" method in order to locate the information, and can report, if > warranted.In what sense do you claim it to be `undocumented'? Methods should do what the generic is documented to do, and that is the case here. [...] -- Brian D. Ripley, ripley at stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595