🚀 go-pugleaf

Hello,

I'm struggling with some results from an auditory perception experiment.   
Basically, I have perceptual magnitude judgments for 120 sounds, and the  
judgments are nicely explained by a regression on five acoustical  
parameters of the sound stimuli.

However, I'm not content to report merely that I had a sucessful  
regression and name the five parameters.  I would like to characterize the  
kinds of interactions that occur among the five parameters (let's call  
them A, B, C, D and E) over the 120 trials (let's call this variable X).

For example, there are an infinite number of values that could be plugged  
into the regression equation to produce any value in the range of X.   
Perhaps the results are such that the values for A, B and C could all be  
zero and D and E could carry the whole wieght...or vice versa.  But the  
fact is, there is only a small set of actual values for A,B,C,D and E  
(120), and there is never any such case.  A meaningful question is, what  
types of ratios of values for A,B,C,D,E exist in the actual data?  For  
example, are there cases where B alone compensates for small values of A?   
Or whenever C is small is it often true that A is small as well but B, D  
and E are large?  Those kinds of questions.

I would be interested in knowing about:

1. methods for posthoc analysis of regression results
2. books on intuitive approaches to explaining regression results (note:   
I'm not interested in intuitive intorductions to statistics concepts)

Thanks in advance,
--
Greg Sandell
Research Fellow, Center for New Music and Audio Technologies (CNMAT)
sandell@cnmat.cnmat.berkeley.edu