[R] compositional data: percent values sum up to 1

[Spencer Graves]

Hi, Christoph:

	  Andy Liaw's suggestion sounds sensible to me, though I don't have 
much experience with this kind of data either.

	  OTHER QUESTIONS:

	  * How big is J?  I'm guessing it might be quite large, but I don't 
know.

	  * Are the spectra relatively smooth?  I wonder if it might be 
appropriate to try to smooth the data some way preliminary to other 
analyses.

	  * How many observations do you have in each of "ill" and "healthy", 
especially relative to J?

	  I might try to do a principal components analysis (or "svd" if 
"princomp" bombed because of singular matrices) on the covariance matrix 
of the spectra.  Then I might want to test how different the spectra were.

hope this helps.  spencer graves

Liaw, Andy wrote:
> Eh?  The original message says it's the design matrix that is perfectly
> collinear after the transformation, not the response.
> 
> I don't know much about this type of data, but seems like you could just fit
> the model w/o intercept to eliminate the collinearity, no?  It's the
> interpretation of the result that may be tricky, I think.
> 
> Andy
> 
> 
> 
>>-----Original Message-----
>>From: Spencer Graves [mailto:spencer.graves at pdf.com]
>>Sent: Monday, June 02, 2003 9:33 AM
>>To: Christoph Lehmann
>>Cc: Spencer Graves; r-help at stat.math.ethz.ch
>>Subject: Re: [R] compositional data: percent values sum up to 1
>>
>>
>>"glm" will do multinomial logistic regression.  However, if J 
>>is large, 
>>I doubt if that will do what you want.  If it were my 
>>problem, I might 
>>feel a need to read the code for "glm" and modify it to do 
>>what I want. 
>>  Perhaps someone else can suggest something better.
>>
>>hth.  spencer graves
>>
>>Christoph Lehmann wrote:
>>
>>>I want to do a logistic regression analysis, and to compare with, a
>>>discriminant analysis. The mentioned power maps are my 
>>
>>exogenous data,
>>
>>>the dependent variable (not mentioned so far) is a diagnosis
>>>(ill/healthy)
>>>
>>>thanks for the interest and the help
>>>
>>>Christoph
>>>
>>>On Sun, 2003-06-01 at 21:01, Spencer Graves wrote:
>>>
>>>
>>>>What are you trying to do?  What I would do with this 
>>>
>>depends on many 
>>
>>>>factors.
>>>>
>>>>spencer graves
>>>>
>>>>Christoph Lehmann wrote:
>>>>
>>>>
>>>>>again, under another subject:
>>>>>sorry, maybe an all too trivial question. But we have 
>>>>
>>power data from J
>>
>>>>>frequency spectra and to have the same range for the data 
>>>>
>>of all our
>>
>>>>>subjects, we just transformed them into % values, pseudo-code:
>>>>>
>>>>>power[i,j]=power[i,j]/sum(power[i,1:J])
>>>>>
>>>>>of course, now we have a perfect linear relationship in 
>>>>
>>our x design-matrix,
>>
>>>>>since all power-values for each subject sum up to 1.
>>>>>
>>>>>How shall we solve this problem: just eliminate one column of x, or
>>>>>introduce a restriction which says exactly that our power 
>>>>
>>data sum up to
>>
>>>>>1 for each subject?
>>>>>
>>>>>Thanks a lot
>>>>>
>>>>>Christoph
>>>>
>>>>______________________________________________
>>>>R-help at stat.math.ethz.ch mailing list
>>>>https://www.stat.math.ethz.ch/mailman/listinfo/r-help
>>>
>>______________________________________________
>>R-help at stat.math.ethz.ch mailing list
>>https://www.stat.math.ethz.ch/mailman/listinfo/r-help
>>
> 
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://www.stat.math.ethz.ch/mailman/listinfo/r-help