12
$\begingroup$

I'm doing linear regression on compositional data using log-ratio transformation with census data. The IVs are compositional (percents summing to 100). The DV is non-compositional and continuous.

The alr and clr results are more easily interpreted. They all produce the same measure of fit. I'm inclined to go with alr (or clr). Aitchison characterizes ilr as the "pure mathematics" approach, but my audience is not statisticians or mathematicians.

If my objective is only to communicate insight from the analysis, why should I go with the much more difficult to interpret ilr (with balances) approach?

I've read heaps of research by Aitchison, Juan Jose Egozcue and Vera Pawlosky-Glahn but not looking to debate.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
7
$\begingroup$

Continuing off of marianess's answer, clr is really not suitable due to the colinearity issue. In words if you try to make inferences with clr transformed data, you may fall in the trap of trying to infer increase/decreases of variables, which you can never never do with proportions in the first place.

The ilr transformation attempts to resolve this by just sticking to ratios of partitions, since ratios are stable quantities. These partitions can be represented as trees, where internal nodes in the tree represents the log ratio of the geometric means of the subtrees. This log ratios of subtrees is known as balances.

I'd also recommend checking out these publications, since they all have nice explanations of how to interpret the ilr transform.

http://msystems.asm.org/content/2/1/e00162-16

https://peerj.com/articles/2969/

https://elifesciences.org/content/6/e21887

Here is an IPython notebook that goes in the details of how to calculate balances given a tree

I also gave a description how to this with the modules in scikit-bio here in case you curious.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Why is the size of the output m - 1? $\endgroup$
    – O.rka
    Commented Jun 13, 2018 at 1:02
  • $\begingroup$ Are you able to directly associate a feature with its a value? $\endgroup$
    – O.rka
    Commented Jun 13, 2018 at 1:18
  • $\begingroup$ its m-1 since it is an isomorphism - you can only have at most m-1 contrasts before you start hitting collinearity issues. And yes, you should be able to link a feature to a specific partition see this answer here: stats.stackexchange.com/a/270203/79569 $\endgroup$
    – mortonjt
    Commented Jun 15, 2018 at 3:05
6
$\begingroup$

There is a problem with clr() transformation. It does preserve the same amount variables after you transform the data, but in case of clr() you get a singular data (actually you get a singular covariance matrix): y1 + ... yD = 0. And as you might know, some statistical analysis cannot be performed on a singular data. ilr() transformation will reduce the number of your variable, so let's say you had D-dimensional space, but after ilr() you will end up with D-1. As a result, your transformed data is nothing more, but ratios. I recommend to read this paper here: http://is.muni.cz/do/rect/habilitace/1431/Hron/habilitace/15_Filzmoser_et_al__2010_.pdf

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ a useful distinction for clr. alr reduces the vector to D-1 as well. ilr ratios are of groups of variables(one or more), whereas alr ratios are of each single variable with the last variable in the vector the common denominator. ilr could provide insight that alr does not, but with my data most of the ilr ratios do not make any intuitive sense, even with balances. My current thinking is that it depends on your data (i.e. sometimes you want to see ratios of groups of variables, sometimes not). $\endgroup$
    – M Kearny
    Commented Nov 2, 2016 at 3:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.