Hi Tim,
Yes you are right this is an issue, BC (and other distance metrics) are
sensitive to sampling intensity, which is often an artefact of the sampling
technique. Transformation is not a great solution to the problem - it works
imperfectly and will have different effects depending on the properties of your
data. There are lots of different types of datasets out there, each with
different properties, and different behaviours under different
transformation/standardisation strategies, so there is no
one-transformation-suits-all solution. An illustration of this (in the case of
row standardisation) is in the below paper:
https://besjournals.onlinelibrary.wiley.com/doi/10.1111/2041-210X.12843
The strategy I would advise here is to go a very different route and build a
statistical model for the data. You can then include row effects in the model
to handle variation in sampling intensity across rows of data (along the lines
of equation 2 of the above paper). Or if the magnitude of the variation in
sampling intensity is known (e.g. it is due to changes in sizes of quadrats
used for sampling, and quadrat size has been recorded), then the standard
approach to handle this is to add an offset to the model. There is plenty of
software out there that can fit suitable statistical models with row effects
(and offsets) for this sort of data, including the mvabund, HMSC, boral, and
gllvm packages on R. Importantly, these packages come with diagnostic tools to
check that the analysis approach adequately captures key properties of your
data - an essential step in any analysis.
All the best
David
Professor David Warton
School of Mathematics and Statistics, Evolution & Ecology Research Centre,
Centre for Ecosystem Science
UNSW Sydney
NSW 2052 AUSTRALIA
phone +61(2) 9385 7031
fax +61(2) 9385 7123
http://www.eco-stats.unsw.edu.au
----------------------------------------------------------------------
Date: Tue, 2 Apr 2019 17:15:45 +0200
From: Tim Richter-Heitmann <trich...@uni-bremen.de>
To: r-sig-ecology@r-project.org
Subject: [R-sig-eco] interpreting ecological distance approaches (Bray
Curtis after various data transformation)
Message-ID: <3834fea1-040a-12b5-c3a3-633e68dc6...@uni-bremen.de>
Content-Type: text/plain; charset="utf-8"; Format="flowed"
Dear list,
i am not an ecologist by training, so please bear with me.
It is my understanding that Bray Curtis distances seem to be sensitive to
different community sizes. Thus, they seem to deliver inadequate results when
the different community sizes are the result of technical artifacts rather than
biology (see e.g. Weiss et al, 2017 on microbiome data).
Therefore, i often see BC distances made on relative data (which seems to be
equivalent to the Manhattan distance) or on data which has been subsampled to
even sizes (e.g. rarefying). Sometimes i also see Bray Curtis distances
calculated on Hellinger-transformed data,
which is the square root of relative data. This again makes sample sizes
unequal (but only to a small degree), so i wondered if this is a valid
approach, especially considering that the "natural" distance choice for
Hellinger transformed data is Euclidean (to obtain, well, the Hellinger
distance).
Another question is what different sizes (i.e. the sums) of Hellinger
transformed communities represent? I tested some datasets, and couldnt find a
correlation between original sample sizes and their hellinger transformed
counterparts.
Any advice is very much welcome. Thank you.
--
Dr. Tim Richter-Heitmann
University of Bremen
Microbial Ecophysiology Group (AG Friedrich)
FB02 - Biologie/Chemie
Leobener Straße (NW2 A2130)
D-28359 Bremen
Tel.: 0049(0)421 218-63062
Fax: 0049(0)421 218-63069
_______________________________________________
R-sig-ecology mailing list
R-sig-ecology@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
