[Vegan-commits] r2203 - in branches/2.0/inst: . doc
noreply at r-forge.r-project.org
noreply at r-forge.r-project.org
Sun May 27 17:15:38 CEST 2012
Author: jarioksa
Date: 2012-05-27 17:15:38 +0200 (Sun, 27 May 2012)
New Revision: 2203
Modified:
branches/2.0/inst/ChangeLog
branches/2.0/inst/doc/FAQ-vegan.texi
Log:
merge FAQ about data size requirements in NMDS
Modified: branches/2.0/inst/ChangeLog
===================================================================
--- branches/2.0/inst/ChangeLog 2012-05-27 15:13:38 UTC (rev 2202)
+++ branches/2.0/inst/ChangeLog 2012-05-27 15:15:38 UTC (rev 2203)
@@ -8,6 +8,7 @@
* merge r2191-2193: standardise handling of 'select' arg in
those plotting functions that support it. Adds non-exported
function .checkSelect().
+ * merge r2182,2182,2199,2201: FAQ about data size in NMDS.
* merge r2178, 2180: ordipointlabel gains 'select' argument.
* merge r2173-2176, 2185: ordihull labels, semintransparent
colours in ordihull & ordiellipse.
Modified: branches/2.0/inst/doc/FAQ-vegan.texi
===================================================================
--- branches/2.0/inst/doc/FAQ-vegan.texi 2012-05-27 15:13:38 UTC (rev 2202)
+++ branches/2.0/inst/doc/FAQ-vegan.texi 2012-05-27 15:15:38 UTC (rev 2203)
@@ -8,7 +8,7 @@
@setfilename FAQ- at pkg{vegan}.info
@settitle @pkg{vegan} FAQ
@setchapternewpage on
- at set FAQ_YEAR 2011
+ at set FAQ_YEAR 2012
@afourpaper
@c %**end of header
@@ -26,7 +26,7 @@
Creative Commons, 543 Howard Street, 5th Floor, San Francisco,
California, 94105, USA.
-Copyright @copyright{} 2008-2011 Jari Oksanen
+Copyright @copyright{} 2008-2012 Jari Oksanen
@end quotation
@end copying
@@ -295,9 +295,10 @@
@menu
* I have only numeric and positive data but @pkg{vegan} still complaints::
-* Can you analyse binary or cover class data?::
+* Can I analyse binary or cover class data?::
* Why dissimilarities in @pkg{vegan} differ from other sources?::
* Why NMDS stress is sometimes 0.1 and sometimes 10?::
+* I get zero stress but no convergent solutions in @code{metaMDS}::
* Zero dissimilarities in isoMDS::
* How the RDA results are scaled?::
* cca fails with ``data.frame expected'' or ``"site.env" missing''::
@@ -315,7 +316,7 @@
* Can I zoom into an ordination plot?::
@end menu
- at node I have only numeric and positive data but @pkg{vegan} still complaints, Can you analyse binary or cover class data?, Ordination, Ordination
+ at node I have only numeric and positive data but @pkg{vegan} still complaints, Can I analyse binary or cover class data?, Ordination, Ordination
@comment node-name, next, previous, up
@section I have only numeric and positive data but @pkg{vegan} still complaints
@@ -328,8 +329,8 @@
the data). Another common reason is that you had empty cells in your
input data, and these were interpreted as missing values.
- at node Can you analyse binary or cover class data?, Why dissimilarities in @pkg{vegan} differ from other sources?, I have only numeric and positive data but @pkg{vegan} still complaints, Ordination
- at section Can you analyse binary or cover class data?
+ at node Can I analyse binary or cover class data?, Why dissimilarities in @pkg{vegan} differ from other sources?, I have only numeric and positive data but @pkg{vegan} still complaints, Ordination
+ at section Can I analyse binary or cover class data?
Yes. Most @pkg{vegan} methods can handle binary data or cover abundance data.
Most statistical tests are based on permutation, and do not make
@@ -337,7 +338,7 @@
analysis) that need count data. These methods check that input data are
integers, but they may be fooled by cover class data.
- at node Why dissimilarities in @pkg{vegan} differ from other sources?, Why NMDS stress is sometimes 0.1 and sometimes 10?, Can you analyse binary or cover class data?, Ordination
+ at node Why dissimilarities in @pkg{vegan} differ from other sources?, Why NMDS stress is sometimes 0.1 and sometimes 10?, Can I analyse binary or cover class data?, Ordination
@section Why dissimilarities in @pkg{vegan} differ from other sources?
Most commonly the reason is that other software use presence--absence
@@ -351,7 +352,7 @@
Another reason may be that indices indeed are defined differently,
because people use same names for different indices.
- at node Why NMDS stress is sometimes 0.1 and sometimes 10?, Zero dissimilarities in isoMDS, Why dissimilarities in @pkg{vegan} differ from other sources?, Ordination
+ at node Why NMDS stress is sometimes 0.1 and sometimes 10?, I get zero stress but no convergent solutions in @code{metaMDS}, Why dissimilarities in @pkg{vegan} differ from other sources?, Ordination
@section Why @acronym{NMDS} stress is sometimes 0.1 and sometimes 10?
Stress is a proportional measure of badness of fit. The proportions can
@@ -363,7 +364,31 @@
100 times higher in @code{isoMDS} than in @code{monoMDS}. Both of these
conventions are equally correct.
- at node Zero dissimilarities in isoMDS, How the RDA results are scaled?, Why NMDS stress is sometimes 0.1 and sometimes 10?, Ordination
+ at node I get zero stress but no convergent solutions in @code{metaMDS}, Zero dissimilarities in isoMDS, Why NMDS stress is sometimes 0.1 and sometimes 10?, Ordination
+ at section I get zero stress but no convergent solutions in @code{metaMDS}
+
+Most common reason is that you have too few observations for your
+ at acronym{NMDS}. For @code{n} observations (points) and @code{k}
+dimensions you need to estimate @code{n*k} parameters (ordination
+scores) using @code{n*(n-1)/2} dissimilarities. For @code{k} dimensions
+you must have @code{n > 2*k + 1}, or for two dimensions at least six
+points. In some degenerate situations you may need even a larger number
+of points. If you have a lower number of points, you can find an
+undefined number of perfect (stress is zero) but different solutions.
+Conventional wisdom due to Kruskal is that you should have @code{n > 4*k
++ 1} points for @code{k} dimensions. A typical symptom of insufficient
+data is that you have (nearly) zero stress but no two convergent
+solutions. In those cases you should reduce the number of dimensions
+(@code{k}) and with very small data sets you should not use @code{NMDS},
+but rely on metric methods.
+
+It seems that local and hybrid scaling with @code{monoMDS} have similar
+lower limits in practice (although theoretically they could differ).
+However, higher number of dimensions can be used in metric scaling, both
+with @code{monoMDS} and in principal coordinates analysis
+(@code{cmdscale} in @pkg{stats}, @code{wcmdscale} in @pkg{vegan}).
+
+ at node Zero dissimilarities in isoMDS, How the RDA results are scaled?, I get zero stress but no convergent solutions in @code{metaMDS}, Ordination
@section Zero dissimilarities in isoMDS
Function @code{metaMDS} uses function @code{monoMDS} as its default
More information about the Vegan-commits
mailing list