NEWS | R Documentation |
Fortran code was modernized to avoid warnings in latest R. The modernization should have no visible effect in functions. Please report all suspect cases as vegan issues.
Several support functions for ordination methods failed if
the solution had only one ordination axis, for instance, if
there was only one constraining variable in CCA, RDA and
friends. This concerned goodness
for constrained
ordination, inertcomp
, fitted
for
capscale
, stressplot
for RDA, CCA (GitHub issue
#189).
goodness
for CCA & friends ignored choices
argument (GitHub issue
#190).
goodness
function did not consider negative
eigenvalues of db-RDA (function dbrda
).
Function meandist
failed in some cases when one of
the groups had only one observation.
linestack
could not handle expressions in
labels
. This regression is discussed in GitHub issue
#195.
Nestedness measures nestedbetajac
and
nestedbetasor
expecting binary data did not cope with
quantitative input in evaluating Baselga's matrix-wide Jaccard
or Sørensen dissimilarity indices.
Function as.mcmc
to cast oecosimu
result to an
MCMC object (coda package) failed if there was only one
chain.
diversity
function returns now NA
if the
observation had NA
values instead of returning
0
. The function also checks the input and refuses to
handle data with negative values. GitHub issue
#187.
rarefy
function will work more robustly in marginal
case when the user asks for only one individual which can only
be one species with zero variance.
Several functions are more robust if their factor arguments
contain missing values (NA
): betadisper
,
adipart
, multipart
, hiersimu
, envfit
and constrained ordination methods cca
, rda
,
capscale
and dbrda
. GitHub issues
#192 and
#193.
Distance-based methods were redesigned and made
consistent for ordination (capscale
, new dbrda
),
permutational ANOVA (adonis
, new adonis2
),
multivariate dispersion (betadisper
) and variation
partitioning (varpart
). These methods can produce
negative eigenvalues with several popular semimetric
dissimilarity indices, and they were not handled similarly by
all functions. Now all functions are designed after McArdle &
Anderson (Ecology 82, 290–297; 2001).
dbrda
is a new function for distance-based
Redundancy Analysis following McArdle & Anderson
(Ecology 82, 290–297; 2001). With metric
dissimilarities, the function is equivalent to old
capscale
, but negative eigenvalues of semimetric indices
are handled differently. In dbrda
the dissimilarities
are decomposed directly into conditions, constraints and
residuals with their negative eigenvalues, and any of the
components can have imaginary dimensions. Function is mostly
compatible with capscale
and other constrained
ordination methods, but full compatibility cannot be achieved
(see issue
#140 in
Github). The function is based on the code by Pierre Legendre.
The old capscale
function for constrained
ordination is still based only on real components, but the
total inertia of the components is assessed similarly as in
dbrda
.
The significance tests will differ from the previous version,
but function oldCapscale
will cast the capscale
result to a similar form as previously.
adonis2
is a new function for permutational ANOVA
of dissimilarities. It is based on the same algorithm as the
dbrda
. The function can perform overall tests of all
independent variables as well as sequential and marginal tests
of each term. The old adonis
is still available, but it
can only perform sequential tests. With same settings,
adonis
and adonis2
give identical results (but
see Github issue
#156 for
differences).
Function varpart
can partition dissimilarities
using the same algorithm as dbrda
.
Argument sqrt.dist
takes square roots of
dissimilarities and these can change many popular semimetric
indices to metric distances in capscale
, dbrda
,
wcmdscale
, adonis2
, varpart
and
betadisper
(issue
#179 in
Github).
Lingoes and Cailliez adjustments change any dissimilarity
into metric distance in capscale
, dbrda
,
adonis2
, varpart
, betadisper
and
wcmdscale
. Earlier we had only Cailliez adjustment in
capscale
(issue
#179 in
Github).
RsquareAdj
works with capscale
and
dbrda
and this allows using ordiR2step
in model
building.
specaccum
: plot
failed if line type
(lty
) was given. Reported by Lila Nath Sharma (Univ
Bergen, Norway)
ordibar
is a new function to draw crosses of
standard deviations or standard errors in ordination diagrams
instead of corresponding ellipses.
Several permustats
results can be combined with a
new c()
function.
New function smbind
binds together null models by
row, column or replication. If sequential models are bound
together, they can be treated as parallel chains in subsequent
analysis (e.g., after as.mcmc
). See issue
#164 in
Github.
Null model analysis was upgraded:
New "curveball"
algorithm provides a fast null model with
fixed row and column sums for binary matrices after Strona et
al. (Nature Commun. 5: 4114; 2014).
The "quasiswap"
algorithm gained argument thin
which can reduce the bias of null models.
"backtracking"
is now much faster, but it is still very
slow, and provided mainly to allow comparison against better and
faster methods.
Compiled code can now be interrupted in null model simulations.
designdist
can now use beta diversity notation
(gamma
, alpha
) for easier definition of beta
diversity indices.
metaMDS
has new iteration strategy: Argument
try
gives the minimum number of random starts, and
trymax
the maximum number. Earlier we only hand
try
which gave the maximum number, but now we run at
least try
times. This reduces the risk of being trapped
in a local optimum (issue
#154 in
Github).
If there were no convergent solutions, metaMDS
will now
tabulate stopping criteria (if trace = TRUE
). This can
help in deciding if any of the criteria should be made more
stringent or the number of iterations increased. The
documentation for monoMDS
and metaMDS
give more
detailed information on convergence criteria.
The summary
of permustats
prints now
P-values, and the test direction (alternative
) can
be changed.
The qqmath
function of permustats
can now plot
standardized statistics. This is a partial solution to issue
#172 in
Github.
MDSrotate
can rotate ordination to show maximum
separation of factor levels (classes) using linear discriminant
analysis (lda
in MASS package).
adipart
, hiersimu
and multipart
expose argument method
to specify the null model.
RsquareAdj
works with cca
and this allows
using ordiR2step
in model building. The code was
developed by Dan McGlinn (issue
#161 in
Github). However, cca
still cannot be used in
varpart
.
ordiellipse
and ordihull
allow setting
colours, line types and other graphical parameters.
The alpha channel can now be given also as a real number in 0 ... 1 in addition to integer 0 ... 255.
ordiellipse
can now draw ellipsoid hulls that
enclose points in a group.
ordicluster
, ordisegments
, ordispider
and lines
and plot
functions for isomap
and
spantree
can use a mixture of colours of connected
points. Their behaviour is similar as in analogous functions in
the the vegan3d package.
plot
of betadisper
is more configurable. See
issues
#128 and
#166 in
Github for details.
text
and points
methods for
orditkplot
respect stored graphical parameters.
Environmental data for the Barro Colorado Island forest plots gained new variables from Harms et al. (J. Ecol. 89, 947–959; 2001). Issue #178 in Github.
Function metaMDSrotate
was removed and replaced
with MDSrotate
.
density
and densityplot
methods for
various vegan objects were deprecated and replaced with
density
and densityplot
for permustats
.
Function permustats
can extract the permutation and
simulation results of vegan result objects.
eigenvals
fails with prcomp
results in
R-devel. The next version of prcomp
will have an
argument to limit the number of eigenvalues shown
(rank.
), and this breaks eigenvals
in vegan.
calibrate
failed for cca
and friends if
rank
was given.
betadiver
index 19
had wrong sign in one of
its terms.
linestack
failed when the labels
were given,
but the input scores had no names. Reported by Jeff Wood (ANU,
Canberra, ACT).
vegandocs
is deprecated. Current R provides better
tools for seeing extra documentation (news()
and
browseVignettes()
).
All vignettes are built with standard R tools and can be
browsed with browseVignettes
. FAQ-vegan
and
partitioning
were only accessible with vegandocs
function.
Dependence on external software texi2dvi
was
removed. Version 6.1 of texi2dvi
was incompatible with R
and prevented building vegan. The FAQ-vegan
that was
earlier built with texi2dvi
uses now knitr. Because
of this, vegan is now dependent on R-3.0.0. Fixes issue
#158 in
Github.
metaMDS
and monoMDS
could fail if input
dissimilarities were huge: in the reported case they were of
magnitude 1E85. Fixes issue
#152 in
Github.
Permutations failed if they were defined as permute
control structures in estaccum
, ordiareatest
,
renyiaccum
and tsallisaccum
. Reported by Dan
Gafta (Cluj-Napoca) for renyiaccum
.
rarefy
gave false warnings if input was a vector
or a single sampling unit.
Some extrapolated richness indices in specpool
needed the number of doubletons (= number of species occurring
in two sampling units), and these failed when only one sampling
unit was supplied. The extrapolated richness cannot be
estimated from a single sampling unit, but now such cases are
handled smoothly instead of failing: observed non-extrapolated
richness with zero standard error will be reported. The issue
was reported in
StackOverflow.
treedist
and treedive
refuse to handle
trees with reversals, i.e, higher levels are more homogeneous
than lower levels. Function treeheight
will estimate
their total height with absolute values of branch
lengths. Function treedive
refuses to handle trees with
negative branch heights indicating negative
dissimilarities. Function treedive
is faster.
gdispweight
works when input data are in a matrix
instead of a data frame.
Input dissimilarities supplied in symmetric matrices or
data frames are more robustly recognized by anosim
,
bioenv
and mrpp
.
Printing details of a gridded permutation design would fail when the grid was at the within-plot level.
ordicluster
joined the branches at wrong coordinates
in some cases.
ordiellipse
ignored weights when calculating standard
errors (kind = "se"
). This influenced plots of cca
,
and also influenced ordiareatest
.
adonis
and capscale
functions recognize
symmetric square matrices as dissimilarities. Formerly
dissimilarities had to be given as "dist"
objects such as
produced by dist
or vegdist
functions, and data
frames and matrices were regarded as observations x variables
data which could confuse users (e.g., issue
#147).
mso
accepts "dist"
objects for the distances
among locations as an alternative to coordinates of locations.
text
, points
and lines
functions for
procrustes
analysis gained new argument truemean
which allows adding procrustes
items to the plots of
original analysis.
rrarefy
returns observed non-rarefied communities
(with a warning) when users request subsamples that are larger
than the observed community instead of failing. Function
drarefy
has been similar and returned sampling
probabilities of 1, but now it also issues a warning. Fixes issue
#144 in
Github.
Permutation tests did not always correctly recognize ties
with the observed statistic and this could result in too low
P-values. This would happen in particular when all predictor
variables were factors (classes). The changes concern functions
adonis
, anosim
, anova
and permutest
functions for cca
, rda
and capscale
,
permutest
for betadisper
, envfit
,
mantel
and mantel.partial
, mrpp
, mso
,
oecosimu
, ordiareatest
, protest
and
simper
. This also fixes issues
#120 and
#132 in
GitHub.
Automated model building in constrained ordination
(cca
, rda
, capscale
) with step
,
ordistep
and ordiR2step
could fail if there were
aliased candidate variables, or constraints that were completely
explained by other variables already in the model. This was a
regression introduced in vegan 2.2-0.
Constrained ordination methods cca
, rda
and
capscale
treat character variables as factors in analysis,
but did not return their centroids for plotting.
Recovery of original data in metaMDS
when computing
WA scores for species would fail if the expression supplied to
argument comm
was long & got deparsed to multiple
strings. metaMDSdist
now returns the (possibly modified)
data frame of community data comm
as attribute
"comm"
of the returned dist
object. metaMDS
now uses this to compute the WA species scores for the NMDS. In
addition, the deparsed expression for comm
is now robust to
long expressions. Reported by Richard Telford.
metaMDS
and monoMDS
rejected dissimilarities
with missing values.
Function rarecurve
did not check its input and this
could cause confusing error messages. Now function checks that
input data are integers that can be interpreted as counts on
individuals and all sampling units have some species. Unchecked
bad inputs were the reason for problems reported in
Stackoverflow.
Scaling of ordination axes in cca
, rda
and
capscale
can now be expressed with descriptive strings
"none"
, "sites"
, "species"
or
"symmetric"
to tell which kind of scores should be scaled by
eigenvalues. These can be further modified with arguments
hill
in cca
and correlation
in rda
. The
old numeric scaling can still be used.
The permutation data can be extracted from anova
results of constrained ordination (cca
, rda
,
capscale
) and further analysed with permustats
function.
New data set BCI.env
of site information for the Barro
Colorado Island tree community data. Most useful variables are the
UTM coordinates of sample plots. Other variables are constant or
nearly constant and of little use in normal analysis.
Constrained ordination functions cca
, rda
and
capscale
are now more robust. Scoping of data set names and
variable names is much improved. This should fix numerous
long-standing problems, for instance those reported by Benedicte
Bachelot (in email) and Richard Telford (in Twitter), as well as
issues #16
and #100 in
GitHub.
Ordination functions cca
and rda
silently
accepted dissimilarities as input although their analysis makes
no sense with these methods. Dissimilarities should be analysed
with distance-based redundancy analysis (capscale
).
The variance of the conditional component was over-estimated
in goodness
of rda
results, and results were wrong
for partial RDA. The problems were reported in an
R-sig-ecology
message by Christoph von Redwitz.
orditkplot
did not add file type identifier to saved
graphics in Windows although that is required. The problem only
concerned Windows OS.
goodness
function for constrained ordination
(cca
, rda
, capscale
) was redesigned. Function
gained argument addprevious
to add the variation explained
by previous ordination components to axes when statistic =
"explained"
. With this option, model = "CCA"
will include
the variation explained by partialled-out conditions, and
model = "CA"
will include the accumulated variation
explained by conditions and constraints. The former behaviour was
addprevious = TRUE
for model = "CCA"
, and
addprevious = FALSE
for model = "CA"
. The argument
will have no effect when statistic = "distance"
, but this
will always show the residual distance after all previous
components. Formerly it displayed the residual distance only for
the currently analysed model.
Functions ordiArrowMul
and ordiArrowTextXY
are
exported and can be used in normal interactive sessions. These
functions are used to scale a bunch arrows to fit ordination
graphics, and formerly they were internal functions used within
other vegan functions.
orditkplot
can export graphics in SVG format. SVG is
a vector graphics format which can be edited with several external
programs, such as Illustrator and Inkscape.
Rarefaction curve (rarecurve
) and species
accumulation models (specaccum
, fitspecaccum
)
gained new functions to estimate the slope of curve at given
location. Originally this was based on a response to an
R-SIG-ecology
query. For rarefaction curves, the function is rareslope
,
and for species accumulation models it is specslope
.
The functions are based on analytic equations, and can also be
evaluated at interpolated non-integer values. In
specaccum
models the functions can be only evaluated for
analytic models "exact"
, "rarefaction"
and
"coleman"
. With "random"
and "collector"
methods you can only use finite differences
(diff(fitted(<result.object>))
). Analytic functions for
slope are used for all non-linear regression models known to
fitspecaccum
.
Species accumulation models (specaccum
) and
non-liner regression models for species accumulation
(fitspecaccum
) work more consistently with weights. In
all cases, the models are defined using the number of sites as
independent variable, which with weights means that observations
can be non-integer numbers of virtual sites. The predict
models also use the number of sites with newdata
,
and for analytic models they can estimate the expected values
for non-integer number of sites, and for non-analytic randomized
or collector models they can interpolate on non-integer values.
fitspecaccum
gained support functions AIC
and deviance
.
The varpart
plots of four-component models were
redesigned following Legendre, Borcard & Roberts Ecology
93, 1234–1240 (2012), and they use now four ellipses instead of
three circles and two rectangles. The components are now labelled
in plots, and the circles and ellipses can be easily filled with
transparent background colour.
This is a maintenance release to avoid warning messages caused by changes in CRAN repository. The namespace usage is also more stringent to avoid warnings and notes in development versions of R.
vegan can be installed and loaded without tcltk
package. The tcltk package is needed in orditkplot
function for interactive editing of ordination graphics.
ordisurf
failed if gam package was loaded due
to namespace issues: some support functions of gam were used
instead of mgcv functions.
tolerance
function failed for unconstrained
correspondence analysis.
estimateR
uses a more exact variance formula for
bias-corrected Chao estimate of extrapolated number of
species. The new formula may be unpublished, but it was derived
following the guidelines of Chiu, Wang, Walther & Chao,
Biometrics 70, 671–682 (2014),
online
supplementary material.
Diversity accumulation functions specaccum
,
renyiaccum
, tsallisaccum
, poolaccum
and
estaccumR
use now permute package for permutations
of the order of sampling sites. Normally these functions only
need simple random permutation of sites, but restricted
permutation of the permute package and user-supplied
permutation matrices can be used.
estaccumR
function can use parallel processing.
linestack
accepts now expressions as labels. This
allows using mathematical symbols and formula given as
mathematical expressions.
Several vegan functions can now use parallel
processing for slow and repeating calculations. All these
functions have argument parallel
. The argument can be an
integer giving the number of parallel processes. In unix-alikes
(Mac OS, Linux) this will launch "multicore"
processing
and in Windows it will set up "snow"
clusters as desribed
in the documentation of the parallel package. If option
"mc.cores"
is set to an integer > 1, this will be used to
automatically start parallel processing. Finally, the argument
can also be a previously set up "snow"
cluster which will
be used both in Windows and in unix-alikes. Vegan vignette
on Design decision explains the implementation (use
vegandocs("decission")
, and parallel package has more
extensive documentation on parallel processing in R.
The following function use parallel processing in analysing
permutation statistics: adonis
, anosim
,
anova.cca
(and permutest.cca
), mantel
(and
mantel.partial
), mrpp
, ordiareatest
,
permutest.betadisper
and simper
. In addition,
bioenv
can compare several candidate sets of models in
paralle, metaMDS
can launch several random starts in
parallel, and oecosimu
can evaluate test statistics for
several null models in parallel.
All permutation tests are based on the permute package
which offers strong tools for restricted permutation. All these
functions have argument permutations
. The default usage of
simple non-restricted permutations is achieved by giving a single
integer number. Restricted permutations can be defined using the
how
function of the permute package. Finally, the
argument can be a permutation matrix where rows define
permutations. It is possible to use external or user constructed
permutations.
See help(permutations)
for a brief introduction on
permutations in vegan, and permute package for the
full documention. The vignette of the permute package can
be read from vegan with command
vegandocs("permutations")
.
The following functions use the permute package:
CCorA
, adonis
, anosim
, anova.cca
(plus
associated permutest.cca
, add1.cca
,
drop1.cca
, ordistep
, ordiR2step
),
envfit
(plus associated factorfit
and
vectorfit
), mantel
(and mantel.partial
),
mrpp
, mso
, ordiareatest
,
permutest.betadisper
, protest
and simper
.
Community null model generation has been completely
redesigned and rewritten. The communities are constructed with
new nullmodel
function and defined in a low level
commsim
function. The actual null models are generated
with a simulate
function that builds an array of null
models. The new null models include a wide array of quantitative
models in addition to the old binary models, and users can plug
in their own generating functions. The basic tool invoking and
analysing null models is oecosimu
. The null models are
often used only for the analysis of nestedness, but the
implementation in oecosimu
allows analysing any
statistic, and null models are better seen as an alternative to
permutation tests.
vegan package dependencies and namespace imports were adapted to changes in R, and no more trigger warnings and notes in package tests.
Three-dimensional ordination graphics using scatterplot3d for static plots and rgl for dynamic plots were removed from vegan and moved to a companion package vegan3d. The package is available in CRAN.
Function dispweight
implements dispersion weighting
of Clarke et al. (Marine Ecology Progress Series, 320,
11–27). In addition, we implemented a new method for
generalized dispersion weighting gdispweight
. Both
methods downweight species that are significantly
over-dispersed.
New hclust
support functions reorder
,
rev
and scores
. Functions reorder
and
rev
are similar as these functions for dendrogram
objects in base R. However, reorder
can use (and defaults
to) weighted mean. In weighted mean the node average is always the
mean of member leaves, whereas the dendrogram
uses always
unweighted means of joined branches.
Function ordiareatest
supplements ordihull
and
ordiellipse
and provides a randomization test for the
one-sided alternative hypothesis that convex hulls or ellipses in
two-dimensional ordination space have smaller areas than with
randomized groups.
Function permustats
extracts and inspects permutation
results with support functions summary
, density
,
densityplot
, qqnorm
and qqmath
. The
density
and qqnorm
are standard R tools that only
work with one statistic, and densityplot
and qqmath
are lattice graphics that work with univariate and
multivariate statistics. The results of following functions can be
extracted: anosim
, adonis
, mantel
(and
mantel.partial
), mrpp
, oecosimu
,
permustest.cca
(but not the corresponding anova
methods), permutest.betadisper
, and protest
.
stressplot
functions display the ordination distances
at given number of dimensions against original distances. The
method functins are similar to stressplot
for
metaMDS
, and always use the inherent distances of each
ordination method. The functions are available for the results
capscale
, cca
, princomp
, prcomp
,
rda
, and wcmdscale
.
cascadeKM
of only one group will be NA
instead
of a random value.
ordiellipse
can handle points exactly on a line,
including only two points (with a warning).
plotting radfit
results for several species failed if
any of the communities had no species or had only one species.
RsquareAdj
for capscale
with negative
eigenvalues will now report NA
instead of using biased
method of rda
results.
simper
failed when a group had only a single member.
anova.cca
functions were re-written to use the
permute package. Old results may not be exactly
reproduced, and models with missing data may fail in several
cases. There is a new option of analysing a sequence of models
against each other.
simulate
functions for cca
and rda
can return several simulations in a nullmodel
compatible
object. The functions can produce simulations with correlated
errors (also for capscale
) in parametric simulation with
Gaussian error.
bioenv
can use Manhattan, Gower and Mahalanobis
distances in addition to the default Euclidean. New helper
function bioenvdist
can extract the dissimilarities
applied in best model or any other model.
metaMDS(..., trace = 2)
will show convergence
information with the default monoMDS
engine.
Function MDSrotate
can rotate a k-dimensional
ordination to k-1 variables. When these variables are
correlated (like usually is the case), the vectors can also be
correlated to previously rotated dimensions, but will be
uncorrelated to all later ones.
vegan 2.0-10 changed the weighted nestednodf
so that weighted analysis of binary data was equivalent to
binary analysis. However, this broke the equivalence to the
original method. Now the function has an argument wbinary
to select the method of analysis. The problem was reported and a
fix submitted by Vanderlei Debastiani (Universidade Federal do
Rio Grande do Sul, Brasil).
ordiellipse
, ordihull
and ordiellipse
can handle missing values in groups
.
ordispider
can now use spatial medians instead of
means.
rankindex
can use Manhattan, Gower and Mahalanobis
distance in addition to the default Euclidean.
User can set colours and line types in function
rarecurve
for plotting rarefaction curves.
spantree
gained a support function as.hclust
to change the minimum spanning tree into an hclust
tree.
fitspecaccum
can do weighted analysis. Gained
lines
method.
Functions for extrapolated number of species or for the size of species pool using Chao method were modified following Chiu et al., Biometrics 70, 671–682 (2014).
Incidence based specpool
can now use (and defaults to)
small sample correction with number of sites as the sample
size. Function uses basic Chao extrapolation based on the ratio of
singletons and doubletons, but switches now to bias corrected Chao
extrapolation if there are no doubletons (species found
twice). The variance formula for bias corrected Chao was derived
following the supporting
online material
and differs slightly from Chiu et al. (2014).
The poolaccum
function was changed similarly, but the small
sample correction is used always.
The abundance based estimateR
uses bias corrected Chao
extrapolation, but earlier it estimated its variance with classic
Chao model. Now we use the widespread
approximate
equation for variance.
With these changes these functions are more similar to EstimateS.
tabasco
uses now reorder.hclust
for
hclust
object for better ordering than previously when it
cast trees to dendrogram
objects.
treedive
and treedist
default now to
match.force = TRUE
and can be silenced with
verbose = FALSE
.
vegdist
gained Mahalanobis distance.
Nomenclature updated in plant community data with the help
of Taxonstand and taxize packages. The taxonomy of
the dune
data was adapted to the same sources and APG
III. varespec
and dune
use 8-character names (4
from genus + 4 from species epithet). New data set on
phylogenetic distances for dune
was extracted from Zanne
et al. (Nature 506, 89–92; 2014).
User configurable plots for rarecurve
.
strata
are deprecated in permutations. It is still
accepted but will be phased out in next releases. Use how
of permute package.
cca
, rda
and capscale
do not return
scores scaled by eigenvalues: use scores
function to
extract scaled results.
commsimulator
is deprecated. Replace
commsimulator(x, method)
with
simulate(nullmodel(x, method))
.
density
and densityplot
for permutation
results are deprecated: use permustats
with its
density
and densityplot
method.
This version is adapted to the changes in permute
package version 0.8-0 and no more triggers NOTEs in package
checks. This release may be the last of the 2.0 series, and the
next vegan release is scheduled to be a major release with
newly designed oecosimu
and community pattern simulation,
support for parallel processing, and full support of the
permute package. If you are interested in these
developments, you may try the development versions of
vegan in
R-Forge or
GitHub and report the
problems and user experience to us.
envfit
function assumed that all external variables
were either numeric or factors, and failed if they were, say,
character strings. Now only numeric variables are taken as
continuous vectors, and all other variables (character strings,
logical) are coerced to factors if possible. The function also
should work with degenerate data, like only one level of a
factor or a constant value of a continuous environmental
variable. The ties were wrongly in assessing permutation
P-values in vectorfit
.
nestednodf
with quantitative data was not
consistent with binary models, and the fill was wrongly
calculated with quantitative data.
oecosimu
now correctly adapts displayed quantiles
of simulated values to the alternative
test direction.
renyiaccum
plotting failed if only one level of
diversity scale
was used.
The Kempton and Taylor algorithm was found unreliable in
fisherfit
and fisher.alpha
, and now the estimation
of Fisher alpha is only based on the number of
species and the number of individuals. The estimation of
standard errors and profile confidence intervals also had to be
scrapped.
renyiaccum
, specaccum
and
tsallisaccum
functions gained subset
argument.
renyiaccum
can now add a collector
curve to
to the analysis. The collector curve is the diversity
accumulation in the order of the sampling units. With an
interesting ordering or sampling units this allows comparing
actual species accumulations with the expected randomized
accumulation.
specaccum
can now perform weighted accumulation
using the sampling effort as weights.
This version is released due to changes in programming interface and testing procedures in R 3.0.2. If you are using an older version of R, there is no need to upgrade vegan. There are no new features nor bug fixes. The only user-visible changes are in documentation and in output messages and formatting. Because of R changes, this version is dependent on R version 2.14.0 or newer and on lattice package.
This is a maintenance release that fixes some issues raised by changed in R toolset for processing vignettes. In the same we also fix some typographic issues in the vignettes.
ordisurf
gained new arguments for more flexible
definition of fitted models to better utilize the
mgcv::gam
function.
The linewidth of contours can
now be set with the argument lwd
.
Labels to arrows are positioned in a better way in
plot
functions for the results of envfit
,
cca
, rda
and capscale
. The labels should no
longer overlap the arrow tips.
The setting test direction is clearer in oecosimu
.
ordipointlabel
gained a plot
method that can
be used to replot the saved result.
tabasco()
is a new function for graphical display
of community data matrix. Technically it is an interface to R
heatmap
, but its use is closer to vegan function
vegemite
. The function can reorder the community data
matrix similarly as vegemite
, for instance, by ordination
results. Unlike heatmap
, it only displays dendrograms if
supplied by the user, and it defaults to re-order the
dendrograms by correspondence analysis. Species are ordered to
match site ordering or like determined by the user.
Function fitspecaccum(..., model = "asymp")
fitted
logistic model instead of asymptotic model (or the same as
model = "logis"
).
nestedtemp()
failed with very sparse data (fill
< 0.38%).
The plot
function for constrained ordination
results (cca
, rda
, capscale
) gained
argument axis.bp
(defaults TRUE
) which can be used
to suppress axis scale for biplot arrays.
Number of iterations in nonmetric multidimensional scaling
(NMDS) can be set with keyword maxit
(defaults
200
) in metaMDS
.
The result objects of cca
, rda
and
capscale
will no longer have scores u.eig
,
v.eig
and wa.eig
in the future versions of
vegan. This change does not influence normal usage,
because vegan functions do not need these items. However,
external scripts and packages may need changes in the future
versions of vegan.
The species scores were scaled wrongly in
capscale()
. They were scaled correctly only when Euclidean
distances were used, but usually capscale()
is used with
non-Euclidean distances. Most graphics will change and should be
redone. The change of scaling mainly influences the spread of
species scores with respect to the site scores.
Function clamtest()
failed to set the minimum
abundance threshold in some cases. In addition, the output was
wrong when some of the possible species groups were missing. Both
problems were reported by Richard Telford (Bergen, Norway).
Plotting an object fitted by envfit()
would fail if
p.max
was used and there were unused levels for one or
more factors. The unused levels could result from deletion of
observations with missing values or simply as the result of
supplying a subset of a larger data set to envfit()
.
multipart()
printed wrong information about the
analysis type (but did the analysis correctly). Reported by
Valerie Coudrain.
oecosimu()
failed if its nestedfun
returned a
data frame. A more fundamental fix will be in vegan 2.2-0,
where the structure of the oecosimu()
result will change.
The plot of two-dimensional procrustes()
solutions
often draw original axes in a wrong angle. The problem was
reported by Elizabeth Ottesen (MIT).
Function treedive()
for functional or phylogenetic
diversity did not correctly match the species names between the
community data and species tree when the tree contained species
that did not occur in the data. Related function
treedist()
for phylogenetic distances did not try to match
the names at all.
The output of capscale()
displays the value of the
additive constant when argument add = TRUE
was used.
fitted()
functions for cca()
, rda()
and
capscale()
can now return conditioned (partial) component
of the response: Argument model
gained a new alternative
model = "pCCA"
.
dispindmorisita()
output gained a new column for
Chi-squared based probabilities that the null hypothesis (random
distribution) is true.
metaMDS()
and monoMDS()
have new default
convergence criteria. Most importantly, scale factor of the
gradient (sfgrmin
) is stricter. The former limit was too
slack with large data sets and iterations stopped early without
getting close to the solution. In addition, scores()
ignore now requests to dimensions beyond those calculated
instead of failing, and scores()
for metaMDS()
results do not drop dimensions.
msoplot()
gained legend
argument for
positioning the legend.
Nestedness function nestednodf()
gained a plot
method.
ordiR2step()
gained new argument R2scope
(defaults TRUE
) which can be used to turn off the criterion
of stopping when the adjusted R-squared of the current
model exceeds that of the scope. This option allows model
building when the scope
would be overdetermined (number of
predictors higher than number of observations).
ordiR2step()
now handles partial redundancy analysis
(pRDA).
orditorp()
gained argument select
to select
the rows or columns of the results to display.
protest()
prints the standardized residual statistic
squared m12 in addition to the squared Procrustes
correlation R-squared. Both were calculated, but only
the latter was displayed.
Permutation tests are much faster in protest()
. Instead
of calling repeatedly procrustes()
, the goodness of fit
statistic is evaluated within the function.
wcmdscale()
gained methods for print
,
plot
etc. of the results. These methods are only used if
the full wcmdscale
result is returned with, e.g., argument
eig = TRUE
. The default is still to return only a matrix of
scores similarly as the standard R function cmdscale()
,
and in that case the new methods are not used.
anova(<cca_object>, ...)
failed with
by = "axis"
and by = "term"
. The bug was reported by
Dr Sven Neulinger (Christian Albrecht University, Kiel, Germany).
radlattice
did not honour argument BIC = TRUE
,
but always displayed AIC.
Most vegan functions with permutation tests have now a
density
method that can be used to find empirical
probability distributions of permutations. There is a new
plot
method for these functions that displays both the
density and the observed statistic. The density
function is
available for adonis
, anosim
, mantel
,
mantel.partial
, mrpp
, permutest.cca
and
procrustes
.
Function adonis
can return several statistics, and it has
now a densityplot
method (based on lattice).
Function oecosimu
already had density
and
densityplot
, but they are now similar to other vegan
methods, and also work with adipart
, hiersimu
and
multipart
.
radfit
functions got a predict
method that
also accepts arguments newdata
and total
for new
ranks and site totals for prediction. The functions can also
interpolate to non-integer “ranks”, and in some models
also extrapolate.
Labels can now be set in the plot
of envfit
results. The labels must be given in the same order that the
function uses internally, and new support function labels
can be used to display the default labels in their correct order.
Mantel tests (functions mantel
and
mantel.partial
) gained argument na.rm
which can be
used to remove missing values. This options should be used with
care: Permutation tests can be biased if the missing values were
originally in matching or fixed positions.
radfit
results can be consistently accessed with
the same methods whether they were a single model for a single
site, all models for a single site or all models for all sites
in the data. All functions now have methods AIC
,
coef
, deviance
, logLik
, fitted
,
predict
and residuals
.
Building of vegan vignettes failed with the latest version of LaTeX (TeXLive 2012).
R versions later than 2.15-1 (including development
version) report warnings and errors when installing and checking
vegan, and you must upgrade vegan to this version.
The warnings concern functions cIndexKM
and
betadisper
, and the error occurs in betadisper
.
These errors and warnings were triggered by internal changes in
R.
adipart
assumed constant gamma diversity in
simulations when assessing the P-value. This could give
biased results if the null model produces variable gamma
diversities and option weights = "prop"
is used. The
default null model ("r2dtable"
) and the default option
(weights = "unif"
) were analysed correctly.
anova(<prc-object>, by = "axis")
and other
by
cases failed due to ‘NAMESPACE’ issues.
clamtest
wrongly used frequencies instead of the
counts when calculating sample coverage. No detectable
differences were produced when rerunning examples from Chazdon
et al. 2011 and vegan help page.
envfit
failed with unused factor levels.
predict
for cca
results with
type = "response"
or type = "working"
failed with
newdata
if the number of rows did not match with the
original data. Now the newdata
is ignored if it has a
wrong number of rows. The number of rows must match because
the results in cca
must be weighted by original row
totals. The problem did not concern rda
or
capscale
results which do not need row weights.
Reported by Glenn De'ath.
Functions for diversity partitioning (adipart
,
hiersimu
and multipart
) have now formula
and default
methods. The formula
method is
identical to the previous functions, but the default
method can take two matrices as input.
Functions adipart
and multipart
can be used for
fast and easy overall partitioning to alpha, beta and gamma
diversities by omitting the argument describing the hierarchy.
The method in betadisper
is biased with small
sample sizes. The effects of the bias are strongest with
unequal sample sizes. A bias adjusted version was developed by
Adrian Stier and Ben Bolker, and can be invoked with argument
bias.adjust
(defaults to FALSE
).
bioenv
accepts dissimilarities (or square matrices
that can be interpreted as dissimilarities) as an alternative to
community data. This allows using other dissimilarities than
those available in vegdist
.
plot
function for envfit
results gained new
argument bg
that can be used to set background colour for
plotted labels.
msoplot
is more configurable, and allows, for
instance, setting y-axis limits.
Hulls and ellipses are now filled using semitransparent
colours in ordihull
and ordiellipse
, and the
user can set the degree of transparency with a new argument
alpha
. The filled shapes are used when these functions
are called with argument draw = "polygon"
. Function
ordihull
puts labels (with argument label = TRUE
)
now in the real polygon centre.
ordiplot3d
returns function envfit.convert
and the projected location of the origin
. Together
these can be used to add envfit
results to existing
ordiplot3d
plots.
Equal aspect ratio cannot be set exactly in ordiplot3d
because underlying core routines do not allow this. Now
ordiplot3d
sets equal axis ranges, and the documents
urge users to verify that the aspect ratio is reasonably equal
and the graph looks like a cube. If the problems cannot be
solved in the future, ordiplot3d
may be removed from
next releases of vegan.
Function ordipointlabel
gained argument to
select
only some of the items for plotting. The
argument can be used only with one set of points.
Added new nestedness functions nestedbetasor
and
nestedbetajac
that implement multiple-site dissimilarity
indices and their decomposition into turnover and nestedness
components following Baselga (Global Ecology and
Biogeography 19, 134–143; 2010).
Added function rarecurve
to draw rarefaction curves
for each row (sampling unit) of the input data, optionally with
lines showing rarefied species richness with given sample size
for each curve.
Added function simper
that implements
“similarity percentages” of Clarke (Australian
Journal of Ecology 18, 117–143; 1993). The method compares
two or more groups and decomposes the average between-group
Bray-Curtis dissimilarity index to contributions by individual
species. The code was developed in
GitHub
by Eduard Szöcs (Uni Landau, Germany).
betadisper()
failed when the groups
was a
factor with empty levels.
Some constrained ordination methods and their support
functions are more robust in border cases (completely aliased
effects, saturated models, user requests for non-existng scores
etc). Concerns capscale
, ordistep
, varpart
,
plot
function for constrained ordination, and
anova(<cca.object>, by = "margin")
.
The scores
function for monoMDS
did not
honour choices
argument and hence dimensions could not be
chosen in plot
.
The default scores
method failed if the number of
requested axes was higher than the ordination object had. This
was reported as an error in ordiplot
in
R-sig-ecology mailing list.
metaMDS
argument noshare = 0
is now
regarded as a numeric threshold that always triggers extended
dissimilarities (stepacross
), instead of being treated
as synonymous with noshare = FALSE
which always
suppresses extended dissimilarities.
Nestedness discrepancy index nesteddisc
gained a
new argument that allows user to set the number of iterations
in optimizing the index.
oecosimu
displays the mean of simulations and
describes alternative hypothesis more clearly in the printed
output.
Implemented adjusted R-squared for partial
RDA. For partial model rda(Y ~ X1 + Condition(X2))
this
is the same as the component [a] = X1|X2
in variance
partition in varpart
and describes the marginal (unique)
effect of constraining term to adjusted R-squared.
Added Cao dissimilarity (CYd) as a new dissimilarity
method in vegdist
following Cao et al., Water
Envir Res 69, 95–106 (1997). The index should be good for
data with high beta diversity and variable sampling
intensity. Thanks to consultation to Yong Cao (Univ Illinois,
USA).
Function capscale
failed if constrained component
had zero rank. This happened most likely in partial models when
the conditions aliased constraints. The problem was observed in
anova(..., by ="margin")
which uses partial models to
analyses the marginal effects, and was reported in an email
message to
R-News
mailing list.
stressplot
and goodness
sometimes failed when
metaMDS
was based on isoMDS
(MASS package)
because metaMDSdist
did not use the same defaults for
step-across (extended) dissimilarities as
metaMDS(..., engine = "isoMDS")
. The change of defaults can
also influence triggering of step-across in
capscale(..., metaMDSdist = TRUE)
.
adonis
contained a minor bug resulting from
incomplete implementation of a speed-up that did not affect the
results. In fixing this bug, a further bug was identified in
transposing the hat matrices. This second bug was only active
following fixing of the first bug. In fixing both bugs, a
speed-up in the internal f.test() function is fully
realised. Reported by Nicholas Lewin-Koh.
ordiarrows
and ordisegments
gained argument
order.by
that gives a variable to sort points within
groups
. Earlier the points were assumed to be in order.
Function ordispider
invisibly returns the
coordinates to which the points were connected. Typically these
are class centroids of each point, but for constrained ordination
with no groups
they are the LC scores.
clamtest
: new function to classify species as
generalists and specialists in two distinct habitats (CLAM test of
Chazdon et al., Ecology 92, 1332–1343; 2011). The test is
based on multinomial distribution of individuals in two habitat
types or sampling units, and it is applicable only to count data
with no over-dispersion.
as.preston
gained plot
and lines
methods, and as.fisher
gained plot
method (which
also can add items to existing plots). These are similar as
plot
and lines
for prestonfit
and
fisherfit
, but display only data without the fitted lines.
raupcrick
: new function to implement Raup-Crick
dissimilarity as a probability of number of co-occurring species
with occurrence probabilities proportional to species
frequencies. Vegan has Raup-Crick index as a choice in
vegdist
, but that uses equal sampling probabilities for
species and analytic equations. The new raupcrick
function uses simulation with oecosimu
. The function
follows Chase et al. (2011) Ecosphere 2:art24
[doi:10.1890/ES10-00117.1],
and was developed with the consultation of Brian Inouye.
Function meandist
could scramble items and give
wrong results, especially when the grouping
was
numerical. The problem was reported by Dr Miguel Alvarez
(Univ. Bonn).
metaMDS
did not reset tries
when a new model
was started with a previous.best
solution from a different
model.
Function permatswap
for community null models using
quantitative swap never swapped items in a 2 by 2
submatrix if all cells were filled.
The result from permutest.cca
could not be
update
d because of a ‘NAMESPACE’ issue.
R 2.14.0 changed so that it does not accept using
sd()
function for matrices (which was the behaviour at
least since R 1.0-0), and several vegan functions were
changed to adapt to this change (rda
, capscale
,
simulate
methods for rda
, cca
and
capscale
). The change in R 2.14.0 does not influence the
results but you probably wish to upgrade vegan to avoid
annoying warnings.
nesteddisc
is slacker and hence faster when trying
to optimize the statistic for tied column frequencies. Tracing
showed that in most cases an improved ordering was found rather
early in tries, and the results are equally good in most cases.
Peter Minchin joins the vegan team.
vegan implements standard R ‘NAMESPACE’. In
general, S3
methods are not exported which means that you
cannot directly use or see contents of functions like
cca.default
, plot.cca
or anova.ccabyterm
. To
use these functions you should rely on R delegation and simply
use cca
and for its result objects use plot
and
anova
without suffix .cca
. To see the contents of
the function you can use :::
, such as
vegan:::cca.default
. This change may break packages,
documents or scripts that rely on non-exported names.
vegan depends on the permute package. This
package provides powerful tools for restricted permutation
schemes. All vegan permutation will gradually move to use
permute, but currently only betadisper
uses the new
feature.
monoMDS
: a new function for non-metric
multidimensional scaling (NMDS). This function replaces
MASS::isoMDS
as the default method in
metaMDS
. Major advantages of monoMDS
are that it
has ‘weak’ (‘primary’) tie treatment which means
that it can split tied observed dissimilarities. ‘Weak’
tie treatment improves ordination of heterogeneous data sets,
because maximum dissimilarities of 1 can be split. In
addition to global NMDS, monoMDS
can perform local and
hybrid NMDS and metric MDS. It can also handle missing and zero
dissimilarities. Moreover, monoMDS
is faster than
previous alternatives. The function uses Fortran
code
written by Peter Minchin.
MDSrotate
a new function to replace
metaMDSrotate
. This function can rotate both metaMDS
and monoMDS
results so that the first axis is parallel to
an environmental vector.
eventstar
finds the minimum of the evenness profile
on the Tsallis entropy, and uses this to find the corresponding
values of diversity, evenness and numbers equivalent following
Mendes et al. (Ecography 31, 450-456; 2008). The code was
contributed by Eduardo Ribeira Cunha and Heloisa Beatriz Antoniazi
Evangelista and adapted to vegan by Peter Solymos.
fitspecaccum
fits non-linear regression models to
the species accumulation results from specaccum
. The
function can use new self-starting species accumulation models
in vegan or other self-starting non-linear regression
models in R. The function can fit Arrhenius, Gleason, Gitay,
Lomolino (in vegan), asymptotic, Gompertz,
Michaelis-Menten, logistic and Weibull (in base R) models. The
function has plot
and predict
methods.
Self-starting non-linear species accumulation models
SSarrhenius
, SSgleason
, SSgitay
and
SSlomolino
. These can be used with fitspecaccum
or
directly in non-linear regression with nls
. These functions
were implemented because they were found good for species-area
models by Dengler (J. Biogeogr. 36, 728-744; 2009).
adonis
, anosim
, meandist
and
mrpp
warn on negative dissimilarities, and
betadisper
refuses to analyse them. All these functions
expect dissimilarities, and giving something else (like
correlations) probably is a user error.
betadisper
uses restricted permutation of the
permute package.
metaMDS
uses monoMDS
as its default ordination
engine. Function gains new argument engine
that can be used
to alternatively select MASS::isoMDS
. The default is not
to use stepacross
with monoMDS
because its
‘weak’ tie treatment can cope with tied maximum
dissimilarities of one. However, stepacross
is the default
with isoMDS
because it cannot handle adequately these tied
maximum dissimilarities.
specaccum
gained predict
method which uses
either linear or spline interpolation for data between observed
points. Extrapolation is possible with spline interpolation, but
may make little sense.
specpool
can handle missing values or empty factor
levels in the grouping factor pool
. Now also checks that
the length of the pool
matches the number of
observations.
metaMDSrotate
was replaced with MDSrotate
that can also handle the results of monoMDS
.
permuted.index2
and other “new” permutation
code was removed in favour of the permute package. This code
was not intended for normal use, but packages depending on that
code in vegan should instead depend on permute.
treeheight
uses much snappier code. The results
should be unchanged.