Wrappers for the unconstrained ordination methods principal components
analysis (PCA), correspondence anslysis (CA), and principal coordinates
analysis (PCO) are now available via pca(), ca(), and pco()
respectively. The underlying methods used are rda(), cca() and dbrda()
respectively. See
#655.
The output from the ordination methods pca(), pco(), ca(),
rda(), cca(), capscale, and dbrda() has changed slightly to
better separate the results from notifications to the user about
issues encountered with the data or the model. Related to changes in
#682.
The constrained ordination functions are now louder at informing users when one or more terms in a model are aliased (linearly dependent) and their effects cannot be estimated. See #682.
cca and rda return centroids for factor levels even when they
are called without formula, for instance, as cca(dune, dune.env).
plot.cca retains default graphical settings also when only one set
of scores was displayed.
ordiplot did not pass character size (cex) to plot.cca. Version
2.7-0 has more extensive changes, but this fixes the immediate issue
#656.
adonis2() now defaults to running an omnibus test of the model
(by = NULL) instead of a sequential test of model terms (by = "terms"). This makes adonis2() more consistent with the default
for related ordination methods. See
#677.
decorana checks now that input data are numeric instead of
confusing error message (see
https://stackoverflow.com/questions/78666646/).
make.cepnames no longer splits names by hyphen: Capsella
bursa-pastoris used to be Capspast but now is Capsburs.
dbrda failed in rare cases when an ordination component had only
negative eigenvalues. Issue
#670.
plot.cca: biplot or regression arrows were not nicely scaled and
drew no arrows when displayed as the only item in graph.
ordipointlabel failed with decorana result. Bounding box for
text could be wrongly estimated with varying values of cex.
vegdist with argument na.rm = TRUE still failed with missing
values. Dissimilarity methods "chisq" (Chi-square distance) and
"mahalanobis" did not implement na.rm = TRUE. Even when missing
values are removed in calculation, dissimilarities may contain NA
depending on the number and pattern of missing values and
dissimilarity method.
decostand standardization method "clr" did not implement
na.rm = TRUE
(issue #661).
Standardization methods "rank" and "rrank" did not retain NA
values but changed them to 0. Original NA values are kept in
decostand, but with na.rm = TRUE they are ignored when
transforming other data values.
metaMDS: half-change scaling failed when maxdist was fixed, but
was not 1.
summary.ordihull (and hence ordiareatest for convex hulls)
failed if input had more than two dimensions.
simulate.rda failed with univariate response.
vegemite returned only the last page of multi-page table in its
(invisible) return object.
do_wcentre (weighted centring) can segfault due to a
protection error. The problem was found in automatic CRAN
checks. do_wcentre is an internal function that is called from
envfit (vectorfit), wcmdscale and varpart (simpleCCA)
Fixes bug #653.vegan depends on R version 4.1.0.
It is possible to build vegan with webR/wasm Fortran compiler. Issue #623.
Permutation tests for CCA were completely redesigned to follow C.J.F ter Braak & D.E. te Beest: Environ Ecol Stat 29, 849–868 (2022) (https://doi.org/10.1007/s10651-022-00545-4). The constraints are now re-weighted for the permuted response data, and in partial model they are also residualized by conditions (partial terms). In vegan (after release 2.4-6) the tests were identical to Canoco, but ter Braak & te Beest demonstrated that the results are biased. In old vegan (release 2.4-2 and earlier) the predictors were re-weighted but not residualized. Re-weighting was sufficient to remove bias with moderate variation of weights, but residualizing of predictors is necessary with strongly varying weights. See discussion in issue #542. The new scheme only concerns CCA which is a weighted method, and RDA and dbRDA permutation is unchanged.
summary of ordination results no longer prints ordination scores
that often are so voluminous that they hide the real summary; see issue
#203. Ordination
scores should be extracted with scores function. This breaks some
CRAN packages that use summary.cca to extract scores. These should
switch to use scores. The maintainers have been contacted and
patch files are suggested to adapt to this change. See
instructions
to fix the packages.
scores function for constrained ordination (CCA, RDA,dbRDA)
default to return all types of scores (display = "all"). Function
can optionally return a single type of scores as a list of one matrix
instead of returning a matrix (new argument droplist).
Constrained ordination objects (cca, rda, dbrda) fitted
without formula interface can have permutation tests (anova) by
"axis" and by "onedf". Models by "terms" and "margin" are only
possible with formula interface.
Permutation tests for constrained ordination objects (cca, rda,
dbrda) with by = "axis" stop permutations of later axis once the
cutoff limit is reached. Earlier cutoff had to be exceeded. The
default is to stop permutations once P-value 1 is reached. The
analysis takes care that P-values of axes are non-decreasing
similarly as in Canoco.
Coefficients of effects in prc models are scaled similarly as they
were scaled in vegan pre 2.5-1. The change was suggested by
Cajo ter Braak.
Handling of negative eigenvalues was changed in the summary of
eigenvals. Negative eigenvalues are given as negative
"explanation", and the accumulated proportions add up over 1 for the
last non-negative eigenvalue, and 1 for the last negative
eigenvalue.
The printed output of capscale shows proportions for real
components only and ignores imaginary dimensions. This is consistent
to summary and other support methods. Issue
#636.
RsquareAdj of capscale is based only on positive eigenvalues,
and imaginary components are ignored.
stressplot.dbrda refuses to handle partial models. Only the first
component of variation can be displayed because dbrda internal
("working") data structures are not additive. For unconstrained
model "CA", for constrained "CCA" and for partial none.
predict for dbrda will return the actual
type = "working". Earlier it returned "lc" scores weighted by
eigenvalues. Both generated same distances and eigenvalues, though.
Parallel processing was inefficiently implemented and could be
slower than non-parallel in permutation tests for constrained
ordination and adonis2.
plot and scores for cca and rda family of methods gave an
error when non-existing axes were requested. Now ignores requests to
axes numbers that are higher than in the result object.
summary of prc ignored extra parameters (such as const).
Over-fitted models with high number of aliased variables caused a
rare failure in adonis2 and permutation tests of constrained
ordination methods (cca, rda, dbrda, capscale) with
arguments by = "margin" or by = "axis". This also concerned
vif.cca and intersetcor. Typically this occurred with high-order
interactions of factor variables. See issues
#452 and
#622
Some methods accept rectangular raw data input as alternative to
distances, but did not pass all arguments to distance
functions. These arguments in vegdist could be binary = TRUE or
pseudocount with Aitchison distance. This concerns dbrda,
capscale and bioenv. See issue
#631
simper gave arbitrary p-values for species that did not occur in
a subset. Now these are given as NA. See
https://stackoverflow.com/questions/77881877/
Rsquare.adj gave arbitrary p-values for over-fitted models with
no residual variation. Now returns NA when R2 cannot
be adjusted. Automatic model building could proceed to such cases,
and this was fixed in ordiR2step which returns R2 = 0
for overfitted cases. The constrained ordination methods issue a
warning if the model has no residual component. See issue
#610
inertcomp(..., display = "sites", proportional = TRUE) gave wrong
values.
adonis is deprecated: use adonis2. There are several CRAN
packages that still use adonis although we have contacted all
their authors in June 2022 and again in April 2024, and printed a
message of forthcoming deprecation since vegan 2.6-2. See issue
#523. See
instructions
to adapt your packages and functions to use adonis2.
orditkplot was moved to CRAN package vegan3d and is deprecated
in vegan. See issue
#585 and
announcement
#632
The use of summary to extract ordination scores is deprecated: you
should use scores to extract scores. This version still allows
extracting scores with summary, but this will fail in next
versions. For summary.cca see
instructions
to change your package.
Support was removed from ancient cca objects (results of cca,
rda, dbrda or capscale) generated before CRAN release 2.5
(2016). If you still have such stray relics, use
newobject <- update(ancientobject) to modernize the result.
as.mcmc.oecosimu and as.mcmc.permat are defunct: use toCoda.
Code of defunct functions was completely removed.
Support of scores for
ggplot2 graphics is
improved and extended for ordination functions. Suitable scores can be
requested with argument tidy = TRUE, and in general all available
types of scores are returned in a data frame with variable score
labelling the type. The option was implemented in default method of
scores and for structured wcmdscale objects, and glitches were
fixed for rda family and decorana. Previously tidy scores were
implemented for cca, rda, dbrda family of methods, metaMDS,
envfit and rarecurve.
adonis2 and anova for constrained ordination results can perform a
sequential test of one-degree-of-freedom effects where multi-level
factors are split to their contrasts. Previously the test was
available only in permutest.
New summary function for varpart for a brief overview. The summary
shows unique and overall contributed variation for each set of
variables. The fractions shared by several sets of variables are
divided equally with all contributing sets following Lai J, Zou Y,
Zhang J, Peres-Neto P (2022) Methods in Ecology and Evolution, 13:
782–788.
decorana estimates orthogonalized eigenvalues and the total inertia
(scaled Chi-square). Orthogonalized eigenvalues can add up to the
total inertia. Together these enabled implementing eigenvals,
bstick and screeplot methods for decorana.
Axis lengths are reported for all decorana methods.
Implemented tolerance method for decorana. This returns the
criterion that was used in rescaling DCA, and can be used to inspect
the success of rescaling: it should be constant 1 over the whole axis.
New toCoda function to transform sequential null model results
from oecosimu to an object that can be analysed with
coda for convergence
and independence as an MCMC model. Function replaces
as.mcmc.oecosimu and as.mcmc.permat.
metaMDS is more informative about finding similar repeated results
with random starts and uses less confusing language when reporting the
results.
Hellinger distance is directly available in vegdist.
vegdist, betadiver and raupcrick set attribute maxdist giving
the numeric value of theoretical maximum of the dissimilarity index.
For many dissimilarities this is 1, but √2 for Chord and
Hellinger distances, for instance. The attribute is NA for open
indices that do not have such a ceiling. betadiver has three
similarity indices and these set maxdist 0.
metaMDS defaults to halfchange scaling when the dissimilarities have
a numeric maxdist attribute, and adapt the threshold to the ceiling
value. For open indices without ceiling, the threshold will be in the
scale of dissimilarities. metaMDS used a simple test to detect index
ceiling 1, but the test is now more robust and can also find other
maximum values. If such inference is made, the function will broadcast
a message of assumed value of the ceiling.
Mountford index in vegdist is now scaled to maximum value log(2).
Earlier Mountford distances were scaled to maximum 1.
hatvalues of constrained ordination objects can sometimes be
practically 1 or above 1, but now these cases will be exactly 1. In
those cases rstandard, rstudent and cooks.distance will be
NaN. The behaviour is similar as in stats::lm.influence functions.
as.rad can handle multi-row data frames or matrices and return a
list of Rank-Abundance data for each row. Earlier only one site was
handled.
decostand returns attribute parameters of settings and variables
used in standardization. New function decobackstand can use
parameters to reconstruct original non-standardized data.
Back-transformation is not exact but has round-off errors, although
there is an attempt to keep original zeros exact. Back-transformation
is not possible for methods pa, rank and rrank and it is not
implemented for alr. Back-transformation queried in
https://stackoverflow.com/questions/73263526/
Rarefaction and rarefaction-based methods make sense only with
original observed counts and give misleading results if data are
multiplied or rare species are removed. Observed counts usually have
singletons (species with count one), and these method issue a
warning if minimum count is higher than one (which may be a false
positive, but inspect your data). Concerns functions rarefy,
drarefy, rrarefy, rarecurve,
specaccum(..., method="rarefy"), rareslope and avgdist.
See github
discussion #537.
avgdist exposes as.dist arguments and can return "dist"ance
objects that appear as lower triangles instead of appearing as
symmetric matrices.
decorana returned wrong results when Hill's piecewise transformation
(arguments before/after) were used, unless downweighting was also
used.
scores failed when metaMDS result had no species scores. Bug was
introduced in release 2.6-2. Issue raised in
https://stackoverflow.com/questions/72483924/
tolerance.cca failed when only one axis (choice) was requested.
decostand(..., method="alr") did not accept name as a reference,
and could fail in some cases.
CRAN package proxy interfered with simper and caused an
obscure error (github issue
#528).
adonis is on way to deprecation. Use adonis2 instead.
as.mcmc.oecosimu and as.mcmc.permat were deprecated: these could
not be used as S3 methods without depending on coda package. Use
toCoda instead.
Compiled code is adapted to the changes in R 4.2.0. See issues #447, #507.
Cross-references to function in other packages were adapted to more stringent tests in CRAN
Aitchison and robust Aitchison distances were added to vegdist.
Similar data transformations were added to decostand.
Several functions can return “tidy” data structures that can be used
in ggplot2 graphics: rarecurve, scores functions for constrained
ordination (cca etc.), decorana, envfit, metaMDS.
scores.envfit gained argument arrow.mul. vegan plot functions
used this automatically, but now it is easier to use envfit in
non-vegan plotting.
Added function simpson.unb for unbiased Simpson diversity that is
more robust to the variation in sample sizes.
diversity gained argument group to calculate indices for pooled
data. Discussed in issue
#393.
simper is much faster even though parallel processing is not
implemented in the new code.
pairs function was added to plot permustats variables against each
other.
varpart accepts dissimilarities given as a symmetric square matrix
instead of "dist" object per wish of issue
#497.
metaMDS adopted a more user-friendly policy, and trymax will
always be the maximum number of tries. See dicussion in
https://stackoverflow.com/questions/66748605/.
adonis2 accepts strata. adonis2 is the new main function that
replaces old adonis. See issue
#427.
Fisher alpha (fisherfit) is badly suited for extreme communities
that do not follow Fisher's model. Now fisherfit returns NA to
communities that have 0 or 1 species, and issues a warning with
communities consisting of singletons and having extreme Fisher alpha.
adipart and multipart formulae will automatically add unique id
and and constant. This will always sandwich the requested grouping
between alpha and gamma diversities, but not change the results for
requested groupings.
anova function failed in marginal tests when constrained partial
ordination model (cca, rda etc.) had interaction terms. Issue
#463.
Constrained ordination (cca etc.) gave misleading results when all
external variables (constraints, condition) were constant and
explained nothing.
decorana could fail when some axes had zero eigenvalues. Issue
#401.
Species accumulation (specaccum) failed when there was only one
species, but several “communities”. Issue
#501.
Parallel processing failed in Windows or with socket clusters in
permutest of betadisper. Issue
#369.
orditorp failed if numeric labels were supplied. Reported in
https://stackoverflow.com/questions/69272366/.
Argument summarize was accidentally dropped from goodness.cca in
2017.
taxa2dist failed if there was only one usable taxonomic level. See
https://stackoverflow.com/questions/67231431/.
Function adonis2 will replace adonis.
humpfit functions are defunct and removed. They are available in
non-CRAN package natto at https://github.com/jarioksa/natto.
commisimulator is defunct. Use simulate for nullmodel objects.
permuted.index is finally defunct (it was deprecated in vegan
2.2-0).
as.mlm is defunct. Use functions documented with influence.cca,
such as hatvalues.cca, rstandard.cca, rstudent.cca,
cooks.distance.cca and others.
Several distance-based functions failed if all distances were zero
(betadisper, capscale, isomap, monoMDS, pcnm, wcmdscale).
Reported in github issue
#372.
Non-linear self-starting regression models SSarrhenius, SSgitay,
SSgleason and SSlomolino failed in future R. The failure was
caused by internal changes in R-devel. Github issue
#382.
Arrow labels were in wrong position in plot.envfit(..., add = FALSE).
rarecurve added unnecessary names to the results. Github issue
#352.
permutest for betadisper failed in parallel processing in Windows
and in other systems when socket clusters were used. Github issue
#369.
Chi-square and Chord distances were added to vegdist. Both of these
distances can be calculated as Euclidean distances of transformed
data, and actually were available earlier, but many users did not
notice this.
monoMDS (and hence metaMDS) uses stricter convergence criteria.
This improves possibilities to find stable solutions. However, users
may still need to tweak convergence criteria with their data. See
discussion in Github issue
#354.
text functions for constrained ordination plots (cca, rda,
dbrda, capscale) accept now expression labels. This allows using
subscripts, superscripts and mathematical expressions. New support
function labels.cca returns the current text labels so that authors
can change the desired ones. See github issue
#374.
vegemite returns invisibly the final formatted table allowing
further processing.
ordiplot passes cex argument to linestack and decorana plots.
vegdist silently accepted missing values (NA) and removed them
from the analysis also with option na.rm = FALSE. The behaviour was
introduced in vegan version 2.5-1. See GitHub issue
#319.
The labels were displaced when the bunch of arrows was not drawn at
the origin of the ordination graph in envfit. See GitHub issue
#315.
Hill scale in coverscale is open-ended and is not limited to percent
data, unlike most traditional cover class scales which are undefined
above 100% cover.
as.rad no longer print the index attribute: the
attribute is still in the object, but printing made the output messy.vegan depends on R 3.4.0 or higher. The next vegan release may increase the dependence to R 3.6.0.
R 3.6.0 improved the method to find random indices for permuting
and sampling data. Vegan relies now on the R functions in its
ecological null models (functions nullmodel, oecosimu, commsim,
permatfull, permatswap and others). Technically this change is
compatible with R 3.4.0 and later, but you can only gain the
benefits of improved code with a current release of R. The null
models may change due to this change, and most certainly they change
in R 3.6.0. See NEWS for the R 3.6.0 release and
discussion in github issue
#312.
Most vegan permutation routines rely on permute, and there you gain similar benefits of improved randomness when you upgrade R.
Thanks to the new R dependence, sigma for constrained ordination
results works without workarounds of vegan 2.5-2. This fixes
completely the issue discussed in
#274.
Vegan test results cannot be reproduced in older versions than R 3.6.0. If you are worried about this, you should upgrade R.
metaMDS failed in scaling results when other engine than monoMDS
was used. However, we recommend you use monoMDS. See github issue
#310.betadisper changed interpretation of negative squared distances
which give complex-valued distances. Now they are regarded as
zero-distances whereas earlier we used their modulus. This will change
the results in cases where you had negative squared distances. For
further discussion, see github issue
#306.The code for interpreting formula will change in R 3.6.0, and
this makes constrained ordination methods (cca, rda, dbrda,
capscale) to fail. See github issue
#299.
R 3.6.0 introduces a new environment variable
_R_CHECK_LENGTH_1_LOGIC2_, and several functions fail if this
variable is set. Changes concern ordiplot, plot and summary for
constrained ordination objects, and ordixyplot. See github issue
#305.
decorana gave incorrect results when downweighting was used
(argument iweigh = 1). The bug was introduced in vegan 2.5-1 and
reported as github issue
#303.
goodness for constrained ordination methods failed when the
constraints had rank = 1 (only one constraining variable). Reported by
Pierre Legendre.
rda and
dbrda) and partial CCA models (function cca) in function
RsquareAdj. The feature was disabled in vegan 2.5-1 for both. For
RDA, the calculation is similar as in vegan 2.4-6 and earlier.
Partial CCA is now consistent with RDA and differs from the earlier
implementation. For both methods, the partial models are consistent
with varpart. See github issue
#295.Constrained ordination gave misleading results when some constraints
or conditions had data with NULL variables. This rarely happens in
normal usage, but could happen in marginal anova as reported in
github issue #291.
Several functions for numerical analysis wrongly accepted non-numeric
data (for instance, factors) and gave either meaningless results or
confusing error messages. Fixed functions include beals,
designdist, diversity, gdispweight, indpower, spantree,
specpool, tsallis, tsallisaccum and vegdist. See github issue
#292.
envfit with vectors could fail with missing data.
The original data were not scaled and centred similarly as simulations
in simulate.rda when several simulations were returned as a simmat
object (which is compatible with nullmodel simulations and can be
used in oecosimu).
anosim checks its input to avoid confusing error messages like that
reported in StackOverflow question
52082743.
Broken-stick distribution (function bstick) is no longer calculated
for distance-based Redundancy Analysis (dbrda) with negative
eigenvalues, because it is not clear how this should be done. Now
dbrda and capscale are similar with this respect.
print function for betadisper results gained new argument neigen
to select the number of eigenvalues shown. The print is more robust
when the number of eigenvalues is lower than the requested neigen.
humpfit was moved to the natto package and is still
available from https://github.com/jarioksa/natto. It is scheduled
for complete removal in vegan 2.6-0.Vegan declares dependence on R version 3.2.0. This dependence
was not yet noticed in the previous vegan release. However, the
generic sigma function was only defined in R-3.3.0, and
therefore sigma.cca of vegan must be spelt out completely when using
R-3.2.x. See discussion in issue
#274.
CRAN package klaR has
function rda, and when loaded together with vegan this clashes with
vegan rda for Redundancy Analysis. Vegan tries to mitigate the
problem. In most cases vegan functions will be used if vegan was
loaded after klaR, and an error message is issued if klaR objects are
handled with vegan functions. klaR is also tricked to print an
informative message if it handles vegan objects. However, vegan
namespace can be attached automatically at the start-up and then klaR
functions will take precedence. This was reported as issue
#277.
Bioconductor package phyloseq has a problem with vegdist function
for dissimilarities. The problem can be fixed by re-installing
phyloseq from its source package. If you cannot do this, you must
either downgrade to vegan version 2.4-6 or wait till Bioconductor
binary packages are upgraded. This was reported in
Stackoverflow, and as
vegan issue #272,
and as phyloseq issues
#918 and
#921.
Plotting betadisper failed if any of the groups had only one
member. Reported in Stackoverflow as “Error: Incorrect no.of
dimensions” when plotting multivariate data in
Vegan.
Permutation tests for constrained ordination (anova.cca,
permutest.cca) could fail in parallel processing with socket
clusters. Socket clusters are always used in Windows and they can also
be used in other operating systems when created with makeCluster.
See issue #276.
This is a major new release with changes all over the package: Nearly 40% of program files were changed from the previous release. Please report regressions and other issues in https://github.com/vegandevs/vegan/issues/.
Compiled code is used much more extensively, and most compiled
functions use .Call interface. This gives smaller memory footprint
and is also faster. In wall clock time, the greatest gains are in
permutation tests for constrained ordination methods (anova.cca) and
binary null models (nullmodel).
Constrained ordination functions (cca, rda, dbrda, capscale)
are completely rewritten and share most of their code. This makes them
more consistent with each other and more robust. The internal
structure changed in constrained ordination objects, and scripts may
fail if they try to access the result object directly. There never was
a guarantee for unchanged internal structure, and such scripts should
be changed and they should use the provided support functions to
access the result object (see documentation of cca.object and github
issue #262). Some
support and analysis functions may no longer work with result objects
created in previous vegan versions. You should use
update(old.result.object) to fix these old result objects. See
github issues #218,
#227.
vegan includes some tests that are run when checking the package installation. See github issues #181, #271.
The informative messages (warnings, notes and error messages) are cleaned and unified which also makes possible to provide translations.
avgdist: new function to find averaged dissimilarities from several
random rarefactions of communities. Code by Geoffrey Hannigan. See
github issues #242,
#243,
#246.
chaodist: new function that is similar to designdist, but uses
Chao terms that are supposed to take into account the effects of
unseen species (Chao et al., Ecology Letters 8, 148-159; 2005).
Earlier we had Jaccard-type Chao dissimilarity in vegdist, but the
new code allows defining any kind of Chao dissimilarity.
New functions to find influence statistics of constrained ordination
objects: hatvalues, sigma, rstandard, rstudent,
cooks.distance, SSD, vcov, df.residual. Some of these could be
earlier found via as.mlm function which is deprecated. See github
issue #234.
boxplot was added for permustats results to display the
(standardized) effect sizes.
sppscores: new function to add or replace species scores in
distance-based ordination such as dbrda, capscale and metaMDS.
Earlier dbrda did not have species scores, and species scores in
capscale and metaMDS were based on raw input data which may not be
consistent with the used dissimilarity measure. See github issue
#254.
cutreeord: new function that is similar to stats::cutree, but
numbers the cluster in the order they appear in the dendrogram (left
to right) instead of labelling them in the order they appeared in the
data.
sipoo.map: a new data set of locations and sizes of the islands in
the Sipoo archipelago bird data set sipoo.
The inertia of Correspondence Analysis (cca) is called “scaled
Chi-square” instead of using a name of a little known statistic.
If elements for Constraints and Conditions are data frames in
non-formula call of rda or cca, these are automatically expanded
to model matrices and can contain factor variables. Earlier they had
to be numerical model matrices and factors could only be used with the
formula interface.
Regression scores for constraints can be extracted and plotted for constrained ordination methods. See github issue #226.
Full model (model = "full") is again enabled in permutations tests
for constrained ordination results in anova.cca and permutest.cca.
permutest.cca gained a new option by = "onedf" to perform tests by
sequential one degree-of-freedom contrasts of factors. This option is
not (yet) enabled in anova.cca.
The permutation tests are more robust, and most scoping issues should have been fixed.
Permutation tests use compiled C code and they are much faster. See github issue #211.
permutest printed layout is similar to anova.cca.
eigenvals gained a new argument (model) to select either
constrained or unconstrained scores. The old argument (constrained)
is deprecated. See github issue
#207.
summary.eigenvals returns a matrix instead of a list containing only
that matrix.
Adjusted R2 is not calculated for partial ordination, because it is
unclear how this should be done (function RsquareAdj).
ordiresids can display standardized and studentized residuals.
Function to construct model.frame and model.matrix for constrained
ordination are more robust and fail in fewer cases.
goodness and inertcomp for constrained ordination result object no
longer has an option to find distances: only explained variation is
available.
inertcomp gained argument unity. This will give “local
contributions to beta-diversity” (LCBD) and “species contribution to
beta-diversity” (SCBD) of Legendre & De Cáceres (Ecology Letters
16, 951-963; 2012).
goodness is disabled for capscale.
prc gained argument const for general scaling of results similarly
as in rda.
prc uses regression scores for Canoco-compatibility.
The C code for swap-based binary null models was made more efficients, and the models are all faster. Many of these models selected a 2 times 2 submatrix, and for this they generated four random numbers (two rows, two columns). Now we skip selecting third or fourth random number if it is obvious that the matrix cannot be swapped. Since most of time was used in generating random numbers in these functions, and most candidates were rejected, this speeds up functions. However, this also means that random number sequences change from previous vegan versions, and old binary model results cannot be replicated exactly. See github issues #197, #255 for details and timing.
Ecological null models (nullmodel, simulate, make.commsim,
oecosimu) gained new null model "greedyqswap" which can radically
speed up quasi-swap models with minimal risk of introducing bias.
Backtracking is written in C and it is much faster. However, backtracking models are biased, and they are provided only because they are classic legacy models.
adonis2 gained a column of R2 similarly as old adonis.
Great part of R code for decorana is written in C which makes it
faster and reduces the memory footprint.
metaMDS results gained new points and text methods.
ordiplot and other ordination plot functions can be chained with
their points and text functions allowing the use of
magrittr pipes. The
points and text functions gained argument to draw arrows allowing
their use in drawing biplots or adding vectors of environmental
variables with ordiplot. Since many ordination plot methods return
an invisible "ordiplot" object, these points and text methods
also work with them. See github issue
#257.
Lattice graphics (ordixyplot) for ordination can add polygons that
enclose all points in the panel and complete data.
ordicluster gained option to suppress drawing in plots so that it
can be more easily embedded in other functions for calculations.
as.rad returns the index of included taxa as an attribute.
Random rarefaction (function rrarefy) uses compiled C code and is
much faster.
plot of specaccum can draw short horizontal bars to vertical error
bars. See StackOverflow question
45378751.
decostand gained new standardization methods rank and rrank
which replace abundance values by their ranks or relative ranks. See
github issue #225.
Clark dissimilarity was added to vegdist (this cannot be calculated
with designdist).
designdist evaluates minimum terms in compiled code, and the
function is faster than vegdist also for dissimilarities using
minimum terms. Although designdist is usually faster than vegdist,
it is numerically less stable, in particular with large data sets.
swan passes type argument to beals.
tabasco can use traditional cover scale values from function
coverscale. Function coverscale can return scaled values as
integers for numerical analysis instead of returning characters for
printing.
varpart can partition Chi-squared inertia of correspondence
analysis with new argument chisquare. The adjusted R2
is based on permutation tests, and the replicate analysis will have
random variation.
The explanatory tables can be data frames with factors or single
factors in varpart and these will be automatically expanded to model
matrices. Earlier factors could only be used with one-sided model
formulae. Based on the code suggested by Daniel Borcard, Univ.
Montréal.
Very long Condition() statements (> 500 characters) failed in
partial constrained ordination models (cca, rda, dbrda,
capscale). The problem was detected in StackOverflow question
49249816.
Labels were not adjusted when arrows were rescaled in envfit plots.
See StackOverflow question
49259747.
ordiArrowMul failed if there was only one arrow to be plotted in
envfit.
as.mlm function for constrained correspondence analysis is
deprecated in favour of new functions that directly give the influence
statistics. See github issue
#234.
commsimulator is now defunct: use simulate for nullmodel
objects.
ade4 cca objects are no
longer handled in vegan: ade4 has had no cca since version 1.7-8
(August 9, 2017).
read.cep function used FORTRAN format to read legacy CEP and Canoco
files. To avoid NOTEs and WARNINGs, the function was re-written in
R. The new read.cep is less powerful and more fragile, and can
only read data in “condensed” format, and it can fail in several cases
that were successful with the old code. The old FORTRAN-based function
is still available in
cepreader. See github
issue #263. The
cepreader package is developed in
https://github.com/vegandevs/cepreader.rrarefy), species abundance
distribution (preston) and species pool (estimateR) need exact
integer data, but the test allowed small fuzz. The functions worked
correctly with original data, but if data were transformed and then
back-transformed, they would pass the integer test with fuzz and give
wrong results. For instance, sqrt(3)^2 would pass the test as 3, but
was interpreted strictly as integer 2. See github issue
#259.ordiresids uses now weighted residuals for cca results.Several “Swap & Shuffle” null models generated wrong number of initial matrices. Usually they generated too many, which was not dangerous, but it was slow. However, random sequences will change with this fix.
Lattice graphics for ordination (ordixyplot and friends) colour the
arrows by groups instead of randomly mixed colours.
Information on constant or mirrored permutations was omitted when
reporting permutation tests (e.g., in anova for constrained
ordination).
ordistep has improved interpretation of scope: if the lower scope
is missing, the formula of the starting solution is taken as the lower
scope instead of using an empty model. See Stackoverflow question
46985029.
fitspecaccum gained new support functions nobs and logLik which
allow better co-operation with other packages and functions. See
GitHub issue #250.
The “backtracking” null model for community simulation is faster. However, “backtracking” is a biased legacy model that should not be used except in comparative studies.
orditkplot should no longer give warnings in CRAN tests.anova(..., by = "axis") for constrained ordination (cca, rda,
dbrda) ignored partial terms in Condition().
inertcomp and summary.cca failed if the constrained component was
defined, but explained nothing and had zero rank. See StackOverflow:
R - Error message in doing RDA analysis - vegan
package.
Labels are no longer cropped in the meandist plots.
The significance tests for the axes of constrained ordination use now forward testing strategy. More extensive analysis indicated that the previous marginal tests were biased. This is in conflict with Legendre, Oksanen & ter Braak, Methods Ecol Evol 2, 269–277 (2011) who regarded marginal tests as unbiased.
Canberra distance in vegdist can now handle negative input entries
similarly as latest versions of R.
vegan registers native C and Fortran routines. This avoids warnings in model checking, and may also give a small gain in speed.
Future versions of vegan will deprecate and remove elements
pCCA$Fit, CCA$Xbar, and CA$Xbar from cca result objects. This
release provides a new function ordiYbar which is able to construct
these elements both from the current and future releases. Scripts and
functions directly accessing these elements should switch to
ordiYbar for smooth transition.
as.mlm methods for constrained ordination include zero intercept to
give the correct residual degrees of freedom for derived statistics.
biplot method for rda passes correlation argument to the scaling
algorithm.
Biplot scores were wrongly centred in cca which caused a small error
in their values.
Weighting and centring were corrected in intersetcor and spenvcor.
The fix can make a small difference when analysing cca results.
Partial models were not correctly handled in intersetcor.
envfit and ordisurf functions failed when applied to species
scores.
Non-standard variable names can be used within Condition() in
partial ordination. Partial models are used internally within several
functions, and a problem was reported by Albin Meyer (Univ Lorraine,
Metz, France) in ordiR2step when using a variable name that
contained a hyphen (which was wrongly interpreted as a minus sign in
partial ordination).
ordispider did not pass graphical arguments when used to show the
difference of LC and WA scores in constrained ordination.
ordiR2step uses only forward selection to avoid several problems
in model evaluation.
tolerance function could return NaN in some cases when it should
have returned 0. Partial models were not correctly analysed.
Misleading (non-zero) tolerances were sometimes given for species that
occurred only once or sampling units that had only one species.
Permutation tests (permutests, anova) for the first axis failed in
constrained distance-based ordination (dbrda, capscale). Now
capscale will also throw away negative eigenvalues when first
eigenvalues are tested. All permutation tests for the first axis are
now faster. The problem was reported by Cleo Tebby and the fixes are
discussed in GitHub issue
#198 and pull
request #199.
Some support functions for dbrda or capscale gave results or some
of their components in wrong scale. Fixes in stressplot, simulate,
predict and fitted functions.
intersetcor did not use correct weighting for cca and the results
were slightly off.
anova and permutest failed when betadisper was fitted with
argument bias.adjust = TRUE. Fixes Github issue
#219 reported by
Ross Cunning, O'ahu, Hawaii.
ordicluster should return invisibly only the coordinates of internal
points (where clusters or points are joined), but last rows contained
coordinates of external points (ordination scores of points).
The cca method of tolerance was returning incorrect values for all
but the second axis for sample heterogeneities and species tolerances.
See issue #216 for
details.
Biplot scores are scaled similarly as site scores in constrained
ordination methods cca, rda, capscale and dbrda. Earlier they
were unscaled (or more technically, had equal scaling on all axes).
tabasco adds argument to scale the colours by rows or columns in
addition to the old equal scale over the whole plot. New arguments
labRow and labCex can be used to change the column or row labels.
Function also takes care that only above-zero observations are
coloured: earlier tiny observed values were merged to zeros and were
not distinct in the plots.
Sequential null models are somewhat faster (up to 10%). Non-sequential
null models may be marginally faster. These null models are generated
by function nullmodel and also used in oecosimu.
vegdist is much faster. It used to be clearly slower than
stats::dist, but now it is nearly equally fast for the same
dissimilarity measure.
Handling of data= in formula interface is more robust, and messages
on user errors are improved. This fixes points raised in Github issue
#200.
The families and orders in dune.taxon were updated to APG IV (Bot J
Linnean Soc 181, 1–20; 2016) and a corresponding classification
for higher levels (Chase & Reveal, Bot J Linnean Soc 161,
122-127; 2009).
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()).browseVignettes. FAQ-vegan and partitioning were only
accessible with vegandocs function.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.
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),
doi:10.1111/biom.12200, 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 on line
material of
doi:10.1111/biom.12200 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 estimate from EstimateS
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.
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
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 α 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.
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.
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
R2 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 R2. 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 R2 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 R2.
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 2x2 submatrix if all cells were
filled.
The result from permutest.cca could not be updated 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.