An Evaluation of the Relative Robustness of Techniques for Ecological Ordination

Author(s): Peter R. Minchin

Source: Vegetatio, Vol. 69, No. 1/3, Theory and Models in Vegetation Science (Apr. 30, 1987),
pp. 89-107
Published by: Springer
Stable URL: http://www.jstor.org/stable/20038106
Accessed: 03/11/2009 19:26

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=springer.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Springer is collaborating with JSTOR to digitize, preserve and extend access to Vegetatio.

http://www.jstor.org
 Vegetatio 69: 89-107, 1987
- 89
? Dr W. Junk Publishers, Dordrecht Printed in the Netherlands

An evaluation of the relative robustness of techniques for ecological

ordination

Peter R. Minchin*
CSIRO Division of Water and Land Resources, G.P.O. Box 1666, Canberra, 2601, Australia

Keywords: Detrended correspondence analysis, Gaussian ordination, Indirect gradient analysis, Non-metric
multidimensional scaling, Ordination, Principal components analysis, Principal co-ordinates analysis,
Robustness, Simulated data

Abstract

Simulatedvegetation data were used to assess the relative robustness of ordination techniques to variations
in the model of community variation in relation to environment. The methods compared were local non
metric multidimensional scaling correspondence
(LNMDS), analysis (DCA), Gaussian
detrended ordination
(GO), principal components analysis (PCA) and principal co-ordinates analysis (PCoA). Both LNMDS and
PCoA were applied to a matrix of Bray-Curtis coefficients. The results clearly demonstrated the ineffective
ness of the linear techniques (PCA, PCoA), due to curvilinear distortion. Gaussian ordination proved very
sensitive to noise and was not robust
departures from a symmetric, unimodal
to marked response model.
The currently popular method displayed a of DCA
lack of robustness to variations in the response model
and the sampling pattern. Furthermore, DCA ordinations of two-dimensional models often exhibited
marked distortions, even when response surfaces were unimodal and symmetric. LNMDS is recommended
as a robust technique for indirect gradient analysis, which deserves more widespread use by community ecol
ogists.

Introduction methodology, ecologists have sought to identify

those techniques which are most appropriate for
Ordination techniques are commonly employed the purpose of indirect gradient analysis.
as research tools in the study of vegetation. A major Three major approaches to the comparative
objective of ordination in vegetation ecology is that evaluation of ordination techniques may be identi
which Whittaker (1967) termed indirect gradient fied:
analysis. When faced with the diversity of available (1) The application of different ordination
* methods to sets of field data (e.g. Prentice, 1977;
I acknowledge the encouragement of Dr M. P. Austin and Dr
B. M. Potts and the co-operation of the staff of the University
Clymo, 1980; Oksanen, 1983; Brown et al, 1984).
of Tasmania computing centre. I also thank Prof. R. S. Clymo, (2) The comparison of vegetational ordinations
Dr I. C. Prentice and Prof. L. Orl?ci for helpful comments on with direct ordinations based on environmental in
the manuscript. This work formed part of a Ph.D. at the dices
project (e.g. Loucks, 1962; Del Moral, 1980).
Botany Department, University of Tasmania, during which I
(3) The use of simulated data, derived from ex
held an Australian Commonwealth Postgraduate Research
Award. plicit models of community variation along en
Present address: Department of and Geomor
vironmental gradients (e.g. Swan, 1970; Austin,
Biogeography
phology, Research School of Pacific Studies, Australian National 1976; Fasham, 1977; review by Whittaker & Gauch,
University, G.P.O. Box 4, Canberra 2601, Australia. 1978; Prentice, 1980).
90

The first strategy suffers from the major limita of species' response along recognised gradients and
tion that there is no precise statement of the under experimental studies with mixed communities, is
lying gradient structure which a successful ordina insufficient to develop a general model of commu
tion is expected to recover. The ordination results nity change along gradients (Austin, 1980). The as
are assessed on the basis of preconceptions about sumptions which Gauch & Whittaker
(1972a, 1976)
the major environmental relationships derived built into their computer programs for the genera
from previous work. It is seldom possible to make tion of artificial data are still only hypotheses, the

quantitative statements about sample positions on critical testing of which is just beginning (Minchin,
the underlying environmental gradients which are 1983; Austin, 1985, 1987). It follows that the most
sufficiently precise to allow a sensitive comparison one can expect to achieve in simulation studies is
of the performance of different ordination the identification of techniques which are robust to
methods and independent of the biases of the for variation in those features of a vegetation-gradient
mal or informal methods of vegetation analysis model which fall within the bounds of current pos
used in previous work. sibilities (cf. Austin, 1976, 1980, 1985). A robust
The second approach assumes that the axes cho method should be capable of achieving an ade
sen for direct ordination do indeed represent the quate recovery of the underlying gradients over a
major gradients to which
vegetational composition range of model properties. This study assesses the
is related. There is usually no way to assess the va comparative robustness of several ordination
lidity of this assumption. Furthermore, it is often methods.
extremely difficult to identify an environmental
variable which is susceptible to measurement and
which adequately expresses the dynamic complexi Methods
ty of inter-related factors which characterises many
environmental gradients. Models
The third, the simulation approach, has a num
ber of advantages: The artificial data matrices used in this comparison were

result of can be an early version of COENOS, a flexible com

(1) The expected ordination generated using
puter program for the simulation of community variation along
specified precisely, using the known co-ordinates of
environmental gradients (Minchin, 1983, 1987), which is availa
sites in the simulated environment space. ble upon request. The program simulates species responses using
(2) Various model
properties may be varied in a generalised beta-function, which can produce unimodal re

dependently in order to study their effects on ordi sponse surfaces of differing skewness. Interaction may be in

nation success, both alone and in combination. troduced between species, leading to ecological responses which
are shouldered, bimodal or multimodal. COENOS can produce
(3) The degree of stochastic variation (noise) can
models with up to six gradients and there are flexible options for
be controlled. introduction of stochastic variation
sampling patterns and the
The major limitation is that the models em In addition, the variation of community
(noise). along gradients

ployed may not be adequate simplifications of the properties such as alpha diversity and sample total can be con

natural situation (Austin, 1980; Greig-Smith, trolled.

1980).
Since the initial work of Swan (1970), most simula The following model properties were examined in

tion studies have used models which assume that this study:

species responses are Gaussian, that sites are spread (1) The number of underlying gradients (1 or 2)

uniformly throughout environment space and that and their beta diversities or compositional lengths.
such as alpha diversity (2) The shape of species' ecological responses
community properties
and sample total unimodal/bimodal/multimo
(number of species per sample) (symmetric/skewed,
(total abundance of all species in a sample) do not dal).
(3) The arrangement of sites in the simulated en
vary systematically along gradients. All these as
sumptions may be questioned. vironment space (regular/random/clumped/re
The available evidence, from direct observations stricted).
 91

Table 1. The structure of the simulation experiments designed to examine the joint effects of model properties on the performance
of ordination techniques. The beta diversity of gradients is expressed in R units, where a gradient of 1/? is equal in length to the mean
range of species occurrence. Quantitative noise levels are expressed in the F units defined by Gauch & Whittaker (1972a).

Models with a single simulated gradient (coenoclines)

Expt. Model properties held constant Model properties varied Replicates Total no.
no. per cell of models

Al Response curve shape (symmetric) Beta diversity (5 levels: 0.25, 0.5, 1, 2, 4R) 1 30
Sampling pattern (regular) Quantitative noise (3 levels: 0.0, 0.1, 0.2F)
Trend in sample totals (not controlled) Qualitative noise (2 levels: absent, present)

A2 Beta diversity (IR) Response curve shape (6 lev.: symmetric, 18

Sampling pattern (regular) slight consistent skewness, ibid, extreme,
Quantitative noise (0.0F) mixed skewness, interaction between
Qualitative noise (absent) symmetric curves, ibid, skewed curves)
Trend in sample totals (not controlleid)

A3 Beta diversity (IR) Response curve shape (6 lev. as in A2) 36

Sampling pattern (regular) Quantitative noise (3 lev.: 0.0, 0.1, 0.2F)
Trend in sample totals (not controlled) Qualitative noise (2 lev.: absent, present)

A4 Beta diversity (IR) Response curve (4 lev.: 16

shape symmetric,
Quantitative noise (0.0F) extreme consistent skewness, interaction
Qualitative noise (absent) between symmetric curves, ibid, skewed
Trend in sample totals (not controlled) curves)
Sampling pattern (4 lev.: regular, random,
concentrated in centre, concentrated
towards ends)

A5 Beta curve
diversity (IR) Response shape (4 lev. as in A4) 12
Sampling pattern (regular) Trend in sample totals (3 lev.: not con
Quantitative noise (0.0F) trolled, linear trend, parabolic trend)
Qualitative noise (absent)

Models with two simulated gradients (coenoplanes)

Bl surface Beta diversities

Response shape (symmetric) (3 lev.: 0.33x0.337?, 18
Sampling pattern (regular) 1x0.33/?, lxl/?)
Trend in sample totals (not controlled) Quantitative noise (3 lev.: 0.0, 0.1, 0.2F)
Qualitative noise (2 lev.: absent, present)

B2 pattern Beta diversities 1x0.33/?,

Sampling (regular) (2 lev.: lxl/?) 30
Quantitative noise (0.0F) Response surface shape (5 lev.: symmetric,
Qualitative noise (absent) slight consistent skewness, ibid, extreme,
Trend in sample totals (not controlled) mixed skewness, interaction between
skewed curves)

B3 pattern Beta diversities 1 x0.33/?,

Sampling (regular) (2 lev.: 1 X IR) 30
Qualitative noise (absent) Response surface (5 lev. as in B2)
shape
Trend in sample totals noise
(not controlled) Quantitative (3 lev.: 0.0, 0.1, 0.2F)

B4 Beta diversities (1 x IR) surface

Response shape (3 lev.: symmetric, 18
Quantitative noise (0.0F) extreme consistent skewness, interaction
Qualitative noise (absent) between skewed curves)
Trend in sample totals (not controlled) Sampling pattern (5 lev.: regular, random,
concentrated in centre, concentrated
around edges, T-shaped pattern, cross
shaped pattern)
 92

(4) The type (quantitative/qualitative) and the frequency distribution of ranges of occurrence. A normal

amount of noise. distribution was accepted as a working hypothesis on the basis

of preliminary, informal analyses of Gauch & Whittaker
(5) The trend of sample total along gradients
(1972a). The use of a random distribution for species modes ac
(none/linear/parabolic). cords with the results of Minchin (1983). Austin (1987) suggests
A number of experiments were designed in which two or three
that the distribution of modes tends to be clumped, but his anal
of the properties were varied factorially while the others were
ysis was restricted to a single functional guild (canopy trees).
held at a constant value. The structure of the experiments is
summarized in Table 1. Computing limitations restricted the
amount of replication. Since the robustness of ordination tech
Techniques compared
niques to variation in response shape was of major interest,
three replicates per treatment cell were employed in the two ex From the range of available ordination tech
periments on response shape (Table 1:A2, B2).
niques, five were selected for comparative evalua
In all, 174 data matrices were produced, 87 with a single un
tion:
derlying gradient and 87 with two gradients. Some data matrices
took part in more than one (e.g. the noiseless,
experiment IR 1. Principal components analysis (PCA) (Hotel
data set in experiment Al was
also used as one of the symmetric ling, 1933).
response shape models in experiment A2). Each data set was 2. Principal co-ordinates analysis (PCoA) (Gow
subjected to ordination by each of the techniques listed below,
er, 1966).
with the exception that Gaussian ordination (GO) was applied
to the unidimensional
3. Detrended correspondence analysis (DCA)
models only: the program used for GO

produces only one ordination axis. (Hill & Gauch, 1980), using the program
Complete details of model construction are given by Minchin DECORANA (Hill, 1979).
(1983). In all models, the total number of species was adjusted 4. Gaussian ordination (GO), (Gauch et al,
to give an average of about 25 species per sample in the data sets
1974), using Cornell Ecology Program 8B (Gauch,
created without noise. The modal abundances of species were al
located from a lograndom distribution, with limits of 1 to 100 1979).
for models with no interspecific interaction and 5 to 100 for in 5. Local non-metric multidimensional scaling
teraction models. Species' modal positions were randomly dis (LNMDS) (Kruskal, 1964a, b; Sibson, 1972), using
tributed. In the models
without interspecific interaction, those the program KYST (Kruskal et al, unpubl.).
15% of species the largest response
with function integrals had PCA has been shown in previous simulation studies (e.g.,
their modes adjusted to a more even spacing (see Minchin, 1983, &
Noy-Meir & Austin, 1970; Austin & Noy-Meir, 1971; Gauch
1987 for more details). Ranges of occurrence on each gradient
Whittaker, 1972b; Fasham, 1977) to produce distorted represen
were allocated from a normal distribution, with the mean value tations of underlying unless beta-diversity is low. The
gradients
determining the compositional length (beta diversity) of the gra reason for the distortion is that the mathematical model of PCA
dient. Minchin (1987) introduced the '/?' unit to express the beta a linear relationship between
implies compositional dissimilarity
diversity of simulated gradients. A beta diversity value of 1/? in (as expressed by the Euclidean metric calculated from species
dicates a gradient whose length is equal to the mean range of and the separation of sites along environmental
data) gradients.
species occurrence. There is no simple relationship between the In fact the relationship is non-linear: as one moves further apart
R unit and other common measures of beta diversity, such as the in environment the rate of increase in compositional dis
space,
half-change (Whittaker, 1960) or the sd unit (Hill & Gauch, similarity tends to decline (Swan, 1970). In order to fit its linear
1980). However, 1/? is approximately equal to 6 sd and 4.5 half PCA must as curved, rather than lin
model, represent gradients
changes, when response functions are unimodal,
fairly sym ear trends.
metrical and not grossly long-tailed or flat-topped. The stan the recognition that the linear model of PCA is inap
Despite
dard deviation of species' ranges was set at 0.3 times the mean propriate unless beta diversity is very low, the method continues
value for models without interaction and 0.5 times
interspecific to be applied to community data by plant ecologists (e.g., Brad
the mean for interaction models. The differences in the simula et al, 1985) and is appar
field & Scagel, 1984; Van der Maarel
tion parameters for interaction models, relative to those for non with animal & Wiens,
ently popular ecologists (e.g., Rotenberry
interaction models, were determined after initial empirical trials. due to a lack of acceptance of the
1980). This may be partly
The aim was to produce interaction models with a reasonable results of simulation because the Gaussian models em
studies,
number of functions but without was
complex ecological response ployed therein were regarded as unrealistic. Thus PCA in
too many 'extinctions' of species due to the interaction adjust cluded in this study to test the expectation that its linear model
ments. would lead to similar distortions with a range of alternative non
The choice of a lograndom distribution for modal abun linear models.
response
dances is based on analyses reported by Minchin (1983) and The of PCA, as a method of indirect
performance gradient
Gauch & Whittaker (1972a). No good evidence is available about varies to the manner in which the data are
analysis, according
 93

standardized (Austin & Noy-Meir, 1971). In this study, PCA was The 'global' variant of NMDS derives a configuration in
applied with each of three standardizations: (1) centred by spe which the distances between all pairs of sample-points are, as far
cies mean; (2) centred by species and standardized by species as possible, in rank order agreement with their compositional
standard deviation
(equivalent to an R-mode PCA of the corre dissimilarities. Any given pair of samples which are less similar
lation matrix between species) and (3) Bray-Curtis successive in composition than some other pair should be placed further
double standardization (i.e. species adjusted to equal maxima, apart than that other pair in the ordination. The 'local' variant
then samples standardized to equal totals), followed by centring of NMDS (Sibson, 1972) has a more relaxed criterion: for each
by species. The abbreviations PCA-C, PCA-CS and PCA-BC sample, the distances from its point to each other sample-point
will be used below when referring to these three variants of in the ordination should be in rank order with the compositional
PCA. dissimilarities between that sample and each other sample. This
PCoA is effectively a generalization of PCA which allows the variant allows for the possibility that the pattern of decline in
use of a much wider range of measures of compositional dis compositional with increasing environmental
dissimilarity sepa
similarity. In this study it was applied to the Bray-Curtis coeffi ration may differ from point to point in environment space
cient which is also known as percentage difference (Gauch, (Prentice, 1977, 1980).
1982) and the Czekanowski coefficient (see, e.g., Greig-Smith, In this study, NMDS was applied in the 'local' form and DCA
1983). This coefficient has been widely used in ecology. Values ordinations were used to provide starting configurations. The
of the Bray-Curtis coefficient have a less curvilinear relationship Bray-Curtis coefficient was used as a measure of compositional
with environmental (the distance between in thus making the NMDS
separation samples dissimilarity, ordinations directly com
environmental space) than do values of the Euclidean distance parable to the ordinations by PCoA. Previous work with Gaussi
metric 1973; Faith et al, It was therefore an models
(Gauch, 1987). expect (Gauch, 1973) has shown that the Bray-Curtis coeffi
ed that the use of this coefficient might increase the effective cient has an approximately monotonie with
relationship
sample
ness of PCoA relative to PCA. Faith et al (1987) used a range of
separation along gradients.
The remaining three techniques, GO, DCA and LNMDS, have models similar to those in this study to compare the rank corre
been introduced to ecology as potential solutions to the problem lations between and
compositional dissimilarity sample separa
of curvilinear distortion in linear ordinations. On the basis of tion for a variety of dissimilarity coefficients. The Bray-Curtis
simulation studies using Gaussian models (Hill & Gauch, 1980; coefficient was among the most effective and robust of the
Gauch et al, 1981) DCA is now commonly regarded as the 'state measures compared. The coefficient attains a maximum value
of the art' method (Gauch, 1982). It is gaining broad acceptance of 1.0 for all pairs of samples which have no species in common.
among ecologists (e.g. Walker & Peet, 1983; Beatty, 1984; Van Values of 1.0 are indeterminate, in the sense that they indicate
der Maarel et al, 1985). However, the sensitivity of DCA to only a lower bound on the gradient of a sample pair.
separation
departures from the Gaussian model has not been assessed. For this reason, all values of 1.0 were coded as in the
'missing'
The underlying model of NMDS is relatively simple: given a present LNMDS ordinations. This results in the exclusion of
matrix of resemblances (similarities or dissimilarities) between such from the monotonie of distance
sample-pairs regressions
pairs of objects, NMDS constructs a configuration of points in on dissimilarity. Stress formula 1 of Kruskal (1964a) was used
a specified number of dimensions, such that the rank order and the 'primary' approach to tied dissimilarities was adopted
agreement between the distances and the resem
inter-point (Kruskal, 1964a): if several pairs of samples have an identical
blance values is maximized. In ecological the ob
applications compositional dissimilarity, they need not be given equal dis
jects are usually samples and the dissimilarities are calculated tances in the ordination. Model data sets with a single gradient
from the compositional data using some chosen coefficient. The were ordinated in both one and two dimensions, while models
epithet 'non-metric' refers to the fact that only the rank order with two gradients were to LNMDS in two and three
subjected
of the input dissimilarities is utilized. This contrasts with dimensions.
methods of metric scaling (e.g. PCA, PCoA, correspondence Gauch et al several available for
(1981) compared programs
analysis) where the distances between points in the derived con NMDS and
recommended ALSCAL (Young & Lewyckyj, 1979)
figuration are proportional to the dissimilarities. as the fastest and most useful. Unfortunately, ALSCAL does
Gauch (1982; see also Gauch et al, 1981), who stated that not perform NMDS as originally proposed by Kruskal (1964a,
NMDS assumes which is a weaker and better as uses an alternating
'monotonicity, b). ALSCAL least-squares algorithm, which
sumption than linearity but is still unrealistic for handling the maximizes the monotonie fit between ordination dis
squared
Gaussian curve, which is ditonic', confused the model of spe tances and squared dissimilarities.
compositional Consequently,
cies' responses to gradients with the model of the relationship the larger dissimilarities receive relatively higher weight in the
between ordination distance and compositional dissimilarity. process. These between
fitting dissimilarities, samples with few
The monotonicity assumption of NMDS refers to the latter. or no in common, are the least informative.
species ALSCAL
NMDS does not make a direct assumption about the form of ordinations therefore tend to represent local structure poorly.
species response functions. In theory, NMDS can accommodate Comparisons based on simulated data sets Faith &
(Minchin,
any type of response function, provided that the resulting rela Belbin, unpublished) have shown that NMDS ordinations by
tionship between compositional dissimilarity (as expressed by KYST recover the gradient structure much more
consistently
some dissimilarity coefficient) and sample in en than ALSCAL ordinations. All
separation successfully LNMDS ordina
vironment space remains approximately monotonie. tions in this study were performed the program KYST.
using
 94

Assessment of fit el properties. The distortion became more severe as

beta diversity increased. Of the three variants of
the purpose of indirect gradient analysis, a
For PCA examined, PCA-CS was the most resistant to
successful ordination is defined as one in which the distortion and PCA-C was the least successful.
relative positions of samples matches their relative These results are consistent with those of earlier
locations in environment space. Consequently, or studies, based on Gaussian models (e.g. Noy-Meir &
dination performance may be assessed by compar Austin, 1970; Austin & Noy-Meir, 1971; Fasham,
ing ordination configurations with the configura 1977). It is evident that the curvilinear distortion of
tion of samples in the simulated environmental gradients by PCA is not due to peculiar properties
space.
of the Gaussian model.
A quantitative measure of the degree of fit between an ordina PCoA with the Bray-Curtis coefficient per
tion and the environmental configuration was obtained using formed similarly to PCA. In general, the degree of
Procrustean analysis (Sch?nemann & Carroll, 1970; Fasham,
curvilinear distortion in PCoA ordinations was
1977). This technique fits one configuration to another using a
combination of origin translation, rigid rotation and reflection
somewhat greater than in those produced by PCA
of reference axes and uniform central dilation or contraction of CS. For unidimensional models, the sequence of
scaling. The combination of transformations is found analyti sites along the simulated gradient appeared as a
cally, so as to minimize the sum of the squared distances be
curved path in the space defined by the first two
tween each point in the fitted configuration and its correspond
axes of a PCA or PCoA ordination. Even when
ing point in the target configuration. The RMS average of these

may be used as a measure of the discrepancy be sites were regularly spaced along the model gra
displacements
tween the configurations (i.e. lower values indicate better fit). In dient they sometimes appeared bunched in PCA or
the results, the Procrustean values are
presenting discrepancy PCoA ordinations, reflecting the curvature of the
denoted by the symbol D, with a subscript indicating the number
gradient into the third (or higher) dimensions.
of dimensions in which the fit was performed.
With two underlying gradients, the consequences
For LNMDS ordinations performed with one more dimension
than the number of gradients in the model,
analysis Procrustean of curvilinear distortion by PCA and PCoA were
was performed in the higher dimensionality. This was achieved more severe (Fig. 1). The planar pattern of sites in
an extra dimension
by adding to the simulated configuration, environment space was twisted and curved into
upon which each sample was given a score of zero. After fitting three or more the recognition
dimensions, making
the ordination to the environmental configuration embedded in
of the two-gradient structure in the resultant ordi
this space, the RMS
displacement (D) was computed in the sub
nation impossible without prior knowledge of the
space defined by the original model gradient(s).
A limitation of D is that it can not distinguish situations underlying model. If the lines joining sets of sam
where the lack of fit is due to a systematic distortion of the tar with on the second
ple points equal co-ordinates
get configuration from those in which each point has been in
simulated gradient were deleted from Fig. 1, it
dependently perturbed. In the first case, the fit is generally much
would be very difficult to perceive the underlying
better in some regions of the configuration than in others. Be
cause of this limitation, the assessment of ordination perfor gradients in the ordinations. The distortion is not
mance was not restricted to comparison of the D values. All difficult to recognize for unidimensional data sets,
configurations were plotted and examined visually. when an arch or horse-shoe is
shaped configuration
For unidimensional models, the degree of rank-order agree
usually obtained in the first two dimensions. How
ment between site locations along the simulated gradient and
site scores on the first ordination axis was also assessed. Ken
ever, the planar structure of a two-dimensional data
set can be twisted and curved into more than three
dall's rank correlation coefficient (r) was used for this purpose.

dimensions, so that the shape of the configuration

cannot be viewed in any two or three dimensional
Results plot.

PCA and PCoA

Gaussian ordination
Curvilinear distortion the underlying
of gra
dients) was evident in all PCA ordinations, ir The only combination of model properties in

respective of response function shape or other mod which GO consistently produced better results than
 95

(a) PCfl-CS (b) PCofl noise. For those noiseless data sets where GO
achieved lower Procrustean discrepancies than
DCA or LNMDS, the advantage of GO was
reduced and usually reversed when noise was ad
ded. The failures of GO with noisy data sets were
occasionally spectacular (e.g., Table 2, interaction
model with noise level = 0.2 F).

DCA versus LNMDS

In view of the curvilinear distortion in PCA and

PCoA ordinations and the restricted
utility of GO,
major interest centred on the relative performance
of DCA and LNMDS. The balance between these
two techniques depended on model properties, in
particular the number and relative beta diversities
of gradients, response shape and sampling pattern.

Fig. 1. PCA-CS and PCoA ordinations of a 1 x 0.33Z? (a, b) Response function shape
and a 1x \R (c, d) coenoplane. Both models had symmetric, Some results for unidimensional models with a
unimodal response surfaces with no noise. Samples were ar
beta diversity of 1R and different response function
ranged on regular 12 x 4 (a, b) and 7 x 7 (c, d) grids, respec

tively. Configurations on ordination axes 1 v 2 are shown, after shapes are given in Table 2. In the absence of noise,
being fitted to the simulated sampling pattern by Procrustean DCA recovered the rank order of samples perfectly
analysis. The lines join samples with equal co-ordinates on the
provided that all response functions were unimo
second simulated gradient. dal. However, DCA failed in rank order recovery on
two of the interaction coenoclines, which included
some species with
shouldered, bimodal or mul
DCA and LNMDS was for data sets with a beta timodal responses. LNMDS proved more robust to
diversity of 1R or less, fairly symmetric response variation in response shape: when performed in
curves and either no noise or qualitative noise only. one dimension, LNMDS achieved perfect rank or
GO is relatively resistant to qualitative noise be der recovery over all shape categories. The Procrus
cause only non-zero values are used in the fitting of tean discrepancy values follow a similar pattern.
Gaussian regressions, however the technique proved LNMDS had slightly worse Procrustean fits than
highly sensitive to quantitative noise. DCA for most of the symmetrical, mixed skewness
Some performance statistics for GO in experi and extreme skewness models, but consistently
ments A2 and A3 are given in Table 2. In the ab achieved better fits than DCA for the interaction
sence of noise, GO consistently recovered the cor models.

rect rank order of samples only when response The effect of variation in response shape for two
curves were symmetrical or of mixed skewness. The dimensional models is exemplified by the results in
Procrustean discrepancy values in these situations Table 3. For 'rectangular' coenoplanes, with a beta
were generally somewhat lower than those for DCA diversity of 1 x 0.33??, neither DCA nor LNMDS
and LNMDS, indicating a better recovery of the had a consistent advantage when response surfaces
inter-sample spacings. were symmetrical or of mixed skewness. Under
The results in Table 2 clearly demonstrate the these conditions, relative performance in terms of
sensitivity of GO to the addition of quantitative Procrustean fit differed between replicates. Howev
96

O
"O C/3 d O^t^m^-non
cd >
2 '-3 o
o *-<m -h <n ^ <Nd ^t

y ?s o
?/->
O On VO OO <N
d d on d ri- d o?
-a >
C e?
e?X,
0\ 0\ 0\ (S o\ ^ o\
o in d v^ o m o
c 2 'S
.2 S (s o\ o
"^ ^ ?^ q
Pi -H d ?i ? ^ tJ-'d rn

q
o -h N -^ M -< <N -h m

q ? oo On O O <NO
d H ^ o ^ h ri H
-2 C
? 2
o? a

?d OmOV0OTfO?O

?m??o?m?Tt
85 On ?o O oo ?
W Pi -H m O ^t ^ ri -h
<
T3
Cas
on r- ^
q
d -H (N -H -h"q i-i ri q-h' ?O

o ? r- on
q
d

O ?-?O m
?d

ON OOOn ^O
? *-?? en d ri O ri

2"S oc ? <N?
O? -H

?I

? ^<~ 2 8 r~ 8
o ?
?Z? o
q on on no
d ?<"d ri

On r- Os ^i" On r~ ON ?Ai
d fi d ^ d ^t d "X
?d

-h on
m O m o Tf O ^t

-h tj- q m On m
q q
C* - -^' m' ?-h'ri ?* en O Tt-'

"Hcd ??
?
m go
Q Q
^ 9- ? <
2 ? U O
Z?Zg
O s Q ?
 97

er, LNMDS consistently achieved more accurate data sets where DCA achieved better results than
recovery of sample positions for the 1 x 033R LNMDS (Fig. 2d, h), the LNMDS configuration
coenoplanes with extremely skewed response sur generally exhibited some systematic curvilinear dis
faces and for the 1 x 0.33R coenoplanes in which tortion.

interaction between species produced more com Figure 3 shows two examples of the consistently
plex response shapes. superior performance of LNMDS for 1 x 0.33??,
Some ordination
configurations for 1 x 0.337? extreme skewness and interaction models. The
coenoplanes are shown in Figs. 2 and 3. In each DCA solution for the extreme skewness model
case, the samples were located on a regular 12 x 4 (Fig. 3b) once again shows compression at one end
grid in the simulated environment space. Figure 2 of the first gradient. For the interaction model il
illustrates the lack of a consistent advantage for lustrated (Fig. 3c, d), the DCA ordination ismarred
either technique with symmetrical and mixed skew by curvilinear distortion at the left-hand side of the
ness models. When DCA gave worse fits than configuration.
LNMDS (Fig. 2b, f)> it was generally due to a com The unequivocal superiority of LNMDS over
pression of variation along the second gradient DCA for 1 x IR coenoplanes is apparent from the
towards one end of the first gradient. For those results presented in Table 3. Irrespective of re
sponse surface shape, LNMDS always achieved
lower Procrustean discrepancies than DCA, often
(a) LNMDS (b) DCA so. Sortie examples of ordination results
markedly
for 1 x IR coenoplanes are shown in Figs. 4 and 5.
For each of the models illustrated, the samples were
located on a regular 7x7 grid in the simulated en
vironment space.

Figure 4 shows the DCA and LNMDS ordina

(c) LNMDS (d) DCfl
tions of all three replicates with symmetrical,
unimodal response regular grid was
surfaces. The
recovered reasonably by DCA wellfor the first
replicate (Fig. 4b), but the DCA configurations for
the other two replicates (Fig. 4d, f) are distorted. A

(e) LNMDS (f) DCfl peculiar feature of the distortion is the displace
ment of the samples in one corner of the grid to

(a) LNMDS (b) DCfl

(g) LNMDS_(h) DCfl

(c) LNMDS (d) DCfl

Fig. 2. LNMDS and DCA ordinations of some 1x 0.337?

coenoplanes: (a) and (b) symmetric responses, replicate 1 (PCA

CS and PCoA ordinations of this model are shown in Fig. la,

b); (c) and (d) symmetric responses, replicate 2; (e) and (f) mixed Fig. 3. LNMDS and DCA ordinations of some 1x 0.33?
skewness, replicate 3; (g) and (h) mixed skewness, replicate 2. All coenoplanes: (a) and (b) extreme skewness, replicate 1; (c) and
models had no noise added and samples were on a
arranged (d) interaction moxlel, replicate 1. In both cases, samples were
regular 12 x 4 grid. arranged on a regular 12 x 4 grid, and no noise was included.
98

? ON On
? ?r> Tt on r-'

? VO? q On ?
? vo ?ri ? o? o?
ir> >?i m vo
VO ^ o?

Tt ro
vo ?o -h' r-'

.2 ? OO ^H .2 ? vo OO
?ri ?ri -h' V?
0?5

? ?ri *-i ?
? NO ?ri ? ?? ON*

? ? vo On
? o? t> ? ? vo

r- oo r?
o? v?' ?

*?3^
O? s?

.55? Tj- OO
Os ?ri
PJ

?o ? q
vo VO ?

O ? rj r ? OO? VO
? o? On ? ON ON ?

en ~h ?
? On ^ rn

g 2 v? ? ^
h ?
?-* o?

.2 ? -H ? -H .2 ? ? m '-h
r^ o? ^

?
x x

q en r; VO <N
? r~* ?r> ^t ON

rn rj en ?t?
c q
ofe
*~ oCu ? ri r-' On o?
5/3 fi oc
3 ^ X Tt vo
X vo ri ?3 ? ? o?

o 2
C
VO ri-' On vo ON
a o E
-a o e
o o ? en m o 00 ^
2 P?5- Z ? VO Tf 0??- On vo

XSl
Q Q Q Q
< %
O S Q hJ es? o e
 99

(a) LNMDS (b) DCfl (a) LNMDS (b) DCfl

(c) LNMDS (d) DCfl (c) LNMDS (d) DCfl

(e) LNMDS (f) DCfl (e) LNMDS if) DCfl

Fig. 4. LNMDS and DCA ordinations of replicate 1x \R Fig. 5. LNMDS and DCA ordinations of some 1x \R

coenoplanes with symmetric, unimodal response surfaces: (a) coenoplanes: (a) and (b) mixed skewness, replicate 1; (c) and (d)
and (b) replicate 1 (PCA-CS and PCoA ordinations of this mod extreme skewness, replicate 1; (e) and (f) interaction model,
el are shown in Fig. lc, d); (c) and (d) replicate 2; (e) and (f) replicate 1. In each case, samples were arranged on a regular
3. In each case, samples were on a regular 7x7
replicate arranged grid and no noise was included.
7x7 grid and no noise was added.

form a narrow 'tongue' extending from one side of (Fig. 5d) shows a 'tongue' distortion similar to that
the configuration. The LNMDS ordinations observed for some of the symmetrical models.
(Fig. 4a, c, e) all display good recovery of the grid LNMDS corrects this distortion and achieves a
structure and there is no sign of the 'tongue' distor tolerable reconstruction of the simulated grid
tion produced by DCA. (Fig. 5c). The performance of LNMDS on this data
Some examples of DCA and LNMDS ordina set was the worst for all of the regularly sampled,
tions of mixed skewness, extreme skewness and in noiseless, 1 x \R coenoplanes examined.
teraction coenoplanes with beta diversities of LNMDS usually achieved better recovery of gra
1x \R are shown in Fig. 5. In each case the regular dients when performed in the correct dimensionali
7x7 grid is recovered better by LNMDS. The ty. When offered an extra dimension, LNMDS
DCA ordination of the extreme skewness model tended to curve the gradient structure in the higher
 100

dimensional space, although the distortion was the effects of variation pattern on both
in sampling
never as severe as that observed in linear ordina DCA and LNMDS were relatively minor. The rela
tions (PCA, PCoA). Consequently, there was tive performance of the techniques with regularly

generally a deterioration in the accuracy with which sampled coenoclines was maintained with the other
the environmental configuration was recovered on sampling patterns. The results do not agree with
the first one or two axes (see Tables 2, 3, 4). The those of Mohler (1981), who reported better grav
problem was more severe for coenocline models dient recovery by DCA when samples were concen
and coenoplanes in which one gradient was longer trated towards the extremes of the gradient.
than the other. For most of the 1 x IR The results for 1x IR coenoplanes were quite

coenoplanes studied, the first two dimensions of a different,illustrating the danger of assuming that
three-dimensional LNMDS still represented sample phenomena observed for unidimensional models

positions more accurately than the corresponding should generalize to the multidimensional case.
DCA solution (Table 3). Some performance statistics for experiment B4 are
givenin Table 4. DCA displayed more variation in

Quantitative noise performance between sampling patterns than did

Both DCA and LNMDS displayed a considerable LNMDS. LNMDS achieved better recovery of sam

degree of resistance to the addition of quantitative ple configurations than did DCA for all data sets,
noise. The effect of noise was to introduce random with the exception of two models
cross-shapedwith
displacements of points in the ordination configu sampling patterns. DCA performed particularly
rations, without altering the overall form of the poorly when samples were randomly distributed.
configurations derived for the corresponding noise Of special interest are the results for 1 x IR
less data sets (cf. Gauch et al., 1981). The introduc coenoplanes in which samples were confined to a
tion of noise did not alter the relative performance restricted region of the simulated environmental
of DCA and LNMDS, as described in the previous space. Two kinds of restricted sampling pattern
section (see Tables 2, 3). Despite the addition of were studied. In the first, sites were confined to a

noise, the overall structure of the simulated en T-shaped region (Fig. 6a), so that the second gra
vironment space is recovered remarkably well. dient was only expressed towards the lower extreme
of the first gradient. The second arrangement was
cross-shaped, with the second gradient being ex

Qualitative noise pressed only near the centre of the first (Fig. 7a).
The effect of qualitative (presence-absence) noise Neither DCA nor LNMDS gave satisfactory results
on ordination performance was severe, especially under this type of sampling, although LNMDS was
for models in which beta diversity was low. Both generally more successful.

DCA and LNMDS were unable to achieve an ac Figure 6 shows

the DCA and LNMDS ordina
along simu tions of the symmetric response coenoplane sam
ceptable recovery of sample positions
lated gradients with beta diversities of less than \R a
pled by T-shaped pattern. In the DCA configura
in the presence of qualitative noise. (It is probable tion (Fig. 6c), the cross-bar of the T has been bent,
that the levels of qualitative noise applied in this giving the appearance of a long gradient with a
study were unrealistically high: many data sets con shorter side branch. In addition, variation in the
tained samples with fewer than five species, against direction of the second
gradient among samples
a mean of ca. 25 in the noiseless data). forming the stem of the T has been suppressed. The
LNMDS ordination (Fig. 6b) represents the T
Sampling pattern shaped pattern reasonably well, although there is

Experiments A4 and B4 (Table 1) revealed marked some curvilinear distortion of the stem of the T.
differences between DCA and LNMDS in their sen Ordinations by DCA and LNMDS of the data set

sitivity to the pattern of sampling in the simulated derived by cross-shaped sampling of the interaction
environmental space. For unidimensional models, coenoplane are shown in Fig. 7. In the DCA ordi
 101

OO ?/^
O? On

^h r^

OO ?o
u

^ o
? C
cd cd
Oh C*

s ?
S* ??0 OO vo
S
VO O?
GO?

xcd Oi-i O O)
^h O)
*-*
co
H O
? Oa

T3

VO -H ^H

T3
O
00 ^h vo
vo ^*-< ?ri
<N ?-h
00
C Vh
ex -g
s &
W &o P<
6

B S o
?

"H*
O <U

t?;
ON
g Q
? e
o o
a
C
a 3
eu c*

S ri
?1 &o VO
3 1?

q q
&o m
> ? *
fi ?'S Q Q
-Ci ;? -5
fi O 6 Q
*S2S
 102

6 (a) SIMULATED 7 (a) SIMULATED

degree of distortion (Fig. 7b).
The poor recovery of T-shaped and cross-shaped
o io ?ScP models
O icp coenoplane by all techniques included in
this study suggests the need for a greater emphasis
? QO o Q ? ?8*0 o in future studies on the implications for ordination
-&Q-TJ ft ^r o
- io ? OO qO o <??>o^ of the shape of the underlying environmental
o?
space. In the past, it has been implicitly assumed
O i
Q that environmental space is completely (and often
o<9
uniformly) covered by samples. However, it is often
the case in nature that certain combinations of en
6 (b) LNMDS 7 (b) LNMDS vironmental conditions do not occur in a particular
landscape (e.g., Austin et al, 1984).

Trend in sample totals

The effect of systematic in sample totals
variation
was studied for unidimensional models only (Ta
ble 1,A5). Both LNMDS and DCA appeared to be
relatively robust to two-fold linear and parabolic
trends in sample totals. These results are prelimi
nary and more work is required on this question,
6 (c) DCA 7 (c) DCA particularly with multidimensional models.

Conclusions

1. The linear ordination techniques PC A and

PCoA are inappropriate for the purpose of indirect
gradient analysis.
2. Gaussian ordination is very sensitive to quan

titative noise and is not robust to departures from

its assumed response model.
Fig. 6. Ordinations of a noiseless, 1x \R symmetric, unimodal 3. The currently popular DCA does not perform
model (replicate 1), with samples confined to a T-shaped region well with more complex response models and non
in environment-space: (a) sample arrangement in model; (b)
regular sampling schemes. The technique often
LNMDS ordination; (c) DCA ordination. The solid and dashed
lines indicate produces distorted ordinations of \ x \R
the direction of the first and second simulated gra
dients, coenoplanes, even with symmetric, unimodal re
respectively.
sponse surfaces.

Fig. 7. Ordinations of a noiseless, 1x \R interaction model 4. Local non-metric multidimensional scaling

(replicate 1), with samples confined to a cross-shaped region in the Bray-Curtis
(LNMDS), using dissimilarity
environment-space: (a) sample arrangement in model; (b) coefficient is the most robust and effective of the
LNMDS ordination; (c) DCA ordination.
methods compared.
5. None of the techniques studied achieved satis
nation (Fig. 7c), both arms of the cross are bent factory recovery of coenoplane models in which
through approximately 90? at their junction. One samples were restricted to T-shaped or cross-shaped
end of the arm marked by a dashed line extends regions of environment space.
into the third dimension of the ordination (not
shown). The LNMDS ordination displays a similar
 103

Discussion special case. There is little justification for the con

tinued application of linear ordination methods to
Linear ordination methods community data for the purpose of indirect gra
dient analysis. That is not to say that such methods
This study confirms and extends the conclusions are not useful for the analysis of other types of eco
of previous work (see above) in showing the ineffec logical data, for which the linear model is appropri
tiveness of PCA due to curvilinear distortion ate.
(the
'type A distortion of Orl?ci, 1974). PCoA with the
Bray-Curtis coefficient is similarly afflicted. The Gaussian ordination and other curve-fitting
distortion is not confined to models withsymmet methods
ric, unimodal responses, but also occurs with the
more complex models examined. Gauch et al. In this
study, Gaussian ordination (GO) was
(1981) refer to the published comments of J. C. shown to lack robustness to departures from its as
Gower on a paper by Sibson (1972). According to sumed symmetric, unimodal response model, thus
Gower (see Gauch et al, 1981) NMDS gives very confirming the preliminary results of Austin (1976).
similar results to metric scaling (i.e. PCoA), but at Furthermore, GO is very sensitive to the addition of
the expense of much more computation. However, quantitative noise, much more so than DCA and
in this study the two approaches only gave similar LNMDS. This result is at variance with Gauch et
ordinations for models with very short gradients, al (1974), who imply that GO is rather resistant to
where the relationship between Bray-Curtis dis noise.

similarities and environmental distances was ap The program used for GO in this work (Gauch,
proximately linear. In all other cases, the LNMDS 1979) produces only one ordination axis and is
ordinations recovered the gradient structure much therefore inadequate for data sets with several un
more effectively than PCoA. derlying gradients. Extensions of GO to the mul
It has been argued (Greig-Smith, 1980; Feoli & tidimensional case are available, but there are theo
Feoli Chiapella, 1980; Van der Maarel, 1980) that retical and computational difficulties which remain
the curvilinear distortion in PCA does not matter, largely unresolved. The approach of Ihm & Van
provided that interpretation of PCA ordinations Groenewoud (1975) makes unacceptably stringent
takes into account its probable occurrence. Howev assumptions about the response model. Orl?ci
er, when there is more than one underlying gra (1978, 1980) has proposed an indirect algorithm,
dient, it is very difficult to perceive that the points based on the application of metric multidimension
in the configurations such as shown in Fig. 1 fall on al scaling to particular measures of compositional
a surface irregularly curved into three or more dissimilarity and ordination distance, derived un
dimensions, let alone to devise a procedure for der specific assumptions about the Gaussian
spe
mapping the points onto a plane. cies response model. Some initial simulation results
Several authors have suggested that linear ordi (Fewster & Orl?ci, 1983) suggest that the perfor
nation methods are useful for indirect gradient mance of he method is undesirably sensitive to the
analysis, provided that their application is restrict assumptions made in the derivation of dissimilari
ed to data sets with low beta diversity (e.g., Austin ties and distances.
& Noy-Meir, 1971). Unfortunately, it is often The maximum likelihood non-linear ordination
difficult to assess a priori whether the beta diversity method of Johnson & Goodall (1979; Goodall &
of a data set is low enough for linear techniques to Johnson, 1982) is similar in philosophy to GO, in
be applied with negligible distortion. In any case, that it attempts to derive a sample configuration
when beta diversity is low, good recovery of the gra which maximizes the fit of symmetric, unimodal re
dients can still be obtained using a robust non sponse functions for each species. Simulation
linear technique, such as LNMDS: the monotonici studies of the method have been restricted to
ty assumption of LNMDS embraces linearity as a models which conform with the assumed model
 104

and no comparison has been made with other non attempt a posteriori to rectify the curvilinear dis
linear ordination techniques, such as LNMDS. tortion of gradients: detrending and rescaling.
Further simulation studies are required to assess Computational details are given by Hill (1979; see
the robustness of variants of GO (and other 'curve also Hill & Gauch, 1980). The procedures were ap
fitting' ordination methods) to variations in the parently developed from a consideration of very

species response model and their sensitivity to simple models of species replacement along a single
noise. The excellent performance of such methods gradient. Both involve an arbitrary choice of the
with data confirming with their assumed
simulated degree of segmentation of axes. From the results of
response models is no basis upon which to predict this study, it appears that the effects of detrending
their effectiveness with real data. and rescaling with non-trivial coenoplane models
are not always desirable. The distortions of the un
derlying gradient structure in some DCA ordina
Poor performance of DCA tions (e.g., Figs. 2b, 4d and 6c) may be attributable
to the behaviour of either or both detrending and
There are at least two factors which may contrib rescaling. For some of the models for which severe
ute to the lack of robustness and erratic perfor flattening of parts of the configuration occurred in
mance of DCA observed in this study; (1) properties DCA ordinations, several other DCA ordinations
of the implied dissimilarity measure, and (2) the be were performed by varying the number of detrend
haviour of the detrending and rescaling processes. ing segments and the number of cycles of rescaling.
Hill & Gauch (1980) claim that correspondence None of these adjustments was successful in reduc
from which DCA is derived, 'makes no ing the degree of distortion. In several published
analysis,
use of the concept of compositional distance'. This applications of DCA, (e.g., Robertson et al, 1984;
is untrue. Correspondence analysis can be formu Gibson & Kirkpatrick, 1985) ordination configura
lated as a particular variety of principal co tions have flattened 'tongues', similar to that
ordinates analysis, in which compositional dis shown in Fig. 4d. These may be artifacts, due to the
dis of detrending or rescaling.
similarity is measured using the Chi-squared operation
tance metric and samples are weighted according to Not all curvilinear structures which may appear
their totals (Chardy et al, 1976). Correspondence in an ordination are distortions, arising from the

analysis derives an ordination in which the dis non-linear relationship between dissimilarity and
tances between pairs of sample points are propor environmental distance. DCA has no way of distin
tional to their Chi-squared distance values. guishing between 'horse-shoe' or 'arch' distortions
The appropriateness of Chi-squared distance as and features of the environmental configuration
a measure of compositional dissimilarity in ecology which happen to be non-linear. There is a danger
may be questioned (Faith et al, 1987). The measure that DCA will introduce new distortions of its own.
accords high weight to species whose total abun A good example is provided by the DCA ordination
dance in the data matrix is low. It thus tends to ex of the 1 x \R symmetric response coenoplane with
aggerate of samples containing
the distinctiveness a T-shaped sampling pattern, shown in Fig. 6c. The
several rare species. Unlike the Bray-Curtis coeffi non-linear structure formed by the upright of the T
cient and related measures, Chi-squared distance and one side of the cross-bar has been flattened
does not reach a constant, maximal value for sam out, as if it were an arch distortion.
ple pairs with no species in common, but fluctuates Reservations about the possible effects of the

according to variations in the representation of spe empirical adjustments in DCA

have been expressed
cies with high or low total abundances. These by some authors (e.g. Fewster & Orl?ci, 1983).
properties of Chi-squared distance may account for Nevertheless, DCA has become probably the most
some of the distortions observed in DCA ordina widely-used ordination technique for community
tions. data. A major contributing factor to the rapid ac
DCA includes two empirical procedures which ceptance of the method has been the distribution of
 105

the program DECORANA (Hill, 1979), which is figurations exist with lower stress. It is possible that
relatively easy to use and economical, in terms of many of the NMDS ordinations obtained by Gauch
both computing time and memory requirements. In et al (1981) were local optima, resulting from the
the light of current results, the status of DCA as a use of random starts or metric scaling solutions
satisfying solution to the problem of curvilinear which exhibited severe curvilinear distortion.
distortion in ordination (cf. Gauch, 1982) must be
questioned. Interpretation of DCA ordinations'
should take into account the possibility of artifac Robustness of non-metric multidimensional
tual distortions, due to the properties of the im scaling
plied dissimilarity measure or the activities of
detrending or rescaling. Published applications of non-metric multidi
The current results suggest a clear preference for mensional scaling (NMDS) in ecology (in either the
LNMDS over DCA. DCA consistently outper 'global' or 'local' form) are relatively rare (e.g.,
formed LNMDS only for coenoclines with simple, Prentice, 1977; Clymo, 1980; Field et al, 1982; Ok
unimodal response curves.
On the basis of very sanen, 1983; Dargie, 1984), but in most cases the
limited simulations
using Gaussian models, Orl?ci technique has been considered effective. Simulation
et al. (1984) have also reported rather poor recovery studies based on Gaussian models have shown that
of simulated gradients by DCA. NMDS can successfully recover gradients of high
Gauch et al(1981) compared DCA with several beta diversity, both in its 'global' (Fasham, 1977)
variants of NMDS, including the LNMDS tech and 'local' forms (Prentice, 1980). Austin's (1976)
nique examined here. They concluded that DCA preliminary examination of some alternative re
was generally more successful than NMDS. Howev sponse models provided an early indication of the
er, their study considered only a small number of relative robustness of 'global' NMDS.
Gaussian response models, all with regular sam This study has identified LNMDS, using the
pling patterns. No replication was apparently per Bray-Curtis dissimilarity coefficient, as a robust
formed within each combination of model proper technique for the analysis of community data when
ties. The in performance
great variation by DCA the aim is to recover the compositional dimensions
among replicate models, observed in this study (e.g. associated with underlying environmental gra
Tables 2 and 3; Fig. 4), highlights the danger of ar dients. The relative merits 'global' versus 'local'
of
riving at misleading conclusions if comparisons are variants of NMDS and the possible effects of the
restricted to a single replicate. choice of dissimilarity measure were not investigat
The use of DCA ordinations as initial configura ed. However, subsequent work has used a similar,
tions for LNMDS is another major factor distin but more extensive, simulation approach to com
guishing the present study from that of Gauch et al pare the robustness of dissimilarity coefficients
(1981). They employed random starting configura (Faith et al, 1987) and several forms of NMDS
tions for LNMDS (using the program SIBSON) (Minchin, Faith & Belbin, unpublished). The
and apparently used metric scaling solutions (simi results suggest a preference for 'local' over 'global'
lar to PCoA) for the other variants of NMDS ex NMDS and the Bray-Curtis measure was among
amined. The choice of a poor initial configuration the most robust of the coefficients compared, in
can reduce the likelihood of NMDS achieving a so terms of its rank correlation with simulated en
lution with the best possible monotonie fit between vironmental distance.
ordination distance and dissimilarity. The iterative Non-metric scaling is recom
multidimensional
procedure may become trapped in a 'local opti mended as a robust technique for indirect gradient
mum' ,where no small change in the configuration analysis which deserves more widespread use by
will decrease the stress, even though different con community ecologists.
 106

References Mar. Ecol. Prog. Ser. 8: 37-52.

Gauch, H. G., 1973. The relationship between sample similarity

and ecological distance.
Ecology 54: 618-622.
Austin, M. P., 1976. Performance of four ordination techniques Gauch, H. G., 1979. Catalog of the Cornell ecology programs
assuming different non-linear species response models. series. 11th ed. Ecology and Systematics, Cornell University,
Vegetado 33: 43-49. Ithaca, New York.
Austin, M. P., 1980. Searching for a model for use in vegetation Gauch, H. G., 1982. Multivariate analysis in community ecolo
analysis. Vegetado 42: 11-21. gy. Cambridge University Press, London and New York.
Austin, M. P., 1985. Continuum concept, ordination methods Gauch, H. G., Chase, G. B. & Whittaker, R. H., 1974. Ordina
and niche theory. Ann. Rev. Ecol. Syst. 16: 39-61. tion of vegetation samples by Gaussian species distributions.
Austin, M. P., 1987. Models for the analysis of species' response Ecology 55: 1382-1390.
to environmental gradients. Vegetatio 69: 35-45. Gauch, H. G. & Whittaker, R. H., 1972a. Coenocline simula
Austin, M.
P., Cunningham, R. B. & Fleming, P. M., 1984. New tion. Ecology 53: 446-451.
approaches to direct gradient analysis using environmental Gauch, H. G. & Whittaker, R. H., 1972b. Comparison of ordi
scalars and statistical curve-fitting procedures. Vegetatio 55: nation techniques. Ecology 53: 868-875.
11-27. Gauch, H. G. & Whittaker, R. H., 1976. Simulation of commu
Austin, M. P. & Noy-Meir, I., 1971. The problem of non nity patterns. Vegetatio 33: 13-16.

linearity in ordination: experiments with two-gradient Gauch, H. G., Whittaker, R. H. & Singer, S. B., 1981. A com
models. J. Ecol. 59: 763-773. parative study of nonmetric ordinations. J. Ecol. 69:
Beatty, S. W., 1984. Influence of microtopography and canopy 135-152.

species on spatial patterns of forest understorey plants. Ecol Gibson, N. & Kirkpatrick, J. B., 1985. Vegetation and flora as

ogy 65: 1406-1419. sociated with localised snow accumulation at Mount Field
Bradfield, G. E. & Scagel, A., 1984. Correlations among vegeta West, Tasmania. Aust. J. Ecol. 10: 91-99.
tion strata and environmental variables in subalpine spruce Goodall, D. W. & Johnson, R. W., 1982. Non-linear ordination
-
fir forests southeast British Columbia. Vegetatio 55: in several dimensions. A maximum likelihood approach.
105-114. Vegetatio 48: 197-208.
Brown, M. J., Ratkowsky, D. A. & Minchin, P. R., 1984. A com Gower, J. C, 1966. Some distance properties of latent root and

parison of detrended correspondence analysis and principal vector methods used in multivariate analysis. Biometrika 53:
co-ordinates analysis using four sets of Tasmanian vegetation 325-338.
data. Aust. J. Ecol. 9: 273-279. Greig-Smith, P., 1980. The development of numerical classifica

Chardy, P., Glemarec, M. & Lauree, A., 1976. Application of in tion and ordination. Vegetatio 42: 1-9.
ertia methods to benthic marine ecology: practical implica Greig-Smith, P., 1983. Quantitative plant ecology. 3rd ed. Black
tions of the basic options. Estuar. Coast. Mar. Sei. 4: well, Oxford.
179-205. -
Hill, M. O., 1979. DECORANA A FORTRAN Program for
Clymo, R. S., 1980. Preliminary survey of the peat-bog Hum Detrended Correspondence Analysis and Reciprocal Averag
mell Knowe Moss using various numerical methods. Vegeta ing. Ecology and Systematics, Cornell University, Ithaca,
tio 42: 129-148. New York.
Dargie, T. C. D., 1984. On the integrated interpretation of in Hill, M. O. & Gauch, H. G., 1980. Detrended correspondence
direct site ordinations: a case study using semi-arid vegeta analysis, an improved ordination technique. Vegetatio 42:
tion in southeastern Spain. Vegetatio 55: 37-55. 47-58.
Del Moral, R., 1980. On selecting indirect ordination methods. Hotelling, H., 1933. Analysis of a complex of statistical varia

Vegetatio 42: 75-84. bles into principal components. J. Ed. Psych. 24: 417-441 &

Faith, D. P., Minchin, P. R. & Beibin, L., 1987. Compositional 498-520.

dissimilarity as a robust measure of ecological distance. Ihm, P. & Van Groenewoud, H., 1975. A multivariate ordering
Vegetatio 69: 57-68. of vegetation data based on Gaussian type gradient response
Fasham, M. J. R., 1977. A comparison of nonmetric mul curves. J. Ecol. 63: 767-777.
tidimensional scaling, principal components analysis and Johnson, R. W & Goodall, D. W., 1979. Maximum likelihood

reciprocal averaging for the ordination of simulated approach to non-linear ordination. Vegetatio 41: 133-142.
coenoclines and coenoplanes. Ecology 58: 551-561. Kruskal, J. B., 1964a. Multidimensional scaling by optimizing
Feoli, E. & Feoli Chiapella, L., 1980. Evaluation of ordination of fit to a nonmetric 29:
goodness hypothesis. Psychometrika
methods through simulated coenoclines: some comments. 1-27.
Vegetatio 42: 35-41. Kruskal, J. B., 1964b. Nonmetric multidimensional a
scaling:
Fewster, P. H. & Orl?ci, L., 1983. On choosing a resemblance numerical method. Psychometrika 29: 115-129.
measure for non-linear ordination. 54:
predictive Vegetatio Loucks, O. L., 1962. Ordinating forest communities by means of
27-35. environmental scalars and phytosociological indices. Ecol.
Field, J. G., Clarke, K. R. & Warwick, R. M., 1982. A practical Monogr. 32: 137-166.
strategy for analysing multispecies distribution patterns. Minchin, P. R., 1983. A comparative evaluation of techniques
 107

for ecological ordination simulated data and Rotenberry, J. T. & Wiens, J. A., 1980. Habitat structure, patchi
using vegetation
an ordination-classification of the alpine ness and avian communities in North American steppe vege
integrated analysis
the Mt Field tation: a multivariate analysis. Ecology 61: 1228-1250.
and subalpine plant communities of Plateau,
Tasmania. Ph.D. of Tasmania. Sch?nemann, P. H. & Carroll, R. M., 1970. Fitting one matrix
thesis, University
P. R., 1987. Simulation of multidimensional commu to another under choice of a central dilation and a rigid mo
Minchin,
towards a comprehensive tion. Psychometrika 35: 245-255.
nity patterns: model. Vegetatio (in
Sibson, R., 1972. Order invariant methods for data analysis. J.
press).
distribution Roy. Statist. Soc. B. 34: 311-349.
Mohler, C. L., 1981. Effects of
sample along gra
on eigenvector ordination. 45: 141-145. Swan, J. M. A., 1970. An examination of some ordination prob
dients Vegetatio
lems by use of simulated vegetational data. Ecology 51:
Noy-Meir, I. & Austin, M. P., 1970. Principal-component ordi
simulated data. 51: 89-102.
nation and vegetational Ecology
551-552. Van der Maarel, E., 1980. On the interpretability of ordination

diagrams. Vegetatio 42: 43-45.

Oksanen, J., 1983. Ordination of boreal heath-like vegetation
Van der Maarel, E., Boot, R., Van Dorp D. & Rijntjes, J., 1985.
with principal component analysis, correspondence analysis
Vegetation succession on the dunes near Oostvoorne, The
and multidimensional scaling. Vegetatio 52: 181-189.
Netherlands; a comparison of the vegetation in 1959 and
Orl?ci, L., 1974. On information flow in ordination. Vegetatio
29: 11-16. 1980. Vegetatio 58: 137-187.

in vegetation research. Walker, J. & Peet, R. K., 1983. Composition and species diversi
Orl?ci, L., 1978. Multivariate analysis
ty of pine-wiregrass savannas of the Green Swamp, North
2nd ed. Junk, The Hague.
ordination. Carolina. Vegetatio 55: 163-179.
Orl?ci, L., 1980. An algorithm for predictive Vegeta
Whittaker, R. H., 1960. Vegetation of the Siskiyou Mountains,
tio 42: 23-25.
N. C. & Fewster, P. H., 1984. Probing simu Oregon and California. Ecol. Monogr. 30: 279-338.
Orl?ci, L., Kenkel,
Whittaker, R. H., 1967. Gradient analysis of vegetation. Biol.
lated vegetation data for complex trends by linear and non
Rev. 42: 207-264.
linear ordination methods. Abstr. Bot. 8: 163-172.
I. C, in ecology. Whittaker, R. H. & Gauch, H. G., 1978. Evaluation of ordina
Prentice, 1977. Non-metric ordination methods
J. Ecol. tion techniques. In: R. H. Whittaker (ed.), Ordination of
65: 85-94.

Prentice, I.C, 1980. Vegetation and order invariant plant communities, pp. 227-336. Junk, The Hague.
analysis
42: 27-34. Young, F. W & Lewyckyj, R., 1979. ALSCAL-4 User's Guide.
gradient models. Vegetatio
P. A., MacKenzie, M. D. & Elliot, L. F., 1984. Gra Data Analysis and Theory Associates, P.O. Box 446, Carr
Robertson,
dient and classification of the woody for boro, North Carolina.
analysis vegetation
four sites in southern Illinois and adjacent Missouri. Vegeta
tio 58: 87-104. Accepted 18.8.1986.