Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Ecological Modelling 157 (2002) 119 /129

www.elsevier.com/locate/ecolmodel

Continuum theory revisited: what shape are species responses


along ecological gradients?
Jari Oksanen a,*, Peter R. Minchin b
a
Department of Biology, University of Oulu, 90014 Oulu, Finland
b
Department of Biological Sciences, Louisiana State University, 202 Life Sciences Building, Baton Rouge, LA 70803, USA

Abstract

The shape of species’ responses along ecological gradients has important implications for both continuum theory and
community analysis. Most current theories and analytical models in community ecology assume that responses are
unimodal and symmetric. However, interactions between species and extreme environmental stress may cause skewed
or non-unimodal responses. To date, statistical tools for evaluating response shapes have been either inappropriate,
inefficient or biased. Using a data set on vascular plant distributions along an elevation gradient, we show that
Huisman /Olff /Fresco (HOF) models are an effective method for this purpose, allowing models of various forms
(skewed, symmetric, plateau, monotonic) to be tested for adequacy. HOF modeling was compared with alternative
methods for response fitting, including Gaussian responses as Generalized linear model (GLMs), Generalized Additive
Models (GAM) and Beta Functions with fixed or estimated endpoints. In our data set, skewed and plateau responses
are less common than symmetric ones. Less than half of the species have skewed or plateau responses that can not be
adequately modeled by Gaussian models. We show that Beta function models with fixed endpoints are biased,
confounding skewness and the location of the mode and should not be used to test response shapes. Beta models with
estimated endpoints are fairly consistent with other models. GAM’s cannot provide clear tests of skewness or kurtosis
of response curves, though we show that GAM’s, in general, confirmed the shapes chosen by HOF modeling. We
provide free software for fitting HOF models and encourage further applications to community data collected along
different types of ecological gradients.
# 2002 Elsevier Science B.V. All rights reserved.

Keywords: Gradient analysis; Beta response function; Gaussian response function

1. Introduction ecological gradients (Gauch and Whittaker, 1972;


Whittaker, 1978; Begon et al., 1996). At the
The canonized view in vegetation ecology is that extreme, this leads into an unrealistic species
species have symmetric, unimodal responses to packing model, where all species are assumed to
have Gaussian response functions with identical
width and height but with uniformly located
* Corresponding author. Tel.: /358-8-553-1526; fax: /358-
optima (Gauch and Whittaker, 1972; ter Braak,
8-553-1061
E-mail addresses: jari.oksanen@oulu.fi (J. Oksanen), 1985). Austin (1999, 2002) and many earlier
minchin@lsu.edu (P.R. Minchin). articles, shows clearly that this view is based on
0304-3800/02/$ - see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 0 4 - 3 8 0 0 ( 0 2 ) 0 0 1 9 0 - 4
120 J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129

very scanty evidence and its canonical status is response is significantly skewed, or whether sym-
unfounded. Indeed, theory predicts that response metric or monotonic responses could be used.
shapes should differ among gradient types (Austin In this paper, we study the suitability of
and Smith, 1989) or gradient locations (Austin and Huisman /Olff /Fresco models (HOF) for detect-
Gaywood, 1994). In addition, species interactions ing the shape of species responses along gradients.
may change the response shape even if the funda- We compare HOF models against symmetric
mental response were symmetric (Austin et al., Gaussian responses, variable Beta responses and
1990). Therefore, the analysis of response shape is non-parametric GAM.
of great theoretical interest (Austin, 1999). More-
over, it has a practical interest for ecologists
because of its implications for ordination metho- 2. Material and methods
dology. Correspondence analysis (CA) and its
derivatives, have been justified using the Gaussian 2.1. Data set
model (ter Braak, 1985), though the validity of
justification is open to debate (Austin, 1999, 2002). We analyzed the response patterns of vascular
The main alternative to CA-based methods, non- plants along an altitude gradient on the Mt. Field
metric multidimensional scaling, appears to be Plateau, Tasmania, which was earlier used to study
robust to variation in response shapes (Minchin, the shapes of response surfaces and some other
1987a). properties of community patterns (Minchin, 1989).
The symmetric Gaussian response function has Vegetation was sampled with equally spaced
been the only ecological response model for which sample plots along transects with random starting
the parameters can be estimated effectively (ter locations. We stratified the data by soil drainage
Braak and Barendregt, 1986; Jongman et al., so that interactions with this second gradient
1987). Although the evidence for the Gaussian would not distort response shapes. We studied
model can be criticized strongly (Austin, 1976, only the well-drained subset in this paper and
1980, 1999, 2002) all the methods proposed so far fitted response models for all species, which
to test the significance of skewness of the response occurred on 20 or more sites in this subset.
have been either inadequate, inefficient or biased: Coverage of sample plots was dense and even in
visual judgment of skewness (Økland, 1986) is this subset, as is obvious in the graphics in this
subjective; third-degree polynomials (Austin et al., article. This gave us a data set of 167 sites and 52
1990) have unrealistic shapes; Beta functions species with an altitude range 930/1380 m. We
(Austin et al., 1994) confound the location of the analyzed the data as presence/absence, or binomial
maximum and the skewness if the response end- with denominator m /1.
points are fixed (Oksanen, 1997); and the shape of
smoothed generalized additive models (GAM) 2.2. Bell-shaped or Gaussian response model
(Hastie and Tibshirani, 1990; Yee and Mitchell,
1991; Bio et al., 1998) must be assessed subjec- The Gaussian model defines a symmetric, bell-
tively. However, Huisman et al. (1993) have shaped response m along gradient x with three
proposed a set of hierarchical models which interpretable parameters: location of the optimum
include a skewed response that can be statistically u, height of the response h and tolerance or width
tested against a symmetric response. They assume of the response t (Fig. 1):
that these models cannot be estimated using  
(x  u)2
maximum likelihood (Huisman et al., 1993), but m h exp  (1)
2t2
we found that this can be done and so it is possible
to use more realistic error distribution for the Instead of direct estimation of Gaussian para-
observations than the originally implied normal meters, it is customary to fit an equivalent poly-
distribution and to test statistically whether a nomial model (ter Braak and Looman, 1986):
J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129 121

Fig. 1. Gaussian response function fitted to Bauera rubioides .


Continuous lines show a polynomial model with logit link (Eq.
(2)) and shading its approximate 95% confidence band. Broken Fig. 2. Beta response models fitted to B. rubioides with
lines show the strict Gaussian model fitted to the same data. endpoints p1 and p2 fixed either 10 or 50 m beyond the extreme
presences and a model where the endpoints were estimated
simultaneously with other parameters (the estimates were p1 /
mexp(b0 b1 xb2 x2 ) (2) /907 and p2 /1239 m, observed range 945 /1220 m).
This can be easily fitted as a generalized linear
model (GLM) with a logarithmic link function m k(xp1 )a (p2 x)g (3)
(McCullagh and Nelder, 1989; Guisan et al., 2002)
This has five parameters and so it can find other
and the maximum likelihood estimates of the
response shapes than the simple Gaussian re-
Gaussian parameters can be easily found from
sponse. The Beta function has been found useful
the polynomial coefficients (ter Braak and Loo-
for simulation of gradient responses, because it
man, 1986; Oksanen et al., 1988, 2001). Quite
can assume a wide range of unimodal shapes
commonly, the polynomial form is simply called a
(Minchin, 1987b). In principle, it should be
Gaussian response without explicit reference to the
possible to fit skewed, unimodal responses with
original form (Eq. (1)) and it is even used with link
various degrees of kurtosis. However, it is cus-
functions other than logarithmic. With other link
tomary to fix parameters p1 and p2 to be the
functions, the general shape of the response is
endpoints of the range of occurrence of species.
similar to strict Gaussian, but the curve still differs
Originally this was seen only as a fitting device,
somewhat from the Gaussian and the width
since in this way Eq. (3) simplifies into a special
parameter (t) has a different interpretation for
case of GLMs (Kay and Little, 1987):
the ecologist (Heegaard, 2002): it no longer gives
the gradient distance at which the expected abun- log(m)log(k)a log(xp1 )g log(p2 x) (4)
dance of species drops to a certain fraction of
expected maximum h (Fig. 1). In this paper, we with explanatory variables log(x/p1) and log(p2/
model the probability of occurrences of species x) and logarithmic link function. However, Austin
using Binomial m /1 error and apply the logistic and Nicholls (1997) regard endpoints, which
link function, as is customary with binomial define the range, as important ecological para-
GLM’s, instead of the strictly correct logarithmic meters. This leaves us only three free parameters,
link. just like in the Gaussian response function (Oksa-
nen, 1997): parameter k is a scaling parameter
adjusting the response height to fit the observa-
2.3. Beta response model tions and parameters a and g determine (1) the
location of the optimum; (2) skewness and (3)
Austin (1976; Austin et al., 1994) proposed kurtosis. Obviously, there is one feature too many
replacing the Gaussian response function with for two parameters and therefore, the model is not
the Beta response function (Fig. 2): as versatile as suggested. However, it is suggested
122 J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129

that a and g are mainly parameters of response cases where either a or g is negative, or where the
skewness (Austin and Nicholls, 1997). If a /g, the optimum lies outside the studied range.
response is symmetric (and the optimum in the
middle of the range), implying the model:
2.4. Huisman /Olff /Fresco or HOF model
log(m) log(k)a[log(xp1 )log(p2 x)] (5)
Huisman et al. (1993) proposed a set of five
with a single explanatory variable log(x1/p1)/ hierarchic models for species response. The most
log(p2/x). Eqs. (4) and (5) define nested GLMs complex model defines a skewed response function
and can be used to test the null hypothesis of and can be written in slightly modified form:
symmetry, a /g .
The selection of endpoints has a great effect on 1 1
m M (6)
fitted response shapes (Fig. 2) and Oksanen (1997) 1  exp(a  bx) 1  exp(c  dx)
suggested that they should be estimated together
Here, m is the expected value, which is depen-
with other model parameters, but suspected this to
dent on the known values of the gradient (x ),
be difficult. However, we found that the simulta-
maximum possible value (M ) and four estimated
neous fitting of all five parameters is relatively
parameters of the function (a, b , c , d ). This model
simple, but has to be done with non-linear max-
imum likelihood regression instead of more stan- can be simplified into four other models by
aliasing some parameters or fixing them to con-
dard GLM.
stant values (Table 1). These simplified models
We fitted Beta responses in two alternative
define symmetric, plateau type or monotone
ways. Firstly, we followed Austin and Nicholls
responses (Fig. 3).
(1997) and fixed the endpoints before fitting the
Huisman et al. (1993) presented only a least-
response as a GLM (Eq. (4)). The Beta function is
squares method for fitting these models and
undefined beyond the endpoints p1 and p2 and
suggested that maximum likelihood estimation
therefore, we pushed these parameters somewhat
outwards from the most extreme occurrences. The cannot be used. However, the use of maximum
likelihood is fairly simple (e.g. Venables and
degree of pushing has a great influence in the
Ripley, 1999). The maximum likelihood estimates
response shapes (Fig. 2). We used 50 m, since it
can be found by maximizing a log-likelihood
produced better fitting curves that were more
function instead of minimizing the squared resi-
similar to other response models than did a smaller
duals. The maximum likelihood function used in
extension. Austin and Meyers (1996), Austin and
the HOF model is non-linear in parameters a, b , c ,
Nicholls (1997) similarly pushed the endpoints
d and so these must be estimated using iterative
outwards by a fixed number of zero-observations.
The second alternative was estimation of p1 and non-linear methods (e.g. Venables and Ripley,
1999).
p2 together with other parameters. In this case, the
parameters p1 and p2 are selected to give a good fit
Table 1
for the observations and they have nothing to do
Simpler HOF models can be derived from the most complex
with the expected range of species (see Fig. 2). In model V (Eq. (6)) by fixing some parameters to constant values
both cases, we used Eq. (4) with logistic link
function instead of the strict form (Eq. (3)), since Model Parameters
we had binary responses. V Skewed a b c d
Beta response function is unimodal only if a /0 ¯ ¯
IV Symmetric a¯ b c¯ b
¯ ¯
and g /0 and so it can define monotone or III Plateau a b c¯ 0
¯
inverted responses (no optimum, but a minimum). II Monotone a¯ b 0¯ 0
¯ ¯
I Flat a 0 0 0
In this article, we interpreted the response shape as ¯
monotone if it was strictly monotone within The estimated parameters are underlined and the fixed
observations. Thus monotone responses include parameters in ordinary print.
J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129 123

(Hastie and Tibshirani, 1990). Consequently, they


may be able to produce shapes that differ from any
parametric model we have studied so far (Fig. 4).
GAM’s have recently become popular in gradient
modeling in vegetation science (e.g. Yee and
Mitchell, 1991; Bio et al., 1998; Leathwick, 1998;
Heegaard and Hangelbroek, 1999; Ejrnæs, 2000;
Lehmann et al., 2002; Zaniewski et al., 2002).
It may be difficult to characterize the shape of
response. If a parametric shape is completely
within the confidence band of a fitted GAM, the
response could be of that parametric shape. In this
Fig. 3. Hierarchic set of HOF models (Table 1) fitted to B.
rubioides . Model IV was rejected and the estimation continued way, we use GAM to verify or falsify our
through model III which was found better than model II. parametric models and in seeking possible para-
metric forms.
For statistical testing, the maximized likelihood We use smoothing splines for the predictor. The
is transformed into a related statistic known as number of degrees of freedom of the smoother is
deviance (Dobson, 1983; McCullagh and Nelder, selected by generalized cross validation (Gu and
1989) which follows asymptotically the x2 distri- Wahba, 1991), with upper limit alternatively 3 or 6
bution. We study only Binomial m /1 models and degrees of freedom. This was only the upper limit
so we can base our testing directly on the deviance, for degrees of freedom: the actual degrees of
since overdispersion can occur only in Binomial freedom were found by cross-validation (Wood,
models where denominator m /1 (McCullagh and 2000; Wood and Augustin, 2002) and were gen-
Nelder, 1989). erally much lower. The use of generalized cross-
The selection of the final model is based on validation is an improvement over the older
backward elimination, starting with the most practice of comparing only some integer alterna-
complex model (V). The critical level used was tives of degrees of freedom.
P /0.05. As the models form a hierarchic series It may be necessary to trim out extreme sites
(Table 1), they are simplified in a pre-determined beyond the species occurrence in multi-gradient
order (McCullagh and Nelder, 1989), V; IV; II and models, because GAM fits the model only for the
I (Table 1). Model III has the same number of main effects instead of local neighborhood and so
estimated parameters as model IV (Table 1) and zeros outside the main range of the species reduce
therefore, it is omitted from this sequence. How- the fitted values within the main range (Hastie and
ever, if model IV is rejected in favor of Model V, Tibshirani, 1990). However, we achieved the same
the latter is compared against model III and model
simplification is continued normally. This intro-
duces a bias against model III, because it may be
about as good- or even better-choice than model
IV and therefore, we will report results for both
models.

2.5. Generalized additive models

All previous models had a strict, parametric


functional form: we forced them to fit the data.
GAM do not specify any strict parametric func- Fig. 4. GAM fitted to B. rubioides with approximate 95%
tion, but instead use non-parametric smoothers confidence intervals.
124 J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129

by stratifying the data for the secondary gradient 3. Results


(drainage) and so we could use an identical set of
sites for all models. According to HOF models, symmetric responses
were the most common response type (Table 2).
These symmetric responses could be adequately
fitted using the Gaussian model. About 21% (11 of
2.6. Software
52) of responses were skewed which is substan-
Most statistical analyses were performed using tially less than the 33% reported by Minchin
the R software (Ihaka and Gentleman, 1996), (1989). In contrast, Beta functions with fixed
which is ‘not unlike S-PLUS’ (Venables and Ripley, endpoints found 37% (19 out of 52) as significantly
1999). Most of the analyses are identical in S-PLUS, skewed. However, this was caused by fixing the
but in some cases the algorithms or user com- endpoints before fitting the Beta model, so that
mands differ slightly. R is available for free down- skewness and location of optima were confounded
load for several platforms and operating systems (Oksanen, 1997) and Beta functions with freely
at http://cran.r-project.org/. estimated ‘endpoints’ were more consistent with
Gaussian and Beta response functions with fixed HOF models (Table 2). Actually, most ‘skewed’
endpoints were fitted as GLMs (function glm). Beta responses with fixed endpoints were very
GAM are not included in base R, but they are in similar to strictly symmetric Gaussian and HOF
the recommended library mgcv (function gam). IV responses and ‘symmetric’ Beta responses fitted
The algorithm used differs somewhat from S-PLUS the data badly and differed from all alternative
(Wood, 2000; Wood and Augustin, 2002). models (Fig. 5). Technically ‘skewed’ Beta re-
Maximum likelihood fitting of non-linear HOF sponses (in the sense that a "/g ) with fixed end-
models and Beta responses with estimated ‘end- points were actually almost symmetric in most
points’ can be performed with non-linear mini- cases: instead of producing an asymmetric peak,
mization (function nlm). This is different from S- different a and g were used for better fitting of
PLUS that uses function ms. However, most HOF tails, or locating the peak in other places than at
models were initially fitted using a stand-alone the average of the endpoints.
program HOF written by Jari Oksanen (JO). All four response models produced very similar
The program HOF is available through web responses for species, which were regarded as
pages of JO as compiled binaries for MS-DOS/ symmetric, unimodal HOF model IV (Fig. 5).
WINDOWS and LINUX. R functions for the analyses Any of the alternative response models tested
and further discussion on analysis methods are seems to be adequate for such species, which are
available at the same address. the largest group in our data (Table 2). Similarly,

Table 2
The most parsimonous HOF models cross-tabulated against the most parsimonous Beta response models

HOF model Total Beta /estimated endpoints Beta /fixed endpoints

Skewed Symmetric Monotone Skewed Symmetric Monotone

V 11 9 1 1 7 3 1
IV 22 4 17 1 12 10
III 4 3 1 4
II 12 1 2 9 4 8
Total 49 14 23 12 19 21 9

The endpoints were fixed 50 m beyond the most extreme presences. One species had an inverted response (no optimum but a
minimum) in all models and two species with flat response (model I) were flat or inverted in all other models and these three are not
tabulated. The total number of species before these omissions was 52.
J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129 125

Fig. 5. Examples of species which were found to have symmetric bell-shaped responses by HOF models: HOF model IV, Gaussian
bell, ‘skewed’ Beta and GAM (df5/3) produced almost similar response, but ‘symmetric’ Beta differs sharply from all others.

species classified as HOF model II (monotone Plateau responses (HOF model III) are uncom-
response) were very similar in all models. mon and were found only for four species. All
Species classified as skewed responses (HOF these are within the confidence intervals of GAM
model V) or plateau responses (HOF model III) responses (Fig. 7). GAM responses were not as
are more interesting. These responses cannot be regular as the simple parametric plateaux of HOF
adequately fitted by Gaussian responses or Beta model III. In some cases the GAM response
responses with fixed endpoints and therefore, we resembles the skewed HOF model V, which was
compare them against GAM responses and Beta not retained in backwards selection.
with estimated endpoints. Our comparison was
biased in favor of model IV, because we studied
model III only if model IV failed. In fact, model 4. Discussion
III was an alternative parsimonious choice for nine
species out of 22 model IV species (Table 2) and Symmetric, bell-shaped responses were the most
for five of these it was indeed a better choice (lower common type at Mt. Field and only about one
deviance) than model IV. We used this rigorous fifth of the responses were clearly skewed. About
but discriminatory strategy because we think that 42% of species could have been fitted using
a plateau cannot be a global response model, but Gaussian response models. At first sight, these
valid only for a limited range on a gradient. conclusions appear to be in contrast with Minchin
It seems that GAM models with df5/3 are (1989) who wrote that ‘‘Although the majority of
usually more symmetric than skewed HOF re- response surfaces appear to be unimodal, only
sponses (Fig. 6). However, GAM models with 45% are symmetric’’. Actually, despite important
df 5/6 usually confirm the main peak of the differences between two studies (the original paper
skewed HOF response, but deviate from it in examined mean cover responses, rather than
other parts, being usually less regular (Fig. 6). probabilities of occurrence and fitted two-dimen-
Either the true response is less regular than sional response surfaces to both altitude and soil
assumed in parametric modeling, or HOF has a drainage), the observed proportion of symmetric
sharper peak than GAM is allowed to have. In bell responses was similar in both analyses. How-
general, HOF V models are within confidence ever, in this paper, we found a larger number of
limits of GAM’s, so that the response looks monotone responses (HOF models III, II and I).
credible (Fig. 6). Similarly, Beta responses with Minchin (1989) used non-parametric smoothing
estimated endpoints were fairly consistent with and had to interpret the shapes subjectively with-
HOF models (Table 2). out help of confidence intervals. Obviously his
126 J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129

Fig. 6. GAM (df5/6) (solid line) with approximate 95% confidence intervals against GAM(df5/3) (dotted line) and HOF V responses
(broken line).

smoothing was not sufficient: in present parlance, that it is hardly possible to say that the Gaussian is
he used too many degrees of freedom in fitting. the universal and general response type. In parti-
Consequently, he emphasized the irregularity of cular, one third (11/33) of unimodal responses
many responses (regarding 22% of species as (models V and IV) are skewed rather than sym-
having multimodal surfaces), whereas our present metric.
analyses did stronger smoothing and found most Austin et al. (1994) have proposed strict rules
responses to be fairly regular. Nevertheless, the for selecting the species that can be studied for
proportion of non-Gaussian responses is so large their response shapes. For instance, there should

Fig. 7. HOF models III responses (broken line) and GAM (df5/3) (solid line) with approximate 95% confidence intervals.
J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129 127

be at least 100 sites with zero observations beyond independently and in our case this really improved
the extreme presence. Such rules may be applicable the applicability of Beta functions.
to huge data sets at the subcontinental scale, but Fixing the endpoints at the location of the most
they are hardly practical for ordinary vegetation extreme presences ignores the fact that the data are
surveys. Moreover, even in subcontinental data only a sample. A unimodal response with tapering
sets, the rule results in a biased selection of species, tails implies that the probability of occurrence near
leaving out those with modes closer to the gradient the extremes will be very low. Even with a very
extremes. Gradient limits may be natural, like the large data set, it is, therefore, unlikely that a
Mt. Field peak at 1380 m in our case, or they may presence will be observed at the real limit of
be arbitrary, like the lower limit of 930 m in our distribution. Beta response models may be very
case. Austin’s rule does not distinguish between effective for simulating realistic response patterns
these two cases. The main difference between our along ecological gradients (Minchin, 1987b), but
approaches is that we are mainly interested in the our results show that they are unsuitable for
shape of the species response at the central area of analyzing response shapes, as also concluded by
species occurrence, whereas Austin seems more Austin and Meyers (1996).
interested in species range and the shapes of the Any of the studied alternative models could be
tails. An essential feature of response shape is the used for symmetric, bell-shaped responses, even
slope of decrease near the species optimum and the Beta response model with estimated ‘end-
this can be studied even with truncated responses. points’. However, for other response types, Beta
functions and Gaussian response models provide
Contrary to Austin and Nicholls (1997), we think
biased or inappropriate models. GAM’s are
that there is no general need to remove extreme
usually very similar to the selected HOF models,
zeros, as most response models are able to predict
but smoothness and regularity of shape is depen-
near-zero occurrence for these. The trimming of
dent on the number of degrees of freedom (Hastie
zero tails is necessary only for Beta responses with
and Tibshirani, 1990). We used generalized cross-
fixed endpoints, since the response is undefined
validation for selecting the effective smoothness
beyond these points (Austin et al., 1994; Austin
(Gu and Wahba, 1991). Hastie and Tibshirani
and Nicholls, 1997; Oksanen, 1997). Trimming of
(1990) found that this fails in several cases and
zeros may be necessary in GAM with several produces responses that are too irregular. There-
gradients, since GAM fit the response for each fore, we defined a fairly low maximum limit for
dimension separately without regarding interac- degrees of freedom. However, for detecting skew-
tions (Hastie and Tibshirani, 1990). For Gaussian ness, a somewhat higher limit must be used,
responses, HOF models or Beta models with freely although this comes at cost of an irregular,
estimated endpoints, there is no need to remove perhaps biased pattern in other parts of the
any extreme zeros, since these functions can fit response.
these adequately. We analyzed a real data set and therefore, we
Beta response functions with fixed endpoints cannot know the true underlying response shapes
failed badly in our analyses. After fixing the and so we do not know if HOF models actually
response endpoints, or the species range (Austin were correct. However, we compared HOF models
and Nicholls, 1997), the responses either fitted the against GAM’s and symmetric and monotone
data very badly or else failed to detect the models against other alternatives and found con-
appropriate shape because they preferred to find sistent patterns. In most cases, the final selected
the location of the optimum instead of skewness HOF model was entirely within the approximate
(Oksanen, 1997). Indeed, Oksanen (1997) sug- confidence limits of a GAM. The most notable
gested that the response endpoints should not be differences were near the extremes of the gradient,
fixed but estimated together with other para- where the GAM often flipped up, whereas the
meters, so that response curve could try to find corresponding HOF model decreased asymptoti-
both the skewness and the location of the optimum cally. However, the confidence bands in fitted
128 J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129

GAM models increase in the margins as well, so References


that obviously this is caused by uncertainty at the
margins and does not reflect real response shapes. Austin, M.P., 1976. On non-linear species responses models in
Another inconsistency was in skewed models when ordination. Vegetatio 33, 33 /41.
Austin, M.P., 1980. Searching for a model for use in vegetation
compared against strongly smoothed GAM’s, but
analysis. Vegetatio 42, 11 /21.
the location and shape of the main peak became Austin, M.P., 1999. On silent clash of paradigms: some
consistent with less strong smoothing in GAM, at inconsistencies in community ecology. Oikos 86, 170 /178.
cost of differences near the gradient margins. Austin, M.P., 2002. Spatial prediction of species distribution:
GAM’s and HOF models were usually consis- an interface between ecological theory and statistical
tent and either could be used in gradient analysis. modelling. Ecol. Model. 157 (2 /3), 101 /118.
Austin, M.P., Smith, T.M., 1989. A new model for the
Recently, GAM’s have enjoyed considerable po-
continuum concept. Vegetatio 83, 35 /47.
pularity and are seen by some authors as an Austin, M.P., Gaywood, M.J., 1994. Current problems of
optimal solution (Yee and Mitchell, 1991; Bio et environmental gradients and species response curves in
al., 1998; Leathwick, 1998; Heegaard and Hang- relation to continuum theory. J. Veg. Sci. 5, 473 /482.
elbroek, 1999; Ejrnæs, 2000). The strongest feature Austin, M.P., Meyers, J.A., 1996. Current approaches to
of GAM’s is that they are indeed non-parametric modelling the environmnetal niche of eucalypts: implica-
tions for management of forest biodiversity. Forest Ecol.
and can take any smooth shape, whereas even
Manage. 85, 95 /106.
HOF models are restricted to certain shapes. For Austin, M.P., Nicholls, A.O., 1997. To fix or not to fix the
instance, multimodal responses may be detected species limits, that is the ecological question: response to
with GAM’s but are beyond the scope of HOF Jari Oksanen. J. Veg. Sci. 8, 743 /748.
models. However, the shape of a fitted GAM Austin, M.P., Nicholls, A.O., Margules, C.R., 1990. Measure-
response must be interpreted subjectively and it ment of the realized qualitative niche: environmental niches
of five Eucalyptus species. Ecol. Monogr. 60, 161 /177.
may be difficult to say whether a response is really
Austin, M.P., Nicholls, A.O., Doherty, M.D., Meyers, J.A.,
skewed or significantly skewed, or could be just as 1994. Determining species response functions to an environ-
adequately described in other terms. Shape is, after mental gradient by means of a b-function. J. Veg. Sci. 5,
all, a parametric concept: if we say that a response 215 /228.
is skewed, the only difference compared with a Begon, M., Harper, J.L., Townsend, C.R., 1996. Ecology:
symmetric response should be the skewness. Individuals, Populations and Communities, third ed.. Black-
GAM’s may be a good choice for general pre- well.
Bio, A.M.F., Alkemade, R., Barendregt, A., 1998. Determining
dictive modeling of species responses, but for alternative models for vegetation response analysis: a non-
parametric questions of shape, parametric HOF parametric approach. J. Veg. Sci. 9, 5 /16.
models appear to provide a better alternative. Dobson, A.J., 1983. An Introduction to Statistical Modelling.
Many more studies of response patterns along Chapman and Hall.
ecological gradients are needed to provide a sound Ejrnæs, R., 2000. Can we trust gradients extracted by detrended
observational basis for continuum theory. HOF correspondence analysis. J. Veg. Sci. 11, 565 /572.
Gauch, H.G., Jr, Whittaker, R.H., 1972. Coenocline simula-
models provide an effective tool for studying the tion. Ecology 53, 446 /451.
shapes of species’ responses and we hope that the Gu, C., Wahba, G., 1991. Minimizing GCV/GML scores with
availability of software for HOF modeling will multiple smoothing parameters via the Newton method.
encourage the reexamination of existing data sets SIAM J. Sci. Stat. Comp. 12, 383 /398.
and stimulate further investigations into the nature Guisan, A., Edwards, T.C., Hastie, T., 2002. Generalized linear
and generalized additive models in studies of species
of community patterns.
distributions: setting the scene. Ecol. Model. 157 (2 /3),
89 /100.
Hastie, T.J., Tibshirani, R.J., 1990. Generalized Additive
Models. Chapman and Hall, London, p. 92.
Acknowledgements Heegaard, E., 2002. The outer border and central border for
species environmental relationships estimated by non-para-
We wish to thank Dr Mike Austin and Dr Alain metric generalized additive models. Ecol. Model. 157 (2 /3),
131 /139.
Zuur for invaluable comments on the manuscript.
J. Oksanen, P.R. Minchin / Ecological Modelling 157 (2002) 119 /129 129

Heegaard, E., Hangelbroek, H.H., 1999. The distribution of Oksanen, J., Läärä, E., Huttunen, P., Meriläinen, J., 1988.
Ulota crispa in relation to both dispersal- and habitat- Estimation of pH optima and tolerances of diatoms in lake
related factors. Lindbergia 24, 65 /74. sediments by the methods of weighted averaging, least-
Huisman, J., Olff, H., Fresco, L.F.M., 1993. A hierarchical set squares and maximum likelihood, and their use for the
of models for species response analysis. J. Veg. Sci. 4, 37 / prediction of lake acidity. J. Paleolimnol. 1, 39 /49.
46. Oksanen, J., Läärä, E., Meriläinen, J., Warner, B.G., 2001.
Ihaka, R., Gentleman, R., 1996. R: a language for data analysis Confidence intervals for the optimum in the Gaussian
and graphics. J. Comput. Graph. Stat. 5, 299 /314. response function. Ecology 82, 1191 /1197.
Jongman, R.H., ter Braak, C.J.F., van Tongeren, O.F.R., 1987. ter Braak, C.J.F., 1985. Correspondence analysis of incidence
Data analysis in community and landscape ecology, Pudoc, and abundance data: properties in terms of a unimodal
Wageningen. response model. Biometrics 41, 859 /873.
Kay, R., Little, S., 1987. Transformations of the explanatory
ter Braak, C.J.F., Barendregt, L.G., 1986. Weighted averaging
variables in the logistic regression model for binary data.
of species indicator values: its efficiency in environmental
Biometrika 74, 495 /501.
calibration. Math. Biosci. 78, 57 /72.
Leathwick, J.R., 1998. Are New Zealand’s Nothofagus species
ter Braak, C.J.F., Looman, C.W.N., 1986. Weighted averaging,
in equilibrium with their environment. J. Veg. Sci. 9, 719 /
logistic regression and the Gaussian response model.
732.
Lehmann, A., Overton, J.M., Leathwick, J.R., 2002. General- Vegetatio 65, 3 /11.
Venables, W.M., Ripley, B.D., 1999. Modern Applied Statistics
ized regression analysis and apatial predictions. Ecol.
Model. 157 (2 /3), 187 /205. with S-PLUS, third ed.. Springer, Heidelberg.
McCullagh, P., Nelder, J.A., 1989. Generalized Linear Models. Whittaker, R.H., 1978. Direct gradient analysis. In: Whittaker,
Chapman and Hall, London, p. 125. R.H. (Ed.), Ordination of Plant Communities. Junk, The
Minchin, P.R., 1987a. An evaluation of relative robustness of Hague, pp. 7 /50.
techniques for ecological ordinations. Vegetatio 69, 89 /107. Wood, S.N., 2000. Modelling and smoothing parameter
Minchin, P.R., 1987b. Simulation of multidimensional commu- estimation with multiple quadratic penalties. J. Roy. Stat.
nity patterns: towards a comprehensive model. Vegetatio 71, Soc. B62, 413 /428.
145 /156. Wood, S.N., Augustin, N.H., 2002. Improving GAMs for
Minchin, P.R., 1989. Montane vegetation of the Mt. Field environmental modelling using GCV and penalized regres-
massif, Tasmania: a test of some hypotheses about proper- sion splines. Ecol. Mod., This Volume, in this issue.
ties of community patterns. Vegetatio 83, 97 /110. Yee, T.W., Mitchell, N.D., 1991. Generalized additive models
Økland, R.H., 1986. Rescaling of ecological gradients. The in plant ecology. J. Veg. Sci. 2, 587 /602.
effect of scale on symmetry of species response curves. Zaniewski, A.E., Lehmann, A., Overton, J.M., 2002. Predicting
Nordic J. Bot. 6, 671 /677. species distribution using presence-only data: A case study
Oksanen, J., 1997. Why the beta-function cannot be used to of native New Zealand ferns. Ecol. Model. 157 (2 /3), 259 /
estimate skewness of species responses. J. Veg. Sci. 8, 147 / 278.
152.

You might also like