Evolution, 54(5), 2000, pp.

1480–1492

TAXON SAMPLING, CORRELATED EVOLUTION, AND INDEPENDENT CONTRASTS

DAVID D. ACKERLY
Department of Biological Sciences, Stanford University, Stanford, California 94305 E-
mail: dackerly@stanford.edu

Abstract. Independent contrasts are widely used to incorporate phylogenetic information into studies of continuous
traits, particularly analyses of evolutionary trait correlations, but the effects of taxon sampling on these analyses have
received little attention. In this paper, simulations were used to investigate the effects of taxon sampling patterns and
alternative branch length assignments on the statistical performance of correlation coefficients and sign tests; ‘‘full-
tree’’ analyses based on contrasts at all nodes and ‘‘paired-comparisons’’ based only on contrasts of terminal taxon
pairs were also compared. The simulations showed that random samples, with respect to the traits under consideration,
provide statistically robust estimates of trait correlations. However, exact significance tests are highly dependent on
appropriate branch length information; equal branch lengths maintain lower Type I error than alternative topological
approaches, and adjusted critical values of the independent contrast correlation coefficient are provided for use with
equal branch lengths. Nonrandom samples, with respect to univariate or bivariate trait distributions, introduce dis-
crepancies between interspecific and phylogenetically structured analyses and bias estimates of underlying evolutionary
correlations. Examples of nonrandom sampling processes may include community assembly processes, convergent
evolution under local adaptive pressures, selection of a nonrandom sample of species from a habitat or life-history
group, or investigator bias. Correlation analyses based on species pairs comparisons, while ignoring deeper relation-
ships, entail significant loss of statistical power and as a result provide a conservative test of trait associations. Paired
comparisons in which species differ by a large amount in one trait, a method introduced in comparative plant ecology,
have appropriate Type I error rates and high statistical power, but do not correctly estimate the magnitude of trait
correlations. Sign tests, based on full-tree or paired-comparison approaches, are highly reliable across a wide range of
sampling scenarios, in terms of Type I error rates, but have very low power. These results provide guidance for
selecting species and applying comparative methods to optimize the performance of statistical tests of trait associations.

Key words. Comparative methods, correlated evolution, independent contrasts, simulations, taxon sampling.

Received November 5, 1999. Accepted April 24, 2000.

In recent years, numerous analytical methods have been
introduced to incorporate phylogenetic information into tests
of evolutionary hypotheses such as rates of phenotypic evo-
lution, lineage diversification, the timing and pattern of co-
evolution among host-parasite or plant-herbivore lineages,
and correlations between phenotypic traits (Harvey and Pagel
1991; Sanderson and Donoghue 1994; Huelsenbeck and Ran-
nala 1997; Pagel 1999). These methods focus on how to
analyze interspecific data given a phylogenetic framework,
but little attention has been paid to the problem of which
taxa to select for analysis and how taxon sampling may in-
fluence the results. Taxon sampling has recently received
increased attention in relation to phylogeny reconstruction
(e.g., Hillis 1996; Graybeal 1998; Poe 1998), with particular
focus on the relative value of more taxa versus more char-
acters for improving phylogenetic accuracy. In studies of
character evolution, it has been shown for discrete traits that
the relative frequency of different character states will influ-
ence ancestral state reconstructions and the analysis of cor-
related trait changes (Frohlich 1987; Maddison 1990; Sille´n-
Tullberg 1993). For continuous traits, the sampling of taxa
and the resulting distribution of character values will also
have significant impacts on ancestral reconstructions (Cun-
ningham et al. 1998). However, the effects of taxon sampling
on the study of correlated evolution in continuous characters
have not been explicitly considered.
In comparative studies, sampling of taxa within lineages
of interest occurs for several reasons. First, all studies face
methodological limitations of time and effort and it is often
difficult or impossible to obtain data on all taxa in a clade.
The resulting samples will frequently reflect geography (e.g.,
taxa that live in the researcher’s immediate vicinity), abun-
1480
© 2000 The Society for the Study of Evolution. All rights reserved.
dance (e.g., common taxa due to ease of study), research
history (taxa that have been the focus of prior studies), and
other practical considerations. These limitations will act more
strongly when the data are expensive and/or time-consuming
to gather (e.g., population genetic parameters, behavior). Sec-
ond, many questions in evolutionary ecology address only a
limited set of taxa that share certain life-history or other
characteristics (Westoby 1999). For example, a study of the
significance of seed size in relation to gap regeneration in
tropical forests (Foster and Janson 1985) will necessarily be
limited to tropical tree species, and as a group these species
do not form a clade. Rather, they represent a very diffuse
sample of taxa drawn from the entire seed plant lineage (>
275,000 species). Any particular study would be limited to
only a small sample, frequently a sample drawn from one
geographic region, and analysis of these data in a phyloge-
netic context will necessarily use a heavily pruned phylogeny
(e.g., Kelly and Purvis 1993; Kelly 1995). The consequences
of this focus on particular life-history groups or communities
and of the associated pruning of phylogenetic trees have re-
ceived little attention in relation to phylogenetic comparative
methods. Finally, in a historical context speciation and ex-
tinction also represent sampling processes, in the sense that
trait-dependent changes in speciation or extinction probabil-
ities will influence the relative frequency of character states
in extant taxa. The resulting patterns may be the focus of
attention for some questions (e.g., patterns of species sorting
during mass extinction events). However, in most cases, stud-
ies are limited to extant taxa because the traits can only be
measured in living organisms or populations, so the potential
effects of nonrandom speciation/extinction processes merit
consideration.
 TAXON SAMPLING AND COMPARATIVE METHODS 1481

FIG. 1. (A) Example of a 256-taxon phylogeny based on a ‘‘time-only’’ Markovian speciation model (Martins 1996). (B) Phylogeny
of a random sample of 32 species drawn from the phylogeny in A, with true branch lengths based on the corresponding segments of the full
tree. (C) Phylogeny for the same set of species, with equal branch lengths. Circles show the pairs of sister taxa that would be selected for
species pair analyses of independent contrasts and sign tests. Tree graphics generated using TREEVIEW (Page 1996).

Westoby et al. (1998; Westoby 1999) have recently ad-
vanced several specific proposals for taxon sampling in com-
parative ecology, specifically considering the role of phy-
logeny. The process of species selection is analogized to the
problem of random sampling in ecological research, where
it has long been recognized as a central component of robust
hypothesis testing. For estimates of properties such as the
mean and variance of a trait in different clades, random sam-
ples of taxa (with respect to the traits of interest) are appro-
priate. In contrast, to detect associations between traits Wes-
toby et al. recommend a nonrandom approach based on sam-
pling pairs of closely related taxa in which the species of each
pair differ by some large amount in the ‘‘independent trait.’’
For example, in studies of seed size and its correlates, they
sampled phylogenetic species pairs in which seed size
differed by at least an order of magnitude (Armstrong and
FIG. 2. Type I error rates under the null model (A) and statistical power under the correlated-changes model (B) for random samples of
different size, based on nominal significance levels at a = 0.05. The last point on each line is based on analysis of the full 256-taxon
phylogenies. Sample size is the number of species or the number of contrasts: N for Rspec ; N —1 for Rpic and Spic; and N/2 for Rpair and Spair.
The ordinate in (A) is log transformed to show differences at low levels. The number of pairs is less than half the number of species in
the tree at each sample size due to tree topology (see Fig. 1).
Westoby 1993; Saverimuttu and Westoby 1996; Swanbor-
ough and Westoby 1996). Comparisons of species pairs,
which invoke only a minimal level of phylogenetic infor-
mation, have been widely used in plant ecology dating back
to Salisbury (1942). In a recent survey of the plant functional
ecology literature, Ackerly (1999) found 28 studies that used
species pairs, compared to 19 papers using independent con-
trasts calculated over a full phylogeny for the species of
interest. In many cases, each pair is chosen to represent con-
trasting states of a discrete life-history trait (e.g., comparisons
of physiology or allocation in annuals vs. perennials; Garnier
1992; Silvertown and Dodd 1996). If the contrasting states of
the discrete trait represent extreme values of an underlying
continuous character, then such comparisons may be similar
to Westoby’s approach of selecting species that exhibit a
large difference in the independent trait. However, the po-
tential advantages and disadvantages of species pairs com-
parisons and of Westoby’s recommendations for statistical
hypothesis testing have not been explicitly investigated (see
Maddison 2000).
In this study, simulations of trait evolution on a phylogeny
were used to address the problem of taxon sampling, focusing
on the method of independent contrasts and its application
to the study of correlated evolution between continuous phe-
notypic traits. Simulations were conducted on a set of ran-
domly generated 256-taxon phylogenies. Subsamples of the
terminal taxa were selected, following various criteria out-
lined below, to address the following questions: For random
species samples, how does sample size influence Type I error
rates (when the true correlation, p = 0) and statistical power
(for p ⁄ 0) of trait correlations based on either independent
contrast or interspecific (nonphylogenetic) analyses? Do non-
random species samples, with respect to one or both char-
acters, lead to biased estimates of trait correlations? How
does nonrandom sampling affect Type I error rates when p
= 0? Similarly, if species are chosen on the basis of an
independent third trait (which is not considered in the cor-
relation analyses), are there effects on parameter estimates
or Type I error rates? How do results based only on sister
taxa pairs compare to full-tree analyses conducted over all
nodes of the phylogeny? Do samples of species pairs that
differ by a predetermined amount in one character (following
Westoby 1999) lead to bias in estimating trait correlations
or influence Type I error or statistical power? Under the dif-
ferent scenarios outlined above, do nonparametric sign tests
of associations between independent contrasts provide reli-
able tests (i.e., appropriate Type I error when p = 0) and
how does their power compare to parametric statistics when
p ⁄ 0?
SIMULATIONS
The simulations presented here are based on a two-step
model of trait evolution followed by taxon sampling and
analysis of character correlations. In the first step, the evo-
lution of two traits was simulated on a large phylogeny, based
either on a null model with no correlation between changes
in the traits ( p = 0) or a correlated-changes model sampling
changes from a distribution with a true correlation of p =
0.5. In the second step, a subsample of species was selected
from the terminal taxa on this phylogeny, based on various
sampling schemes described below, and the phylogeny was
pruned to show relationships among this subset (‘‘species’’
and ‘‘taxa’’ are used interchangeably in this paper; Fig. 1).
For each sample, correlations between the two traits were
estimated based on species values and independent contrasts
(including species pairs analyses, sign tests, and alternative
branch lengths) to estimate: (1) the sample correlation be-
tween the two traits to determine if the taxon sampling led
to bias in estimates of the true correlation ( p = 0 or 0.5); (2)
for the null model, Type I error rates at the a = 0.05 and
0.01 levels, based on standard significance testing; and (3)
TABLE 1. Summary of correlation statistics (see abbreviations in text) for simulated evolution of two traits on a set of 256 random phylogenies
(full tree) and for subsamples of taxa chosen based on six different sampling algorithms; for NR2 the sampling bias coefficient = 0.75, and
for NR3 the probability of change in Z per time step = 0.001 (see text). N, number of taxa sampled for Rspec, number of contrasts for Rpic and
Spic, and number of species pairs for Rpair and Spair (the mean and range of values is shown for sampling schemes in which sample size varied);
mean and SD of correlation coefficients (R) or proportion of same sign contrasts (S) are based on 10,000 replicates. Type I error and statistical
power are the proportion of nominally significant results at a S 0.05 or 0.01. Results for p = 0 are for the null model of no true correlation;
results for p = 0.5 are for an alternative model with a true correlation between evolutionary trait changes.

Null model: p = 0 I error rates

Type Correlated-changes model: p = 0.5
Statistical power
Sample Statistic N Range R or S SD (a = 0.05) (a = 0.01) R or S SD (a = 0.05) (a = 0.01)
Full tree Rspec 256 0.003 0.213 0.568 0.447 0.488 0.171 0.971 0.960
Rpic.true 255 0.002 0.070 0.058 0.014 0.497 0.060 0.995 0.994
Rpic.equal 255 0.002 0.084 0.122 0.045 0.497 0.071 0.995 0.994
Rpair 84.9 (78–94) 0.003 0.152 0.155 0.067 0.493 0.121 0.981 0.954
Spic 255 0.501 0.033 0.053 0.014 0.666 0.033 0.994 0.992
Spair 84.9 (78–94) 0.501 0.056 0.055 0.014 0.666 0.054 0.877 0.700
Random Rspec 32 0.003 0.263 0.195 0.089 0.486 0.209 0.768 0.623
Rpic.true 31 0.000 0.180 0.051 0.010 0.494 0.139 0.850 0.657
Rpic.equal 31 —0.001 0.202 0.083 0.023 0.493 0.157 0.825 0.651
Rpair 10.7 (7–15) —0.005 0.352 0.095 0.026 0.478 0.283 0.418 0.213
Spic 31 0.499 0.091 0.073 0.013 0.666 0.085 0.529 0.243
Spair 10.7 (7–15) 0.497 0.155 0.025 0.010 0.665 0.145 0.111 0.047
NR1 Rspec 32 —0.001 0.246 0.165 0.067 0.318 0.229 0.481 0.312
Rpic.true 31 0.000 0.179 0.049 0.009 0.385 0.158 0.611 0.364
Rpic.equal 31 0.000 0.198 0.075 0.019 0.359 0.177 0.554 0.332
Rpair 10.6 (6–15) —0.004 0.346 0.084 0.024 0.388 0.308 0.303 0.131
Spic 31 0.501 0.091 0.074 0.011 0.624 0.089 0.340 0.126
Spair 10.6 (6–15) 0.500 0.154 0.022 0.009 0.639 0.150 0.084 0.035
NR2 Rspec 32 0.573 0.117 0.958 0.855 0.651 0.102 0.994 0.959
Rpic.true 31 0.362 0.147 0.557 0.292 0.588 0.108 0.975 0.891
Rpic.equal 31 0.392 0.161 0.622 0.390 0.591 0.123 0.959 0.872
Rpair 10.7 (6–15) 0.253 0.336 0.182 0.065 0.540 0.258 0.510 0.276
Spic 31 0.621 0.086 0.329 0.109 0.703 0.081 0.702 0.398
Spair 10.7 (6–15) 0.566 0.153 0.037 0.014 0.689 0.142 0.144 0.063
NR3 Rspec 32 0.002 0.295 0.252 0.130 0.482 0.231 0.744 0.610
Rpic.true 31 0.000 0.179 0.050 0.009 0.494 0.137 0.850 0.659
Rpic.equal 31 0.000 0.218 0.106 0.035 0.491 0.168 0.810 0.639
Rpair 10.5 (6–15) 0.002 0.390 0.138 0.050 0.473 0.317 0.437 0.243
Spic 31 0.501 0.090 0.073 0.013 0.668 0.086 0.540 0.255
Spair 10.5 (6–15) 0.502 0.155 0.021 0.009 0.669 0.146 0.111 0.054
DP Rspec 66.9 (42–94) 0.003 0.210 0.258 0.139 0.525 0.157 0.952 0.900
Rpic.true 65.9 (41–93) —0.001 0.122 0.048 0.009 0.635 0.069 1.000 1.000
Rpic.equal 65.9 (41–93) 0.001 0.133 0.066 0.015 0.617 0.078 1.000 0.999
Rpair 33.4 (21–47) —0.001 0.184 0.061 0.015 0.688 0.091 0.998 0.990
Spic 65.9 (41–93) 0.499 0.062 0.048 0.010 0.745 0.053 0.987 0.943
Spair 33.4 (21–47) 0.498 0.087 0.049 0.008 0.824 0.066 0.985 0.924
RP Rspec 66.9 (42–94) —0.005 0.232 0.309 0.188 0.490 0.183 0.901 0.834
Rpic.true 65.9 (41–93) 0.000 0.124 0.051 0.012 0.497 0.094 0.992 0.962
Rpic.equal 65.9 (41–93) 0.000 0.143 0.094 0.028 0.496 0.111 0.982 0.939
Rpair 33.4 (21–47) 0.003 0.213 0.113 0.038 0.490 0.166 0.829 0.671
 Spic 65.9 (41–93) 0.500 0.062
Spair 33.4 (21–47) 0.501 0.087
for the correlated-changes model, statistical power at the a
= 0.05 and 0.01 levels. All simulations and analyses of in-
dependent contrasts were conducted using ACAP, Version 5,
written by the author (see Ackerly 1997; Ackerly and
Donoghue 1998).
Simulations for this study were conducted on a set of 100
random phylogenies of 256 taxa, generated using a ‘‘time-
only’’ speciation model in COMPARE 4.1 (see Martins
1996). The trees differed from each other in topology and
branch lengths. Branch lengths generated by COMPARE
were multiplied by 1000 and rounded to the nearest integer
to allow stepwise simulation of trait evolution along each
branch, as described below. Across all trees, mean branch
0.048 0.009 0.666 0.059 0.775 0.548
0.051 0.010 0.666 0.081 0.491 0.245
length was 50.2 (range = 1–571), and the total height of the
trees (sum of branch lengths from the root to each terminal
taxon) averaged 623 (range = 422–960 across trees). Within
each tree all taxa were at the same height from the root, so
the trees can be interpreted as a set of contemporaneous taxa
in which branch lengths are proportional to time and the rate
of character evolution is constant over the tree.
Evolution of two continuous traits (X,Y) was simulated by
Brownian motion with initial character states = (0,0) at the
root of the tree. In each time step (one time step equals one
unit of branch length), evolutionary changes in the two char-
acters were drawn from a bivariate normal distribution with µ
= 0, s2 =X s2 =Y1, and correlation coefficient (the input
correlation, following Martins and Garland 1991) p = 0 for
the null model and p = 0.5 for the correlated-changes model.
For a particular sampling algorithm or sample size, each run
of the model was based on a total of 10,000 replicates (100
character simulations on each of the 100 trees). From each
simulation one taxon sample was selected for analysis, guar-
anteeing that each replicate was independent both in terms
of the distribution of character states in the full tree and in
the sampling of taxa for analysis.
Taxon Sampling Algorithms
Six different taxon sampling algorithms were evaluated to
address the questions presented in the introduction.
Random (R). A specified number of species was selected
at random. A series of N = 16, 32, 64, and 128 were used
to examine the effects of sample size. For comparison of this
method with the next three nonrandom algorithms, the results
for N = 32 were used.
Nonrandom on character 1 (NR1). A specified number of
species (N = 32) was selected at random from those in which
the value of character X was greater than the mean value for
all taxa. This represents a case in which the species chosen
for study exhibit a larger (or smaller) than average trait value
for one of the traits of interest. For example, an analysis of
seed size versus height in trees would represent a nonrandom
sample of species with respect to height, compared to the
distribution of height in seed plants as a whole. Because there
are numerous evolutionary transitions between ‘‘tree’’ and
‘‘nontree’’ in the seed plants, any particular sample of trees
would be broadly distributed across the underlying seed plant
phylogeny.
Nonrandom on both characters (NR2). A specified num- ber
of species (N = 32) was chosen with an enforced cor-
relation (the ‘‘sampling bias’’ coefficient) between character
values. This was accomplished by randomly selecting co-
ordinates from a bivariate distribution with a specified cor-
relation coefficient (0, 0.25, 5, 0.75, or 0.9), and mean and
standard deviations equal to those observed in the simulated
data, and then selecting the species in the data that were
closest to the randomly chosen points in the X-Y space. Bi-
ological or investigator-driven scenarios that might result in
bivariate nonrandom sampling are discussed in detail below
(see Discussion).
Nonrandom with respect to a third character (NR3). Evo-
lution of a binary character (Z) was simulated along the tree,
starting with state 0 and changing from 0 to 1 or 1 to 0 with
probability 0.001 at each time step (simulations resulting in
fewer than 33 taxa with Z = 1 were discarded). This low rate
of change was selected such that less than half the ter- minal
taxa would end up in state 1 (on average) and they would be
clustered at the tips of the tree. Over the 10,000 simulations,
the number of taxa with Z = 1 averaged 84.6 (range = 33–
231). A random sample of 32 taxa was then selected from
those in which Z = 1 for analysis of correla- tions between
continuous traits X and Y. This scheme rep- resents a
situation in which sampling is concentrated in a certain
habitat or life-history group (e.g., grassland species or
annual plants), but the distribution and bivariate relation- ship
between the phenotypic traits of interest is not biased
by membership in that group. It differs from random sampling
because selected taxa are significantly clustered in groups at
the tips of the tree.
Species pairs with a large difference (DP). Phylogenetic
species pairs were selected in which the value of character X
differed by at least a specified amount (following the ap-
proach of Westoby 1999; see introduction). This was accom-
plished by starting with the first terminal taxon (at the left
side of the tree) and searching down the tree until a node
was reached in which the descendant taxa with the minimum
and maximum values for character X differed by more than
the specified amount; these two taxa were then chosen for
inclusion in the study, all other taxa in the clade excluded,
and the process was repeated in the next clade in the tree.
Only one species pair was drawn from each monophyletic
clade (paraphyletic pairs, or pairs nested within other pairs
were not considered; cf. Burt 1989; Purvis and Rambaut
1995). The minimum allowable difference was set arbitrarily
(at a value of 14.5) such that approximately 64 taxa (= 32
pairs) would be chosen, thus facilitating comparison of re-
sults with the methods above.
Random pairs (RP). Species pairs were chosen at random.
This algorithm was run in tandem with the DP selection pro-
cess above, and the random pairs were chosen from the same
set of clades in each simulation run. This approach guaranteed
that the number of pairs were the same for each replicate, so
that comparisons of the DP and RP methods were not affected
by the distribution of sample sizes.
Branch Lengths
Relative branch lengths, in terms of expected variance in
character evolution, are critical for correct application of in-
dependent contrasts (Felsenstein 1985; Martins and Garland
1991; Purvis et al. 1994). Incorrect branch lengths inevitably
raise Type I error rates, although they do not cause biased
estimates of the true correlations (Martins and Garland 1991;
Purvis et al. 1994). In empirical studies, fossil or molecular
clock data may be available to estimate branch lengths in
units of time, which would be appropriate if the characters
under study evolve at a constant rate. Alternatively, if the
rate of evolution of phenotypic characters along each branch
is assumed to be proportional to the corresponding rates for
characters used in phylogeny reconstruction (molecular or
morphological), branch lengths may be estimated based on
number of character changes in the phylogenetic dataset. In
either case, when a subsample of taxa are included in a study
of character evolution, the ‘‘correct’’ branch lengths can be
calculated for the pruned tree based on sums of branch
lengths in the original phylogeny (Fig. 1A, B).
In many (maybe most) cases, there is no basis to assume
that phenotypic traits evolve at constant or proportional rates
(Martins and Garland 1991). In addition, some methods of
phylogeny reconstruction do not provide biologically mean-
ingful branch lengths (e.g., supertree analyses; Sanderson et
al. 1998). In these situations, all branches may be assigned
equal lengths (because only relative branch lengths are im-
portant, lengths would normally be set to one; Fig. 1C). Equal
branch lengths are sometimes interpreted in terms of a punc-
tuational model of evolutionary change because there is one
‘‘bout’’ of evolution associated with each speciation event.
However, the punctuational interpretation is weaker when
considering taxon sampling because each branch represents
an unknown number of speciation events in the full phylog-
eny (even if all extant taxa in a clade are sampled, this is
true due to the unknown distribution of speciation and ex-
tinction events in the past). Because only relative, and not
absolute, branch lengths influence independent contrast anal-
yses, equal branch lengths may also be effective if the true
branch lengths are randomly distributed with respect to depth
in the tree. This is true for trees generated by time-only spe-
ciation models (although the resulting branch length distri-
butions are right-skewed; Martins 1996), and for such trees
equal branch lengths are fairly effective (Purvis et al. 1994).
Purvis and Webster (1999) have also argued that equal branch
lengths may be appropriate when there is significant error in
the estimation of species trait values, to avoid dispropor-
tionate weighting of contrasts between closely related spe-
cies, which are particularly sensitive to these errors.
Alternatively, branch lengths may be constructed using
topological approaches that lengthen branches selectively,
placing all terminal taxa at the same total height from the
root (e.g., Grafen 1989; or a ‘‘minimum-extension’’ method,
as in the graphical output from PAUP, Swofford 1993). If
all taxa under study are extant and rates of evolution are
assumed to be constant, then the placement of all terminal
taxa at the same height is intuitively appealing. However,
there is no a priori basis to expect that the relative branch
lengths constructed from these topological rules are biolog-
ically appropriate. Simulations have found that Type I error
rates are higher using Grafen’s topological algorithm in com-
parison with equal branch lengths when the true branch
lengths are derived from a random speciation model (Purvis
et al. 1994).
For the simulations reported here, branch lengths in the
pruned trees were calculated from the true lengths in the full
phylogeny or were set to equal length. All analyses were also
conducted using the two topological approaches mentioned
above (the Grafen and minimum extension algorithms), but in
almost all cases these resulted in higher Type I error rates
compared to analysis on equal branch lengths, so the results
are not shown (full results are available from the author on
request).
Statistical Analyses of Correlations
For each simulation run, the following correlations were
calculated for all taxa over the full phylogeny and for the
subset of sampled taxa over the pruned tree: (1) species cor-
relations (Rspec), based on trait values among terminal taxa (i.e.,
TIPS; Martins and Garland 1991); (2) correlations of
phylogenetically independent contrasts (R pic), which are sta-
tistically equivalent to correlations of evolutionary changes
(Pagel 1993); Rpic.true is based on contrasts calculated from true
branch lengths, and Rpic.equal is based on equal branch lengths
(equivalent to FLG and FLP, respectively, in Martins and
Garland 1991); (3) correlations of contrasts between pairs of
sister taxa at terminal nodes (R pair), thereby excluding contrasts
based on internal nodes (Fig. 1C); these contrasts were
unstandardized for branch lengths (equivalent to as-
signment of equal branch lengths between all pairs) because
the species pairs approach is employed to minimize necessary
phylogenetic information; (4) sign tests based on independent
contrasts over the full phylogeny (S pic), calculated as the
proportion of nodes with contrasts of same vs. different sign;
and (5) sign tests based on sister taxa pairs only (S pair), anal-
ogous to Rpair above. The Pearson correlation coefficient was
used for Rspec and the correlation coefficient calculated
through the origin was used for Rpic and Rpair (for details, see
Garland et al. 1992). Degrees of freedom were N — 2 for
Rspec and Rpic and NP — 1 for Rpair, where NP is the number of
sister taxa pairs (note that the total number of species in
paired tests = 2NP). Significance at the a = 0.05 and 0.01
levels was determined by a t-test for df S 30, and by the z*-
transformation for df S 31 (Sokal and Rohlf 1995, pp. 575–
579). For sign tests, significance at the a = 0.05 level was
determined from critical values of the binomial distribution
for N or NP S 22 (for Spic and Spair, respectively) and by a t-test
approximation for N or NP S 23 (Sokal and Rohlf 1995,
p. 444). The mean and standard deviation of the correlation
coefficients were calculated from each distribution of 10,000
replicates to test for bias in estimating the true evolutionary
correlation (= 0 or 0.5 in the null and correlated-changes
models, respectively). Type I error rates and statistical power
were calculated as the proportion of significant results (at P
S 0.05 and S 0.01) under the null model and the correlated-
changes model, respectively.
RESULTS
Full Trees
For the full 256-taxon trees, the mean values of Rspec and
Rpic (analyzed using true or equal branch lengths) were in-
distinguishable from 0.0 under the null model and 0.5 under
the correlated-changes model, demonstrating that both mea-
sures provide unbiased estimates of the evolutionary corre-
lation (Table 1). However, the standard error of the corre-
lations was very high for Rspec and somewhat elevated for
Rpic.equal, compared to Rpic.true. As a result, the Type I error rate
for Rspec was 0.57 (all values for Type I error and sta- tistical
power discussed in the text are based on a = 0.05; values for
a = 0.01 are also reported in Table 1). Type I error rates for
Rpic.true were close to nominal levels, whereas rates for Rpic.equal
were somewhat elevated at 0.12. Statistical power for Rpic
when p = 0.5 was close or equal to one under both branch
length algorithms, but slightly lower for Rspec (Table 1). Spic,
the sign test based on contrasts at all nodes, also had
appropriate Type I error rate and high power at this large
sample size (Table 1).
The set of 100 random phylogenies used in this study had an
average of 84.9 pairs of sister taxa (range = 78–94; the
remaining taxa were allied with clades of two or more species
so paired comparisons were not possible). Contrast correla-
tions based on these sister taxa pairs (R pair) provided unbiased
estimates of true correlations, but elevated Type I error rates
of 0.15 (Table 1). Spair, the sign test based on sister taxa pairs
only, had appropriate Type I error rates but relatively low
power (0.88) under the correlated-changes model, presum-
ably due to the reduced sample size and the conservative
nature of sign tests.
 Random Samples
For random samples ranging from N = 16 to 128, Rpic.true
was unbiased with Type I error rates close to nominal levels
in all cases (Fig. 2A). Under a true correlation of p = 0.5,
statistical power ranged from 0.54 at N = 16 to 1.0 at N =
128 (Fig. 2B). Rspec was also unbiased at all sample sizes, but
Type I error rates increased from 0.12 at N = 16 to 0.44 at N
= 128, whereas statistical power increased from ap-
proximately 0.5 to 0.95 (Fig. 2A). Thus, the probability of
incorrectly rejecting the null hypothesis using species cor-
relations increases with larger sample sizes. Sign tests based
on all comparisons (S pic) maintained appropriate Type I error
rates, but had substantially reduced statistical power at low
sample sizes (Fig. 2A, B).
The use of equal branch lengths (Rpic.equal) resulted in
slightly elevated Type I error rates, increasing from 0.074 at
N = 16 to 0.1 at N = 128, with slightly lower power relative
to Rpic.true (Table 1, Fig. 2B). For random samples, Type I
error rates were lower using equal branch lengths compared
to results based on topological alternatives (results not
shown; see Simulations: Branch Lengths). In all cases, es-
timates of the true correlation were unbiased.
Analyses of sister taxa pairs (R pair and Spair) maintained
appropriate Type I error rates at all samples sizes. However,
these samples had greatly reduced sample size because only
a portion of the species selected in each sample were paired FIG. 3. An example of the NR2 sampling algorithm. Character
evolution on a 256-taxon phylogeny was simulated with p = 0
with another species (see Fig. 1C). Based on the random tree (open circles) and then 32 taxa were chosen that fell closest to
topologies in this study, the number of pairs in random sam- points randomly chosen from a bivariate distribution with a
ples averaged approximately two-thirds the maximum pos- sampling bias coefficient of 0.5 (filled squares). For this example,
sible number (e.g., for N = 32, the average number of pairs Rspec for the full tree = —0.24 and in the sample Rspec = 0.59 and
Rpic.true = 0.39.
was 10.7 of a maximum of 16). In addition, each pair of
species provides only one comparison. Thus, although the
statistical power of Rpair and Spair were similar to correspond-
ing values for Rpic and Spic, for the same degrees of freedom, sampling bias was similar in magnitude to the underlying
they were much lower when compared on the basis of the evolutionary trait correlation.
number of species studied (Table 1, Fig. 2B). For a range of sampling bias coefficients from 0.0 to 0.9,
mean values of Rspec increased rapidly under the null model,
Nonrandom Samples whereas Rpic and Rpair increased less markedly (Fig. 4A). Type
I error rates also increased somewhat for Rspec, Rpic, and Spic,
NR1. (See sampling methods under Simulations: Taxon reaching 0.65 for Rpic.true under the strongest sam- pling bias
Sampling Algorithms.) Under the null model, sampling of 32 coefficient of 0.9. Rpair had a stable but elevated Type I error
taxa in which X > X¯ resulted in unbiased estimates of of < 0.2, and only Spair maintained appropriate Type I error
the true correlation for Rpic and Rpair, and approximately cor- rates across all values of the sampling bias coefficient (results
rect Type I error rates for Rpic.true, Spic, and Spair. However, not shown). Under a true correlation of p
under the correlated-changes model, the true correlation of
= 0.5, Rspec increased from 0.14 to 0.69 for sampling bias
0.5 was underestimated by Rpic and Rpair (both averaged ap-
coefficients increasing from 0.0 to 0.9 (Fig. 4B). Rpic and Rpair
proximately 0.38) and statistical power was diminished rel-
were less sensitive across this range. All parameters
ative to random samples of the same size (Table 1).
converged at a sampling bias of 0.5, where the sampling
NR2. Nonrandom sampling relative to both characters
pattern matched the underlying evolutionary correlation (Fig.
imposed a correlation among the two traits in the sampled
4B). Thus, if sampling processes impose a weaker or stronger
taxa relative to all taxa in the clade (e.g., Fig. 3). This was
correlation of trait values in the sample, relative to the cor-
the only sampling method, among those tested here, that re-
relation among all species in the lineage, then interspecific
sulted in biased estimates of the true correlation under the
and independent contrast correlations will underestimate or
null model; for p = 0 and a sampling bias coefficient of 0.75, overestimate, respectively, the evolutionary correlation. In-
mean Rspec was 0.59 and mean Rpic.true was 0.36 (Table 1). dependent contrast correlations are less sensitive to these
Type I error rates were elevated correspondingly for all cor- biases, but the use of a phylogenetically structured analysis
relation statistics, except for Spair (which also had very low does not eliminate these effects.
statistical power). For p = 0.5, mean correlations for a sam- NR3. When taxa were sampled based on the character state
pling bias coefficient of 0.75 were slightly elevated using of an independent binary trait, all statistical indicators were
Rspec and Rpic, although the effect was small because the similar to random samples of the same size, except that
FIG. 4. Mean correlations under the NR2 sampling regime as a function of the sampling bias coefficient. (A) Null model, p = 0. (B) Correlated-
changes model, p = 0.5. N = 32 for all samples.
Type I error rates of Rpic.equal and Rpair were very slightly higher
and power was correspondingly lower (Table 1).
Sampling of Species Pairs
For the DP sampling algorithm, the minimum difference
between taxa in character X was set at 14.5 units, which was
equivalent to approximately 0.73 standard deviations, based
on the trait distributions over the 256 species. Mean sample
sizes using this criterion were 66.8 species or 33.4 pairs, and
the mean difference in character X between species in each
pair was 21.4 units. Under the null model, Type I error rates
were approximately 0.05 for Rpic, Rpair, Spic, and Spair, but, as
in other cases, they were elevated for Rspec (Table 1).
However, for p = 0.5, mean values of Rpic and Rpair were
substantially elevated (0.64 and 0.69, respectively). Thus,
this sampling regime maintained appropriate Type I error,
but resulted in biased estimates of evolutionary trait corre-
lations when the true correlation was not equal to zero.
Random species pairs (RP), drawn from the same clades
as the DP samples (and thus with the same sample sizes),
exhibited a mean difference in character one of 7.9 units.
These samples had appropriate Type I error rates for Rpic.true,
Spic, and Spair and slightly elevated values for Rpic.equal and Rpair;
all estimates of the evolutionary correlation were un- biased
under the null and correlated-changes models. How- ever, the
statistical power of Rpair and Spair was much lower than the
corresponding values using DP sampling (0.83 vs.
0.99 for Rpair; 0.49 vs. 0.99 for Spair; Table 1).
DISCUSSION
Recent reviews of empirical studies using comparative
methods have noted that correlation coefficients (and re-
gression slopes) based on interspecific analyses and on in-
dependent contrast analyses are usually quite similar (Rick-
lefs and Starck 1996; Price 1997; Ackerly and Donoghue
1998; Ackerly 1999). This pattern is consistent with theo-
retical predictions based on the Brownian motion model, be-
cause the expected values of the interspecific and the contrast
correlations are both equal to the true evolutionary correlation
describing the underlying patterns of change (Pagel 1993).
Thus, as expected, estimates of trait correlations based on
random samples of different sizes are also unbiased with
respect to the true correlation, using interspecific or contrast
correlations. However, as has been noted in all simulation
studies on this problem, the variance in the outcome of in-
terspecific (i.e., nonphylogenetically adjusted) correlations is
much greater than for independent contrast correlations, so
Type I error rates for interspecific correlations are greatly
inflated (Martins and Garland 1991; Garland et al. 1993;
Purvis et al. 1994; Diaz-Uriarte and Garland 1996). Another
aspect of this result, which has not been previously noted,
is that Type I error rates using the interspecific correlation
increase with sample size (Fig. 2A); this occurs because the
standard deviations of Rspec declined only slightly with in-
creasing N (SD = 0.35 for N = 12; SD = 0.21 for N = 256),
but the critical values for significance testing decline rapidly
and therefore more of the outcomes are judged significant
using standard criteria. For independent contrast analyses,
the maintenance of appropriate Type I error rates for random
samples (using true branch lengths) is also consistent with
theoretical expectations. The method of independent con-
trasts is specifically designed to estimate the independent
events along each branch of the phylogeny, and pruning some
branches will not alter these estimates for the remaining taxa
and their underlying divergences, as long as the correct
branch length information is available.
Nonrandom Sampling
Three approaches to nonrandom sampling were examined
here: (1) NR1 sampled species in which the value of character
X was greater than the mean value across all species; (2)
NR2 sampled species from a nonrandom bivariate distribu-
tion (Fig. 3); and (3) NR3 sampled species with state 1 for
a discrete character Z which evolved independently along the
phylogeny with a low probability of change. NR3 resulted in
a clustering of species in groups at the tips of the phy-
logeny, but there was no bias in the distribution of character
values for the traits under study. The results for NR3 were
essentially the same as random sampling for an equivalent
sample size, except for a very slight elevation of Type I error
rates and reduction in power when using equal branch
lengths. Thus, clustering of taxa does not affect the use of
independent contrast analyses to assess trait correlations. In
practice, clustering of taxa will occur if a study focuses on
species that share a particular life-history trait or habitat af-
finity (e.g., tropical trees) and if this trait evolves slowly so
that closely related species tend to exhibit the same state. If
the trait evolves more quickly, the distribution of taxa with
a particular state will be random with respect to the phylog-
eny. In either case, this has no effect on correlations of con-
tinuous traits provided that the traits under study evolve in-
dependently of the sampling character. It is not clear whether
the same conclusion would hold for associations of discrete
trait changes based on Maddison’s (1990) concentrated
changes test.
In contrast, the NR1 and NR2 schemes represent nonran-
dom sampling with respect to the distributions of the traits
of interest. NR1 resulted in unbiased estimates and appro-
priate Type I error rates under the null model. However, for
an evolutionary correlation of 0.5, the sample correlation
coefficients underestimated the true value, ranging from 0.32
for Rspec to 0.39 for Rpic.true, and statistical power was reduced
compared to a random sample of the same size. Thus, for
example, a study of relationships between body size and me-
tabolism based only on the large species in a particular clade
would slightly underestimate the correlation between these
two traits. The underestimate of the correlation probably re-
sults from the reduced range of variation in the data, because
contrasts with large differences in X and Y would be rep-
resented with lower frequency than expected by chance. The
fact that the bias appears in Rspec and Rpic demonstrates that it
is a general result of biased sampling, although the use of
independent contrasts ameliorates the effect slightly.
Species samples that are nonrandom with respect to both
characters (NR2, Fig. 3) influence estimates of trait corre-
lations based on species values or independent contrasts, and
this was the only sampling scheme in which estimates were
biased under the null model (Fig. 4). Under this scenario the
interspecific correlation provided the best estimate of the co-
efficient underlying the sampling process, but the worst es-
timate of the underlying correlation of evolutionary changes in
the traits. Trait correlations based on contrasts of species pairs
(Rpair) were the most robust for estimation of the evo- lutionary
correlation, and only the sign test based on species pairs
provided appropriate significance testing.
In light of these results, it is important to consider the
biological processes that may result in nonrandom trait dis-
tributions or correlations in a set of species relative to the
evolutionary correlation in the history of the clade. In an
ecological context, community assembly processes may act as
filters resulting in taxon sampling patterns analogous to the
nonrandom schemes examined here. As mentioned in the
introduction, many comparative studies in plant ecology fo-
cus on species sampled from particular communities or geo-
graphic areas (e.g., the Indiana Dunes, Mazer 1989; the Shef-
field, U.K., flora, Rees 1996). The distributions and associ-
FIG. 5. Hypothetical scenario of a ‘‘grade shift’’ in which two traits
are strongly correlated within communities, but the elevations of
the relationships are displaced, leading to a much weaker overall
correlation for all species combined. See discussion in text.
ations of traits in these communities reflect both the evolu-
tionary history of the traits in the larger, encompassing
lineages and the sorting and assembly of the resulting species
into local communities. For example, Tofts and Silvertown
(2000) recently demonstrated that ‘‘environmental structur-
ing’’ led to greater similarity of coexisting species on a local
scale than expected by chance (as opposed to the prediction
of greater dissimilarity resulting from competitive exclusion).
Thus, the species in a community may exhibit a narrower
range of trait variation than would a random sample from
the local species pool or the encompassing lineage. Based on
the NR1 results, this reduction in variability would result in
underestimates of trait correlations in analyses based on this
sample of species.
It is also conceivable that the bivariate trait combinations
that are viable in a particular community will fall into a
narrow envelope, leading to tighter associations between
traits within communities than in the set of all species across
communities (Fig. 5). Associations among leaf functional
traits in species of contrasting environments illustrate this
scenario. Reich et al. (1997, 1999) found that several traits
(e.g., leaf life span, leaf mass per area, assimilation rate) are
strongly correlated across species from a broad range of hab-
itats. However, in comparisons of different communities, the
elevation of these relationships shift with climate; for ex-
ample at any particular value of leaf life span, leaves from
drier habitats are thicker. Consequently, the overall relation-
ship observed across multiple communities arises from a se-
ries of tighter relationships within communities with grade
shifts across communities. In the case of leaf life span and
leaf mass per area, the correlation across 110 species from
six communities was 0.75, whereas the correlations within
communities ranged from 0.85 to 0.92, and the elevations of
the regression lines were significantly displaced (Reich et al.
1999). When this relationship was analyzed using phyloge-
netic independent contrasts (and equal branch lengths), Rpic-
values were 0.64 for the full sample, but ranged from 0.69
to 0.93 within communities (Ackerly and Reich 1999; unpubl.
analyses). While all of these correlations are large and bio-
logically significant, it is important to note that both the
species correlations and contrast correlations were higher
within communities. These values presumably reflect the con-
straints on ‘‘allowable’’ trait combinations or fine-tuning by
convergent evolution due to environmental conditions within
each community (for a model of trait evolution with analo-
gous results, see Price 1997).
If data are available from multiple communities in con-
trasting environments, patterns of grade shifts can be easily
detected. A phylogenetically structured multiple regression or
analysis of covariance could be used to test for shifts in
patterns of correlated evolution within versus among com-
munities (for a related example, see Garland et al. 1993).
However, the situation that arises more commonly in plant
ecology is that data from species in a single community are
used to test a general evolutionary hypothesis and comparable
data are not available to test for differences among com-
munities. In a literature review of comparative plant ecology,
Ackerly (1999) found 22 studies (of 92 employing phylo-
genetic comparative methods) in which species were sampled
from a geographic region or ecological community, without
regard to phylogenetic affiliation. Thus, if trait correlations
are stronger within communities, as in the case described
above, these studies may result in overestimation of the
strength of correlated evolution in the traits under consid-
eration. Furthermore, the interspecific correlation may pro-
vide a more accurate measure of the biological process re-
sponsible for patterns within communities (e.g., community
assembly or convergent evolution).
Paired Comparisons
In comparative analyses, the method of paired comparisons
has several advantages (see Maddison 2000). It requires min-
imal phylogenetic information, and it may be applied simply
by comparing congeners (e.g., Salisbury 1942). Paired com-
parisons are also used in testing relationships between a dis-
crete, independent trait and a continuous, dependent trait
(e.g., growth analysis of congeneric annuals vs. perennials;
Garnier 1992; also see Grafen 1989; Purvis and Rambaut
1995; Martins and Hansen 1997). Another circumstance in
which the paired comparison method may be valuable is when
each pair represents a recent divergence that has occurred in
a particular habitat or under certain environmental conditions.
For example, if several wind-dispersed plant lineages colo-
nize a newly formed island, correlated changes in the size of
the seed and the dispersal structure would demonstrate pos-
sible adaptations of dispersal to island conditions. Indepen-
dent contrasts for these two traits deeper in the phylogeny
would reflect evolution under other circumstances that would
not be relevant to the question at hand. This scenario high-
lights the critical assumption of homogeneity in the evolu-
tionary process when independent contrasts are calculated
over the full phylogeny for species of interest. Methods for
detecting heterogeneity in trait correlations, using a priori or
a posteriori tests, merit further consideration (for related ex-
amples, see Garland et al. 1993; Hansen 1997).
For analyses of continuous traits, the primary disadvantage
of the paired comparisons method is the loss of statistical
power, due to the elimination of the deeper nodes in the
phylogeny and the resulting reduction in sample size from
N — 1 to N/2 comparisons. For example, the power to detect
true correlations of 0.5 (at a = 0.05) based on data for 32
species is 0.85 using 31 contrasts over a full phylogeny
(Rpic.true) compared to approximately 0.53 if the 32 species
were arranged as 16 pairs (Rpair). For sign tests, the corre-
sponding values were 0.53 for Spic compared to 0.21 for Spair.
In most cases, there is further loss of power because some
species that might be available for study are not directly allied
with a sister taxon for comparison (this limitation can be
circumvented by using paraphyletic or nested pairs; see Burt
1989; Purvis and Rambaut 1995; for algorithms for optimiz-
ing the selection of species pairs, see Maddison 2000). Wes-
toby’s method, based on the selection of pairs of species that
differ by a large amount in the independent trait (e.g., order
of magnitude differences in seed size; Armstrong and Wes-
toby 1993), increases the power of the paired comparisons
method to levels comparable to full-tree analyses (Table 1),
and Type I error rates are correct under the null model. How-
ever, the method has a significant drawback in that estimates
of nonzero correlations are inflated. Thus, the method can
only be used to determine the presence or absence of a sig-
nificant association between traits, but not to estimate the
magnitude of a correlation (or presumably the slope of a
regression, although this was not explicitly examined here).
On balance, if phylogenetic relationships among species are
known and the assumption of homogeneity in evolutionary
processes over the full tree is acceptable, it appears that there
is little to be gained from the paired comparisons method if
one is interested in both the significance and magnitude of
trait associations.
Branch lengths. Previous studies have clearly demon-
strated that appropriate branch lengths are critical for appro-
priate Type I error when using independent contrasts (Martins
and Garland 1991; Garland et al. 1992; Purvis et al. 1994).
However, there are many situations in which there is no
meaningful source of branch length information; e.g., when
only the phylogenetic topology is known or when there is no
correspondence between rates of evolution in continuous
traits and in the systematic traits used to construct the phy-
logeny. As discussed in the introduction, topological algo-
rithms may be applied to generate branch lengths that will
raise all terminal taxa to the same height (e.g., Grafen 1989);
these methods have some intuitive appeal in the study of
extant taxa because the terminal taxa appear contempora-
neous. Alternatively, all branch lengths may be set equal, in
the absence of other information. For traits evolving on a
tree generated by a Markovian branching process, both this
study and Purvis et al. (1994) found that equal branch lengths
performed better than the topological lengths based on Gra-
fen’s algorithm, although Type I error rates were inflated in
both cases compared to true branch lengths.
Garland et al. (1992) have recommended that the appro-
priateness of a branch length distribution (e.g., from molec-
ular clock or fossil data) should be tested by the regression
of standardized contrasts on the square root of the summed
branch lengths at each divergence (the slope should not be
different from zero). If this test fails (and branch lengths
cannot be adequately transformed) or if no information is
available at the outset, then it appears that the use of equal
branch lengths may be a reasonable alternative. However,
equal branch lengths do inflate Type I error rates, and this
effect has not been taken into account in empirical studies. To
correct this tendency, I have calculated critical values of the
independent contrast correlation coefficient at a = 0.1, 0.05,
0.02, and 0.01, based on the assumption that branch lengths
on the true tree resemble those generated by random
speciation models (Martins 1996) and that the analysis was
then conducted on a random species sample using equal
branch lengths (see Ackerly 2000). The use of these adjusted
critical values for the random samples in this study reduced
statistical power (under the correlated-changes model of p =
0.5) from 0.53 to 0.46 at N = 16, whereas there was virtually
no effect on statistical power at N S 64.
Sign Tests
A further conclusion of these studies is that the use of sign
tests, which do not use branch length information at all, pro-
vides a very robust and conservative approach to significance
testing for independent contrasts. Across all sampling re-
gimes examined here (except NR2), sign tests based on all
contrasts or on terminal species pairs alone (S pic and Spair)
maintained Type I error rates S 0.05. As with all nonpara-
metric tests, this robust performance is gained at the price
of greatly reduced statistical power for detecting nonzero
correlations. Sign tests are therefore very conservative and a
significant result can be confidently interpreted as indicating
a true relationship, but a nonsignificant result does not pro-
vide a strong basis for accepting the null hypothesis (cf. Kelly
and Purvis 1993). Furthermore, sign tests do not provide
estimates of the magnitude of a correlation or the sign of the
slope in regression analyses. As in all statistical analysis, the
problems of significance testing are most critical for results
close to P = 0.05 (or other relevant cutoffs). Highly signif-
icant results will usually be significant under any of the al-
ternatives, and any result that is dependent on a particular
branch length distribution or that is significant using para-
metric tests but not a sign test, should be accepted with cau-
tion.
Conclusions
The problem of taxon sampling focuses attention on the
nature of inference in comparative biology (see Westoby
1999). Scientific inference has several components: the spec-
ification of a statistical universe to which a hypothesis ap-
plies, the selection of an appropriate sample from that uni-
verse, and the choice of a statistical model for analysis. The
introduction of independent contrasts and other quantitative
methods incorporating phylogenies (Harvey and Pagel 1991;
Huelsenbeck and Rannala 1997) has led to much more ex-
plicit consideration of the statistical models underlying com-
parative analyses. However, there has been much less atten-
tion to the problem of taxon sampling, and the explicit spec-
ification of the ‘‘universe’’ of taxa to which inferences apply.
This universe may be all taxa in a particular clade, the species
of a particular ecological community or geographic region,
or taxa that share a morphological or ecological character-
istic. In each case, the appropriate sample of species will
differ, and it is necessary to be explicit about the nature of
the hypothesis and the scope of inference to determine the
appropriate sampling protocol. Westoby et al. (1998) have
provided very useful guidance in this direction by suggesting
several different forms of questions that arise in comparative
ecology and the appropriate species selection designs in each
case. The following conclusions and recommendations re-
garding taxon sampling and the implementation of indepen-
dent contrast analyses emerge from the results of this sim-
ulation study.
For studies of associations between evolutionary changes
in continuous traits, random samples of species within clades
provide robust estimates of correlation coefficients (and pre-
sumably regression slopes as well, although regression anal-
yses were not examined in this study). Therefore, there is no a
priori reason why comparative studies of correlated evo-
lution should attempt to include all known taxa in a clade
(thereby restricting the scope of inference to that lineage),
rather than subsamples drawn from larger clades.
Biological or methodological factors that lead to nonran-
dom taxon samples with respect to univariate or bivariate trait
distributions may introduce systematic discrepancies be-
tween interspecific and phylogenetically structured analyses
and bias estimates of evolutionary correlations. Examples
may include community assembly processes, convergent evo-
lution under adaptive constraints in a particular environment,
or a focus by investigators on species that exhibit the trait
combinations of interest (for an example with discrete traits,
see Maddison 1990; Sille´n-Tullberg 1993). If the bias is
in- troduced by biological causes, then it is important to rec-
ognize that two or more processes may be at work (e.g.,
correlated character evolution and nonrandom community as-
sembly processes), and comparative analyses must be care-
fully designed to isolate the effects of each process. Large
discrepancies between the results of interspecific and inde-
pendent contrast analyses may be one indication of hetero-
geneous processes influencing trait distributions (see Price
1997; Ackerly and Donoghue 1998). Methodological bias, in
contrast, should be corrected by an effort to sample taxa
randomly with respect to the traits or trait associations of
interest.
Analyses based on related species pairs entail a significant
loss of statistical power relative to the investment in data
collection, because N species provide only N/2 contrasts. Spe-
cies pairs sampling may be appropriate if patterns of evo-
lutionary divergence or functional relationships are expected to
be different in recent divergences, compared to deeper nodes
of the phylogeny, but this question remains to be ex- plored.
Barring specific rationales to the contrary or the ab- sence of
phylogenetic information at deeper nodes, indepen- dent
contrast analyses should be conducted over all available and
appropriate species for a particular question and not re- stricted
to samples of related species pairs.
If species pairs are employed, Westoby’s (1999) method
of choosing species pairs that differ by a large amount in the
independent character may be adequate for determining
whether a relationship exists between two traits. This ap-
proach maintains appropriate Type I error rates (using cor-
relation or sign tests) and statistical power is comparable to
full-tree analyses based on similar total numbers of species.
However, a major drawback of this method is that the mag-
nitude of the correlation (and possibly of regression slopes
as well) is not estimated accurately, so only presence/absence
of a relationship can be established.
Exact significance tests are highly dependent on appro-
priate branch length information (see Garland et al. 1992).
In the absence of other information, equal branch lengths
apparently maintain the lowest Type I error rates under a
variety of conditions. A table of adjusted critical values is
provided to obtain correct Type I error rates on the assump-
tion that the true branch lengths correspond to a time-only
speciation model but equal branch lengths are used for anal-
ysis (Ackerly 2000).
Alternatively, the use of sign tests provides a very con-
servative approach to significance testing. The nonparametric
sign test maintains appropriate Type I error rates under a
wide range of circumstances, but the cost is a marked re-
duction in statistical power. Thus, significant results based
on sign tests may be presented with great confidence, but
nonsignificant results do not provide a strong basis for ac-
cepting the null hypothesis.
ACKNOWLEDGMENTS
I thank M. Donoghue, J. Gittleman, J. Losos, E. Martins,
A. Purvis, T. Price, D. Schwilk, and M. Westoby for com-
ments and discussions that greatly improved this paper. I also
thank A. Purvis for suggesting the NR3 sampling scheme.
W. Maddison provided critical support in development of the
ACAP simulation and analysis software. Financial support
from the National Science Foundation (DEB 94-03252) and
a Terman Fellowship from Stanford University is gratefully
acknowledged.
LITERATURE CITED
Ackerly, D. D. 1997. ACAP2: another comparative analysis pro-
gram. Available at http://www.stanford.edu/~dackerly/
ACAP.html.
———. 1999. Phylogeny and the comparative method in plant func-
tional ecology. Pp. 391–413 in M. Press, J. D. Scholes, and M.
G. Barker, eds. Physiological plant ecology. Blackwell Scien- tific,
Oxford, U.K.
———. 2000. Table of critical values of independent contrast cor-
relations based on equal branch lengths. Available at http://
www/stanford.edu/~dackerly/archive/Rcritvals.html.
Ackerly, D. D., and M. J. Donoghue. 1998. Leaf size, sapling al-
lometry, and Corner’s rules: a phylogenetic study of correlated
evolution in maples (Acer). Am. Nat. 152:767–791.
Ackerly, D. D., and P. B. Reich. 1999. Convergence and correlations
among leaf size and function in seed plants: a comparative test
using independent contrasts. Am. J. Bot. 86:1272–1281.
Armstrong, D. P., and M. Westoby. 1993. Seedlings from large seeds
tolerate defoliation better: a test using phylogenetically
independent contrasts. Ecology 74:1092–1100.
Burt, A. 1989. Comparative methods using phylogenetically in-
dependent contrasts. Evol. Biol. 6:33–53.
Cunningham, C. W., K. E. Omland, and T. H. Oakley. 1998. Re-
constructing ancestral character states: a critical reappraisal.
Trends Ecol. Evol. 13:361–366.
Diaz-Uriarte, R., and T. Garland Jr. 1996. Testing hypotheses of
correlated evolution using phylogenetically independent con-
trasts: sensitivity to deviations from Brownian motion. Syst. Biol.
45:27–47.
Felsenstein, J. 1985. Phylogenies and the comparative method. Am.
Nat. 125:1–15.
Foster, S. A., and C. H. Janson. 1985. The relationship between
seed size and establishment conditions in tropical woody plants.
Ecology 66:773–780.
Frohlich, M. W. 1987. Common-is-primitive: a partial validation
by tree-counting. Syst. Bot. 12:217–237.
Garland, T., Jr., P. H. Harvey, and A. R. Ives. 1992. Procedures for
the analysis of comparative data using phylogenetically in-
dependent contrasts. Syst. Biol. 41:18–32.
Garland, T., Jr., A. W. Dickerman, C. M. Janis, and J. A. Jones. 1993.
Phylogenetic analysis of covariance by computer simu- lation.
Syst. Biol. 42:265–292.
Garnier, E. 1992. Growth analysis of congeneric annual and pe-
rennial grass species. J. Ecol. 80:665–675.
Grafen, A. 1989. The phylogenetic regression. Phil. Trans. R. Soc.
Lond. Ser. B 326:119–157.
Graybeal, A. 1998. Is it better to add taxa or characters to a difficult
phylogenetic problem. Syst. Biol. 47:9–17.
Hansen, T. F. 1997. Stabilizing selection and the comparative anal-
ysis of adaptation. Evolution 51:1341–1351.
Harvey, P. H., and M. Pagel. 1991. The comparative method in
evolutionary biology. Oxford Univ. Press, Oxford, U.K.
Hillis, D. M. 1996. Inferring complex phylogenies. Nature 383:130.
Huelsenbeck, J. P., and B. Rannala. 1997. Phylogenetic methods come
of age: testing hypotheses in an evolutionary context. Sci-
ence 276:227–232.
Kelly, C. K. 1995. Seed size in tropical trees: a comparative study
of factors affecting seed size in Peruvian angiosperms. Oecol-
ogia 102:377–388.
Kelly, C. K., and A. Purvis. 1993. Seed size and establishment
conditions in tropical trees: on the use of taxonomic relatedness in
determining ecological patterns. Oecologia 94:356–360.
Maddison, W. P. 1990. A method for testing the correlated evolution
of two binary characters: are gains or losses concentrated on
certain branches of a phylogenetic tree? Evolution 44:539–557.
———. 2000. Testing character correlations using pairwise com-
parisons on a phylogeny. J. Theor. Biol. 202:195–204.
Martins, E. P. 1996. Conducting phylogenetic comparative studies
when the phylogeny is not known. Evolution 50:12–22.
Martins, E. P., and T. Garland Jr. 1991. Phylogenetic analyses of
the correlated evolution of continuous characters: a simulation
study. Evolution 45:534–557.
Martins, E. P., and T. F. Hansen. 1997. Phylogenies and the com-
parative method: a general approach to incorporating phyloge-
netic information into the analysis of interspecific data. Am. Nat.
149:646–667.
Mazer, S. J. 1989. Ecological, taxonomic, and life history correlates of
seed mass among Indiana dune angiosperms. Ecol. Monogr.
59:153–175.
Page, R. D. M. 1996. TREEVIEW: an application to display phy-
logenetic trees on personal computers. Comp. Appl. Biosci. 12:
357–358.
Pagel, M. D. 1993. Seeking the evolutionary regression coefficient: an
analysis of what comparative methods measure. J. Theor. Biol.
164:191–205.
———. 1999. Inferring the historical patterns of biological evo-
lution. Nature 401:877–884.
Poe, S. 1998. Sensitivity of phylogeny estimation to taxonomic
sampling. Syst. Biol. 47:18–31.
Price, T. 1997. Correlated evolution and independent contrasts. Phil.
Trans. R. Soc. Lond. Ser. B 352:519–529.
Purvis, A., and A. Rambaut. 1995. Comparative analysis by inde-
pendent contrasts (CAIC): an Apple Macintosh application for
analysing comparative data. Comp. Appl. Bios. 11:247–251.
Purvis, A., and A. J. Webster. 1999. Phylogenetically independent
comparisons and primate phylogeny. Pp. 44–70 in P. C. Lee,
ed. Comparative primate socioecology. Cambridge Univ. Press,
Cambridge, U.K.
Purvis, A., J. L. Gittleman, and L. Hang-Kwang. 1994. Truth or
 consequences: effects of phylogenetic accuracy on two com-
parative methods. J. Theor. Biol. 167:293–300.
Rees, M. 1996. Evolutionary ecology of seed dormancy and seed size.
Phil. Trans. R. Soc. Lond. Ser. B 351:1299–1308.
Reich, P. B., M. B. Walters, and D. S. Ellsworth. 1997. From tropics
to tundra: global convergence in plant functioning. Proc. Natl.
Acad. Sci. 94:13730–13734.
Reich, P. B., D. S. Ellsworth, M. B. Walters, J. M. Vose, C. Gresh-
am, J. C. Volin, and W. D. Bowman. 1999. Generality of leaf trait
relationships: a test across six biomes. Ecology 80: 1955–1969.
Ricklefs, R. E., and J. M. Starck. 1996. Applications of phyloge-
netically independent contrasts: a mixed progress report. Oikos
77:167–172.
Salisbury, E. J. 1942. The reproductive capacity of plants. Bell,
London.
Sanderson, M. J., and M. J. Donoghue. 1994. Shifts in diversifi- cation
rate with the origin of angiosperms. Science 264: 1590–1593.
Sanderson, M. J., A. Purvis, and C. Henze. 1998. Phylogenetic
supertrees: assembling the trees of life. Trends Ecol. Evol. 13:
105–109.
Saverimuttu, T., and M. Westoby. 1996. Seedling longevity under
deep shade in relation to seed size. J. Ecol. 84:681–689.
Sille´n-Tullberg, B. 1993. The effect of biased inclusion of taxa on
the correlation between discrete characters in phylogenetic trees.
Evolution 47:1182–1191.
Silvertown, J., and M. Dodd. 1996. Comparing plants and con-
necting traits. Phil. Trans. R. Soc. Lond. Ser. B 351:1233–1239.
Sokal, R. R., and F. J. Rohlf. 1995. Biometry. 3rd ed. W. H. Free-
man, New York.
Swanborough, P., and M. Westoby. 1996. Seedling relative growth
rate and its components in relation to seed size: phylogenetically
independent contrasts. Funct. Ecol. 10:176–184.
Swofford, D. 1993. PAUP: phylogenetic analysis using parsimony.
Smithsonian Institution, Washington, DC.
Tofts, R., and J. Silvertown. 2000. A phylogenetic approach to
community assembly from a local species pool. Proc. R. Soc.
Lond. Ser B 267:363–369.
Westoby, M. 1999. Generalization in functional plant ecology: the
species-sampling problem, plant ecology strategy schemes, and
phylogeny. Pp. 847–872 in F. I. Pugnaire and F. Valladares, eds.
Handbook of functional plant ecology. M. Dekker, New York.
Westoby, M., S. A. Cunningham, C. M. Fonseca, J. M. Overton, and
I. J. Wright. 1998. Phylogeny and variation in light capture area
deployed per unit investment in leaves: designs for selecting study
species with a view to generalizing. Pp. 539–566 in H. Lambers, H.
Poorter, and M. M. I. Van Vuuren, eds. Inherent variation in plant
growth: physiological mechanisms and eco- logical consequences.
Backhuys Publishers, Leiden, the Neth-
erlands.
Corresponding Editor: J. Losos