OIKOS

16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Research article
Measuring individual-level trait diversity: a critical assessment
of methods

Oluwafemi D. Olusoji ✉1,2, György Barabás5,6,7, Jurg W. Spaak2 , Simone Fontana ,

4,8,9

Thomas Neyens1,3, Frederik De Laender2 and Marc Aerts1

1
Center for Statistics, Data Science Inst., Hasselt Univ., Hasselt, Belgium
2
Research Unit in Environmental and Evolutionary Biology (URBE), Inst. of Life-Earth-Environment (ILEE), Namur Inst. for Complex Systems
(NAXYS), Univ. de Namur, Namur, Belgium
3
L-BioStat, Dept of Public Health and Primary Care, Faculty of Medicine, KU Leuven, Leuven, Belgium
4
Nature Conservation and Landscape Ecology, Univ. of Freiburg, Freiburg, Germany
5
Division of Ecological and Environmental Modeling, Linköping Univ., Linköping, Sweden
6
ELTE-MTA Theoretical Biology and Evolutionary Ecology Research Group, Budapest, Hungary
7
Inst. of Evolution, Centre for Ecological Research, Budapest, Hungary
8
Biodiversity and Conservation Biology, Swiss Federal Research Inst. WSL, Birmensdorf, Switzerland
9
Abteilung Natur und Landschaft, Amt für Natur, Jagd und Fischerei, Kanton St. Gallen, St. Gallen, Switzerland
Correspondence: Oluwafemi D. Olusoji (oluwafemi.olusoji@uhasselt.be)

Oikos
2022: e09178
doi: 10.1111/oik.09178
Subject Editor: Simon Peter Hart
Editor-in-Chief: Pedro Peres-Neto
Accepted 31 October 2022
Individual-level trait diversity has been identified as an essential component of trait
diversity (TD), influencing community assembly and structure. Traditionally, one
employs trait diversity indices to measure facets of individual-level trait diversity (diver-
gence, richness and evenness). However, the application of species-level trait diversity
indices to individual-level traits data and their implications have not been adequately
studied. Thus, we examined the possible challenges of using four commonly used
multi-trait TD indices: Rao’s quadratic entropy (Rao), functional dispersion (FDis),
functional evenness (FEve) and functional richness (FRic); two indices primarily
developed to measure individual-level trait diversity: trait evenness distribution (TED-
for evenness) and trait onion peeling (TOP-for richnness); and a modified version of
TED (TEDM-for evenness). Additionally, we considered an index that integrates both
evenness and richness by generalizing ordinary Hill indices for traits (coined HIT). We
measured individual-level trait diversity with these indices using simulated traits data
and experimental data from a growth experiment with cyanobacteria. Comparing the
observed trends from the indices with the expected trends, we observed that only the
trait divergence indices (FDis and Rao) produced the expected trends in the simulation
scenarios and experimental data. TED and TEDM are not robust against the number
of individuals used, and FEve is not sensitive to some changes in the location of indi-
viduals in the trait space. Also, TOP proved to be a discontinuous function dependent
on the number of individuals, and FRic did not produce the anticipated trend when
changes in the trait space did not affect the edges of the trait space. HIT did produce
the anticipated changes, but it was only reliable when many individuals were sampled.
In summary, applying these individual-level trait diversity indices to quantify anything
except trait divergence may lead to misinterpretation of the original situation of trait
distribution in the trait space if their specific properties are not adequately considered.

© 2022 Nordic Society Oikos. Published by John Wiley & Sons Ltd
www.oikosjournal.org

Page 1 of 15

Keywords: divergence, evenness, Hill numbers, individual-level trait diversity, intraspecific diversity, richness, statistical
ecology
Introduction
Quantifying biodiversity is crucial because of its effect on
ecosystem functions and services (Cardinale et al. 2012).
Biodiversity comes in various forms, including phylogenic and
taxonomic diversity and trait diversity. Trait diversity (TD) is
of prime interest because it connects structure to functioning
(Tilman 1997, Hillebrand and Matthiessen 2009, Reiss et al.
2009, Cardinale et al. 2012, Gagic et al. 2015), and co-
determines ecosystem responses to environmental change
(Violle et al. 2007, de Bello et al. 2010, De Laender et al. 2016).
Often, one assesses TD using averages across the trait values
of species (Cianciaruso et al. 2009, Violle et al. 2012). This
reflects the idea that interspecific diversity is more important
than intraspecific diversity (Westoby et al. 2002, McGill et al.
2006, Carmona et al. 2015). However, over the last decades,
intraspecific variation of traits has been found both highly
prevalent and ecologically important (Bolnick et al. 2011, Des
Roches et al. 2018). For instance, intraspecific variation of
traits influences coexistence (Bolnick et al. 2011, Albert et al.
2012, Barabás and D’Andrea 2016, Hart et al. 2016) and
helps predicting species abundance, the establishment of
alien species and community resource use and biomass pro-
duction (González-Suárez et al. 2015, Wohlrab et al. 2016,
Fontana et al. 2018). For this reason, TD has been increas-
ingly studied at the individual level, in which case it is referred
to as individual-level trait diversity (ITD). Acknowledging the
potential contribution of taxonomy to trait diversity, ITD can
also be calculated within species, or integrate species abun-
dance data (Carmona et al. 2016).
Three facets of ITD exist: divergence, evenness and rich-
ness (Carmona et al. 2016, Fontana et al. 2016). These
facets are identical to TD metrics applied to species means
(Mason et al. 2005). For example, ITD evenness measures
how evenly individuals are distributed within a trait space.
ITD richness focuses on the space occupied by individuals in
the trait space, while ITD divergence focuses on trait differ-
entiation amongst individuals in the trait space. From a com-
putational standpoint, it is possible to apply available TD
indices to measure ITD (de Bello et al. 2010, Fontana et al.
2016). However, the conditions under which they can be
applied to individual-level trait data and their subsequent
interpretation are less straightforward (Mason et al. 2005,
Ricotta 2005, Petchey and Gaston 2006).
It is essential for ITD indices to signal trait changes
across different gradients (Grether 2005, Mason et al. 2005,
Petchey and Gaston 2006). This is important because envi-
ronmental change, for example, will often cause trait changes
(De Laender et al. 2016, Pacifici et al. 2017). Several authors
have carried out simulation exercises to evaluate the ability
of different TD indices to capture different trait changes
in a simulated trait space (Laliberté and Legendre 2010,
Schleuter et al. 2010, Violle et al. 2012, Fontana et al. 2016).
However, environmental change can – apart from affecting
Page 2 of 15
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
traits – also alter the number of individuals in a community.
Mason et al. (2005) therefore noted that TD indices should
be independent of the number of units on which we mea-
sure the trait(s) of interest. Else, it would become difficult to
conclude what drives changes in ITD: changes of abundance
or of traits. Walker et al. (2008), Carmona et al. (2016) and
Fontana et al. (2016) also noted this potential dependence
between TD indices, especially trait richness indices and the
number of individuals or species. Another criterion concerns
comparative TD indices. Such indices use standardized trait
data as a benchmark against which to compare the observed
trait data (Carmona et al. 2016, Fontana et al. 2016). These
standardized data should be consistent with the observed
traits data. For example, using discrete standard trait data
as a reference for continuous observed trait data is inconsis-
tent. Finally, the last property often desired for TD indices
is that the contribution of each unit (species or individual)
to the index should be weighted. When applied to spe-
cies, this weight corresponds to abundance (Ricotta 2005,
Villéger et al. 2008). Applying these indices in the context of
ITD implies that each individual contributes 1/N (the empir-
ical distribution of the traits), N being the number of indi-
viduals in the trait space, to the computation of the index.
Existing examination of TD indices in the context of
measuring ITD has been testing if a given index gives the
expected change in the light of specific discrete changes in
the trait space (two levels) (Mason et al. 2005, Villéger et al.
2008, Schleuter et al. 2010). Other examinations of TD
indices involve comparison of their performances with newly
developed indices for measuring individual-level diver-
sity (Laliberté and Legendre 2010, Carmona et al. 2016,
Fontana et al. 2016). While these examinations provide cru-
cial insights into the behaviour of TD indices, they mostly
consider specific and often substantial changes in a trait space
and not gradual changes. Because trait changes along envi-
ronmental gradients can be expected to be gradual rather
than abrupt, it is needed to evaluate the behaviour of TD
indices along such gradients.
To fill the gap identified in the paragraph above, we quan-
tified ITD along simulated gradients of trait change using
eight TD indices. To this end, we extended the scenarios con-
tained in Schleuter et al. (2010) and Fontana et al. (2016)
with scenarios of continuous changes in trait space (changes
along a grid). We critically assess seven indices that are capable
of dealing with multivariate traits, four of which were initially
developed for species mean trait data. Functional dispersion
(FDis, Laliberté and Legendre 2010) and Rao’s quadratic
entropy (Q, Rao 1982) are two divergence indices. Functional
evenness (FEve, Villéger et al. 2008) is an existing evenness
index, while trait even distribution (TED, Fontana et al.
2016) is an evenness index that was specifically designed to
analyze individual-level trait data. We considered two richness
indices: trait onion peeling (TOP, Fontana et al. 2016, devel-
oped for individual trait data) and functional richness (FRic,

Villéger et al. 2008). In addition, we also propose and evalu-
ate a modified version of the TED (TEDM), which accounts
partly for the limitations of the TED index. Finally, we also
evaluate a recently developed index based on Hill numbers
applied on individual-level trait data (Leinster and Cobbold
2012, Scheiner et al. 2016, Barabás et al. 2022), coined ‘HIT’.
For each of those indices, we assess how continuous changes
in the bi- and trivariate distribution of traits (along a contin-
uum) within a population alter these indices and if these changes
match expectations. We achieve this aim by simulating individ-
ual-level trait data in a 2-dimensional (2D) and 3-dimensional
(3D) trait space, either uniformly or normally distributed. The
uniformly distributed trait space assumes that trait combina-
tions have an equal probability of occurrence throughout the
trait space. In contrast, the normally distributed trait space
allows extreme trait values to have a lower probability of occur-
rence than trait values closer to the center of the trait space.
We also examined the performances of these indices using an
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
experiment in which we measured the growth of two strains
of cyanobacteria. This experiment offers the chance to examine
the trend in the performance of these indices beyond simulated
data. Also, our study setup enables us to examine the effects of
the number of individuals, the use of inconsistent reference trait
distribution and empirical distribution (1/N) as weights on the
capacity of these indices to measure the respective facets of ITD
they were designed to measure.
Material and methods
Trait diversity indices
Divergence indices
FDis (functional dispersion, Laliberté and Legendre 2010)
measures species-level trait divergence as the average dis-
tance of species to the center of gravity of the trait space.

(A) (B)

yi
Trait 2

Trait 2

Trait 1 Trait 1

(C) (D)
Observed
Discrete−Uniform (TED)
Continuous−Uniform (TEDM)
Density

Trait 2

Distance Trait 1

(F)
Trait 2

Trait 1

Figure 1. A hypothetical population of individuals in a 2D trait space. (A) FDis is the average of the yi’s, and each yi is the distance of a trait
value yj to that center c. (B) FEve examines the regularity of the branch length from one individual to the other in the minimum spanning
tree (the dotted lines connecting all individuals). (C) TED compares the density distribution of the Euclidean distances of the observed trait
data with the density distribution of Euclidean distances from a discrete uniform. TEDM does the same, but it uses a continuous uniform
distribution instead. (D) The volume of the black polygon (minimum convex-hull encompassing all individuals) gives the FRic. (E) TOP
is the sum of the area of all peel-able convex-hulls (black polygons).

Page 3 of 15
 16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Although developed independently based on the multivari- Evenness indices
ate dispersion of Anderson et al. (2006), it is very similar to FEve (Villéger et al. 2008) measures evenness by considering
Rao’s Q (Champely and Chessel 2002). FDis is computed by the regularity of the distances between species in a trait space.
weighing the Euclidean distance or other distance measure Consider a branch  with length d  from a minimum span-
(Manhattan) to the center by the relative abundance of the ning tree connecting the trait values of two species 1 and 2;
species: species with relative abundances w and w respectively;
1 2
and define:
é
å ù
S
ê w j xij ú d
j =1
c = [c i ] = ê ú EW =
w + w
å
S
ê wj ú 1 2
êë j =1 úû
The total number of branches is S − 1, with S the number of
å
S
wj yj species. FEve is then defined as:
j =1
FDis = , yj = x j - c
å
S
wj S -1
æ ö
ç 1 ÷
å
j =1
S -1 EW 1
FEve = min ç , ÷-
where i is the counter for the traits, wj is a vector of relative S -2
å - -1
S -1
=1 ç EW S 1 ÷ S
abundances, xj is a vector representing the position of the è =1 ø
jth species in the trait space, S is the number of species and
FEve ensures that each species contributes proportionally to
c is the centroid of the observed positions in the trait space
the computation of the indices by using the relative abun-
(Fig. 1A). When applied to individuals, FDis translates to:
dance of the species as weights. When applied to individuals,
FEve measures the regularity of the distance between indi-
é1 N ù viduals in a trait space. Since every single individual has an
c = [c i ] = ê
êN
ë
åx úúû
j =1
ij abundance equal to 1, FEve becomes:

N -1
æ ö
ç 1 ÷
å
N N -1 d 1
÷-
åy
1 FEve = min ç , (3)
å
FDis = (1) N -2 d N - 1 ÷ N - 1
N -1
N
j
=1 ç
j =1 è =1 ø
with N the number of individuals and d  denoting the length
where yj the Euclidean distance of the jth individual to the of branch  (panel B in Fig. 1).
center of the (observed) trait space and N the number of indi- TED (Fontana et al. 2016) is a relative comparison
viduals in the trait space. between the distribution of Euclidean distances among indi-
Rao’s quadratic entropy (Q) (Rao 1982) expresses inter- viduals in an observed trait space to that of a hypothetical or
specific divergence as the average difference between species. reference trait space filled with evenly distributed individu-
It is computed by weighing all pairwise distances between als (a discrete regular grid). TED was explicitly designed to
species in the trait space by their relative abundances: quantify ITD and is essentially a measure of how close the
distribution of the Euclidean distances of the observed traits
Q = w ¢Dw is to that of the Euclidean distances of the reference distribu-
tion (Fig. 1C). This is achieved by using the Kullback–Leibler
D is a S × S matrix of squared distances between species, S divergence (Kullback and Leibler 1951), measuring a statisti-
being the number of species and w is the vector of relative cal difference between two distributions:
abundances. When applied to individuals, Q measures the
average difference between individuals in a trait space and 1
TED = (4)
can be rewritten as: 1 + KLDiv
where KLDiv is the asymmetric Kullback–Leibler divergence
N N
of the observed trait data and reference distribution.
N
1
Q= 2 åå || x k - x j ||2 (2) We also introduce TEDM, similar to the TED index,
k =1 j = k +1 but with two modifications. The first modification involves
replacing the even distribution based on a regular grid with
since each individual has a relative abundance of 1/N; xk and simulated trait values from a continuous uniform distribution
xj are trait vectors of the kth and jth individual, N is the total (Gregorius and Gillet 2021). We do so because the continu-
number of individuals in the trait space. ous uniform distribution better characterizes the definition

Page 4 of 15

of trait evenness. Moreover, the comparison in TEDM is
made between two continuous distributions (in contrast to
TED, which uses a discrete reference distribution). The sec-
ond modification involves limiting the comparison in even-
ness to the space covered by the observed trait values. This
modification ensures that the comparison does not include
values of the reference distribution outside the observed trait
space (see the Supporting information for more explanation
on these modifications). It is important to note that different
references could be generated for the same dataset. To avoid
this, we encourage computing TEDM over several simulated
reference (this is potentially computationally intensive) or set
a seed before generating the reference distribution.
Richness indices
FRic (Cornwell et al. 2006) measures species trait richness as
the (hyper-)volume or area covered by individuals in a trait
space (Fig. 1D). It is obtained by computing the volume of
the smallest convex hull containing all individuals in the
trait space. It can be applied to measure individual-level trait
richness without modifications. Note that this index only
involves species/individuals at the edges of the trait space.
TOP (Fontana et al. 2016) is developed to assess ITD
and measures trait richness as the total area covered by all
individuals in the trait space (not only those spanning the
outer convex hull). It is obtained by summing areas of con-
vex hulls peeled off from the individuals in the trait space
(Fig. 1E). By summing all successive areas of the peeled-off
convex hulls, TOP can detect trait changes at and within the
edges of the trait space. It is conceptually similar to the FRic,
and its computation involves all individuals in the trait space.
Consequently, TOP is sensitive to changes involving individ-
uals across the whole trait space:
H
TOP = åA k =1
k (5)
where Ak is the area of the kth peelable convex hull from the
trait space, and H is the number of hulls.
A joint measure of richness and evenness based on
Hill indices
For categorical data where each category is represented as a
fraction, Hill numbers offer a flexible and widely used way
of measuring diversity. Given C distinct categories with frac-
tions pi in category i (for instance, the categories could be
species and the pi their relative abundances), the Hill number
D(q) of order q is obtained as
1
æ C ö 1- q
å
D( q ) = ç piq ÷ (6)
ç ÷
è i =1 ø vector of values, but raising a vector to some power is inter-
(Leinster and Cobbold 2012). Various choices for q return
different well-known diversity indices. For q = 2, one recovers
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
the inverse Simpson index, while the exponential of the
Shannon index obtains in the limit of q→1.
This index does not immediately generalize to measuring
trait diversity, for the following reason. A good measure of
diversity ought to reflect the fraction of trait space covered
by individuals, as well as the evenness of this cover. But any
number of discrete individuals, each of them a single point
in trait space, will form a set of volume zero and thus techni-
cally cover a fraction zero of that space. To get around this
problem, we must treat the observed individuals as samples
from an underlying trait distribution. This distribution can
be estimated either using regression based on theoretical
expectations (e.g. one might expect some trait to be approxi-
mately normally distributed), or via more agnostic methods
such as kernel density estimation (Carmona et al. 2016,
Barabás et al. 2022). Here we opt for the latter procedure.
This will yield a trait probability density function D(x) ,
whose value multiplied by an infinitesimal dx0 at a given
point x0 is the probability that a randomly sampled individ-
ual from the underlying population has a trait value falling
within an infinitesimal radius dx0 of x0.
Once D(x) is obtained, it is tempting to directly apply
Eq. 6 by replacing the summation with an integral. However,
it is known that this procedure does not work; among other
problems, it leads to a diversity metric that is not dimension-
less (Barabás et al. 2022). Instead, the way to go is to subdi-
vide the trait space into C equally-sized grid cells, evaluate
D(x) in the center of each, obtain the population’s diversity
using Eq. 6, and then take the C→∞ limit. To prevent the
index from diverging to infinity (which it would, given the
ever-larger number of categories C introduced by refining the
grid resolution), one can normalize the metric by dividing
with C. This will yield a finite result, and corresponds to com-
paring the diversity of the observed population with that of
another whose trait distribution is uniform throughout the
trait space (Barabás et al. 2022). The diversity index formula
of order q thus reads
1
é æ ö
q ù 1- q
1ê
C
ç ÷ ú
å
Di
D(q ) = lim ê ç ÷ ú (7)
C ®¥ C ê
å ú
C
i =1 ç Dk ÷
ê è ø ú
ë k =1
û
where Di is the value of D(x) evaluated at the center of cell
i. We call this class of metrics Hill indices for traits (HIT).
The same formula can be written in an alternative form by
introducing the vector p whose ith entry is Di . Then, with a bit
1/(1- q )
é qù
of algebra, Eq. 7 can be expressed as D(q ) = ê p / p ú
êë
( )
úû
(Barabás et al. 2022), in the limit of p having infinitely many
entries. Here the overbars denote the arithmetic mean of a
preted entrywise. Keeping the number of entries in p large
but finite, this alternative form of D(q) offers a straightfor-
ward numerical implementation of Eq. 7.
Page 5 of 15

(A) (B)
Trait 2
Trait 2
Trait 1
Density
0 5
Figure 2. Trait diversity for a two-dimensional trait distribution formed by 200 individuals (A, points). The individual data are used to
estimate the underlying trait probability density function (B) via kernel density estimation (Carmona et al. 2016). The trait space is then
subdivided into a grid (C); the density is approximated as constant within each grid cell. The HIT index of functional diversity can now be
evaluated using Eq. 7. The finer the grid resolution, the more accurately the computed index converges to its true value. In this example,
setting q = 2 and using 25 grid points along both dimensions leads to C = 252 and D(q) = 0.142. In turn, C = 1002 yields D(q) = 0.151,
while C = 10002 results in D(q) = 0.153.
HIT inherits the properties of Hill numbers, now applied
to individual-level trait data. It yields higher values if more
of the trait space is covered, and if that cover is more even.
The parameter q can be used to adjust the sensitivity of the
metric; here we fix q to be 2, leading to the usual inverse
Simpson diversity index. See Fig. 2 for an illustration of how
one computes this metric.
Simulation study
We evaluated the behaviour of these indices following contin-
uous changes of individuals in trait space using simulations.
We considered three scenarios: for the first two scenarios, we
simulated populations composed of 1000 individuals located
in a 2D and 3D trait space, using 2D and 3D uncorrelated
normally or uniformly distributed trait values. For the third
scenario, we simulated between 50 and 4000 individuals. The
three scenarios are defined as follows:
Scenario one. Here, we shift the trait values of one part of
the population, akin to disruptive selection, i.e. the formation
of a second cluster in a trait space. We started with a bivariate
normal trait space (two uncorrelated traits with a mean value
of 5 and 10 on axis one and axis two respectively, with a vari-
ance of 0.85 for both dimensions), and gradually shifted the
trait values of 20% of randomly selected individuals. (addi-
tion of δ = 0.5 to their trait values along all trait axes). We
repeat the same action in a 3D trait space, but with a mean
value of 5 on each trait axis, the same variance as in the two-
dimensional trait space, and a correlation of 0 between the
traits. Figure 3A–C show three sample cases from scenario 1.
Scenario two. Here, we (re)move individuals that are
0–50% (in steps of 10%) closest to the centroid of the trait
space. This represents selection against trait combinations at
the center of the trait space. We consider trait values from a
bivariate and trivariate uniform trait space. In scenario two
Page 6 of 15
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
(C)
Trait 2
Trait 1 Trait 1
10
A, we kept the number of individuals constant and only sam-
pled individuals away from the distribution’s center. We do
this in the following steps:
1) take a sample of size n individuals from the desired trait
space (2D or 3D uniform)
2) calculate the mean trait along each trait axis to form a vec-
tor of mean trait values
3) obtain the Euclidean distance of each individual in step 1
to the mean trait vector in step 2 and order these distances
4) obtain the 10th–50th quantiles (q) of the euclidean dis-
tance distribution. Per quantile, discard all observations
with a Euclidean distance lesser than the current quantile
5) per quantile, sample another individual from the trait
space as in step 1 and check if its Euclidean distance to
the mean vector is greater than the current quantile. If
it is, keep such sample; if it is not, discard it. Continue
sampling until the number of individuals is again n.
In scenario two B, we followed the same algorithm but
without the last step. Hence, we do not re-sample individuals
from the trait space, but we simply remove the individuals
with Euclidean distances less than the particular q quantile.
We did this to evaluate the case where the sample size is not
constant (scenario two B) and where it is constant (scenario
two A). Figure 3E–F show three sample cases from the first
version of this scenario.
Scenario three. Here, we randomly added individuals to a
bivariate and trivariate trait space in which individuals were
normally (scenario three A) or uniformly (scenario three B)
distributed with a mean value of 5 and 10 on axis one and
axis two respectively. The variance was 0.85 for all trait axes.
For the trivariate trait space, we used exactly the same setup
as in scenario 1, i.e. a trait mean of five on each axis, a vari-
ance of 0.85, and a correlation of 0. We kept the trait space
unchanged (by fixing its mean and variance) as individuals
 16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
A B C D
Richness
Divergence

Index value

Scenario One
Trait 2

Evenness

Trait 1 Location Shift

E F G H

Divergence

Index value

Scenario Two
Trait 2

Evenness
Richness
Trait 1 % Shifted
I J K L

Divergence
Evenness

Scenario Three
Index value

Richness
Trait 2

Trait 1 Number

Figure 3. Illustrations of the three simulated scenarios and the expected trend for each component of individual-level trait diversity (ITD).
Scenario 1 (row 1) represents trait expansion. (A), (B) and (C) show different degrees of expansion. Scenario 2 (row 2) represents selection
against trait combinations near the center of the trait space. (E), (F) and (G) show varying degrees of selection. Scenario 3 (row 3) represents
a scenario with constant trait probability distribution, but with varying number of individuals. (I), (J) and (K) shows the trait spaces with
200, 500, 1000 individuals respectively. (D), (H) and (L) present the expected trends of ITD in each scenario.

randomly entered into the trait space such that population
size grows from 50 to 4000 individuals. This situation repre-
sents immigration or reproduction events. We present three
example cases from scenario three A in Fig. 3I–K.
Expected trends from the three scenarios
In scenario one, we expect trait richness and divergence to
increase because shifting 20% of the individuals away from
the centroid should result in increasing trait differentiation
(divergence) and trait space expansion (richness). Since we
are sampling individuals with normally distributed trait val-
ues, we expect an initial increase in evenness because indi-
viduals moving away from the centroid of the trait space will
initially make the trait space more even (i.e. the distribution
more uniform) compared to the initial trait distribution.
However, we expect evenness to decrease as soon as clusters
start forming (Fig. 3B–C).
In scenario two, we expect trait divergence to increase
while trait evenness and richness would decrease. This is

Page 7 of 15

because the movement of individuals away from the cen-
tre of the trait space (scenario two A) or increasing loss of
individuals at the centre of the trait space (scenario two B)
increases trait dissimilarity. This makes individuals no longer
evenly spread across the trait space (with an increasingly sized
empty centre of the trait space) (Fig. 3E–G). For trait rich-
ness, we expect a decreasing trend in scenario two A because
the movement of individuals away from the centre of the trait
space results in an overall loss of trait combinations at the
centre. While one can argue that the new individuals entering
into the trait space compensate for trait combinations lost at
the centre, these new individuals would impact trait richness
less since they introduce trait combinations similar to those
already present in trait space. In scenario two B, we expect a
decreasing trend since we increasingly remove individuals at
the centre of the trait space and do not add new individuals
to compensate for lost individuals.
In scenario three, we expect the indices to remain stable
as the number of individuals increases (Fig. 3I–K). That is
because, ideally, a trait index should not change with the
number of individuals in the trait space. In case a trend
would appear, we expect this trend to be asymptotic with
respect to the number of individuals, since the impact of
adding new individuals drops as the trait space becomes
more populated.
Furthermore, we expect FDis and Rao to perform simi-
larly in all the simulation scenarios based on the relationship
between both indices (Champely and Chessel 2002). Also,
we expect FRic to not produce any trend in both cases of
scenario two since we leave the edges of the trait space intact.
For each scenario, we performed fifteen replicates. Thus,
per scenario, we computed the indices fifteen times. Larger
numbers of replicates were not feasible as the calculation of
some indices (notably FDis, Rao, FEve and FRic) was com-
putationally too intensive. However, this number of repli-
cates gives us an idea of the variability of these indices within
each scenario.
Cyanobacteria invasion experiment
To apply these indices to individual-level data, we used data
from a flow cytometry (FCM) experiment involving two
strains of Synechococcus cyanobacteria (BS4 and BS5 with
accession numbers RCC2434 and RCC2385 respectively).
The samples were grown under 12H/12H day/night cycles at
20°C. The cyanobacteria remained in monoculture samples
until they achieved carrying capacity. Afterward, they were
invaded with the other strain. Thus, BS4 was invaded with
BS5 at carrying capacity and vice-versa. FCM was used to
measure the presence of chlorophyll a (488 nm excitation,
695 nm fluorescence), phycoerythrin (488 nm excitation, 583
nm fluorescence), phycocyanin (642 nm excitation, 664 nm
fluorescence), cell size (forward scatter) and cell granularity
(side scatter). We refer to Spaak et al. (2020) for more details
on this experiment. We computed the indices using data in
which BS5 is the resident species and we used three traits: cell
size, chlorophyll a and phycoerythrin.
Page 8 of 15
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Expected trends from the cyanobacteria invasion
experiment
The experimental setup resembles simulation scenarios
one and three. The pre-invasion part of the experiment
resembles scenario three since only changes in the number
of individuals is expected during this phase of the experi-
ment. The post-invasion part of the experiment resembles
scenario one since the invading species will expand the trait
axis, unless it shares trait values with the resident species.
Thus, we expect that the invasion of the resident strain
should increase trait richness and divergence but decrease
trait evenness. While both strains have similar cell size, the
resident BS5 strain is known to have higher levels of phy-
coerythrin and chlorophyll a than BS4 (Stomp et al. 2004).
Thus, the invading BS4 will expand the trait space along
this trait axis. At the same time, we expect no change along
the cell size axis (see the Supporting information for posi-
tions of the strains along the phycoerythrin and chlorophyll
a channels).
Calculation of indices and settings
We performed all simulations using R (www.r-project.org),
ver. 4.0.2. We used the dmvnorm function in the mvt-
norm package (Mi et al. 2009) to simulate trait values in
the bi- and trivariate normal case, and the runif function
in base R to simulate trait values for the uniform case.
To compute the TED, we simulated the reference trait
matrix from a 2-dimensional square using the geozoo
package (Schloerke et al. 2016). For TEDM, we simu-
lated the reference trait matrix using the runif function.
We used the FD package (Laliberté and Legendre 2010) to
compute FEve, FRic, FDis and Rao, and the geometry
package (Barber et al. 2015) to compute convex hulls for
TOP and TEDM. We also computed all kernel densities
for the TED and TEDM using the density function in the
stats package (Duong 2007). To compute D(q), we used
the ks package to compute the required kernel density
(Duong 2021).
Results
ITD divergence
Both trait divergence indices (Rao and FDis) reproduced
the expected positive effect of disruptive selection (scenario
one) and selection against average trait values (scenario two)
(row one-three, column one in Fig. 4) on individual-level
trait divergence. Increasing the sample size (scenario 3) did
not affect the divergence indices (row four-five, column one
in Fig. 4). Also, the differences in the number of individu-
als in scenarios two A and two B did not disrupt the per-
formance of the two divergence indices. As expected, both
indices have a larger standard deviation when the number
of individuals is small.
 16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ITD evenness index (FEve) did not exhibit this response; it remained con-
stant for minor location shifts and then decreased mono-
Two evenness indices (TED and TEDM) produced the tonically. Selection against central trait values (scenario
expected unimodal response of trait evenness to disruptive two) led to the expected decline for TED and TEDM in
selection (scenario one). Both indices initially increased but an almost identical way. In contrast, FEve did not show
then decreased with selection, but more outspokenly for the expected decreasing trend (row two-three, column two
TEDM (row one, column two in Fig. 4). The third evenness in Fig. 4). Also, in scenario two B, the averaged values of

FDis TPDiv TOP TED HIT

Index: Rao FRic FEve TEDM

Scenario One Scenario One Scenario One Scenario One

Divergence Evenness JRE Richness
1.00
0.75
0.50
0.25
0.00
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Location shift
Scenario Two A Scenario Two A Scenario Two A Scenario Two A
Divergence Evenness JRE Richness
1.00

0.75

0.50

0.25

0.00
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Percentage Replaced
Rescaled Index Value

Scenario Two B Scenario Two B Scenario Two B Scenario Two B

Divergence Evenness JRE Richness
1.00
0.75
0.50
0.25
0.00
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Percentage Removed
Scenario Three A Scenario Three A Scenario Three A Scenario Three A
Divergence Evenness JRE Richness
1.00
0.75
0.50
0.25
0.00
50 100 200 500 1000 2000 4000 50 100 200 500 1000 2000 4000 50 100 200 500 1000 2000 4000 50 100 200 500 1000 2000 4000
Number of individuals
Scenario Three B Scenario Three B Scenario Three B Scenario Three B
Divergence Evenness JRE Richness
1.00
0.75
0.50
0.25
0.00
50 100 200 500 1000 2000 4000 50 100 200 500 1000 2000 4000 50 100 200 500 1000 2000 4000 50 100 200 500 1000 2000 4000
Number of individuals

Figure 4. Mean ± standard deviation of the indices’ values from the 2D trait space. The plots are obtained by re-scaling the computed
indices to the 0–1 interval before averaging, i.e. each point is an average of fifteen re-scaled index values, while the vertical lines represent
the standard deviation of the indices across the fifteen replicates. The trend line in each plot panel is obtained from a generalised additive
model (GAM) fit. Formal statistical test for trend in these indices is reported in the Supporting information. We computed HIT with q = 2.
JRE = joint measure for richness and evenness.

Page 9 of 15

FEve showed some limited decrease. Adding individuals
did affect all three evenness indices, contrasting our expec-
tations. These effects were both positive (scenario three B)
and negative (scenario three A) and were smallest for FEve
(Fig. 4). Nevertheless, the results also suggest an asymptotic
behaviour for scenario three B. The responses of these indi-
ces to abundance changes seizes at high abundance (around
500 and 1000 individuals for FEve and TEDM, respec-
tively; around 2000 individuals for TED). FEve varied most
across all scenarios because it has the largest variance (dis-
played by the length of the vertical lines in Fig. 4). Also, the
similarity in the performance of the TED and its modified
version suggests that using different types of uniform refer-
ence distributions does not affect much the performance
of the index. Across all the scenarios, HIT produced the
expected richness trends. While this index also produced
the expected evenness trend in scenario two A and B, this
was largely due to the strong relationship between evenness
and richness in these scenarios.
ITD richness
The two richness indices reproduced the expected increase
of trait richness following disruptive selection (scenario
one). However, removing individuals at the centre of the
trait space (scenario two) only led to the expected trait rich-
ness decrease for the TOP index, in particular for scenario
two B. This difference between 2a (two A in Fig. 4) and 2b
(two B in Fig. 4) suggested some sensitivity to the number
of individuals. Indeed, both indices increased when abun-
dance increased (scenario three). In scenario three A, the
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
increase was linear for FRic and quadratic for TOP, but
the difference between the curve was limited. In scenario
three B, the dependence on the number of individuals dif-
fers among richenss indices, i.e. convex for TOP and con-
cave for FRic; also, FRic seemed to reach an asymptote after
2000 individuals, which, as expected, can be addressed to
the bounded space of the uniform distribution. HIT cap-
tured the increasing richness patterns in scenario one. Also,
the HIT index is less influenced by the number of individu-
als in scenario three A.
Cyanobacteria invasion experiment
The two divergence indices captured the expected change
in divergence since the overall divergence (squares in Fig. 5)
is consistently larger than the divergence for either strain.
Furthermore, the invading BS4 strain has larger trait diver-
gence compared to the resident BS5 strain despite having
a lower number of individuals. This is due to the higher
co-variance in the trait values of the BS4 strain across the
two trait channels (Supporting information). The three
evenness indices captured the expected decrease in even-
ness. However, TED and TEDM appear to be the more
sensitive to the changes in the trait space caused by the
invading strain (FEve has the shortest distance between the
squares and either of the circles and triangles in Fig. 5).
This aligns with the findings from simulation scenario
one, where the slope of FEve is much smaller than that of
TED and TEDM. The richness indices also captured the
expected increase in trait richness caused by the invading
strain (Fig. 5).

Indices
FDis FRic FEve TEDM Strain BS4 BS5 Combined
Rao TOP TED HIT

Divergence Evenness JRE Richness

1.00
Rescaled Index Value

0.75

0.50

0.25

0.00
4 10 15 44 47 51 58 4 10 15 44 47 51 58 4 10 15 44 47 51 58 4 10 15 44 47 51 58

Time(Days) Resident Strain = B5

Figure 5. Individual-level trait diversity (ITD) measured in a cyanobacteria growth experiment. The ‘combined’ values are values of ITD
when computed across the two strains, i.e. at the community level. Instead, the circles and triangles represent ITD computed within each
strain. We computed HIT with q = 2. Note that the invading B4 has fewer individuals compared to the resident species B5 during the inva-
sion. Lines are for clarity and do not reflect any trend.

Page 10 of 15

Discussion
Trait divergence indices should reflect dissimilarity in trait
combinations across an entire trait space (Mason et al. 2005),
i.e. a trait space with more different trait combinations is
considered more diverse. The trait divergence indices FDis
and Rao, which we examined, captured continuous trait dis-
similarity amongst the simulated individuals and those from
the experiments. Furthermore, these indices proved robust
against the number of individuals used in the simulations and
the number of individuals counted in the experiments. While
the trait divergence values for the combined case with more
individuals are larger than those from either of the strains,
this does not imply that more individuals mean higher diver-
gence because the trait divergence indices for BS4 are con-
sistently larger than those of BS5. This is true despite BS5
having more individuals.
As earlier discussed, both trait divergence indices weigh
for abundance, a property that has been criticized because
such indices are likely to give small weights to rare species
(Walker et al. 2008), but this is not the case when we use these
indices to measure ITD. Results from simulation scenarios
three A and three B support this conclusion. Both indices
give the weight 1/N for all trait values or vectors, regardless of
the distribution of individuals in the trait space. This implies
that extreme trait values have the same weight on both indi-
ces as the less extreme ones. It is essential to mention that,
provided the distance metric employed in FDis is Euclidean,
FDis is fundamentally the same as Rao’s quadratic entropy
(Champely and Chessel 2002). Thus, it is not surprising that
both indices behaved similarly across all the scenarios consid-
ered in this study.
Trait evenness indices should measure how evenly indi-
viduals are distributed within a trait space (Mason et al.
2005, Villéger et al. 2008). The closer the distribution of
individuals in a trait space is to a uniform distribution, the
more evenly individuals are distributed across that trait space
(Mason et al. 2005). In essence, evenness is about uniformity.
Two of the three evenness indices considered in this study
are relative indices (TED and TEDM), i.e. they compare the
observed trait data to a reference. We already established that
TED uses simulated trait values from a discrete uniform trait
space as its reference. Thus, it will only be entirely consistent
in its comparison if the trait data under comparison comes
from a discrete trait space. However, using a discrete uniform
distribution as a reference instead of a continuous trait space
does not hinder TED’s performance. TED and TEDM did
capture continuous changes in trait evenness in situations of
trait space expansion in scenario one, and both cases of sce-
nario two involving loss of individuals in the center of the
trait space. Also, these results point out that the distribution
of distance between trait vectors in the discrete uniform refer-
ence of TED is a sufficiently good approximation to that of
TEDM (with simulated reference from a continuous uniform
distribution). Also, the results from the cyanobacteria experi-
ment confirmed TED and TEDM to be more sensitive to
the changes caused by the invading strain as compared to the
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
FEve. Schleuter et al. (2010), Fontana et al. (2016), Legras
and Gaertner (2018), Kosman et al. (2020) also documented
other scenarios where FEve did not produce the expected
changes in a manipulated trait space. However, FEve offers
the advantage of being less influenced by the number of
individuals. The simulations show that this effect however
reaches an asymptote at about 2000 individuals in case of
the uniform distribution. This is true in both the simulated
individual-level data and the cyanobacteria experiments.
The difference in performance between TED, TEDM and
FEve is explainable by their approach to measuring evenness
in a trait space. FEve responds to changes in the regularity of
the distances (branch length) between individuals in the trait
space (Villéger et al. 2008), while TED and TEDM respond
to changes in distance distributions (Fontana et al. 2016).
FEve is highly dependent on changes in the branch length
of the minimum spanning tree employed in computing it.
If the changes in branch lengths are minimal, for example,
in the case of individuals becoming more evenly distributed
due to initial trait expansion in scenario one, FEve will most
likely miss such changes. On the contrary, TED and TEDM
employ a reference distribution to which they compare the
observed traits data. Thus, it is more sensitive and does
pick up the small changes in trait evenness missed by FEve.
Indeed, it is important to stress that the performance of TED
and TEDM might depend on the metric used for compar-
ing trait distributions since different comparison metrics can
be used (Fontana et al. 2016). The choice of which index to
employ to measure individual-based trait evenness is a choice
between sensitivity and the number of individuals. TED and
TEDM are more sensitive than FEve, whereas FEve is more
robust against the number of individuals used in computing
it. However, bootstrapping or refraction techniques can be
employed to correct for the dependence of TED and TEDM
on the number of individuals in the trait space. Also, although
TED performs similarly to TEDM, we cannot exclude that
in other scenarios, the discrete reference might be less suitable
as an approximation.
Fontana et al. (2016) introduced the TOP index and
showed that it always increases with the appearance of unique
trait combinations in any part of the trait space (monotonic-
ity). TOP did perform as expected in all scenarios considered.
However, we demonstrated in the Supporting information
that this index is discontinuous at points where individuals
move from a smaller convex-hull to a larger one (Supporting
information). At these discontinuous points, the red point
in Supporting information is on a convex-hull with a larger
area than the previous convex-hull it belonged to, which dis-
continues the increasing TOP trend observed initially. This
feature might make the TOP potentially problematic for
studying trends in individual-level trait richness. However,
simulation results and the cyanobacteria invasion experiment
do not show the discontinuity property as a problem. It is
important to note that the discontinuity is proportional to
the distance between the convex hulls. Thus, if there are many
individuals and hulls, for example, in phytoplankton systems,
these discontinuities may not result in significant jumps.
Page 11 of 15

TOP producing no clear trend in scenario two A implies
that the effect of the new individuals entering the trait space
to replace those lost at the centre of the trait space is insignifi-
cant. The newly simulated individuals brought in trait com-
binations that were similar to those already existing in the
trait space, resulting in similar trait richness. As the effects of
these new individuals with similar trait values to existing ones
wear off, trait richness reduces, which explains the reduction
in TOP when we replaced 40–50% of the individuals at
the centre of the trait space. It is well established that FRic
remains unchanged when unique trait combinations appear
within a trait space, as it is susceptible to extreme trait values
only (Cornwell and Ackerly 2009, Schleuter et al. 2010,
Legras et al. 2018). This explains the inability of this index
to produce the expected trend in both cases of scenario two.
While the index does not produce any trend in scenario two A,
which might suggest good performance as in the case of TOP,
this is because the edges of the trait space remain unchanged
in this scenario. This feature makes FRic less suitable for
measuring individual-level trait richness. It should instead be
defined as a multidimensional range (Schleuter et al. 2010,
Fontana et al. 2016). Despite these shortcomings, results
from scenario one and the cyanobacteria experiment confirm
that FRic can be applied in studies that address trait-based
niche expansion and trait ranges (Pigot et al. 2016).
It is difficult to imagine a trait richness index that does
not change with the number of individuals in the trait space.
More specifically, it is essential to note that the dependence
on the number of individuals is implicit in the definition
of individual-based trait richness. Thus, it is no surprise
that both TOP and FRic increase with increasing individu-
als. However, the impact of the number of individuals on
ITD, in general, is conditioned on the uniqueness of their
trait combinations. As we argued in the introductory part
of this paper, the strong association between the number of
individuals and ITD richness implies that disentangling the
effect of pure trait richness from the effect of sample size
becomes difficult. This becomes clearer when we consider
the results from the cyanobacteria experiment. It is difficult
to conclude that the higher trait richness in the combined
case of Fig. 5 is due to trait expansion and not due to the
increasing number of individuals because the trait richness
indices increase exponentially with increasing number of
individuals. To remedy for this, bootstrapping procedures
(Fontana et al. 2016), which are also employed in rarefac-
tion (Walker et al. 2008), can be applied to make trait
richness independent of the number of individuals. FRic
reaching an asymptote in scenario three B results from the
uniform trait space we employed in this case. Since we were
sampling individuals from a uniform distribution between
0 and 1, there are fewer changes at the edges of the trait
space with increasing sample sizes, implying no change in
FRic. However, this is not the case for TOP as it can detect
changes everywhere in the trait space (Fontana et al. 2016).
The choice of which index to employ to measure individ-
ual-based trait richness depends on the research question.
However, such research should ideally standardize for the
Page 12 of 15
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
number of individuals since the two richness indices exam-
ined will increase with increasing abundance.
HIT balanced between expected richness and evenness
patterns throughout the entire simulation scenarios. For
example, in scenario one, the sharp initial increase in HIT
is due to the initial increase in evenness and continuous
increase in richness. The following lesser increase is a result
of a balance between the decreasing evenness and the increas-
ing richness. This is expected of the HIT since it is a joint
measure of richness and evenness. HIT, in contrast to indices
of pure richness (FRic, TOP) or pure evenness (FEve, TED,
TEDM), provides a joint measure of richness and even-
ness (Eq. 7). For q = 2, this metric reduces to a normalized
inverse Simpson diversity index (Leinster and Cobbold 2012,
Barabás et al. 2022). However, instead of applying the index
to a set of categories, such as different morphs of the same
species, we apply it to the fitted trait probability density func-
tion (Carmona et al. 2016) inferred from the observed data.
As the HIT index measures both richness and evenness (a
population with two morphs is more diverse than a single-
morph one, but only marginally so if the difference in relative
abundance between the two morphs is large), this property
is inherited when applied to the trait distribution. In other
words, a population will be more diverse if its individuals
occupy a larger fraction of the trait space, and still more so if
they are evenly distributed, as opposed to being heavily con-
centrated in one segment of trait space while leaving others
empty. In summary, this metric fulfills the original purpose
of diversity metrics: it is a dimensionless number simultane-
ously quantifying richness and evenness.
HIT’s ability to measure both evenness and richness is
especially visible in scenario one of Fig. 4. Compared with
the separate expectations for richness and evenness (Fig. 3D),
HIT first produces the evenness trend, but then starts fol-
lowing the expected trends for richness. In doing so, it does
not increase as fast as the richness indices FRic and TOP,
maintaining some of its ‘in-between’ behavior that balances
both diversity aspects in a single metric. Scenario two is sim-
pler: the predicted trends are declining for both richness and
evenness, which is duly observed in Fig. 4 as well. In sce-
nario three, where richness and evenness theoretically ought
to stay constant, we observe a slight increase in the metric
as the number of points increases (the only exception is for
scenario three A in a two-dimensional trait space). While
not necessarily desirable, this behavior was expected, due
to the fact that we inferred the underlying trait probability
density function by performing kernel density estimation on
our data. Similarly to the case of TED and TEDM, having
only a few data points means that the estimated trait prob-
ability will be concentrated around those points and won’t
necessarily reflect the true underlying distribution. As an
extreme example: no matter how large the variance of a trait
is, having a single observed individual makes it impossible to
estimate this variance. Indeed, when switching to a different
method of estimating the trait distribution whereby we fit a
multivariate normal distribution to the individual trait values
(Barabás et al. 2022), the increasing trend largely disappears,
 16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 1. Summary of the performance of all indices tested.
Type Indices Performance Recommendations
Divergence FDis Measured trait divergence as expected in all None
scenarios considered
Rao Measured trait divergence as expected in all None
scenarios considered
Evenness FEve Can miss changes in evenness, not correlated with N Not advised to use
TED Measures evenness as expected, approximates Bootstrap with fixed N to correct for correlation with N
expected evenness despite discrete reference
TEDM Measures evenness as expected, correlated with N Bootstrap with fixed N to correct for correlation with N
Richness FRic Detect changes at the edge of the trait space, Only suitable for studying trait space expansion, bootstrap
correlated with N with fixed N to correct for correlation with N
TOP Detect changes everywhere in the trait space, Bootstrap with fixed N to correct for correlation with N
correlated with N
Integrated HIT Performed as expected, long computation time Use bootstrap with fixed N < 1000 individuals to cut
computation time

and we are left with an index that does not depend on the
number of individuals. In summary, this way of measuring
diversity combines the good properties of richness and even-
ness indices, is close in spirit to more traditional diversity
measures, and behaves in a consistent way. Importantly how-
ever, comparing the index across different populations only
makes sense if the trait space over which diversity is evaluated
is identical for all of them.
Several studies have either carried out simulations to eval-
uate and understand the performance of diversity indices or
proposed criteria they must satisfy to correctly measure dif-
ferent aspects of biodiversity (Botta-Dukát 2005, Grün and
Leisch 2007, Schleuter et al. 2010, Carmona et al. 2016,
Fontana et al. 2016, McPherson et al. 2017). While most
of these indices correctly measure the envisaged aspect of
ITD, caution is required in the use of some of these indices
especially with respect to the effect of the number of indi-
viduals. We provide a summary of important recommenda-
tions on each of the indices tested in Table 1. FRic is limited
to applications involving trait space expansion while FEve
is not suitable for measuring trait evenness in a number of
scenarios (Schleuter et al. 2010, Fontana et al. 2016, Legras
and Gaertner 2018, Kosman et al. 2020) including the one
identified in this article. Thus, we recommend not using
FEve to measure ITD evenness. Rather, TED or TEDM
computed over several bootstraps with a fixed number of
individuals (N) should be used. The same bootstrap tech-
nique can be used in computing TOP for measuring trait
richness. HIT is computationally intensive, especially for
a trait space with more than 3 dimensions and 1000 indi-
viduals (see figure in Supporting information). Bootstrap
with a fixed number of individuals lesser than 1000 indi-
viduals can be employed to reduce the computational time.
However, we must stress that this is not guarantee to lead
to significant reduction in computational time in trait space
with 4 or more traits. It is important to stress that we con-
sider strictly indices designed for continuous data in this
article, especially TED and TEDM. HIT is based on Hill
indices which were originally invented for categorical data,
so this particular index can be applied to both discrete and
continuous traits. In principle, the indices considered in
this article could be extended for binary and categorical
traits. Such extensions require distance metrics for binary
and categorical data (Gower distance: Gower 1971, Jaccard
or Tanimoto distance: Jaccard 1912, Chung et al. 2019).
Also, computing HIT for binary data requires a multivari-
ate binary density (Teugels 1990, Qaqish 2003). One idea
presented in the article that might be a challenge for cat-
egorical and binary traits is implementing the concept of
continuous trait change. This concept is abstract for said
traits. However, computing the value of a TD index for all
possible permutations of binary or categorical trait values
in a (multivariate) trait space can provide valuable insights
into the performance of such TD index. However, these
extensions fall outside the scope of this article.
In summary, we recommend either of FDis or Rao for
measuring individual-level trait divergence. Also, we recom-
mend using TED and TEDM (bootstrap with fixed number
of individuals) instead of FEve for measuring individual-level
trait evenness, and TOP (bootstrap with fixed number of
individuals) instead of FRic for measuring individual-level
trait richness. FRic (bootstrap with fixed number of individu-
als) should only be considered if the focus of the study is to
measure trait space expansion.
Funding – ODO received funding support from the Bijzonder
Onderzoeksfonds (BOF) co-operation between Hasselt University
and Université de Namur. FDL acknowledges funding from
the Special Research Fund’s Concerted Research Action (ARC)
(Convention 18/23-095). TN gratefully acknowledges funding by
the Internal Funds KU Leuven (project number 3M190682).
Author contributions
Oluwafemi D. Olusoji: Conceptualization (equal); Data
curation (lead); Formal analysis (lead); Funding acquisition
(supporting); Investigation (equal); Methodology (equal);
Project administration (supporting); Resources (equal);
Software (equal); Validation (supporting); Visualization
(lead); Writing – original draft (lead); Writing – review and
editing (supporting). György Barabás: Conceptualization
(supporting); Formal analysis (equal); Investigation

Page 13 of 15

(supporting); Methodology (equal); Project administra-
tion (equal); Resources (equal); Validation (supporting);
Visualization (supporting); Writing – review and edit-
ing (supporting). Simone Fontana: Conceptualization
(equal); Investigation (equal); Methodology (support-
ing); Writing – review and editing (equal). Jürg W. Spaak:
Conceptualization (supporting); Data curation (lead);
Formal analysis (supporting); Investigation (supporting);
Methodology (supporting); Validation (supporting).
Thomas Neyens: Conceptualization (supporting); Data
curation (supporting); Formal analysis (supporting); Funding
acquisition (equal); Investigation (equal); Methodology
(equal); Resources (equal); Supervision (equal); Writing
– review and editing (equal). Frederik De Laender:
Conceptualization (lead); Data curation (equal); Formal
analysis (equal); Funding acquisition (lead); Investigation
(equal); Methodology (equal); Project administration (lead);
Resources (equal); Supervision (lead); Writing – review and
editing (lead). Marc Aerts: Conceptualization (lead); Data
curation (equal); Formal analysis (equal); Funding acqui-
sition (lead); Investigation (lead); Methodology (equal);
Project administration (lead); Resources (equal); Supervision
(lead); Writing – review and editing (lead).
Data availability statement
Data are available from the figshare Digital Repository:
https://doi.org/10.6084/m9.figshare.17372006.v3
(Olusoji et al. 2022).
Supporting information
The Supporting information associated with this article is
available with the online version.
References
Albert, C. H., de Bello, F., Boulangeat, I., Pellet, G., Lavorel, S.
and Thuiller, W. 2012. On the importance of intraspecific
variability for the quantification of functional diversity. – Oikos
121: 116–126.
Anderson, M. J., Ellingsen, K. E. and McArdle, B. H. 2006. Mul-
tivariate dispersion as a measure of beta diversity. – Ecol. Lett.
9: 683–693.
Barabás, G. and D’Andrea, R. 2016. The effect of intraspecific
variation and heritability on community pattern and robust-
ness. – Ecol. Lett. 19: 977–986.
Barabás, G., Parent, C., Kraemer, A., Van de Perre, F. and De
Laender, F. 2022. The evolution of trait variance creates a ten-
sion between species diversity and functional diversity. – Nat.
Commun. 13: 2521.
Barber, C., Habel, K., Grasman, R., Gramacy, R., Stahel, A. and
Sterratt, D. C. 2015. Mesh generation and surface tesselation:
geometry, https://davidcsterratt.github.io/geometry/.
Bolnick, D. I., Amarasekare, P., Araújo, M. S., Bürger, R., Levine,
J. M., Novak, M., Rudolf, V. H., Schreiber, S. J., Urban, M. C.
and Vasseur, D. A. 2011. Why intraspecific trait variation mat-
ters in community ecology. – Trends Ecol. Evol. 26: 183–192.
Page 14 of 15
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Botta-Dukát, Z. 2005. Rao’s quadratic entropy as a measure of func-
tional diversity based on multiple traits. – J. Veg. Sci. 16: 533–540.
Cardinale, B. J., Duffy, J. E., Gonzalez, A., Hooper, D. U., Perrings,
C., Venail, P., Narwani, A., Mace, G. M., Tilman, D., Wardle,
D. A., Kinzig, A. P., Daily, G. C., Loreau, M., Grace, J. B.,
Larigauderie, A., Srivastava, D. S. and Naeem, S. 2012. Biodi-
versity loss and its impact on humanity. – Nature 486: 59–67.
Carmona, C. P., de Bello, F., Mason, N. W. and Lepš, J. 2016.
Traits without borders: integrating functional diversity across
scales. – Trends Ecol. Evol. 31: 382–394.
Carmona, C. P., Rota, C., Azcárate, F. M. and Peco, B. 2015. More
for less: sampling strategies of plant functional traits across local
environmental gradients. – Funct. Ecol. 29: 579–588.
Champely, S. and Chessel, D. 2002. Measuring biological diversity
using Euclidean metrics. – Environ. Ecol. Stat. 9: 167–177.
Chung, N. C., Miasojedow, B., Startek, M. and Gambin, A. 2019.
Jaccard/Tanimoto similarity test and estimation methods for
biological presence–absence data. – BMC Bioinform. 20: 644.
Cianciaruso, M. V., Batalha, M. A., Gaston, K. J. and Petchey, O.
L. 2009. Including intraspecific variability in functional diver-
sity. – Ecology 90: 81–89.
Cornwell, W. K. and Ackerly, D. D. 2009. Community assembly
and shifts in plant trait distributions across an environmental
gradient in coastal California. – Ecol. Monogr. 79: 109–126.
Cornwell, W. K., Schwilk, D. W. and Ackerly, D. D. 2006. A trait-
based test for habitat filtering: convex hull volume. – Ecology
87: 1465–1471.
de Bello, F., Lavorel, S., Díaz, S., Harrington, R., Cornelissen, J.
H. C., Bardgett, R. D., Berg, M. P., Cipriotti, P., Feld, C. K.,
Hering, D., Martins da Silva, P., Potts, S. G., Sandin, L., Sousa,
J. P., Storkey, J., Wardle, D. A. and Harrison, P. A. 2010.
Towards an assessment of multiple ecosystem processes and ser-
vices via functional traits. – Biodivers. Conserv. 19: 2873–2893.
De Laender, F., Rohr, J. R., Ashauer, R., Baird, D. J., Berger, U.,
Eisenhauer, N., Grimm, V., Hommen, U., Maltby, L., Meliàn,
C. J., Pomati, F., Roessink, I., Radchuk, V. and Van den Brink,
P. J. 2016. Reintroducing environmental change drivers in bio-
diversity–ecosystem functioning research. – Trends Ecol. Evol.
31: 905–915.
Des Roches, S., Post, D. M., Turley, N. E., Bailey, J. K., Hendry,
A. P., Kinnison, M. T., Schweitzer, J. A. and Palkovacs, E. P.
2018. The ecological importance of intraspecific variation. –
Nat. Ecol. Evol. 2: 57–64.
Duong, T. 2007. ks: kernel density estimation and kernel discrimi-
nant analysis for multivariate data in R. – J. Stat. Softw. 21: 1–16.
Duong, T. 2007.7 ks: Kernel Density Estimation and Kernel Dis-
criminant Analysis for Multivariate Data in R Journal of Sta-
tistical Software .10.18637/jss.v021.i07
Fontana, S., Petchey, O. L. and Pomati, F. 2016. Individual-level
trait diversity concepts and indices to comprehensively describe
community change in multidimensional trait space. – Funct.
Ecol. 30: 808–818.
Fontana, S., Thomas, M. K., Moldoveanu, M., Spaak, P. and
Pomati, F. 2018. Individual-level trait diversity predicts phyto-
plankton community properties better than species richness or
evenness. – ISME J. 12: 356–366.
Gagic, V., Bartomeus, I., Jonsson, T., Taylor, A., Winqvist, C., Fis-
cher, C., Slade, E. M., Steffan-Dewenter, I., Emmerson, M.,
Potts, S. G., Tscharntke, T., Weisser, W. and Bommarco, R.
2015. Functional identity and diversity of animals predict eco-
system functioning better than species-based indices. – Proc. R.
Soc. B 282: 20142620.

González-Suárez, M., Bacher, S. and Jeschke, J. M. 2015. Intraspe-
cific trait variation is correlated with establishment success of
alien mammals. – Am. Nat. 185: 737–746.
Gower, J. C. 1971. A general coefficient of similarity and some of
its properties. – Biometrics 27: 857.
Gregorius, H.-R. and Gillet, E. M. 2021. The concept of evenness/
unevenness: less evenness or more unevenness? – Acta Biotheor.
70: 3.
Grether, G. F. 2005. Environmental change, phenotypic plasticity
and genetic compensation. – Am. Nat. 166: E115–E123.
Grün, B. and Leisch, F. 2007. FlexMix: an R package for finite
mixture modelling. – R News 7: 8–13.
Hart, S. P., Schreiber, S. J. and Levine, J. M. 2016. How variation
between individuals affects species coexistence. – Ecol. Lett. 19:
825–838.
Hillebrand, H. and Matthiessen, B. 2009. Biodiversity in a complex
world: consolidation and progress in functional biodiversity
research. – Ecol. Lett. 12: 1405–1419.
Jaccard, P. 1912. The distribution of the flora in the alpine zone.
– New Phytol. XI: 37–50.
Kosman, E., Scheiner, S. M. and Gregorius, H. 2020. Severe limi-
tations of the FEve metric of functional evenness and some
alternative metrics. – Ecol. Evol. 11: 123–132.
Kullback, S. and Leibler, R. A. 1951. On information and suffi-
ciency. – Ann. Math. Stat. 22: 79–86.
Laliberté, E. and Legendre, P. 2010. A distance-based framework
for measuring functional diversity from multiple traits. – Ecol-
ogy 91: 299–305.
Legras, G. and Gaertner, J.-C. 2018. Assessing functional evenness
with the FEve index: a word of warning. – Ecol. Indic. 90:
257–260.
Legras, G., Loiseau, N. and Gaertner, J.-C. 2018. Functional rich-
ness: overview of indices and underlying concepts. – Acta
Oecol. 87: 34–44.
Leinster, T. and Cobbold, C. A. 2012. Measuring diversity: the
importance of species similarity. – Ecology 93: 477–489.
Mason, N. W. H., Mouillot, D., Lee, W. G. and Wilson, J. B. 2005.
Functional richness, functional evenness and functional diver-
gence: the primary components of functional diversity. – Oikos
111: 112–118.
McGill, B., Enquist, B., Weiher, E. and Westoby, M. 2006. Rebuild-
ing community ecology from functional traits. – Trends Ecol.
Evol. 21: 178–185.
McPherson, J. M., Yeager, L. A. and Baum, J. K. 2017. A simula-
tion tool to scrutinise the behaviour of functional diversity
metrics. – Methods Ecol. Evol. 9: 200–206.
Mi, X., Miwa, T. and Hothorn, T. 2009. New numerical algorithm
for multivariate normal probabilities in package mvtnorm. – R
J. 1: 37.
Olusoji, O. D., Barabás, G., Spaak, J. W., Simone, F., Neyens, T.,
De Laender, F. and Aerts, M. 2022. Data from: Measuring
individual-level trait diversity: a critical assessment of methods.
– Figshare Digital Repository, https://doi.org/10.6084/m9.fig-
share.17372006.v3.
Pacifici, M., Visconti, P., Butchart, S. H. M., Watson, J. E. M.,
Cassola, F. M. and Rondinini, C. 2017. Species’ traits influ-
enced their response to recent climate change. – Nat. Clim.
Change 7: 205–208.
16000706, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/oik.09178 by Lib4ri Library Of Eawag Empa P, Wiley Online Library on [11/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Petchey, O. L. and Gaston, K. J. 2006. Functional diversity: back
to basics and looking forward. – Ecol. Lett. 9: 741–758.
Pigot, A. L., Trisos, C. H. and Tobias, J. A. 2016. Functional traits
reveal the expansion and packing of ecological niche space
underlying an elevational diversity gradient in passerine birds.
– Proc. R. Soc. B 283: 20152013.
Qaqish, B. F. 2003. A family of multivariate binary distributions
for simulating correlated binary variables with specified mar-
ginal means and correlations. – Biometrika 90: 455–463.
Rao, C. 1982. Diversity and dissimilarity coefficients: a unified
approach. – Theor. Popul. Biol. 21: 24–43.
Reiss, J., Bridle, J. R., Montoya, J. M. and Woodward, G. 2009.
Emerging horizons in biodiversity and ecosystem functioning
research. – Trends Ecol. Evol. 24: 505–514.
Ricotta, C. 2005. A note on functional diversity measures. – Basic
Appl. Ecol. 6: 479–486.
Scheiner, S. M., Kosman, E., Presley, S. J. and Willig, M. R. 2016.
Decomposing functional diversity. – Methods Ecol. Evol. 8:
809–820.
Schleuter, D., Daufresne, M., Massol, F. and Argillier, C. 2010. A
user’s guide to functional diversity indices. – Ecol. Monogr. 80:
469–484.
Schloerke, B., Wickham, H., Cook, D. and Hofmann, H. 2016.
Escape from Boxland. – R J. 8: 243.
Spaak, J. W., De Laender, F. and Metcalf, J. 2020. Intuitive and
broadly applicable definitions of niche and fitness differences.
– Ecol. Lett. 23: 1117–1128.
Stomp, M., Huisman, J., de Jongh, F., Veraart, A. J., Gerla, D.,
Rijkeboer, M., Ibelings, B. W., Wollenzien, U. I. A. and Stal,
L. J. 2004. Adaptive divergence in pigment composition pro-
motes phytoplankton biodiversity. – Nature 432: 104–107.
Teugels, J. L. 1990. Some representations of the multivariate Ber-
noulli and binomial distributions. – J. Multivar. Anal. 32:
256–268.
Tilman, D. 1997. Community invasibility, recruitment limitation
and grassland biodiversity. – Ecology 78: 81–92.
Villéger, S., Mason, N. W. H. and Mouillot, D. 2008. New
multidimensional functional diversity indices for a multifac-
eted framework in functional ecology. – Ecology 89:
2290–2301.
Violle, C., Enquist, B. J., McGill, B. J., Jiang, L., Albert, C. H.,
Hulshof, C., Jung, V. and Messier, J. 2012. The return of the
variance: intraspecific variability in community ecology. –
Trends Ecol. Evol. 27: 244–252.
Violle, C., Navas, M.-L., Vile, D., Kazakou, E., Fortunel, C., Hum-
mel, I. and Garnier, E. 2007. Let the concept of trait be func-
tional! – Oikos 116: 882–892.
Walker, S. C., Poos, M. S. and Jackson, D. A. 2008. Functional
rarefaction: estimating functional diversity from field data. –
Oikos 117: 286–296.
Westoby, M., Falster, D. S., Moles, A. T., Vesk, P. A. and Wright,
I. J. 2002. Plant ecological strategies: some leading dimensions
of variation between species. – Annu. Rev. Ecol. Syst. 33:
125–159.
Wohlrab, S., Tillmann, U., Cembella, A. and John, U. 2016. Trait
changes induced by species interactions in two phenotypically
distinct strains of a marine dinoflagellate. – ISME J. 10:
2658–2668.
Page 15 of 15