Behavioral Ecology and Sociobiology (2018) 72: 128

https://doi.org/10.1007/s00265-018-2538-y

METHODS PAPER

Circular data in biology: advice for effectively implementing

statistical procedures
Lukas Landler 1 & Graeme D. Ruxton 2 & E. Pascal Malkemper 1,3

Received: 4 March 2018 / Revised: 19 June 2018 / Accepted: 22 June 2018 / Published online: 11 July 2018
# The Author(s) 2018

Abstract
Circular data are common in biological studies. The most fundamental question that can be asked of a sample of circular data is
whether it suggests that the underlying population is uniformly distributed around the circle, or whether it is concentrated around at
least one preferred direction (e.g. a migratory goal or activity phase). We compared the statistical power of five commonly used tests
(the Rayleigh test, the V-test, Watson’s test, Kuiper’s test and Rao’s spacing test) across a range of different unimodal scenarios. The
V-test showed higher power for symmetrical distributions, Rao’s spacing performed worst for all explored unimodal distributions
tested and the remaining three tests showed very similar performance. However, the V-test only applies if the hypothesis is restricted
to one (pre-specified) direction of interest. In all other unimodal cases, we recommend using the Rayleigh test. Much less explored
is the multimodal case with data concentrated around several directions. We performed power simulations for a variety of
multimodal situations, testing the performance of the widely used Rayleigh, Rao’s, Watson, and Kuiper’s tests as well as the more
recent Bogdan and Hermans-Rasson tests. Our analyses of alternative statistical methods show that the commonly used tests lack
statistical power in many of multimodal cases. Transformation of the raw data (e.g. doubling the angles) can overcome some of the
issues, but only in the case of perfect f-fold symmetry. However, the Hermans-Rasson method, which is not yet implemented in any
software package, outcompetes the alternative tests (often by substantial margins) in most of the multimodal situations explored. We
recommend the wider uptake of the powerful but hitherto neglected Hermans-Rasson method. In summary, we provide guidance
for biologists helping them to make decisions when testing circular data for single or multiple departures from uniformity.

Keywords Animal navigation . Emlen funnel . Magnetoreception . Biostatistics . Circadian . Chronobiology

Introduction measurements across scientific disciplines. Common data are

angles (e.g. compass bearings) or measurements over time where
Circular data are characterized by an inherent periodicity absent an inherent periodicity is relevant: e.g. time of day, seasonality
from measurements made on a linear scale (such as mass or and point in the lunar cycle. Circular data need special treatment
length). Such data are generated by a range of common in data analysis: consider that an angle of 355° is much nearer to
an angle of 5° than it is to an angle of 330°, and so a simple
Communicated by L. Z. Garamszegi arithmetic mean for example can be quite misleading. Many texts
Electronic supplementary material The online version of this article have been dedicated to describing statistical treatment of circular
(https://doi.org/10.1007/s00265-018-2538-y) contains supplementary data: e.g. Batschelet (1981), Fisher (1995), Jammalamadaka and
material, which is available to authorized users. SenGupta (2001), Ley and Verdebout (2017), Mardia and Jupp
(2000), Pewsey et al. (2013), and some commonly used general
* E. Pascal Malkemper statistics texts aimed at biologists also cover this (e.g. Zar 2013).
[email protected]
However, there is a lack of clear guidance on how to most effec-
1
Research Institute of Molecular Pathology (IMP), Vienna Biocenter tively test for departure from circular uniformity.
(VBC), 1030 Vienna, Austria
2
School of Biology, University of St Andrews, St Andrews KY16 Unimodal departure from uniformity
9TH, UK
3
Department of General Zoology, Faculty of Biology, University of The most common statistical exploration of circular data in-
Duisburg-Essen, 45117 Essen, Germany volves testing to see if there is a unimodal bias in the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

128 Page 2 of 10 Behav Ecol Sociobiol (2018) 72: 128

distribution around the circle (i.e. concentration of the data in a application of the test. Recently, Ruxton (2017) recommended
certain region of the circle) or whether the null hypothesis that increased use of this test in large part because it ought to offer
the underlying population involves a uniform spread around increased power over the Rayleigh test. However, again, the
the circle is supported. For example, one might test whether quantitative performance of this test relative to the alternatives
the bearings of the initial flights of released homing pigeons mentioned above remains unclear. As we will discuss later, the
are random (uniform) or show a bias towards the direction of alternative hypothesis for the V-test is identical to a Rayleigh
the home loft of the pigeons. In this section, we focus on this test, and statistical significance alone does not indicate orien-
situation where, if there is departure from circular uniformity, tation in the a priori direction. Interpretation of this test needs
we expect that departure to be unimodal with increased con- care, but evaluation of a confidence interval around the esti-
centration in a single region of the circle (in contrast to mul- mated population mean direction can help.
timodal departures such as axial distributions, see next sec-
tion). By far, the most common statistical test applied to test Multimodal departure from uniformity
for departure from uniformity is the Rayleigh test (introduced
by Lord Rayleigh in 1880, but also defined in the four In most cases, researchers might be interested in a unimodal
monographs listed above). The test statistic for this test is departure from uniformity; however, there are many examples
derived from the ‘mean resultant length’ (r), which is a mea- where we expect more than one cluster around the circle. For
sure of the concentration of data points around a circle. The example, we might have data on the times of activity in domes-
test supports rejection of uniformity for high values of this tic cats. We might test the null hypothesis that the activity is
statistic. Its theoretical foundation lies in the von Mises distri- uniformly distributed throughout the 24 h/day. That is, that the
bution (e.g. Watson and Williams 1956). The Rayleigh test is rate of activity is constant and shows no daily variation. From
known to be the most powerful invariant test against a von your general knowledge of cats, you might expect that this null
Mises alternative to uniformity (Watson and Williams 1956), hypothesis will be false since cats are crepuscular animals,
but its performance against other plausible unimodal alterna- active around the times of sunrise and sunset. Indeed, what
tives, such as a skew normal or wrapped Cauchy distribution, you might expect to see is that there is a peak of activity in
remains underexplored. This is a significant gap in our knowl- the early morning and another peak in the evening. Therefore,
edge since although the von Mises distribution is often de- you might predict that the data will show a multimodal distri-
scribed as the circular analogue of the normal distribution, it bution. As discussed above, the Rayleigh test is known to be
does not have a similar mechanistic basis (analogous to the the most powerful invariant test against a von Mises alternative
central limit theorem). Thus, there is no theoretical reason to to uniformity (Watson and Williams 1956). However, it is well-
expect von Mises distributions to be commonplace in real known that the Rayleigh test (which was designed to detect
datasets. The Rayleigh test is a parametric test, and its original unimodal von Mises departures from uniformity) has very
derivation involved the assumption that if there was a devia- low power to detect multimodal departures from uniformity
tion from uniformity, that deviation would be of the von Mises (see Batschelet 1981), that is situations with more than one
form. The robustness of the performance of the test to viola- concentration of data around the circle. As argued above, we
tions of this assumption are under-explored but will be illumi- might expect our imaginary cat-activity data to feature two
nated further in our study. As well as this essentially paramet- peaks. The unimodal limitation to the Rayleigh test can be
ric test, there is also a range of non-parametric, so-called om- overcome, if we expect that the alternative hypothesis has f-
nibus tests that make no assumptions about the nature of pos- fold symmetry for some f that can be specified, i.e. multiple
sible deviations from uniformity and might be used for peaks are of the same size, symmetrical and evenly spaced
unimodal departures (see Batschelet 1981 for an overview), around the circle are expected. In this special case, multiplying
but the relative performance of these to each other and to the the raw angles by f, and taking their modulus with respect to the
Rayleigh test is also underexplored (see Stephens 1969). Here, maximum allowable angle (2π in radians, 360° in angles), will
we present simulation analyses that close this knowledge gap convert f-fold symmetric data into a unimodal distribution.
for unimodal distributions. However, the requirement for f-fold symmetry is restrictive.
It is important to keep in mind that the tests discussed so far In the case of our cat-activity example, our two activity peaks
make no assumption about where on the circle any concentra- are in most situations not expected to be 12 h apart.
tion of data might lie if the underlying distribution is non- As for the unimodal case discussed above, there is a wide
uniform. In some situations (such as the case of the homing range of omnibus tests of circular uniformity that can be used
pigeons mentioned above), the researcher may have a priori for multimodal deviations from uniformity without the assump-
expectation of where any concentration might occur. For these tion of a von Mises distribution. The three most commonly
cases, there is a variant of the Rayleigh test (often called the V- used of these are Watson’s test, Kuiper’s test and Rao’s spacing
test; due to Durand and Greenwood 1958) that allows the test (see Batschelet 1981 for an overview), but the relative
mean of the alternative distribution to be specified prior to performance of these tests against multimodal departures from

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Behav Ecol Sociobiol (2018) 72: 128 Page 3 of 10 128

uniformity has been little explored. The investigation that has parameters corresponding to the mean value, and a measure of
been undertaken (Stephens 1969; Bergin 1991; Bogdan et al. the concentration around that mean. For both distributions, we
2002; Pycke 2010) has highlighted potential for low power (as used a mean of zero (360°), and a range of different values for the
discussed above). For this reason, alternative tests have been concentration parameters (see Chapter 4 of Pewsey et al. (2013)
proposed that were designed specifically to have better power for more on the properties of these distributions). For specified
against multimodal alternatives: the two most developed that values of the sample size and concentration parameter, we drew
we could find in the literature were due to Hermans and Rasson samples from these distributions using functions from the ‘circu-
(1985) and Bogdan et al. (2002); see Pycke (2010) for theoret- lar’ package in R (Agostinelli and Lund 2013). We also consid-
ical underpinning of general classes to tests for circular unifor- ered an asymmetric distribution: the skew normal distribution—
mity. Both have been subject to very little published evaluation, the properties of this are described in Azzalini (1985) and
and neither is available in any software package that we know Pewsey (2000). This distribution is controlled by the values of
of. Our aim in this section of the paper is to evaluate these two three parameters (a location parameter (ε) set to zero (360°)
methods alongside the four more established methods throughout, a shape parameter (α) set to 30 to give a right skew
(Rayleigh, Kuiper’s, Watson’s and Rao’s spacing tests) for a and a (positive) dispersion parameter (ω). For specified sample
range of multimodal distributions. This investigation should size, distribution and parameter values (‘model’ was set to null),
offer researchers clear advice for best practice in the common we generated samples using the rcircmix function from the R
case where circular data have been gathered and there is a package ‘NPCirc’ (Oliveira Pérez et al. 2014). Statistical power
desire to explore evidence for departure from uniformity in to reject the null hypothesis of a uniform distribution is defined as
situations where knowledge of the system gives reason to sus- the fraction of 10,000 samples of a given size drawn from a given
pect that such departures might be multimodal. Such advice is distribution generating a p value less than 0.05. This is by no
urgently required given the multiplicity of alternative tests and means a thorough exploration of theoretical distributions for cir-
given that the limited existing published explorations warn of cular data (see Chapter 4 of Pewsey et al. (2013) for an overview
potential for very low statistical power in some cases (e.g. of the diversity of these). However, we consider that these three
Stephens 1969; Bogdan et al. 2002). distributions between them capture key features of many empir-
ical unimodal distributions. The von Mises distribution is sym-
Overview metric about its central tendency and shows a similar ‘bell’ shape
to the Normal distribution; the wrapped Cauchy is similarly
Considering the wide use of circular data and statistics not unimodal and symmetric but generally has the potential to offer
only in animal behaviour, ecology and behavioural ecology a sharper distribution with more concentration near the central
(Wiltschko and Wiltschko 1972; Shimatani et al. 2012), but tendency than the von Mises. Just as with linear data, there are
also in branches of science that range from neurobiology many circular processes that are not symmetric about the central
(Taube 2007), chronobiology (Gustafson and Partch 2014), tendency, and our final distribution, the skew normal distribution,
atmospheric sciences (Gaumond et al. 2014) to astrophysics allows exploration of such resulting asymmetric distributions.
(Archibald et al. 2015), it is concerning that the essential pow- We considered a range of different tests, the Rayleigh test,
er comparisons are still missing. Our aim in this paper, there- the V-test and three omnibus tests: Watson’s test, Kuiper’s test,
fore, is to test the performance of all these alternative tests by and Rao’s spacing test. For the V-test, we included three dif-
simulation over a broad range of different sampling situations. ferent scenarios: one where the predicted mean direction co-
We then use these results to offer researchers clear guidance incides with the actual mean of the distribution, one where it is
on how to select the optimal test for the most commonly off by 20° and one where the means differ by 45°. All these
encountered situations in circular statistics. tests are defined, and their properties discussed in standard
textbooks on circular statistics (e.g. Batschelet 1981; Fisher
1995; Mardia and Jupp 2000; Pewsey et al. 2013). These tests
Methods are also all available as functions in the circular package in R.

All simulations were performed in R (R core team 2013), and our Power analyses of tests for multimodal distributions
code is provided as an electronic supplement (Online Resources
4–7). Graphs were prepared in Prism (GraphPad). We analysed the power of six different statistical tests for mul-
timodal distributions: Rayleigh test, Watson test, Rao’s spacing
Power analyses of tests for unimodal distributions test, Kuiper test and two recently proposed Hermans-Rasson
and Bogdan tests. The first four tests were calculated using the
We considered three different unimodal circular distributions: R library circular (Agostinelli and Lund 2013), using the func-
von Mises, wrapped Cauchy and skew normal. The first two tions rayleigh.test, watson.test, rao.spacing.test and kuiper.test,
are symmetrical distributions fully characterized by two respectively. The latter two tests were calculated using our own

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

128 Page 4 of 10 Behav Ecol Sociobiol (2018) 72: 128

functions, based on descriptions in the original papers (see eight different sample sizes used (10, 15, 20, 25, 30, 40, 80
Online Resources 2, 3). We applied the Rayleigh test both to and 100). The proportion of significant results was then plot-
the raw data and to data manipulated as described in the intro- ted for each sample size.
duction to deal with symmetric multimodal situations (multi-
plying the raw angles by f). We conducted power analyses by
drawing 10,000 samples from the distribution of interest, Results
performing the statistical test on each of the drawn samples
and calculating the proportion of significant test results. We Performance of tests on uniform distributions,
set the alpha levels for all analyses to p = 0.05. and introduction to other distributions considered
In our first analysis set, multimodal von Mises distributions
were generated using the rcircmix function from the R library When testing the performance of a statistical test, it is impor-
NPCirc (Oliveira Pérez et al. 2014). tant to not only consider the power to detect departures but
We generated a wide variety of multimodal distributions also the likelihood of falsely detecting departure from unifor-
(from two modes up to six modes). For each number of mity that does not exist, i.e. the type I error rate. We note that
modes, we used two different clustering situations: symmetri- all the tests that we consider control type I error rate close to
cally distributed modes (e.g. four modes at 360°, 90°, 180° the nominal 5% value (Fig. 1). For some statistical tests, it is
and 270°) or asymmetrically distributed modes (e.g. four important to know the underlying distribution from which the
modes at 360°, 60°, 120° and 180°). We then simulated the sample has been drawn. For example, the Rayleigh test as-
distributions over the concentration parameter κ in 0.1 incre- sumes a von Mises distribution and is most powerful to detect
ments ranging from 0 to 6. This parameter controls the con- such unimodal departures. To address how the Rayleigh and
centration of points around each mode in the underlying dis- other tests perform with different underlying distributions, we
tribution from which samples were drawn. performed all power analyses with three different distribu-
In a second analysis, we analysed the effects on statistical tions: von Mises, skew normal and wrapped Cauchy.
power of differences between the means (a parameter we call Figure 2 shows examples of the three different distributions
Δ) of two distributions, in the case of bimodal data. For this, with varying concentrations around the mean direction.
we generated samples of bimodal distributions where Δ
ranged from 0° (when the two modes are coincident) to 180° Performance of tests on unimodal distributions
(when the two modes are diametrically opposite each other on
the circle). This analysis was repeated for a range of sample When the sample is drawn from a von Mises distribution, the
sizes: 15, 25, 40 and 100. power to reject the null hypothesis of uniformity increases both
Next, we examined the effect of unequal proportions of two with sample size and with how concentrated the underlying dis-
underlying von Mises distributions that combine to give our tribution is around the central value (how high the concentration
underlying bimodal distribution. For this, we generated sam- parameter κ is). We find a consistent pattern among tests in
ples from bimodal distributions (either symmetrical, 0° and Fig. 3: Rao’s spacing test has the lowest power, and then there
180° or asymmetrical, 360° and 90°) and changed the propor- is a group of tests of intermediate power (Rayleigh, Watson’s,
tion of data points belonging to one of the peaks ranging from Kuiper’s and the V-test where the pre-specified test and popula-
0.05 to 0.5 (in 0.05 steps). This simulation was performed for tion mean values differ by 45°), then a group with highest power
κ = 3 and 2; our exploration of sensitivity to this parameter (V-test with coincident and with 20° different means). Although
(see immediately below) suggests that these two values effec- the V-test with coincident means has higher power than when the
tively capture the variety of conditions. means differ by 20°, the effect is slight (less than 5%). The
Finally, we tested the effects of differing κ values between the difference in power between these two tests and that of the in-
two von Mises distributions, in a bimodal situation (this analysis termediate group of four tests is strongest when sample size is
was performed for both symmetrical and asymmetrically distrib- low and powers are far from either 0.05 or 1, and generally, this
uted modes). The κ for the first distribution was fixed at 3, but κ difference is of the order of 5–10%. Similar results can be seen
for the second distribution was changed in 0.1 steps over a range for the wrapped Cauchy distribution (Online Resource 1, Fig.
of κ values from 0 to 6 (10,000 iterations each). A1). There are slight differences when the skew normal distri-
bution is considered (Online Resource 1, Fig. A2). Here, the
Type 1 error calculations shape of the graphs flips because the spread of values for this
distribution is traditionally described by a dispersion parameter
We complimented our power analyses with type 1 error cal- ω, such that spread of values increases for increasing ω and
culation. Drawing 100,000 samples from uniform distribu- power decreases accordingly. Interestingly, with the skew normal
tions (using rcircularuniform function in the circular package) distribution, there is a change in the relative performance of the
and applying the tests described above, this was done for the different tests. The V-test performs badly in this situation (apart

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Behav Ecol Sociobiol (2018) 72: 128 Page 5 of 10 128

Fig. 1 The estimated statistical 0.06

type 1 error, i.e. proportion of

type I errorprobability
tests generating p < 0.05. 0.05 Rayleigh
Estimates are based on 100,000
V-Test
uniform samples for each of a 0.04
range of sample sizes: 10, 15, 20, Watson
25, 30, 40, 80 and 100. We 0.03 Rao's spacing
compare the Rayleigh test, the
three omnibus tests (Kuiper’s,
Kuiper's
0.02
Watson’s and Rao’s spacing HRT2
tests), the V-test, the Bogdan, and 0.01 Bogdan
Hermans-Rasson tests
0.00

10

15

20

25

30

40

80

0
10
sample size n

from superior performance at small sample sizes and 2 < ω < 3), are concentrated about the two modes, and the higher power we
and Rao’s spacing test again gives relatively poor performance, have to reject the null hypothesis. It is already known that the
while there is little difference between the Rayleigh, Kuiper’s Rayleigh test struggles in this situation, and we see this in our
and Watson’s tests, which give best performance of all. simulations. In this case, where the underlying distribution has
However, Kuiper’s and Watson’s tests slightly outperform the perfect f-fold symmetry (f = 2 in our case), then multiplying the
Rayleigh test in this special situation, especially at high samples measured values by f and applying a modulus of the maximum
sizes, but never offering more than 1–2% more power. value on the circle (360° or 2π) transforms the data into a
unimodal distribution. Performing this transformation prior to
Performance of tests on multimodal distributions application of the Rayleigh test then greatly improves the per-
formance of this test. Given that the Rayleigh test is the most
An obvious underlying multimodal population is one with two efficient test for detecting a von Mises alternative, it is unsur-
modes of identical shape and size at opposite ends of the circle prising that our simulation results show that this ‘Rayleigh (2×)’
(i.e. symmetrically 180° apart). We explore the ability of the procedure offers the best power. However, we note that the
different tests to reject the null hypothesis of uniformity from performance of the Hermans-Rasson test data is almost as good,
samples of size N = 25 from such a distribution in Fig. 4a. We without the need to transform the data. The powers of the
explore performance over a wide range of local shapes of dis- Bogdan and Rao’s tests are similar, but a little lower than
tribution around the two modes, controlled by the concentration Rayleigh (2×) and Hermans-Rasson. The power curves for
parameter (κ). The higher the value of κ, the more closely data the Kuiper and Watson tests are similar and lower again.

0 0 0 0

Fig. 2 Examples for the

distributions used in this analysis.
The von Mises distribution,
Von Mises
wrapped Cauchy and skewed 270
κ=0 90 270
κ=1 90 270
κ=2 90 270
κ=3 90

normal distribution, over a range

of concentration/dispersion pa-
rameters (values given in the 180 180 180 180

plots, n = 100 for each)

0 0 0 0

Wrapped 270
ρ=0 90 270
ρ = 0.25 90 270
ρ = 0.5 90 270
ρ = 0.75 90

Cauchy

180 180 180 180

0 0 0 0

Skew Normal
Distribution
270
ω=3 90 270
ω=2 90 270
ω= 1 90 270
ω = 0.5 90

180 180 180 180

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

128 Page 6 of 10 Behav Ecol Sociobiol (2018) 72: 128

a) 1.0
b) 1.0

Power

Power
0.5 0.5

Von Mises, n = 15 Von Mises, n = 25

0 0
1 2 3 1 2 3

c) 1.0
d) 1.0
Power

Power
0.5 0.5

Von Mises, n = 40 Von Mises, n = 100

0 0
1 2 3 1 2 3
Concentration Parameter κ Concentration Parameter κ

Rayleigh Kuiper's Watson V-Test V-Test +45deg V-Test +20deg Rao's spacing
Fig. 3 The estimated statistical power to reject the null hypothesis of situations for the V-test (where the test mean value and mean value of
uniformity based on a sample from a von Mises distribution. Estimates the underlying distribution either exactly coincide, differ by 20° or differ
are based on 10,000 samples for each of four sample sizes: a 15, b 25, c by 45°). We obtain estimates for a range of different values of the
40 and d 100. We compare the Rayleigh test, the three omnibus tests parameter κ that defines the concentration of values for a von Mises
(Kuiper’s, Watson’s and Rao’s spacing tests) and three different distribution

It is possible to imagine f-fold symmetry to the underlying application of the Rayleigh test continues to offer the best power,
distribution for any value of f (i.e. any number of modes). It is although this power still drops off with increasing number of
intuitive that the greater the number of modes, the harder the modes (Online Resource 1; Fig. A3). The only other investigated
challenge is to detect departure from uniformity on the basis of test that offers valuable power in this situation of perfect f-fold
small samples, and our simulations bear this out. Regardless of symmetry is Rao’s spacing test for f = 3. The challenge of this
the number of modes, the transformation approach prior to situation is obviously reduced by increasing sample size, but we

a) 1.0
b) 1.0
Power

Power

0.5 0.5

n=25 n=25

0.0 0
0 2 4 6 8 10 0 2 4 6 8 10
concentration parameter κ concentration parameter κ
Rayleigh Rayleigh (2x) Bogdan HR Rao's Watson Kuiper's
Fig. 4 The estimated statistical power to reject the null hypothesis of Rasson tests. We obtain estimates for a range of different values of the
uniformity based on a sample with two modes each from a von Mises parameter κ that defines the concentration of values for a von Mises
distribution. Estimates are based on 10,000 samples with a sample size of distribution. a Power estimates for two symmetric modes 180° apart. b
25. We compare the Rayleigh (2× = modification for two symmetrical Power estimates for two asymmetric modes 90° apart
modes), Kuiper’s, Watson’s, Rao’s spacing, Bogdan and Hermans-

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Behav Ecol Sociobiol (2018) 72: 128 Page 7 of 10 128

have observed that power remains low even for large samples if a)
the number of modes is also high (Online Resource 1; Fig. A4).
In Fig. 4b, we break the symmetry of the situation underlying
Fig. 4a by moving one of the modes a quarter of the way around
the circle, so the two modes are now a quarter of the circle apart
in one direction (three quarters in the other direction); otherwise,
the situation is unchanged from that of Fig. 4a. We notice that
breaking symmetry makes departure from uniformity generally
easier to detect. Importantly, breaking the symmetry greatly im- n=25
proves the performance of the Rayleigh test, which now per-
forms very well, but all the explored tests perform well, with
only Rao’s spacing test having slightly lower power than the
others. These general trends hold for higher numbers of asym-
Δ
metrically distributed modes (Online Resource 1; Fig. A5). b)
Next, we investigated the relative position on the circle for two
otherwise-identical modes. We hold the spread of points around
these modes constant and use the x-axis to vary the smallest n=25
distance between the two modes (Fig. 5a). When this distance
is zero, then the modes are coincident (and we have a unimodal
underlying distribution), when it is 90°, we have the situation
considered in Fig. 4b, and when it is 180°, we have the perfectly
symmetric case considered in Fig. 4a. We can see that the
Hermans-Rasson method offers the best all-around performance;
all other tests show a much more substantial drop in power close
to the situation of perfect symmetry. The only other test that
shows robustness to this challenge is Rao’s spacing test, but it κ
generally has lower power than the Hermans-Rasson method. c)
Another way to break the f-fold symmetry would be to have
different spreads of points about the modes. In Fig. 5b, we have
two modes at opposite points on the circle again, and the two
models have equal weight (in that the data-points in the sample
are just as likely to be associated with one mode as the other),
but the spread of data-points about the two modes can differ.
We hold the concentration about one mode constant and vary n=25
the concentration parameter associated with the other mode
over a range of concentrations as described along the x-axis.
The figure emphasizes several themes already seen: the chal-
lenge of detecting perfect f-fold symmetry, the poor perfor-
mance of the Rayleigh test when modes are symmetrically Δ
distributed and the broadly good performance of the Herman-
Rasson test. If we break symmetry by varying the weights of Rayleigh Bogdan HR
the two distributions instead of the spread (Fig. 5c), then this Rao's Watson Kuiper's
greatly improves the performance of the Rayleigh test, but we Fig. 5 The estimated statistical power to reject the null hypothesis of
did not find circumstances where it (or any of the others) con- uniformity based on a sample with two modes each from a von Mises
sistently substantially outperformed the Hermans-Rasson test. distribution. Estimates are based on 10,000 samples with a sample size of
25. We compare the Rayleigh, Kuiper’s, Watson’s, Rao’s spacing,
Bogdan and Hermans-Rasson tests. We obtain estimates for a different
values of the distance (Δ) between the two modes, b different values of
Discussion the concentration κ of one of the modes while holding the concentration
of the other mode constant at 3.0, and c different weight ratios of the
modes (equal weight = a delta proportion of 0.5)
Recommendations for unimodal distributions

Based on our simulation results, we can offer very clear guid- Simply put, at present, there seems no compelling evidence to
ance on testing for unimodal departure from circular uniformity. depart from adoption of the commonly used Rayleigh test. Only

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

128 Page 8 of 10 Behav Ecol Sociobiol (2018) 72: 128

in a situation where a single a priori specified mean direction is Recommendations for multimodal distributions
of interest to the researcher can the V-test offer superior power
to the more general Rayleigh test. This advantage must be set Our results emphasize that detecting symmetrically distributed
against the need for care in interpreting a significant departure multimodal departures from uniformity when the number of
from uniformity by this test. The V-test on its own does not modes is greater than two is a challenge. If researchers expect
provide compelling evidence of departure concentrated about to be in such a position, then our advice would be to maximize
the specified direction, and our simulations demonstrate that the size of their sample and then apply the Hermans-Rasson
even departures centred quite far from the specified direction test but still expect relatively low power to reject the null
can trigger significant results using this test (Fig. 3). However, if hypothesis. Although the ‘trick’ of transforming a multimodal
there is only one direction of interest, the V-test is the best distribution can be effective, it relies on three requirements: (i)
choice, especially if sample sizes are low and power becomes The number of modes can be identified correctly (model-
an issue. In such cases, there may also be a philosophical reason fitting approaches might be useful for helping identify this
for adopting the V-test rather than the Rayleigh test, because the number (e.g. Fitak and Johnsen 2017), (ii) these modes are
alternative hypothesis provides a better match to the fundamen- symmetrically distributed around the circle and (iii) the local
tal question that is being explored. It is important, however, that distribution around each mode is itself symmetrical. We sus-
a significant test result from a V-test must always be interpreted pect that these criteria will only be met in special cases in
in combination with a confidence interval for the population practice, but in these cases, the Rayleigh test offers superior
mean value. The confidence interval can easily be calculated, performance. Multimodal departures that are not symmetrical-
see Pewsey et al. (2013)—who also provide suitable R code. If ly distributed can readily be detected by all tests considered
the confidence interval does not include the pre-specified value, here. Our results allow several conclusions to be drawn about
then the researcher should conclude that the data do not support the relative merits of the different multimodal tests. The
their alternative hypothesis, but if the confidence interval is Rayleigh test is the most commonly used test for departure
narrow and includes this value, then there is evidence to support from circular uniformity, and its performance against
it. Uncertainty arises, however, if the confidence interval in- unimodal departures is very good. Thus, when researchers
cludes the pre-specified value but is relatively wide. In this case, suspect (or are only interested in) unimodal departures from
the conclusion must be very tentative: The data could be uniformity, the Rayleigh test can be recommended, but in the
interpreted as being in accord with the alternative hypothesis, multimodal case, they should seek an alternative offering bet-
but it is also in accord with other explanations. In brief, there is ter performance.
a price to be paid for the extra power offered by the V-test: When researchers suspect that departure from uniformity
Having pre-specified one possible position of concentration of might be multimodal, then we recommend the method of
the data to gain extra power to test this possibility, the researcher Hermans-Rasson as offering the best performance across a
loses the ability to consider any other potential departures from wide range of situations. There is no particular multimodal
uniformity. In fairness, once the V-test has been applied, re- situation in which any other test performs substantially better
searchers must avoid any speculation about support in the data than this test. Although the test is not available in any software
for alternatives to uniformity other than the one they pre-specify package, it is relatively simple to code and we offer a full
in their V-test. We would also emphasize that the mean value to description of the procedure and its implementation in R in
be tested in a V-test should be decided before any preliminary Online Resources 2 and 3.
inspection of the data; if this value is influenced in any way by Although the modern method of Bogdan et al. (2002) of-
inspection of the data and then the V-test is applied to that same fers generally improved performance over older methods, it
data, then the type I error rate will be substantially increased. does not perform as well as that of Hermans-Rasson in the
Of all the unimodal situations that we explored, we did situations that we explore and is conceptually and computa-
not find one where any of the commonly used alternative tionally more complex. Hence, we do not strongly recom-
tests offered substantially better performance than the mend its wider uptake—although for completeness, we pro-
Rayleigh test. The Kuiper’s and Watson’s tests give almost vide a full description of the procedure and its implementation
similar power, but Rao’s spacing test offers consistently in R in Online Resources 2 and 3.
lower power. Given the relative unfamiliarity of Kuiper’s For the researcher who would rather avoid these modern
and Watson’s tests, we see no reason to recommend their methods and use one of the three omnibus methods provided
use over the Rayleigh test. It is of course possible to further in existing software, it is difficult to offer clear guidance. We
expand the results presented here by studying different found no circumstances where there was substantial difference
sample sizes, or different unimodal distributions, but we between Watson’s and Kuiper’s tests. In situations where all
do not see strong reason to expect that such explorations tests perform relatively well, Rao’s spacing test tends to have
will produce results that significantly affect our recommen- lower power than either Watson’s or Kuiper’s tests, but it is
dations above. more resistant than the other two to drop in power in situations

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Behav Ecol Sociobiol (2018) 72: 128 Page 9 of 10 128

Table 1 Summary of our recommendations based on the simulations Authors’ contributions LL, GDR and EPM conceptualized the problem
presented. In all cases, confidence intervals need to be calculated in order and discussed the analytic approaches. LL and GDR prepared the code.
to provide evidence for clustering around a defined direction. Expected LL, GDR and EPM interpreted the results. GDR, LL and EPM wrote the
directions (dir.) must be set a priori; similarly if an f-fold transformation is manuscript.
used, it has to be decided on or before data inspection
Funding Information Open access funding provided by Research
Expected distribution Test recommended Institute of Molecular Pathology (IMP) / IMBA - Institute of Molecular
Biotechnology / Gregor Mendel Institute of Molecular Plant Biology.
von Mises (unimodal)-expected dir. V-test
Other circular dist. (unimodal)-expected dir. V-test
Compliance with ethical standards
von Mises (unimodal)-no expected dir. Rayleigh test
Other circular dist. Rayleigh test Competing interests The authors declare that they have no competing
(unimodal)-no expected dir. interests.
von Mises (bimodal-modes mirrored) Rayleigh test (after f-fold
transformation) Open Access This article is distributed under the terms of the Creative
von Mises Hermans-Rasson test Commons Attribution 4.0 International License (http://
(bimodal-mode-symmetry unknown) creativecommons.org/licenses/by/4.0/), which permits unrestricted use,
von Mises (> 2 modes-modes mirrored) Rayleigh test (after f-fold distribution, and reproduction in any medium, provided you give
transformation) appropriate credit to the original author(s) and the source, provide a link
von Mises (> 2 mode-symmetry unknown) Hermans-Rasson test to the Creative Commons license, and indicate if changes were made.

References
where testing is difficult. Pending further research, researchers
should consider themselves free to select whichever of these Agostinelli C, Lund U (2013) R package circular: Circular Statistics (ver-
sion 0.43). https://r-forge.r-project.org/projects/circular
three tests is traditionally used in their discipline. However, we
Archibald AM, Bogdanov S, Patruno A, Hessels JWT, Deller AT, Bassa
hope that the results presented here will encourage the imple- C, Janssen GH, Kaspi VM, Lyne AG, Stappers BW, Tendulkar SP,
mentation of the Hermans-Rasson test in easy-to-use software D’Angelo CR, Wijnands R (2015) Accretion-powered pulsations in
packages. Pending such developments, R users may find our an apparently quiescent neutron star binary. Astrophys J 807:62
implementation of some value. Azzalini A (1985) A class of distributions which includes the normal
ones. Scand J Stat 12:171–178
Batschelet E (1981) Circular statistics in biology. Academic press,
London
Bergin TM (1991) A comparison of goodness-of-fit tests for analysis of
Conclusion nest orientation in western kingbirds (Tyrannus verticalis). Condor
93:164–171
We explored the power of a wide range of approaches to test Bogdan M, Bogdan K, Futschik A (2002) A data driven smooth test for
non-uniformity of distributions (see Table 1 for a summary of circular uniformity. Ann Inst Stat Math 54:29–44
Durand D, Greenwood JA (1958) Modifications of the Rayleigh test for
our recommendations). When researchers expect a unimodal
uniformity in analysis of two-dimensional orientation data. J Geol
departure and cannot a priori define the expected mean direc- 66:229–238
tion, the Rayleigh test should be used. In case there exists a Fisher NI (1995) Statistical analysis of circular data. Cambridge
predefined mean direction, the V-test offers most power. For University Press, Cambridge
multimodal departures, the Hermans-Rasson test performed Fitak RR, Johnsen S (2017) Bringing the analysis of animal orientation
data full circle: model-based approaches with maximum likelihood.
very well. However, in cases of symmetrical multiple modes, J Exp Biol 220:3878–3882
all tests perform poorly. The most powerful approach in such Gaumond M, Réthoré PE, Ott S, Peña A, Bechmann A, Hansen KS
cases is a transformation of the f-fold symmetry into a (2014) Evaluation of the wind direction uncertainty and its impact
unimodal situation and applying a Rayleigh test. In this case, on wake modeling at the Horns Rev offshore wind farm. Wind
Energy 17:1169–1178
the number of symmetrical modes must be defined a priori in
Gustafson CL, Partch CL (2014) Emerging models for the molecular
order to avoid alpha error inflation. Our paper provides up-to- basis of mammalian circadian timing. Biochemistry 54:134–149
date and comprehensive comparisons of available tests, with Hermans M, Rasson J (1985) A new Sobolev test for uniformity on the
clear guidelines which approach to use in which situation. We circle. Biometrika 72:698–702
note, however, that we have considered only continuously Humphreys RK, Ruxton GD (2017) Consequences of grouped data for
testing for departure from circular uniformity. Behav Ecol Sociobiol
distributed data, and for aggregated (grouped) circular data,
71:167
we recommend Humphreys and Ruxton (2017) as a useful Jammalamadaka SR, SenGupta A (2001) Topics in Circular Statistics.
starting point. World Scientific, Singapore
Ley C, Verdebout T (2017) Modern directional statistics. Chapman &
Acknowledgments We thank two reviewers for very insightful Hall/CRC Press, Boca Raton
comments. Mardia KV, Jupp PE (2000) Directional statistics. Wiley, Chichester

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

128 Page 10 of 10 Behav Ecol Sociobiol (2018) 72: 128

Oliveira Pérez M, Crujeiras Casais RM, Rodríguez Casal A (2014) Shimatani IK, Yoda K, Katsumata N, Sato K (2012) Toward the quanti-
NPCirc: An R package for nonparametric circular methods. https:// fication of a conceptual framework for movement ecology using
www.jstatsoft.org/article/view/v061i09 circular statistical modeling. PLoS One 7:e50309
Pewsey A (2000) Problems of inference for Azzalini’s skewnormal dis- Stephens M (1969) A goodness-of-fit statistic for the circle, with some
tribution. J Appl Stat 27:859–870 comparisons. Biometrika 56:161–168
Pewsey A, Neuhäuser M, Ruxton GD (2013) Circular statistics in R. Taube JS (2007) The head direction signal: origins and sensory-motor
Oxford University Press, Oxford integration. Annu Rev Neurosci 30:181–207
Pycke JR (2010) Some tests for uniformity of circular distributions pow- Watson GS, Williams EJ (1956) On the construction of significance tests
erful against multimodal alternatives. Can J Stat 38:80–96 on the circle and the sphere. Biometrika 43:344–352
Rayleigh L (1880) On the resultant of a large number of vibrations of the Wiltschko W, Wiltschko R (1972) Magnetic compass of European robins.
same pitch and of arbitrary phase. Philos Mag 10:73–78 Science 176:62–64
Ruxton GD (2017) Testing for departure from uniformity and estimating Zar JH (2013) Biostatistical Analysis: Pearson New International Edition.
mean direction for circular data. Biol Lett 13:20160756 Pearson Higher Education, New York

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]