You are on page 1of 15

Environmental and Ecological Statistics 8, 237±251, 2001

Spectral tests of randomness for spatial


point patterns
M O I R A A . M U G G L E S TO N E 1 and E R I C R E N S H AW 2
1
MRC Institute for Environment and Health, University of Leicester, 94 Regent Road, Leicester
LE1 7DD, UK
E-mail: mam14@le.ac.uk
2
Department of Statistics and Modelling Science, University of Strathclyde, Livingstone Tower,
26 Richmond Street, Glasgow G1 1XH, UK

Received July 1999; Revised May 2000

We develop a spectral framework for testing the hypothesis of complete spatial randomness (CSR)
for a spatial point pattern. Five formal tests based on the periodogram (sample spectrum) proposed in
Mugglestone (1990) are considered. A simulation study is used to evaluate and compare the power of
the tests against clustered, regular and inhomogeneous alternatives to CSR. A subset of the tests is
shown to be uniformly more powerful than the others against the alternatives considered. The
spectral tests are also compared with three widely used space-domain tests that are based on the
mean nearest-neighbor distance, the reduced second-order moment function (K-function), and a
bivariate CrameÂr-von Mises statistic. The test based on the scaled cumulative R-spectrum is more
powerful than the space-domain tests for detecting clustered alternatives to CSR, especially when the
number of events is small.
Keywords: clustering, heterogeneity, inhibition, point spectrum, power, spatial point processes, tests
of hypotheses

1352-8505 # 2001 Kluwer Academic Publishers

1. Introduction
The need to analyze spatial point patterns features prominently in disciplines as diverse as
archaeology, astronomy, ecology, forestry, geography, medicine and meteorology (see, for
example, Neymann and Scott, 1958; Graham, 1980; Haining, 1982; Harkness and Isham,
1983; Diggle et al., 1991; Penttinen et al., 1992; and Mugglestone and Taylor, 1994). Until
recently, most available methods of analysis have been ``space-domain'' techniques that
involve the examination of inter-point distances (see, for example, Diggle, 1983). The
potential of applying spectral techniques in the study of spatial point patterns is gaining
recognition, at least in relation to the exploratory stages of analysis (see Mugglestone and
Renshaw, 1996a, b). Formal procedures for constructing tests of hypotheses using the
periodogram (sample spectrum) are now needed.
The central hypothesis in the analysis of spatial point patterns is that of complete spatial
randomness (CSR). Alternative hypotheses can range from small-scale local interaction,
such as clustering or inhibition, to large-scale environmental effects, such as those related

1352-8505 # 2001 Kluwer Academic Publishers


238 Mugglestone, Renshaw

to wind conditions or variations in fertility, that may induce wave-like changes in spatial
intensity. In this paper we consider ®ve spectral tests of CSR proposed by Mugglestone
(1990). We evaluate their power relative to each other and relative to three of the best
available space-domain tests.
The paper is organized as follows. Section 2 reviews the steps involved in calculating
the periodogram for a spatial point pattern. Section 3 describes the ®ve spectral tests of
CSR and the three space-domain tests. In Section 4 we use a simulation study to examine
the power of the tests against clustered, regular and inhomogeneous alternatives to CSR.
The results of the simulation study are discussed in Section 5.

2. The periodogram for a spatial point pattern


Mugglestone and Renshaw (1996a) provide a detailed account of both the methodology
and the application of spectral techniques for spatial point processes, and they present
illustrative examples involving both simulated and real patterns. Readers unfamiliar with
the spectral approach and who are interested in speci®c details should consult that paper;
here we provide a summary of the underlying concepts that is suf®cient for our current
needs.
Suppose that an observed pattern consists of N events ( points) in a rectangle with sides
of length lx and ly . The ®rst step in calculating the periodogram is to scale the coordinates
of the events to the unit square; this reduces bias in the periodogram at low frequencies
(see Bartlett, 1964). Let f…xi ; yi †g…i ˆ 1; . . . ; N† denote the coordinates of the events
relative to the unit square. The periodogram is given by
( )2 ( )2
X
N XN
f^…wp ; wq † ˆ cosfN…wp xi ‡ wq yi †g ‡ sinfN…wp xi ‡ wq yi †g ; …1†
iˆ1 iˆ1

which is calculated for the frequencies …wp ; wq † ˆ …2pp=N; 2pq=N† where


p ˆ 0; 1; . . . ; pmax and q ˆ qmax ; qmax ‡ 1; . . . ; qmax 1 for suitable values of pmax
and qmax (see below).
Calculation of the periodogram amounts to a transformation from coordinate data to a
…pmax ‡ 1†62qmax matrix of periodogram values. The ordinates (elements) of the
periodogram hold information about the strength of periodicities in the intensity of the
point pattern (or, more formally, in the covariance density function of the underlying
process). The term f^…wp ; wq † corresponds to a periodic pattern with p repeats in the x
direction and q repeats in the y direction. Care must be taken when drawing inferences
from the periodogram about scales of pattern present in the data, since the periodogram
relates to the unit square and inverse scaling is needed to convert back to the scale of the
original coordinates.
Completely random spatial point processes are characterized by ``¯at'' periodograms;
the values are roughly constant at all frequencies because no one frequency dominates the
process. The periodograms of cluster processes and inhomogeneous processes are typi®ed
by large values at low frequencies; those of inhibition processes are typi®ed by small
values at low frequencies, compensated by larger values at high frequencies. Extensive
examples which illustrate the spectral properties of cluster and inhibition processes may be
found in Mugglestone and Renshaw (1996a).
Spectral tests of randomness for spatial point patterns 239

A fundamental difference exists between point spectra and the more familiar lattice
spectra. A lattice spectrum is obtained from a two-dimensional extension of standard time
series analysis, and relates to the spatial structure of a measured variable (for example, tree
crown height or pixel intensity) at ®xed locations on an Mx 6My grid or lattice. In contrast,
a point spectrum describes the spatial structure of the locations themselves (for example,
tree positions in a naturally regenerating forest) and does not involve measured variables
such as tree height. For lattice spectra, the highest frequency that can be resolved (the
Nyquist frequency) is completely controlled by the grid spacing; it corresponds to
…p; q† ˆ …Mx =2; My =2† and represents a ``saw-tooth'' pattern on the grid. The events of a
point pattern may, however, be located arbitrarily close together so that there is no
mathematical constraint on the upper limits for p and q. There is, nevertheless, a very real
problem in determining which values of p and q we should choose. Given N events, it
follows from (1) that we should consider at most N/2 frequencies, f…wp ; wq †g, in order to
maintain some semblance of independence between the corresponding spectral estimates.
Exceeding this limit involves the calculation of more than N transformations of the
coordinate data (counting cosine and sine terms separately) and con¯icts with the notional
N 1 degrees of freedom in the data. For an initial exploratory analysis we recommend
regarding the N events as though they correspond to an N 0 6N 0 lattice process, where
N 0 &HN. We can then evaluate (1) for p ˆ 0; . . . ; N 0 =2 and q ˆ N 0 =2; . . . ; N 0 =2 1 to
produce a periodogram with N 1 ordinates. In order to conduct formal tests of CSR we
must select at most N/2 of these periodogram ordinates to use as a basis for the tests.
Mugglestone (1990) and Mugglestone and Renshaw (1996a) compared periodograms
for simulated realizations of cluster and inhibition processes with those of completely
random processes. The results suggest that the most useful techniques for detecting
clustering and regularity using the periodogram are likely to be those which consider
power at low frequencies. Mugglestone (1990) therefore proposed the following method
of selecting periodogram ordinates for use in formal tests of CSR. First, arrange the
ordinates f f^…wp ; wq †g in order of increasing frequency magnitude, that is,
r ˆ H… p2 ‡ q2 †. Next, select the ®rst n ˆ ‰N=2Š ordinates from the list, where ‰N=2Š
denotes the integer part of N/2. The selected ordinates are referred to collectively as the
reduced periodogram, denoted by f f~…wp ; wq †g. Whether or not this approach will
generally produce an optimal choice of periodogram ordinates on which to base tests of
CSR will, of course, be highly dependent on the exact nature of any departures from CSR
present in a given data-set.
Renshaw and Ford (1983, 1984) de®ne two one-dimensional summaries of the
periodogram. The R-spectrum, f^R …r†, summarizes average periodogram values for
ordinates with similar values of r ˆ H… p2 ‡ q2 †; it is used to investigate scales of pattern.
If a spatial pattern is isotropic (that is, if it does not contain any directional structure), then
the R-spectrum fully captures the second-order structure of the pattern. The Y-spectrum,
f^Y …y†, summarizes average periodogram values for ordinates with similar values of
y ˆ tan 1 … p=q†; it is used to investigate directional features. Formally, we de®ne
XX XX
f^R …r† ˆ f^…wp ; wq †=nr and f^Y …y† ˆ f^…wp ; wq †=ny : …2†
r0 y r y0

Here nr denotes the number of periodogram ordinates for which r 15r 0  r and
0  y5180 ; ny denotes the number of periodogram ordinates for which r  1 and
240 Mugglestone, Renshaw

y 5 5y0  y ‡ 5 . Experience shows that the r- and y-increments of one and 10
implied above are useful ®rst tries. Note that the ordinate for p ˆ q ˆ 0 is excluded from
the averaging procedure (since it has different sampling properties to the rest of the
periodogram); ordinates corresponding to positive values of q on the row for p ˆ 0 are also
excluded (since these are exact repeats of ordinates for negative values of q). These
ordinates are discarded before calculating nr and ny .
The statistical properties of periodogram ordinates center around the w2 distribution.
Under the hypothesis of CSR, the theoretical spectrum is constant and equal to l, the
intensity of the process (the expected number of events per unit area). The distribution of
individual periodogram ordinates is given by
2f^…wp ; wq †=l*w22 : …3†
Since we have already scaled to the unit square we may estimate l by N, so that
2f^…wp ; wq †=N*w22 . It immediately follows that

f^R …r†=N*…2nr †
1 2
w2nr ; …r ˆ 1; 2; . . .†
and

f^Y …y†=N*…2ny † 1 2
w2ny ; …y ˆ 0 ; 10 ; . . . ; 170 †:
We refer to f^R …r†=N and f^Y …y†=N as scaled R- and Y-spectra, respectively.
Several ways of constructing tests of CSR using the n ˆ ‰N=2Š ordinates of the reduced
periodogram are presented in the next section. The ®rst test is a direct extension of a test
used in time series analysis. The remaining spectral tests are presented in order of
increasing complexity in terms of the extent to which they incorporate detailed
information about the distribution of periodogram ordinates under CSR and the likely
impact of particular alternatives to CSR on the form of the underlying spectrum. The ®rst
four tests draw on the equivalence between the w22 distribution and the exponential
distribution with mean two.

3. Formal tests of CSR


3.1 Test based on the maximum periodogram ordinate

Fisher (1929) introduced a test for the maximum ordinate of the periodogram,
f f^…wp †g… p ˆ 0; 1; . . . ; N=2†, of a one-dimensional time series, fXt g. Under the
assumption that the Xt are independently, identically distributed as N…0; s2 † random
variables, the distribution of periodogram ordinates is given by
2f^…wp †=s2 *w22 : …4†
Note the similarity between (3) and (4). The denominator on the left hand side of each
expression represents the variance of the null process: for a completely random spatial
point process with intensity l, the number of events per unit area follows a Poisson
distribution with mean and variance both equal to l (see, for example, Diggle, 1983).
Spectral tests of randomness for spatial point patterns 241

The test for the maximum ordinate of the periodogram is based on the statistic
g ˆ maxf2f^…wp †=^
s2 g:
p

Fisher (1929) determined an exact expansion for Pr…g4z† under the null process. The ®rst
term of the expansion yields the approximation
Pr…g4z†^1 f1 exp… z=2†gn ; …5†
where n ˆ ‰N=2Š. This result was noted independently by Walker (1914). The
approximation in (5) arises from estimating the variance of the process, s2 ; when the
variance is known the result is exact.
To apply Fisher's (1929) test to the reduced periodogram of a spatial point pattern we
calculate
g ˆ max f2f~…wp ; wq †=Ng
…wp ;wq †

and evaluate Pr…g4z† using (5). Conditional on the intensity of the process (estimated by
N) the variance under CSR is known, and so the result in (5) is exact when applied to a
spatial point pattern.
As periodograms for cluster processes tend to have large values at low frequencies and
those for inhibition processes tend to have very small values at low frequencies a two-
tailed test based on (5) is generally required. For a test at the …100 a†% signi®cance
level, we refer g to the upper and lower …100 a=2†% critical values given by
z ˆ 2 lnf1 …a=200†1=n g and z ˆ 2 ln‰1 f…200 a†=200g1=n Š, respectively.

3.2 Test based on the coef®cient of variation

Greenwood (1946) introduced a goodness-of-®t test for an exponential distribution in the


context of spacings between observations from a uniform distribution. Greenwood's test is
equivalent to a test derived by Kimball (1947) for studying the dispersion of observations
from an exponential random variable. Both tests are based on the fact that the coef®cient of
variation (CV) of an exponentially distributed random variable is one.
It follows from (3) that, when CSR holds, the CV of the values in the reduced
periodogram is one. In order to apply Greenwood's test to the reduced periodogram we
need to calculate the sample mean and variance of the ordinates of the reduced
periodogram given by
XX XX
f ˆ f~…wp ; wq †=n and s2 ˆ f f~…wp ; wq † fg =…n 1†;
2

p q p q

respectively. Then
Gn 1 ˆ s2 =f…n 1†f2 g ‡ 1=…n 1† ˆ …CV2 ‡ 1†=…n 1†
is known as Greenwood's statistic, and is referred to tables of critical values of
…n 1†Gn 2 (see, for example, Table 8.3 of d'Agostino and Stephens, 1986).
The variance of the reduced periodogram will tend to be in¯ated by both cluster and
inhibition processes, while the mean will increase for cluster processes and decrease for
242 Mugglestone, Renshaw

inhibition processes. Thus the CV will be less than or greater than one depending on the
exact form of non-randomness, and a two-sided test is required in order to detect both
clustered and regular alternatives to CSR.
The test based on the CV of the reduced periodogram ignores the fact that the
distribution of periodogram ordinates under CSR is fully speci®ed by (3). Tests that
explicitly incorporate knowledge of the mean and variance of the periodogram ordinates
may be more powerful than the test based only on their ratio. We shall, therefore, now
explore non-parametric tests for a fully speci®ed exponential distribution.

3.3 Test based on the Kolmogorov-Smirnov statistic

A Kolmogorov-Smirnov test for an exponential distribution with mean two may be applied
to the ordinates of the reduced periodogram as follows. First, calculate the probability
integral transformation (PIT) given by
Zi ˆ Hf2f~…wp ; wq †=Ng; …i ˆ 1; . . . ; n†;
where H…x† ˆ 1 exp… x=2† is the cumulative distribution function (CDF) of the
exponential distribution with mean two. Next, calculate
D ‡ ˆ max fi=n Z…i† g and D ˆ max fZ…i† …i 1†=ng;
1in 1in

where fZ…i† g …i ˆ 1; . . . ; n† are the order statistics of the PIT. The statistics D ‡ and D
represent maximum positive and negative deviations of the empirical distribution function
(EDF) from the null CDF, respectively. Cluster processes will tend to produce large values
of D ‡ , whereas inhibition processes will tend to produce large values of D . The two-
sided Kolmogorov-Smirnov statistic, given by D ˆ max…D ‡ ; D †, is therefore needed to
detect both clustered and regular alternatives to CSR. Critical values of D ‡ ; D and D are
given in, for example, Table 4.2 of d'Agostino and Stephens (1986).

3.4 Test based on the Anderson-Darling statistic


The Kolmogorov-Smirnov test is good for detecting one or two outliers in an otherwise
good ®t but deviations in the tails of the distribution are detected more effectively by the
Anderson-Darling statistic (Pearson and Hartley, 1976). The Anderson-Darling statistic is
given by
X
n
A2 ˆ n …2i 1†‰lnfZ…i† g ‡ lnf1 Z…n ‡ 1 i† gŠ=n;
iˆ1

where fZ…i† g …i ˆ 1; . . . ; n† are the order statistics of the PIT as de®ned in Section 3.3. The
test based on A2 is naturally two-sided. Critical values of A2 are given in, for example,
Table 4.2 of d'Agostino and Stephens (1986).
Spectral tests of randomness for spatial point patterns 243

3.5 Test based on the cumulative R-spectrum

All the tests considered so far use a wide range of frequencies constrained only by the
number of degrees of freedom available in the data. Tests based on more precise bands of
frequency magnitudes might be even more powerful, since certain frequency ranges will
be more affected by departures from CSR than others. One way of achieving this more
sensitive form of test is to determine appropriate ``cut-off'' levels r1  r  r2 in the
R-spectrum that are especially tailored for testing against speci®c alternatives to CSR.
A natural extension to the R-spectrum is the cumulative R-spectrum de®ned by
X
r
f^CR …r† ˆ ns f^R …s†=Nr ;
sˆ1
Pr
where Nr ˆ s ˆ 1 ns . This indicates the total power present in the spectrum up to a scale
of pattern, r. The scaled cumulative R-spectrum is analogous to the scaled R-spectrum and
has the distribution

f^CR …r†=N*…2Nr † 1 2
w2Nr : …6†
Invoking (6) for a given value of r allows us to calculate critical values of f^CR …r†=N for a
test of CSR. A two-sided test is need to detect both clustered and regular alternatives to
CSR.
The simulation experiments of Mugglestone (1990) and Mugglestone and Renshaw
(1996a) suggest that tests of CSR should be based on ordinates of the cumulative R-
spectrum for which r  5. At these scales, sample spectra for cluster and inhibition
processes tend to be well above and well below the ¯at spectrum expected under CSR,
respectively. Since the ordinates of the cumulative R-spectrum for r  2; 3; 4; 5 are based
on Nr ˆ 6; 14; 24; 40 periodogram values, these form appropriate test statistics for patterns
containing at least N ˆ 12; 28; 48; 80 events, respectively.
In Section 4 we will use a power study to compare the proposed spectral tests of CSR
with each other and with three existing space-domain tests. We shall, therefore, now
brie¯y describe the space-domain tests.

3.6 Test based on the mean nearest-neighbor distance

For a spatial point pattern consisting of NP events observed in the unit square, the mean
nearest-neighbor distance is given by d ˆ Niˆ 1 di =N, where di …i ˆ 1; . . . ; N† represents
the distance from the ith event to its nearest neighbor. Under CSR, the distribution of d is
approximately normal with mean and variance given by
1=2 1 3=2
md ˆ 0:500N ‡ 0:206N ‡ 0:164N and s2d ˆ 0:070N 2
‡ 0:148N 5=2
;
respectively (Donnelly, 1978). For cluster processes, d tends to be greater than md, whilst
for inhibition processes, it tends to be less than md. Thus a two-sided test for CSR is
obtained by comparing
T ˆ …d md†=sd
244 Mugglestone, Renshaw

with critical values of the standard normal distribution. This is a modi®cation of a test
proposed by Clark and Evans (1954) that takes account of dependencies amongst the
nearest-neighbor distances and incorporates a correction for edge effects. It is known to be
powerful against clustered and regular alternatives to CSR (Diggle, 1979; Ripley, 1979).

3.7 Test based on the reduced second-order moment function

The reduced second-order moment function, or K-function, represents the expected


number of events within distance t of a randomly chosen event. Under CSR, K…t† is equal
to pt2 , whereas for cluster processes it exceeds pt2 and for inhibition processes it is smaller
than pt2 . A test of CSR is typically based on
^
Lm ˆ sup jHfK…t†=pg tj;
t5t0

^ is Ripley's (1976) estimator for K…t† and t0 is some maximum distance of


where K…t†
interest. We shall use t0 ˆ 1:25=HN, so that as N increases the test concentrates on smaller
inter-event distances.
The critical value of the test based on Lm is obtained by Monte Carlo simulation (see, for
example, Dwass, 1957) as follows. Let L1m denote the value of Lm for the sample pattern,
and calculate fLim g…i ˆ 2; . . . ; j†, the values of Lm for … j 1† realizations of CSR with
intensity N. The attained signi®cance level of the test of CSR is given by rank…L1m †=j. We
shall use j 1 ˆ 99 realizations of CSR, these being obtained by simulating from a
homogeneous Poisson process with intensity N (see Section 4). This test is more powerful
than that based on the mean nearest-neighbor distance for detecting simple sequential
inhibition processes and Strauss inhibition processes (Diggle, 1979; Ripley, 1979; see also
Section 4).

3.8 Test based on the bivariate CrameÂr-von Mises statistic


This test, which is due to Zimmerman (1993), measures deviations between the EDF of the
two-dimensional coordinates of a point pattern and the bivariate uniform distribution. For
data in the unit square, the test statistic is
N X
X N
2 ˆ
o …1 jxi xj j†…1 jyi yj j†=…4N†
iˆ1 jˆ1

X
N
‡ …x2i xi 0:5†…y2i yi 0:5†=2 ‡ N=9:
iˆ1

Clustering and inhomogeneity produce large values of o,  whereas inhibition produces


 2 under CSR is that of an in®nite sum of w21
small values. The limiting distribution of o
random variables, and is tabulated in Zimmerman (1994). This test is more powerful than
those based on the mean nearest-neighbor distance and the reduced second-order moment
function for detecting inhomogeneous Poisson processes with planar trends (Zimmerman,
1993; see also Section 4).
Spectral tests of randomness for spatial point patterns 245

4. Simulation study
In order to evaluate the performance of the tests of CSR we simulated four distinct types of
spatial point process in the unit square. The processes considered were as follows.

* Homogeneous Poisson process (HPP) constrained to produce N ˆ 12, 25, 50, 75 or


100 events. The coordinates are pairs of independent uniform random variables on
the interval [0,1].
* Poisson cluster process (PCP) constrained to produce N ˆ 25 or 75 events. The
cluster centers form an HPP with lP ˆ 5 or 10 events when N ˆ 25, and lP ˆ 10 or
15 events when N ˆ 75. The number of events per cluster is a Poisson random
variable with mean m ˆ N=lP . Events of the process are randomly allocated to
cluster centers, and displacements relative to cluster centers are independently and
uniformly distributed in circles of radius u ˆ 0:15 or 0.30 for each combination of N
and lP . Toroidal edge correction prevents events from leaving the unit square.
* Simple sequential inhibition process (SSIP) constrained to produce N ˆ 25 or 75
events. The ®rst event is simulated uniformly on the unit square. Subsequent events,
also distributed uniformly, are discarded if any ``older'' event lies within a distance d
when the unit square is wrapped onto a torus. The minimum inter-event distance is
given by d ˆ 0:04, 0.08 or 0.12 when N ˆ 25, and d ˆ 0:03, 0.04 or 0.05 when
N ˆ 75.
* Inhomogeneous planar trend Poisson process (IPP) with intensity

l…x; y† ˆ flx …lx 1†xgfly …ly 1†yg

at location …x; y†. This process is constrained to produce N ˆ 25 or 75 events for all
possible combinations of lx ˆ 1, 4 and ly ˆ 4, 10, and also for lx ˆ ly ˆ 10.
Simulation is based on rejection sampling: a potential event at location (x,y) is
generated uniformly on the unit square but is rejected unless l…x; y†= maxfl…x; y†g is
less than or equal to a further random variable drawn from the uniform distribution
on the interval ‰0; 1Š.

Simulated realizations of each type of process are shown in Fig. 1. More detailed
descriptions of the processes can be found in, for example, Diggle (1983).
Each combination of parameter values was replicated 1000 times. The simulations of
the HPP allow us to estimate the size of nominal 4% tests. The estimated size is given by
the proportion of the 1000 replicates for which CSR was rejected at the 4% level (see Table
1). The simulations of the other processes allow us to estimate the power of the tests
against different alternatives to CSR (see Tables 2, 3 and 4). For these processes, and for a
particular combination of parameter values, the estimated power is given by the proportion
of the 1000 replicates for which CSR was rejected at the 4% level.
The processes, the parameter values, and the size of the tests used for the power
calculations were chosen to match those of Zimmerman (1993) so that we could compare
our results directly with his. We implemented all the spectral tests, and as a check on our
calculations we also implemented the space-domain tests based on the mean nearest-
neighbor distance and the reduced second-order moment function. The number of
realizations of CSR used in the test based on the reduced second-order moment function
246 Mugglestone, Renshaw

Figure 1. Simulated realizations of spatial point processes with N ˆ 75 events in the unit square:
(a) HPP; (b) PCP with lP ˆ 10, m ˆ 7:5 and u ˆ 0:15; (c) SSIP with d ˆ 0:05; (d) IPP with
lx ˆ ly ˆ 10. The dashed line indicates the boundary of the unit square.

was set at j 1 ˆ 99 to match Zimmerman's (1993) calculations. Note that Zimmerman's


(1993) implementation of the test based on the reduced second-order moment function
uses t0 ˆ 0:20 when N ˆ 25 and t0 ˆ 0:15 when N ˆ 75. These are almost the same as the
values given by the rule t0 ˆ 1:25=HN. The results for the test based on the bivariate
CrameÂr-von Mises statistic in Tables 2, 3 and 4 are reproduced from Zimmerman (1993).

Table 1. Estimated size of nominal 4% tests of CSR against N events of an HPP.

Model Parameters

Test Statistic N 12 25 50 75 100

g 0.03 0.01 0.02 0.02 0.01


Gn 1 0.07 0.05 0.04 0.04 0.04
D 0.02 0.02 0.02 0.02 0.02
A2 0.05 0.05 0.05 0.05 0.04
f^CR …r†=N 0.03 0.04 0.04 0.04 0.04

T 0.05 0.04 0.03 0.03 0.03


Lm 0.05 0.05 0.04 0.04 0.06
Spectral tests of randomness for spatial point patterns 247

Table 1 shows that the tests based on the maximum periodogram ordinate and the
Kolmogorov-Smirnov statistic are always conservative (the estimated size should be about
0.04 for nominal 4% tests). The test based on the coef®cient of variation is liberal when the
number of events is small, and that based on the Anderson-Darling statistic is slightly
liberal for patterns with as many as 75 events. The test based on the scaled cumulative R-
spectrum is slightly conservative when the number of events is very small, but is closer to
achieving the nominal size than any other test considered. The test based on the mean
nearest-neighbor distance is slightly liberal when the number of events is very small but
becomes conservative as the number of events increases. The test based on the reduced
second-order moment function is generally liberal.
Table 2 shows that the tests based on the maximum periodogram ordinate and the
Anderson-Darling statistic are reasonably powerful for detecting clustering when the
cluster radius is small. The test based on the coef®cient of variation has virtually no power
unless the number of events is large. The Kolmogorov-Smirnov statistic is never
particularly powerful for the range of parameter values considered. The test based on the
scaled cumulative R-spectrum is very powerful when the cluster radius is small, and
reasonably powerful when the cluster radius is large. The tests based on the mean nearest-
neighbor distance and the reduced second-order moment function are also powerful when
the cluster radius is small but never more so than the test based on the scaled cumulative R-
spectrum. The power of the test based on the bivariate CrameÂr-von Mises statistic is also
greater when the cluster radius is small. This test is slightly more powerful than the test
based on the scaled cumulative R-spectrum when the number of events and the cluster
radius are both large.
Table 3 shows that the tests based on the maximum periodogram ordinate, the
coef®cient of variation and the bivariate CrameÂr-von Mises statistic have almost no power
to detect hard-core inhibition. The tests based on the Kolmogorov-Smirnov statistic, the
Anderson-Darling statistic and the scaled cumulative R-spectrum perform better. These

Table 2. Estimated power of tests of CSR against N events of a PCP with cluster intensity
lP , mean number of events per cluster m, and cluster radius u.{

Model Parameters

N 25 25 25 25 75 75 75 75
lP 5 5 10 10 10 10 15 15
m 5.0 5.0 2.5 2.5 7.5 7.5 5.0 5.0
Test Statistic u 0.15 0.30 0.15 0.30 0.15 0.30 0.15 0.30

g 0.82 0.22 0.40 0.06 0.92 0.29 0.71 0.17


Gn 1 0.06 0.15 0.07 0.08 0.57 0.23 0.39 0.14
D 0.46 0.03 0.15 0.01 0.41 0.04 0.24 0.04
A2 0.96 0.19 0.60 0.11 0.89 0.14 0.71 0.11
f^CR …r†=N 0.95 0.31 0.61 0.13 0.99 0.31 0.93 0.19

T 0.86 0.15 0.43 0.05 0.93 0.10 0.68 0.07


Lm 0.87 0.17 0.42 0.08 0.98 0.29 0.88 0.17
2
o 0.60 0.27 0.32 0.13 0.79 0.36 0.61 0.25

 2 are reproduced from Zimmerman (1993).


{Results for o
248 Mugglestone, Renshaw

Table 3. Estimated power of tests of CSR against N events of an SSIP with minimum
inter-event distance d.{

Model Parameters

N 25 25 25 75 75 75
Test Statistic d 0.04 0.08 0.12 0.03 0.04 0.05

g 0.02 0.05 0.40 0.00 0.00 0.00


Gn 1 0.06 0.06 0.07 0.04 0.04 0.05
D 0.02 0.12 0.70 0.11 0.30 0.72
A2 0.05 0.22 0.85 0.18 0.42 0.86
f^CR …r†=N 0.03 0.10 0.56 0.14 0.32 0.74

T 0.06 0.45 1.00 0.36 0.84 1.00


Lm 0.20 1.00 1.00 1.00 1.00 1.00
2
o 0.04 0.10 0.43 0.05 0.10 0.19

 2 are reproduced from Zimmerman (1993).


{Results for o

tests all behave in a similar way but they are not at all powerful unless the minimum inter-
event distance is very large. The best of the spectral tests is that based on the Anderson-
Darling statistic. The tests based on the mean nearest-neighbor distance and the reduced
second-order moment function are much more powerful than the spectral tests. The test
based on the reduced second-order moment function is by far the most powerful of all the
tests considered.
Table 4 shows that the test based on the Kolmogorov-Smirnov statistic has virtually no
power to detect inhomogeneity. All the other spectral tests, together with the tests based on
the mean nearest-neighbor distance and the reduced second-order moment function
behave in a very similar way: their power is weak against weak trends, is moderate for
stronger trends, and improves as the number of events increases. The best test in this group

Table 4. Estimated power of tests of CSR against N events of an IPP with planar trend
parameters lx and ly .{

Model Parameters

N 25 25 25 25 25 75 75 75 75 75
lx 1 1 4 4 10 1 1 4 4 10
Test Statistic ly 4 10 4 10 10 4 10 4 10 10

g 0.03 0.07 0.05 0.07 0.12 0.10 0.29 0.18 0.35 0.54
Gn 1 0.07 0.11 0.05 0.08 0.09 0.09 0.24 0.12 0.23 0.33
D 0.02 0.01 0.02 0.02 0.03 0.02 0.03 0.05 0.05 0.09
A2 0.06 0.10 0.10 0.14 0.19 0.07 0.10 0.17 0.25 0.33
f^CR …r†=N 0.05 0.10 0.12 0.21 0.29 0.10 0.22 0.31 0.51 0.69

T 0.04 0.07 0.06 0.08 0.10 0.05 0.09 0.08 0.13 0.20
Lm 0.05 0.09 0.12 0.22 0.33 0.11 0.34 0.42 0.68 0.87
2
o 0.17 0.38 0.40 0.61 0.78 0.70 0.93 0.98 0.99 1.00

 2 are reproduced from Zimmerman (1993).


{Results for o
Spectral tests of randomness for spatial point patterns 249

is that based on the reduced second-order moment function. The test based on the scaled
cumulative R-spectrum comes a close second. All the spectral tests in this group
outperform the test based on the mean nearest-neighbor distance. However, the test based
on the bivariate CrameÂr-von Mises statistic has substantially greater power than any other
test, whatever the combination of parameter values.

5. Discussion
The simulation study reveals that the most powerful spectral tests are those based on the
scaled cumulative R-spectrum and the Anderson-Darling statistic. The test based on the
scaled cumulative R-spectrum is to be preferred for detecting clustering and
inhomogeneity, whereas the test based on the Anderson-Darling statistic is to be preferred
for detecting regularity. The remaining spectral tests can be safely discarded. We also
examined the power of Kolmogorov-Smirnov and Anderson-Darling tests of uniformity
over all scales of pattern in the reduced periodogram. Not one of the 1000 replications of
any process under any combination of parameter values resulted in rejection of CSR,
although this lack of power might be a function of the way in which we de®ned the reduced
periodogram.
Comparing the spectral tests with the space-domain tests reveals that the most powerful
method for detecting clustering is the test based on the scaled cumulative R-spectrum. This
result holds for almost all the parameter values considered. The main exceptions occur
when the number of events and cluster radius are both so large that the process verges on
inhomogeneity. In these circumstances, the test based on the bivariate CrameÂr-von Mises
statistic is slightly more powerful. For detecting regularity, the test based on the reduced
second-order moment function is unsurpassed. The test based on the mean nearest-
neighbor distance is also more powerful than the spectral tests against this form of non-
randomness. Against simulations of planar inhomogeneity, the test based on the bivariate
CrameÂr-von Mises statistic is uniformly most powerful. The next best tests are those based
on the scaled cumulative R-spectrum and the reduced second-order moment function,
these being almost equally powerful when the number of events is small. The test based on
the Anderson-Darling statistic is less powerful but it outperforms the test based on the
mean nearest-neighbor distance.
Clearly, no one test is uniformly most powerful against the entire range of alternatives to
CSR considered here. Nevertheless, we have demonstrated that the spectral tests based on
the the scaled cumulative R-spectrum and the Anderson-Darling statistic are useful
additions to the existing collection of space-domain tests of CSR. Our calculations of the
estimated size of nominal 4% tests con®rm that the stated null distributions for the two
periodogram-based test statistics are valid, even for patterns with as few as 12 events. The
estimated size of the test based on the cumulative R-spectrum is closer to the nominal size
than for any of the other spectral tests. This is explained by the fact that periodogram
ordinates require smoothing to achieve consistency, and that ordinates of the cumulative
R-spectrum incorporate a greater degree of smoothing than any other spectral test statistic.
The tendency for tests based on the Anderson-Darling statistic and the reduced second-
order moment function to be liberal is not a major problem, since the additional power of
these tests against inhibition processes more than outweighs the loss of one or two
percentage points in size. The Anderson-Darling statistic does, however, suffer from the
250 Mugglestone, Renshaw

disadvantage that the test statistic must be referred to tables of critical values, and so only
approximate signi®cance values can be obtained.

Acknowledgments
We thank Craig Corbett for running the programs that performed the simulations and the
referees for their constructive comments and suggestions for improving the original
manuscript.

References
d'Agostino, R.B. and Stephens, M.A. (1986) Goodness-of-Fit Techniques, Dekker, New York.
Bartlett, M.S. (1964) The spectral analysis of two-dimensional point processes. Biometrika, 51, 299±
311.
Clark, P.J. and Evans, F.C. (1954) Distance to nearest neighbor as a measure of spatial relationships
in populations. Ecology, 35, 23±30.
Diggle, P.J. (1979) On parameter estimation and goodness-of-®t testing for spatial point patterns.
Biometrics, 35, 87±101.
Diggle, P.J. (1983) Statistical Analysis of Spatial Point Patterns, Academic Press, London.
Diggle, P.J., Lange, N., and BenesÆ, F.M. (1991) Analysis of variance for replicated spatial point
patterns in clinical neuroanatomy. Journal of the American Statistical Association, 86, 618±25.
Donnelly, K. (1978) Simulations to determine the variance and edge-effect of total nearest-neighbor
distance. In Simulation Studies in Archaeology, Hodder I. (ed), Cambridge University Press,
pp. 91±5.
Dwass, M. (1957) Modi®ed randomization tests for nonparametric hypotheses. Annals of
Mathematical Statistics, 28, 181±7.
Fisher, R.A. (1929) Tests of signi®cance in harmonic analysis. Proceedings of the Royal Society of
Edinburgh, Series A, 125, 54±9.
Graham, I. (1980) Spectral analysis and distance methods in the study of archaeological
distributions. Journal of Archaeological Science, 7, 105±29.
Greenwood, M. (1946) The statistical study of infectious disease. Journal of the Royal Statistical
Society, Series A, 109, 85±110.
Haining, R. (1982) Describing and modeling rural settlement maps. Annals of the Association of
American Geographers, 72, 211±23.
Harkness, R.D. and Isham, V. (1983) A bivariate spatial point pattern of ants nests. Applied Statistics,
32, 293±303.
Kimball, B.F. (1947) Some basic theorems for developing tests of ®t for the case of the non-
parametric probability distribution function. Annals of Mathematical Statistics, 18, 540±8.
Mugglestone, M.A. (1990) Spectral Analysis of Spatial Processes, Ph.D. thesis, University of
Edinburgh.
Mugglestone, M.A. and Renshaw, E. (1996a) A practical guide to the spectral analysis of spatial
point processes. Computational Statistics and Data Analysis, 21, 43±65.
Mugglestone, M.A. and Renshaw, E. (1996b) The exploratory analysis of bivariate spatial point
patterns using cross-spectra. Environmetrics, 7, 361±77.
Mugglestone, M.A. and Taylor, P.C. (1994) Investigating spatial patterns of convective storms in the
Sahel Region of Africa: Tests for random labelling of storm initiation sites relative to storm
duration, Technical Report Number 94/10, Department of Applied Statistics, The University of
Reading.
Spectral tests of randomness for spatial point patterns 251

Neymann, J. and Scott, E.L. (1958) Statistical approach to problems of cosmology (with discussion).
Journal of the Royal Statistical Society, Series B, 20, 1±43.
Penttinen, A., Stoyan, D., and Henttonen, H.M. (1992) Marked point processes in forest statistics.
Forest Science, 38, 806±24.
Pearson, E.S. and Hartley, H.O. (1976) Biometrika Tables for Statisticians, Volume 2. Cambridge
University Press.
Renshaw, E. and Ford, E.D. (1983) The interpretation of process from pattern using two-dimensional
spectral analysis: methods and problems of interpretation. Applied Statistics, 32, 51±63.
Renshaw, E. and Ford, E.D. (1984) The description of spatial pattern using two-dimensional spectral
analysis. Vegetatio, 56, 75±85.
Ripley, B.D. (1976) The second-order analysis of stationary point processes. Journal of Applied
Probability, 13, 255±66.
Ripley, B.D. (1979) Tests of ``randomness'' for spatial point patterns. Journal of the Royal Statistical
Society, Series B, 41, 368±74.
Walker, G.T. (1914) Correlation in seasonal variation of weather. III. On the criterion for the reality
of relationships or periodicities. Memoranda of the Indian Meteorological Department, 21,
13±5.
Zimmerman, D.L. (1993) A bivariate CrameÂr-von Mises type of test for spatial randomness. Applied
Statistics, 42, 43±54.
Zimmerman, D.L. (1994) On the limiting distribution of and critical values for an origin-invariant
bivariate CrameÂr-von Mises-type statistic. Statistics and Probability Letters, 20, 189±95.

Biographical Sketches
Moira Mugglestone is a statistician with many years' experience of advising scientists,
social scientists, and civil servants on statistical aspects of their work. She is also
experienced in teaching statistics, one of her previous posts being that of Lecturer in
Applied Statistics at The University of Reading. The main responsibilities of her current
post are to provide statistical and epidemiological support within IEH, to develop
multidisciplinary research programmes in the area of environment and health, and to
engage in methodological research in statistics. She has a Ph.D. in spatial statistics, and her
current research interests include spatial statistics, computational statistics, and
environmental epidemiology.
Eric Renshaw is a Professor of Statistics at the University of Strathclyde, with more than
thirty years' experience in stochastic processes and spatial analysis. His main interests lie
in the development of new methodologies and subsequent application over a wide range of
®elds, including biology, forestry, geology, meteorology and physics. He is a Fellow of the
Royal Society of Edinburgh, and is active in several professional statistical societies.

You might also like