Stopping Rules in Principal Components Analysis: A Comparison of Heuristical and Statistical

Approaches
Author(s): Donald A. Jackson
Source: Ecology, Vol. 74, No. 8 (Dec., 1993), pp. 2204-2214
Published by: Ecological Society of America
Stable URL: http://www.jstor.org/stable/1939574 .
Accessed: 17/08/2013 15:45

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Ecological Society of America is collaborating with JSTOR to digitize, preserve and extend access to Ecology.

http://www.jstor.org

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
Ecology, 74(8), 1993, pp. 2204-2214
? 1993 by the Ecological Society of America

STOPPING RULES IN PRINCIPAL COMPONENTS ANALYSIS:

A COMPARISON OF HEURISTICAL AND
STATISTICAL APPROACHES'

DONALD A. JACKSON2
Departmentof Zoology, Universityof Toronto,Toronto,Ontario,CanadaM5S JA]

Abstract. Approaches to determining the number of components to interpret from

principal components analysis were compared. Heuristic procedures included: retaining
components with eigenvalues (Xs) > 1 (i.e., Kaiser-Guttman criterion); components with
bootstrapped Xs > 1 (bootstrapped Kaiser-Guttman); the scree plot; the broken-stick model;
and components with Xs totalling to a fixed amount of the total variance. Statistical ap-
proaches included: Bartlett's test of sphericity; Bartlett's test of homogeneity of the cor-
relation matrix, Lawley's test of the second X;bootstrapped confidence limits on successive
Xs (i.e., significant differences between Xs); and bootstrapped confidence limits on eigen-
vector coefficients (i.e., coefficients that differ significantly from zero). All methods were
compared using simulated data matrices of uniform correlation structure, patterned ma-
trices of varying correlation structure and data sets of lake morphometry, water chemistry,
and benthic invertebrate abundance. The most consistent results were obtained from the
broken-stick model and a combined measure using bootstrapped Xs and associated eigen-
vector coefficients. The traditional and bootstrapped Kaiser-Guttman approaches over-
estimated the number of nontrivial dimensions as did the fixed-amount-of-variance model.
The scree plot consistently estimated one dimension more than the number of simulated
dimensions. Bartlett's test of sphericity showed inconsistent results. Both Bartlett's test of
homogeneity of the correlation matrix and Lawley's test are limited to testing for only one
and two dimensions, respectively.
Key words: bootstrap;eigenvalues;multivariate;ordination;principalcomponentsanalysis;sta-
tistics;stoppingrules.

INTRODUCTION
Although ecologists have compared the results of
various ordination methods (see Gauch 1982a, Pielou
1984, Digby and Kempton 1987, Minchin 1987 for
comparisons), few guidelines exist to evaluate how many
ordination axes should be considered nontrivial and
interpretable. An implicit assumption in the use of
ordination methods is that the experienced ecologist
can separate meaningful patterns from random noise
(i.e., ecologically meaningful information vs. sampling
variation or measurement error; Gauch 1982b). An
ability to distinguish "signal" from "noise" is essential
and in a statistical sense, these decisions provide "stop-
ping rules." The failure to distinguish between signal
and noise may lead to the rejection of useful infor-
mation or the interpretation of ecologically meaning-
less information. In the former case, a loss of infor-
mation may limit our understanding of ecological
processes. In the latter case erroneous conclusions may
result as we would be interpreting essentially mean-
ingless patterns (Jackson et al 1992). Rexstad et al.
(1988) questioned the value of multivariate analyses
' Manuscript received 10 March 1992; revised 11 March
1993; accepted 15 March 1993.
2 Present address: Department of Zoology, University of
Western Ontario, London, Ontario, Canada N6A 5B7.
after their study demonstrated "significant" results from
randomly collected data.
My study examines the issue of assessing multivar-
iate data dimensionality using both heuristic and sta-
tistical approaches. I restrict this study to principal
components analysis (PCA) because it represents the
simplest and most commonly used multivariate meth-
od. In many instances, these results can be extrapolated
to related multivariate techniques. I used a parallel
analysis of field and simulated data to examine the
implications of different: (1) degrees or strength of in-
tervariable correlations; (2) numbers of variables; and
(3) structure within correlation matrices (i.e., blocks of
correlated variables that are uncorrelated with other
variables). These three conditions are important fac-
tors in determining the success of a multivariate anal-
ysis. The types of data ecologists analyze often lead to
different degrees of correlations. For example, vari-
ables like the morphology of organisms or lakes often
show strongly correlated variables whereas correlations
of the abundance of organisms may be weaker. Studies
often differ in their ratio of the number of observations
relative to variables in the analysis. Although some
researchers recognize the importance of having a ratio
of 3: 1 or greater to provide a stable solution (see
Grossman et al. 1991), it is not uncommon to find
studies having a lower observation: variable ratio than

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
December 1993 SIGNIFICANT PRINCIPAL COMPONENTS
recommended. Often the implication for this latter sit-
uation remains unrecognized. Within data sets, there
may be groups of variables (e.g., species) that are highly
correlated with one another, but uncorrelated with oth-
er groups. The within- and between-group structure
also contributes to substantial effects within principal
components analysis. The use of simulated data per-
mits data to be generated having an a priori underlying
dimensionality such that the various methods may be
compared relative to known values (i.e., the true num-
ber of nontrivial components; see Lambert et al. 1990
for a discussion).
METHODS
Data matrices-ecological
Three ecological data sets were used for comparisons
with the simulated data. The first data set was based
on four lake morphological variables from 40 lakes
from south-central Ontario. These variables were as-
sociated strongly with one another and generally had
correlations between 0.6 and 0.9. The second matrix
included measurements on 12 chemical elements or
compounds from the lakes. Correlations for this matrix
ranged from near 0 to 0.7. The third matrix comprised
abundance measurements on 32 benthic invertebrate
taxa from the same 40 lakes. The correlations between
these taxa varied between 0 and 0.7 with most corre-
lations between 0.3 and 0.4. All data in these ecological
data sets were transformed to linearize intervariable
relationships and approximate normal distributions (see
Jackson 1992 for details).
Simulated data matrices- uniform correlation
Normally distributed data were simulated to match
the number of variables in the three ecological data
sets. Population data matrices were constructed with
4, 12, and 32 variables and 1000 observations. Ma-
trices were simulated having three levels of overall
correlation structure. For each population, the corre-
lations were uniformly generated to be 0, 0.3, or 0.8
for all off-diagonal correlations (i.e., RO, R3, and R8).
This approach generated matrices having no inter-
pretable dimensions (i.e., RO),a weak one-dimensional
structure (R3), and a strong one-dimensional structure
(R8). Analyses were done on three replicate samples
of 40 observations each drawn from the population of
1000 individuals. This approach was used to assess the
ability of the methods to correctly resolve the dimen-
sionality of the population from the analysis of a sam-
ple. This parallels the same problems confronting ecol-
ogists when analyzing field data.
Structured correlation
For comparison, matrices were simulated with three-
dimensional structure. These three-dimensional ma-
trices also varied as to whether submatrices were un-
correlated or weakly correlated with one another. In
the first set of simulations, 12-variable matrices were
This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM
All use subject to JSTOR Terms and Conditions
2205
divided into groups of four variables each. Within-
group correlations were equal to either 0.3 or 0.8
whereas the between-group correlations were equal to
either 0 or 0.3 (Fig. 1 for correlation matrices from
simulations S-I to S-IV). This approach was used to
simulate S-I to S-III in order to examine the effect of
differing degrees of strength in correlation structure.
To study the effect of groups having different numbers
of constituent variables, the matrices for S-IV used
groups containing 5, 4, and 3 variables, respectively.
Within-group correlations were set to 0.8 and between-
group correlations to 0. From each population, three
replicates of 40 observations were sampled.
The simulation resulted in a 3 x 3 x 3 design based
on the number of variables, level of correlation, and
replication for the uniform matrices and 4 matrices x
3 replicates for the structured-correlation matrices.
Matrix names are coded such that the first number
indicates the number of variables and the subsequent
alphanumeric indicates the degree of correlation within
the matrix. For example, 12R3 is a 12-variable matrix
with all intervariable correlations equal to 0.3.
Statistical analysis
Principal components analyses were conducted us-
ing the PRINCOMP procedure in SAS (SAS 1989).
Methods of component assessment-heuristic ap-
proaches. -1. Kaiser-Guttman. -The most common
stopping rule in principal components analysis (PCA)
is based on the average value of the eigenvalues (i.e.,
the Kaiser-Guttman criterion; Guttman 1954, Cliff
1988; H. Kaiser, unpublished manuscript). Because
variables are often measured in different units, most
ecologists use a correlation matrix in PCA, thereby
giving each variable equal weight in the analysis. As a
result, the sum of the eigenvalues equals the number
of variables. In the Kaiser-Guttman method, eigen-
values greater than the average eigenvalue (i.e., X >
1.0) are retained because these axes summarize more
information than any single original variable. There-
fore, only components with X > 1.0 are interpreted.
Unfortunately, a PCA of randomly generated, uncor-
related data will produce eigenvalues exceeding one.
As a result, this method has been criticized (e.g., Karr
and Martin 1981, Stauffer et al. 1985, Rexstad et al.
1986, 1988, Grossman et al. 1991); however the Kai-
ser-Guttman criterion remains the most popular stop-
ping rule in ecology.
2. Bootstrapped Kaiser-Guttman. -The bootstrap
resampling technique (Efron 1979) was proposed as a
means of determining the interpretability of eigenval-
ues by Lambert et al. (1990). They argued that the
Kaiser-Guttman criterion was arbitrary and it ignored
error associated with each X due to sampling. Conse-
quently, eigenvalues of 0.99 would be discarded,
whereas an eigenvalue of 1.01 would be retained even
though an eigenvalue of 1.01 may have 95% confidence
limits ranging from 0.9 to 1.1. As a result, they pro-
2206 DONALD A. JACKSON Ecology, Vol. 74, No. 8

S-I S-Il

1-4 0.8 0.0 1-4 0.3 0.0

5-8 0.8 5-8 0.3

9-12 0.0 0.8 9-12 0.0 0.3

1-4 5-8 9-12 1-4 5-8 9-12

S-l1l S-Iv

1-4 0.8 0.3

1-5 0.8 0.0

5-8 0.8
6-9 0.8

9-12 0.3 0.8

10-12 0.0 0.8

1-4 5-8 9-12 1-5 6-9 10-12

FIG. 1. Correlation structure for the patterned matrices used in S-I to S-IV. The values presented are the off-diagonal
intervariable correlations. For example in S-I, variables 1-4 were correlated with one another at r = 0.8, as were variables
5-8, and 9-12. However, correlations between variables from different submatrices were equal to 0.

posed that the bootstrap should be used to determine
how many eigenvalues had confidence limits encom-
passing the 1.0 criterion (i.e., a bootstrap Kaiser-Gutt-
man approach).
3. Scree plot.-Another common method (although
used infrequently by ecologists; e.g., Zebra and Collins
[1992]) is the scree plot. To apply the scree method,
one plots the value of each successive eigenvalue against
the rank order (Fig. 2; the log of the eigenvalues also
can be used with covariance-based PCAs). The smaller
eigenvalues, representing random variation, tend to lie
along a straight line. The point where the first few
eigenvalues depart from the line distinguishes the "in-
terpretable" and trivial components. Cattell (1966)
originally proposed that points to the left of the straight-
line segment should be considered important (i.e., three
components in the structured data of Fig. 2), but sub-
sequently concluded (Cattell and Vogelmann 1977) that
the first eigenvalue to the right of this point should be
included also (i.e., four interpretable components in
Fig. 2). Often the scree approach is complicated by
either the lack of any obvious break or the possibility
of multiple break points.
Horn (1965, Horn and Engstrom 1979) recognized
that with matrices composed of random data, the scree
plot would show a stable negative slope. Horn argued
that distinguishing eigenvalues in scree plots remained
quite arbitrary. As a result, he proposed a modification
to the scree plot (Horn 1965). After analyzing a given
data set and plotting the eigenvalues in a traditional
scree plot, numerous matrices of rank equal to the
observed data, but with uncorrelated variables, are
generated and eigenvalues are calculated. These eigen-
values from the random data are tabulated and the
mean values plotted on the scree plot of the original
data. The point where the two lines cross indicates the
4
Structured Data
3
cm~
1 _ I _ Random Data
2 4 6 8 10 12
Component Number
FIG. 2. Eigenvalues from a principal components analysis
of a 12-variable data set of randomly generated, uncorrelated
data and for a data set with underlying structure.

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
December 1993 SIGNIFICANT PRINCIPAL COMPONENTS 2207

maximum limit where eigenvalues are considered in-
terpretable. Further variations of this method have in-
cluded regression or Monte Carlo approaches (e.g., Al-
len and Hubbard 1986, Lautenschlager 1989).
4. Broken-stick. -Frontier (1976) proposed a bro-
ken-stick method that is based on eigenvalues from
random data. Frontier's model assumes that if the total
variance (i.e., sum of the eigenvalues) is divided ran-
domly amongst the various components, then the ex-
pected distribution of the eigenvalues will follow a bro-
ken-stick distribution (i.e., the random data in Fig. 2).
Observed eigenvalues are considered interpretable if
they exceed eigenvalues generated by the broken-stick
model. Frontier (1976) and Legendre and Legendre
(1983) provide a table of eigenvalues based on the
broken-stick distribution, but the solution is easily cal-
culated as:
P 1
bk = . and the test has p(p - 1)/2 degrees of freedom. The
i=k 1
where p is number of variables and bk is the size of the
eigenvalue for the kth component under the broken-
stick model.
5. Proportion of total variance. -Another simple
criterion for estimating the number of nontrivial com-
ponents is to include all components up to some ar-
bitrary proportion of the total variance. This method
typically includes components comprising 95% of the
total variance. Although this method is advocated by
some statisticians (Jolliffe 1986), Jackson (1991)
variance matrix and several studies recommend its use
only in covariance-based analyses (Dillon and Gold-
stein 1984, Morrison 1990, Grossman et al. 1991,
Jackson 1991). However, the test can be used with a
correlation matrix where such results are considered
to be conservative estimates of the number of non-
trivial components (Pimentel 1979, Kendall 1980).
7. Bartlett's test of the equality of X,.-Bartlett (1954)
also developed a statistical test of whether the first
eigenvalue of a correlation matrix is equal to the re-
maining set of eigenvalues (i.e., correlation matrix ho-
mogeneity). A modified Bartlett's test (Box 1949, Krza-
nowski 1988) is calculated as
- 1(2p + 11) InIR 1,
x2=-n
where IR Iis the determinant of the correlation matrix,
test is limited because it only examines the first eigen-
value. However it provides an assessment ofthe overall
PCA (i.e., if the null hypothesis is not rejected, it is
pointless to interpret the PCA).
8. Lawley's test of X2.-Lawley (1956, 1963) pro-
posed a method to test for the equality of the p - 1
eigenvalues (i.e., all but the first eigenvalue). It is based
on the following
X2 X2 (rij - r, u 2; (rk )2,
i=k+1 k=1 k=1

strongly recommended against its application as being where ri is the correlation between variable i and vari-
unfounded and unreliable. able j and
Statistical approaches. -Some data analysts retain
2
components with significant correlations (e.g., P < .05)
-
between the component scores and the original vari- P(P 1) i=k+l k=1

ables. Statistically this approach is flawed because the

X= 1-r,
PCA solution and original variables are not indepen-
dent, and as a result, the attributed significance is in- (p _ 1)2(1 - X2)
appropriate. In addition, components with only a sin- p -
(p -
2)X2
gle "significant" correlation suggest that the axis is not
1 P
a satisfactory multivariate summary. Tk rik, k = ,...p
6. Test of sphericity. -Bartlett's test of sphericity P- 1 i=1,ink

(Cooley and Lohnes 1971, Pimentel 1979) evaluates

whether each sequential eigenvalue is significantly dif-
ferent from the remaining eigenvalues. Conceptually,
the test attempts to reveal the point where the PCA
summarizes a spherical distribution of points. The test
statistic is calculated as
P ~~~~~P
(p - k) In L
,i=k+ I
Xi/(p - k)1 2i,+
i=k+
where p is the number of variables, k represents a
specific component, Xiis the eigenvalue of component
i, and n is the number of observations. If the resultant
statistic is multiplied by n - k, the product is x2 dis-
tributed with 0.5(p - k - l)(p - k + 2) degrees of
freedom. This test was originally developed for a co-
with (p + 1) (p -2)/2 degrees of freedom. One limi-
tation of this approach is that the test only evaluates
the second eigenvalue. Subsequent eigenvalues are not
compared when the null hypothesis is rejected. This
approach has been applied in a recent study of principal
components analysis with ecological data (Grossman
I
etal. 1991).
9. Bootstrap eigenvalue-eigenvector. -In this study,
each principal components analysis was bootstrapped
100 times. Eigenvalues for each bootstrap sample and
the associated eigenvector coefficients were retained.
Means, minima, maxima, and 95% confidence limits
were calculated from the distribution of the eigenval-
ues. Where the confidence intervals overlapped be-
tween pairs of successive eigenvalues, these eigenval-

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
2208 DONALD A. JACKSON Ecology, Vol. 74, No. 8

ues were considered to be indistinguishable from one Guttman method of interpreting Xs > 1.0 indicated all
another. However, if the ranges did not overlap, the PCAs from S-I to S-IV contained three interpretable
eigenvalues were assumed to be different. This latter components (i.e., within each analysis there were three
condition was considered to represent the break-point eigenvalues exceeding 1.0; Table 4). For the S-I ma-
between "meaningful" or nontrivial components and trices the approach indicated retaining five compo-
those associated with sampling and random noise. nents although only three dimensions were constructed
Similarly, the eigenvector coefficients were evaluated
in the simulations.
using a bootstrap approach. Coefficients that did not 2. Bootstrapped Kaiser-Guttman. -The bootstrap
differ significantly from zero were categorized as trivial
of the eigenvalues using the Kaiser-Guttman approach
or nonsignificant. However, if zero fell outside the 95 %
resulted in only one component being considered non-
confidence limits, then the coefficient was considered trivial with each of the 4-variable matrices (i.e., 4RO,
to be relatively stable and informative. Only boot- 4R3, and 4R8). Four components were retained with
strapped components having two or more coefficients 12RO PCAs and 2-3 components for the 12R3 matri-
different from zero were considered meaningful. Com- ces. The 32RO matrices had 9-10 components retained
ponents with only a single nonzero coefficient repre- and 8 components for the 32R3 matrices. For all R8
sented only a single variable, hence the component matrices (i.e., 4R8, 12R8, and 32R8), only the first
does not provide a true multivariate summary. component was considered nontrivial using this meth-
As a means of evaluating the overall similarity amongod. In the structured matrices, 3-4 components were
the different approaches with the different data sets, aidentified as nontrivial in the low-correlation S-II ma-
multivariate summary was done. The number of non- trices, whereas only two components were retained from
trivial components for each method from each data set S-III, and three components from the other analyses
was used as an input matrix. For example, the Kaiser- (i.e., S-I and S-IV).
Guttman method had a value for each of 4R0-A, 4R0- Both versions of the Kaiser-Guttman approach in-
B,.. .S-IV-C. The number of dimensions that were dicated a single interpretable dimension with the
simulated for each data set was included as an addi- 4-variable matrix of lake morphometry. The PCA based
tional observation. A Euclidean distance matrix was on the 12-variable matrix of water chemistry revealed
calculated between the methods and a principal co- three nontrivial components with Xs > 1.0 retained.
ordinates analysis done on the distance matrix to However, the bootstrapped evaluation suggested that
graphically integrate the differences among approaches only the first eigenvalue was significantly greater than
across all the data sets (e.g., see Jackson and Somers 1.0. Both the traditional and bootstrapped approaches
1991). indicated that nine and eight components, respectively,
were nontrivial in the 32-variable matrix of benthic
RESULTS
invertebrates, similar to results for the 32RO and 32R3
1. Kaiser-Guttman approach (X > 1.0). -For ma- matrices.
trices with correlations of RO or R3, the Kaiser-Gutt- 3. Scree plot. -Results from the scree plot based on
man method retained ; 50% of the components for the 4-variable matrices were difficult to interpret. In some
4- and 12-variable matrices and 30-40% of the com- cases, trends were apparent, but in other cases it is
ponents in the 32-variable matrices (Tables 1-3). For difficult to discern any pattern in the plot because only
each R8 matrix, only one eigenvalue exceeded 1.0, four points were available. Where a trend was apparent,
indicating a single interpretable gradient. The Kaiser- the approach advocated by Cattell and Vogelmann

TABLE 1. Number of nontrivial components indicated by various methods. Simulated data matrices have four variables and
40 observations with uniform correlations as follows: 4R0 has uniform correlation structure of r = 0, 4R3 has uniform
correlation structure of r = 0.3, and 4R8 has uniform correlation structure of r = 0.8. The letters A-C indicate replicates
drawn from a simulated population having that correlation structure. The morphology data set comprised four lake
morphometric variables.

4R0 4R3 4R8 Mor-

A B C A B C A B C phology
Kaiser-Guttman 2 2 2 1 1 1 1 1 1 1
Bootstrap Kaiser-Guttman 1 1 1 1 1 1 1 1 1 1
Scree plot 0 2 2 2 2 2
Broken-stick 0 0 0 0 1 1 1 1 1 1
95% variance 4 4 4 4 4 4 3 2 3 3
Sphericity test 0 0 0 1 1 1 1 1 1 3
Bartlett's first eigenvalue 0 0 0 1 1 1 1 1 1 1
Lawley's <2 <2 <2 <2 <2 <2 <2 <2 <2 2+
Bootstrap eigenvalue 0 0 0 1 1 1 1 1 1 2
Bootstrap eigenvector 0 0 0 1 1 1 1 1 1 3
Known 0 0 0 1 1 1 1 1 1

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
December 1993 SIGNIFICANT PRINCIPAL COMPONENTS 2209

TABLE 2. Number of nontrivial components indicated by various methods. Data matrices having 12 variables and 40
observations with uniform correlations as follows: 12R0 has uniform correlation structure of r = 0, 12R3 has uniform
correlation structure of r = 0.3, and 12R8 has uniform correlation structure of r = 0.8. The letters A-C indicate replicates
drawn from a simulated population having that correlation structure. The chemistry data set comprised 12 lake water
chemistry variables.

12R0 12R3 12R8

Chem-
A B C A B C A B C istry
Kaiser-Guttman 5 6 6 3 5 3 1 1 1 3
Bootstrap Kaiser-Guttman 4 4 4 2 3 2 1 1 1 1
Scree plot 4 6 7 2 2 3 2 2 2 4
Broken-stick 0 0 0 1 1 1 1 1 1 3
95% variance 11 10 11 10 11 10 6 7 7 8
Sphericity test 0 0 0 1 1 1 1 1 1 3
Bartlett's first eigenvalue 0 1 0 1 1 1 1 1 1 1
Lawley's <2 <2 <2 <2 <2 2+ 2+ 2+ 2+ 2+
Bootstrap eigenvalue 0 0 0 1 1 1 1 1 1 1
Bootstrap eigenvector 0 0 0 1 1 1 1 1 1 2
Known 0 0 0 1 1 1 1 1 1

(1977) suggested that two components were nontrivial.
With the 12RO analyses, the scree indicated 4-7 com-
ponents should be interpretable. The number of com-
ponents dropped to 2-3 when 12R3 matrices were used
and 2 components would be retained with 12R8 ma-
trices (Fig. 3). In the 32RO analyses, the scree plot
results suggested from 5 to 15 components should be
interpreted, 2-6 components with 32R3, and 2 com-
ponents with 32R8 analyses.
For the structured matrices, the scree plot suggested
that there were four nontrivial components in the S-I,
S-Ill, and S-IV matrices. For PCAs based on S-Il, scree
results indicated that between two and four compo-
nents would be considered interpretable. As with the
simulated 4-variable matrices, no estimate of the num-
ber of dimensions for the lake morphometry data could
be made because no obvious trend was apparent. With
the water chemistry data, there were three interpretable
components, but a second break is evident (Fig. 3). If
this latter point was considered, then a total of five
components were nontrivial.
4. Broken-stick model. -The broken-stick method
correctly identified the dimensionality of all uniform-
correlation matrices (a single exception being one of
the replicates from 4R3). For the RO matrices, the
method indicated that the underlying dimensionality
was 0, and one component as nontrivial with R3 or
R8 matrices. This method revealed three interpretable
components for matrices from S-I and S-IV, a single
component from S-Il, and 2-3 components from S-Ill.
A single component would be retained from the lake
morphometry data, three from the water chemistry
data, and two components from the benthic inverte-
brate data. (The application of the broken-stick model
to the eigenvalues presented in Rexstad et al.'s [1988]
criticism of PCA showed no nontrivial components in
contrast to 7 of the 15 being considered useful from
the Kaiser-Guttman method.)
5. 95% of the total variance. -The approach of re-
taining components until 95% of the total variance was
achieved would result in all components being inter-
preted for the 4RO or 4R3 analyses, and 2-3 compo-

TABLE 3. Number of nontrivial components indicated by various methods. Data matrices having 32 variables and 40
observations with uniform correlations as follows: 32R0 has uniform correlation structure of r = 0, 32R3 has uniform
correlation structure of r = 0.3, and 32R8 has uniform correlation structure of r = 0.8. The letters A-C indicate replicates
drawn from a simulated population having that correlation structure. The benthic invertebrate data set comprised abun-
dances for 32 lake benthic invertebrate taxa.

32R0 32R3 32R8 Benthic

inverte-
A B C A B C A B C brates
Kaiser-Guttman 12 14 13 10 10 11 1 1 1 9
Bootstrap Kaiser-Guttman 10 10 9 8 8 8 1 1 1 8
Scree plot 9 15 5 3 2 6 2 2 2 4
Broken-stick 0 0 0 1 1 1 1 1 1 2
95% variance 22 22 22 21 21 21 14 14 14 19
Sphericity test 0 0 0 1 1 11 1 2 6 19
Bartlett's first eigenvalue 1 1 1 1 1 1 1 1 1 1
Lawley's <2 <2 <2 2+ 2+ 2+ 2+ 2+ 2+ 2+
Bootstrap eigenvalue 0 0 0 1 1 1 1 1 1 1
Bootstrap eigenvector 0 0 0 1 1 1 1 1 1 1
Known 0 0 0 1 1 1 1 1 1

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
2210 DONALD A. JACKSON Ecology, Vol. 74, No. 8

nents for the 4R8 analyses. Application of the method 10

with the 12-variable PCAs led to 10-11 eigenvalues
being considered important for 12RO and 12R3 ma-
trices, and 6-7 eigenvalues for the 12R8 matrices. A
total of 22 components would be retained from the
32RO analyses, 21 from the 32R3, and 14 components
from the 32R8 PCAs. The fixed-variance approach
also led to more components being retained with the
S-I to S-IV matrices than any other method. In all
cases, the method would lead to 7-1 1 components be-
ing considered important and interpreted. With the
lake morphometry data, this method identified three
components as nontrivial and eight components from
the water chemistry data. This procedure retained three-
quarters of the components from each PCA. Nineteen
of a possible 32 components would be kept from the
benthic invertebrate PCA.
6. Bartlett's test of sphericity. -Bartlett's test cor-
rectly identified the dimensionality of the PCAs based
on four variables (i.e., 0 for RO, 1 for R3 and R8). The
method correctly estimated the dimensionality of the
12-variable matrices, but was more erratic with the 32-
variable matrices. The method correctly indicated no
significant eigenvalues with 32RO. However, one of the
32R3 matrices led to 11 eigenvalues being identified
as significant and two PCAs from 32R8 matrices led
to two and six components being retained. The test
indicated three significant components for S-I and S-Ill
matrices, two components for the S-I analyses, and
3-4 for those from S-IV although all these matrices
were constructed with three-dimensional structure.
Bartlett's test of sphericity yielded 3 significant eigen-
values from the 4-variable lake morphometry data and
the 12-variable water chemistry, plus 19 significant
eigenvalues from the PCA of lake benthic invertebrate
data.
7. Bartlett's test of the equality ofA,. Bartlett's meth-
od testing whether the first X from a correlation matrix
is equal to all others led to correct identification of the
minimum number of interpretable dimensions in all
4R data sets. With the 12- and 32-variable data sets,
r =0
8
8__ _ _ _ _ _ _. r = 0.3
......... r = 0.8
M 6 .......... Chemistry
cdg
Mj
4
\
2 4 6 8 10 12
Component Number
FIG. 3. Eigenvalues from principal components analyses
(PCA) from data sets that contained uncorrelated, weakly
correlated, or strongly correlated variables. Each matrix com-
prised 12 variables and 40 observations. Eigenvalues from a
PCA of 12 lake water chemistry variables are plotted also.
the test erroneously indicated AXin one 12RO PCA and
all of the 32RO PCAs as being significantly different
from the remaining eigenvalues. With the patterned
matrices (i.e., S-I to S-IV) the test indicated at least
one significant eigenvalue in each PCA. Bartlett's test
indicated that there was at least one significant com-
ponent in each of the lake morphometry, chemistry,
and invertebrate PCAs.
8. Lawley's test ofX,2-Lawley's test for correlation-
based PCAs correctly identified a maximum of one
underlying dimension with the four-variable solutions.
With 12RO matrices, the test indicated <2 significant
Xs for the PCAs. However, with 12R3 matrices, one
result suggested a minimum of two or more significant
eigenvalues, and all 12R8 matrices had two or more
significant eigenvalues. Results based on 32RO matri-
ces led to <2 eigenvalues being considered interpret-
able, whereas all 32R3 and 32R8 analyses indicated
>-2 significant Xs. Lawley's method indicated that each

TABLE 4. Number of nontrivial components indicated by various methods. Patterned data matrices having 12 variables and
40 observations. The intervariable correlations were generated following Fig. 1. Letters A-C represent replicate samples
drawn from each simulated population.

S-I S-u1 S-I11 S-IV

A B C A B C A B C A B C
Kaiser-Guttman 3 3 3 5 5 5 3 3 3 3 3 3
Bootstrap Kaiser-Guttman 3 3 3 4 3 3 2 2 2 3 3 3
Scree plot 4 4 4 3 4 2 4 4 4 4 4 4
Broken-stick 3 3 3 1 1 1 3 2 2 3 3 3
95% variance 7 7 8 10 10 10 7 7 8 8 7 8
Sphericity test 3 3 3 2 2 2 3 3 3 3 3 4
Bartlett's first eigenvalue 1 1 1 1 1 1 1 1 1 1 1 1
Lawley's 2+ 2+ 2+ 2+ 2+ 2+ 2+ 2+ 2+ 2+ 2+ 2+
Bootstrap eigenvalue 3 3 3 0 0 0 3 3 1 3 3 2
Bootstrap eigenvector 3 3 3 0 2 1 0 3 3 3 3 3
Known 3 3 3 3 3 3 3 3 3 3 3 3

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
December 1993 SIGNIFICANT PRINCIPAL COMPONENTS
PCA from S-I to S-IV contained a minimum of two
significantly different eigenvalues. Although neither
Lawley's nor Bartlett's test is capable of accurately as-
sessing the true dimensionality of these matrices (i.e.,
three dimensions), they can establish that a minimal
number of components is interpretable (i.e., either one
or two components, respectively). Lawley's test indi-
cates that a minimum of two components should be
considered interpretable in each of the field data sets.
9. Bootstrapped eigenvalue-eigenvector.-When the
bootstrap approach was used to determine overlap be-
tween eigenvalues, the number of components retained
was lower. For the 4R, 12R, and 32R matrices, the
method correctly identified that no components were
interpretable with ROand only one component for PCAs
based on R3 and R8 matrices. Identical results oc-
curred with the estimations based on the bootstrapped
eigenvector coefficients. When the bootstrap was used
to distinguish between overlapping eigenvalues with
the patterned matrices, results suggested that three
components be retained for S-I, no components from
S-Il, 1-3 components from S-Ill, and 2-3 components
from S-IV. Bootstrapped eigenvector coefficients in-
dicated 3 interpretable components for S-I and S-IV,
0-2 components for S-Il, and 0-3 components for S-Ill.
These approaches suggested two and three dimensions,
respectively, for the PCA of lake morphometric data.
For the water chemistry matrix, the bootstrapped ei-
genvector coefficients indicated that two components
were interpretable, although the bootstrapped-eigen-
value version indicated only a single nontrivial com-
ponent. With the benthic invertebrate data, only a sin-
gle component would be retained based on either
approach.
The multivariate summary using ordination analysis
provided an overall assessment of the similarity
amongst the methods. Results from the 4-variable
analyses were not included because some of the eigen-
values could not be interpreted. Results from Bartlett's
test of X1 and Lawley's test of X2 were not included
because they are limited to evaluating only one or two
eigenvalues, respectively. The first PCoA Axis sum-
marized most of the variation among the methods (Fig.
4). The 95% approach differed most from the other
methods. The Kaiser-Guttman, bootstrapped Kaiser-
Guttman, and scree plot provided relatively similar
results, but differed from the true number of dimen-
sions in each data set. The bootstrapped eigenvalue,
bootstrapped eigenvector, and broken-stick methods
all led to similar numbers of dimensions for each data
set, and each was very similar to the number of un-
derlying simulated dimensions.
DISCUSSION
The standard Kaiser-Guttman approach of inter-
preting X > 1.0 led to the retention of too many com-
ponents except with matrices having strong correlation
structure. Although such correlation structures often
This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM
All use subject to JSTOR Terms and Conditions
2211
12
Scree
.e KG
BKG
0
BVec
LA no Bck
Known
Sphere 95%
-12 l l
-16 0 16 32 48
Axis 1 (84.4%)
FIG. 4. Principalcoordinatesanalysisbased on a Euclid-
ean distance matrix. Data in the analysis are based on the
numberof nontrivialcomponentsresultingfromeachmethod
with each data set. Codes are: KG is the Kaiser-Guttman;
BKG is the bootstrappedKaiser-Guttman;Screeis the scree
plot; BStickis the broken-stickmodel; 95%is the 95%of the
total variancecriterion;Sphereis Bartlett'stest of sphericity;
BVal is the bootstrappedeigenvalue method; BVec is the
bootstrappedeigenvectorcoefficientmethod; and Known is
the truenumberof dimensionsincorporatedinto the datasets
duringsimulation.
are found in morphometric analyses, ecologists often
analyze matrices of more limited structure (e.g., weakly
correlated data). Jolliffe (1972, 1986) suggested that
the choice of X > 1.0 was too conservative and that
components with X > 0.7 should be considered useful.
Clearly, this level of selection would only exacerbate
a bad situation. Given that the Kaiser-Guttman ap-
proach is frequently used by ecologists, it would be
wise to use an alternative method to choose nontrivial
components. The use of eigenvalues in a modified Kai-
ser-Guttman approach does not change the results sub-
stantially. The bootstrapped Kaiser-Guttman did re-
duce the number of nontrivial components, but the
number still exceeded the number of underlying di-
mensions except for the R8 matrices.
The scree plot (as applied as Cattell and Vogelmann
[1977]) provided poor resolution of the underlying di-
mensionality. The method invariably overestimated
the number of interpretable components. If Cattell's
(1966) original criterion was used, the method was
more conservative. This means including the eigen-
values up to, but not including, the first eigenvalue on
the straight-line portion of the plot (Fig. 2). However
this modification still overestimates the number of in-
terpretable components in analyses of matrices of
weakly correlated data. Surprisingly, the original ap-
proach (Cattell 1966) provided a better estimate of the
correct number of dimensions with the simulated data
matrices.
The broken-stick method correctly assessed the di-
mensionality of the data matrices. It did underestimate
the number of interpretable components for S-I ma-
trices. This method provided a good combination of
2212 DONALD A. JACKSON
simplicity of calculation and accurate evaluation of
dimensionality relative to the other statistical ap-
proaches.
The 95%-variance-threshold method provided un-
satisfactory results. Although the choice at 95% of the
total variance is relatively high, any level is arbitrary.
No matter what cumulative percentage level is select-
ed, this approach does not appear promising because
there is the high risk that many of the components that
are retained will summarize noise or nontrivial com-
ponents will not be included.
Bartlett's test of sphericity correctly identified the
dimensionality in many of the data sets, but in some
cases indicated up to 11 significant eigenvalues al-
though only a single dimension was simulated (Table
3). Despite the statement by Kendall (1980) that this
test is overly conservative when applied to correlation
matrices, it appears to correctly identify the number
of dimensions with many data sets, but it is too liberal
a test with matrices having a low observation-to-vari-
able ratio (e.g., less than the 3: 1 ratio advocated by
Grossman et al. 1991). With the ecological data, the
test also retained large numbers of the components,
i.e., 19 of 32 components were considered significant
with the benthic invertebrate data.
Bartlett's approach to test for homogeneity of the
correlation matrix (i.e., whether the first X equalled all
others) appeared to identify the correct minimal di-
mensionality except with the 32-variable analyses. Here
the method indicated significant structure with ran-
dom, uncorrelated data. Likewise, Lawley's test con-
sistently overestimated the dimensionality of the 12-
and 32-variable matrices having uniform correlations.
Because the test is designed to evaluate only whether
X2 is the same as successive eigenvalues, the method
is rather limited. As a result of its limited utility and
relatively poor performance in this set of comparisons,
the method is not recommended.
The combination of testing for overlap in ranges of
bootstrapped eigenvalues and for eigenvector coeffi-
cients differing from 0 appears more promising. With
simple matrices either lacking structure or having a
single dimension, both approaches consistently re-
vealed the underlying dimensionality of the simula-
tion. However with patterned matrices, the results were
less reliable for either approach individually. Both
methods worked well with S-I and S-IV matrices hav-
ing strong inter-variable correlations. However, in S-I1
where there were three underlying, but weak dimen-
sions, there were no differences between the boot-
strapped eigenvalues. The eigenvector coefficient ap-
proach produced inconsistent results and frequently
underestimated the correct dimensionality. With ma-
trices from S-Ill, both methods correctly identified two
dimensions, but from different replicated matrices.
The poor showing of the eigenvector approach for
S-I1 and S-Ill is easily explained and similar to situ-
ations discussed elsewhere (Oksanen 1988). In both
This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM
All use subject to JSTOR Terms and Conditions
Ecology, Vol. 74, No. 8
simulations, three dimensions were created by having
three sets of four variables, each set having identical
correlations. Due to this condition and chance selection
of observations in the bootstrap, any specific dimen-
sion could be expressed on the first component of one
PCA, but on the second or third component from an-
other PCA. This is similar to the re-ordering of com-
ponents or solution instability found by Oksanen (1988)
with detrended correspondence analysis. When the or-
der of expression of the underlying dimensions varies
between components for different analyses, the eigen-
vector coefficient approach will fail. With these same
data characteristics, the first three eigenvalues also
overlap in their 95% confidence limits, but are signif-
icantly different from the fourth eigenvalue. The prob-
lem with the eigenvector coefficients is particularly ev-
ident when the initial correlation structure is weak (e.g.,
S-Il). However, if the dimensions differ in: (1) the
strength of correlation structure (i.e., several high cor-
relations vs. low or medium correlations); (2) the num-
ber of constituent variables; or (3) have strong corre-
lations, this method provides more accurate results.
Overall, it appears that the combination of these two
approaches, i.e., the bootstrapped eigenvalue and ei-
genvector coefficients, provides a better measure of the
dimensionality than either approach alone. The max-
imum value obtained with either approach was close
to the true dimensionality, except with the S-I1 data.
An additional consideration of the bootstrapped ei-
genvector method is that it assists with the evaluation
of whether or not each variable contributes to a given
component. If a specific variable is not significantly
weighted on any nontrivial component, then that vari-
able could be removed from the analysis. For example,
in the PCA of the lake morphology data, three variables
had eigenvector coefficients that differed from zero.
However, lake volume coefficients included 0 in the
95% confidence limits on each component. Therefore,
lake volume did not contribute to the analysis and
added little information to the PCA.
With the use of any of the methods employing formal
statistical tests, e.g., Bartlett's test of sphericity, it is
important to recognize the increased probability of re-
jecting the null hypothesis when many components are
evaluated sequentially. When such tests are used, re-
searchers may remove this increased risk of a Type I
error by employing some form of a correction for mul-
tiple comparisons such as Bonferroni's adjustment.
The most promising approaches to component eval-
uation are the broken-stick model and the boot-
strapped eigenvalue-eigenvector method. The broken-
stick approach has the advantage of being simple to
calculate. Within the scope of this study, both methods
led to similar conclusions about the dimensionality of
the simulated data sets. The matrices simulated in this
study all represented relatively well-conditioned data
(e.g., from a normal distribution, independent sam-
pling). However, many data used in ecological studies
December 1993 SIGNIFICANT PRINCIPAL COMPONENTS 2213

do not meet formal assumptions of classical statistical Grossman, G. D., D. M. Nickerson, and M. C. Freeman.
approaches. The extension of this comparison to sim- 1991. Principal component analyses of assemblage struc-
ture data: utility of tests based on eigenvalues. Ecology 72:
ulated data varying in departure from the statistical 341-347.
"ideal" would be of considerable value (e.g., Davis Guttman, L. 1954. Some necessary conditions for common
1977). Approaches such as the bootstrapped eigenval- factor analysis. Psychometrika 19:149-161.
ue-eigenvector method would likely prove more useful Horn, J. L. 1965. A rationale and test for the number of
with such data conditions than the relatively sensitive factors in factor analysis. Psychometrika 30:179-185.
methods based on idealized distributions and formal Horn, J. L., and R. Engstrom. 1979. Cattell's scree test in
relation to Bartlett's chi-square test and other observations
tests (e.g., both of Bartlett's and Lawley's methods). on the number of factors problem. Multivariate Behavioral
Research 14:283-300.
ACKNOWLEDGMENTS Jackson, D. A. 1992. Fish and benthic invertebrates: ana-
This studywas greatlyassistedby the criticalcommentsof lytical approaches and community-environment relation-
K. P. Burnham,M. Dennison, R. H. Green, H. H. Harvey, ships. Dissertation. University of Toronto, Toronto, On-
K. M. Somers,and D. F Stauffer.Fundingwas providedby tario, Canada.
a NaturalSciencesandEngineering ResearchCouncil(NSERC) Jackson, D. A., and K. M. Somers. 1991. Putting things in
GraduateScholarshipand OntarioGraduateScholarshipto order: the ups and downs of detrended correspondence
D. A. Jackson,an NSERC Operatinggrantto H. H. Harvey, analysis. American Naturalist 137:704-712.
and OntarioMinistryof Environmentand Ontario Renew- Jackson, D. A., K. M. Somers, and H. H. Harvey. 1992.
able ResourcesResearchGrantsto H. H. Harvey and D. A. Null models and fish communities: evidence of nonrandom
Jackson. patterns. American Naturalist 139:930-951.
Jackson, J. E. 199 1. A user's guide to principal components.
LITERATURE CITED John Wiley & Sons, New York, New York, USA.
Allen, S. J., and R. Hubbard. 1986. Regression equations Jolliffe, I. T. 1972. Discarding variables in a principal com-
for the latent roots of random data correlation matrices ponents analysis. I. Artificial data. Applied Statistics 23:
with unities on the diagonal. Multivariate Behavioral Re- 160-173.
search 21:393-398. 1986. Principal components analysis. Springer-Ver-
Bartlett, M. S. 1950. Tests of significance in factor analysis. lag, New York, New York, USA.
British Journal of Psychology (Statistical Section) 3:77-85. Karr, J. R., and T. E. Martin. 1981. Random numbers and
2 1954. A note on the multiplying factors for various principal components: further searches for the unicorn. Pages
X2 approximation. Journal of the Royal Statistical Society, 20-24 in D. E. Capen, editor. The use of multivariate sta-
Series B 16:296-298. tistics in studies of wildlife habitat. United States Forest
Box, G. E. P. 1949. A general distribution theory for a class Service General Technical Report RM-87.
of likelihood criteria. Biometrika 36:317-346. Kendall, M. 1980. Multivariate analysis. Second edition.
Cattell, R. B. 1966. The scree test for the number of factors. Charles Griffin, London, England.
Journal of Multivariate Behavioral Research 1:245-276. Krzanowski, W. J. 1983. Cross-validatory choice in prin-
Cattell, R. B., and S. Vogelmann. 1977. A comprehensive cipal components analysis: some sampling results. Journal
trial of the scree and KG criteria for determining the num- of Statistical Computation and Simulation 18:299-314.
ber of factors. Multivariate Behavioral Research 12:289- 1988. Principles of multivariate analysis: a user's
325. perspective. Oxford University Press, London, England.
Cliff, N. 1988. The eigenvalues-greater-than-one rule and Lambert, Z. V., A. R. Wildt, and R. M. Durand. 1990. As-
the reliability of components. Psychological Bulletin 103: sessing sampling variation relative to number-of-factors
276-279. criteria. Educational and Psychological Measurement 50:
Cooley, W. W., and P. R. Lohnes. 1971. Multivariate data 33-49.
analysis. John Wiley & Sons, New York, New York, USA. Lautenschlager, G. J. 1989. A comparison of alternatives
Davis, A. W. 1977. Asymptotic theory for principal com- to conducting Monte Carlo analyses for determining par-
ponents analysis: non-normal case. Australian Journal of allel analysis criteria. Multivariate Behavioral Research 24:
Statistics 19:206-212. 365-395.
Digby, P. G. N., and R. A. Kempton. 1987. Multivariate Lawley, D. N. 1956. Tests of significance for the latent roots
analysis of ecological communities. Chapman and Hall, of covariance and correlation matrices. Biometrika 43:128-
New York, New York, USA. 136.
Dillon, W. R., and M. Goldstein. 1984. Multivariate anal- 1963. On testing a set of correlation coefficients for
ysis: methods and applications. John Wiley & Sons, New equality. Annals of Mathematical Statistics 34:149-151.
York, New York, USA. Legendre, L., and P. Legendre. 1983. Numerical ecology. El-
Dudzifiski, M. L., J. T. Chmura, and C. B. H. Edwards. 1975. sevier, Amsterdam, The Netherlands.
Repeatability of principal components in samples: normal Morrison, D. F. 1990. Multivariate statistical methods. Third
and non-normal data sets compared. Multivariate Behav- edition. McGraw-Hill, New York, New York, USA.
ioral Research 10: 109-118. Oksanen. J. 1988. A note on the occasional instability of
Efron, B. 1979. Bootstrap methods: another look at the detrending in correspondence analysis. Vegetatio 74:29-32.
jackknife. Annals of Statistics 7:1-26. - Orloci, L. 1978. Multivariate analysis in vegetation re-
Frontier, S. 1976. Etude de la decroissance des valeurs propres search. Second edition. Dr. W. Junk, The Hague, The Neth-
dans une analyze en composantes principales: comparison erlands.
avec le module de baton bris6. Journal of Experimental Pielou, E. C. 1984. The interpretation of ecological data.
Marine Biology and Ecology 25:67-75. John Wiley & Sons, New York, New York, USA.
Gauch, H. G., Jr. 1982a. Multivariate analysis in com- Pimentel, R. A. 1979. Morphometrics: the multivariate
munity ecology. Cambridge University Press, New York, analysis of biological data. Kendall-Hunt, Dubuque, Iowa,
New York, USA. USA.
1982b. Noise reduction by eigenvector ordination. Rexstad, E. A., D. D. Miller, C. H. Flather, E. M. Anderson,
Ecology 63:1643-1649. J. W. Hupp, and D. R. Anderson. 1988. Questionable

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions
2214 DONALD A. JACKSON Ecology, Vol. 74, No. 8

multivariate statistical inference in wildlife habitat and
community studies. Journal of Wildlife Management 52:
794-798.
Rexstad, E. A., D. D. Miller, C. H. Flather, E. M. Anderson,
J. W. Hupp, and D. R. Anderson. 1990. Questionable
multivariate statistical inference in wildlife habitat and
community studies: a reply. Journal of Wildlife Manage-
ment 54:189-193.
SAS. 1989. SAS/STAT user's guide. Version 6. SAS Insti-
tute, Cary, North Carolina, USA.
Stauffer, D. F., E. 0. Gordon, and R. K. Steinhorst. 1985.
A comparison of principal components from real and ran-
dom data. Ecology 66:1693-1698.
Taylor, J. 1990. Questionable multivariate statistical infer-
ence in wildlife habitat and community studies: a comment.
Journal of Wildlife Management 54:186-189.
Zebra, K. E., and J. P. Collins. 1992. Spatial heterogeneity
and individual variation in diet of an aquatic predator.
Ecology 73:268-279.

This content downloaded from 130.240.43.43 on Sat, 17 Aug 2013 15:45:17 PM

All use subject to JSTOR Terms and Conditions