Austral Ecology (2001) 26, 3246

A new method for non-parametric multivariate analysis of variance

MARTI J. ANDERSON Centre for Research on Ecological Impacts of Coastal Cities, Marine Ecology Laboratories A11, University of Sydney, New South Wales 2006, Australia

Abstract Hypothesis-testing methods for multivariate data are needed to make rigorous probability statements about the effects of factors and their interactions in experiments. Analysis of variance is particularly powerful for the analysis of univariate data. The traditional multivariate analogues, however, are too stringent in their assumptions for most ecological multivariate data sets. Non-parametric methods, based on permutation tests, are preferable. This paper describes a new non-parametric method for multivariate analysis of variance, after McArdle and Anderson (in press). It is given here, with several applications in ecology, to provide an alternative and perhaps more intuitive formulation for ANOVA (based on sums of squared distances) to complement the description provided by McArdle and Anderson (in press) for the analysis of any linear model. It is an improvement on previous non-parametric methods because it allows a direct additive partitioning of variation for complex models. It does this while maintaining the exibility and lack of formal assumptions of other non-parametric methods. The teststatistic is a multivariate analogue to Fishers F-ratio and is calculated directly from any symmetric distance or dissimilarity matrix. P-values are then obtained using permutations. Some examples of the method are given for tests involving several factors, including factorial and hierarchical (nested) designs and tests of interactions. Key words: ANOVA, distance measure, experimental design, linear model, multifactorial, multivariate dissimilarity, partitioning, permutation tests, statistics.

INTRODUCTION The analysis of multivariate data in ecology is becoming increasingly important. Ecologists often need to test hypotheses concerning the effects of experimental factors on whole assemblages of species at once. This is important for core ecological research and in studies of biodiversity or environmental impacts in many habitats, including marine subtidal environments (Warwick et al. 1988; Gray et al. 1990; Chapman et al. 1995; Glasby 1997), mangroves (Skilleter 1996; Kelaher et al. 1998), freshwater systems (Faith et al. 1995; Quinn et al. 1996) and terrestrial systems (Oliver & Beattie 1996; Anderson & Clements, in press). Univariate analysis of variance (ANOVA) provides an extremely powerful and useful tool for statistical tests of factors and their interactions in experiments (Underwood 1981, 1997). Partitioning variation, as in multifactorial ANOVA, is particularly important for testing hypotheses in complex ecological systems with natural temporal and spatial variability. This partitioning is also needed to test multivariate hypotheses in ecology for experimental designs involving several factors.
*Present address: Department of Statistics, University of Auckland, Private Bag 92019, Auckland, New Zealand (Email: mja@stat.auckland.ac.nz). Accepted for publication March 2000.

This paper describes a new non-parametric test of the general multivariate hypothesis of differences in the composition and/or relative abundances of organisms of different species (variables) in samples from different groups or treatments. This test is a signicant advance on previous methods because it can be based on any measure of dissimilarity and can partition variation directly among individual terms in a multifactorial ANOVA model. The test is applicable to any situation where the simultaneous responses of many potentially non-independent variables (usually abundances of species in an assemblage) have been measured in samples from a one-factor or multifactorial ANOVA design. Powerful multivariate statistical methods, such as the traditional multivariate analysis of variance (MANOVA), have existed for decades (Hotelling 1931; Wilks 1932; Fisher 1936; Bartlett 1939; Lawley 1939; Pillai 1955), but tests using these statistics rely on assumptions that are not generally met by ecological data. The assumption that the data conform to a multivariate normal distribution is particularly unrealistic for most ecological data sets. This is because the distributions of abundances of individual species are usually highly aggregated or skewed (e.g. Gaston & McArdle 1994). Also, abundances take discrete values rather than being continuous, species with small means often have asymmetric distributions because they are necessarily

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

33

truncated at zero, and rare species contribute lots of zeros to the data set. MANOVA test statistics are not particularly robust to departures from multivariate normality (Mardia 1971; Olson 1974; Johnson & Field 1993). Finally, many of these test statistics are simply impossible to calculate when there are more variables than sampling units, which often occurs in ecological applications. Many non-parametric methods for tests of differences among a priori groups of observations (as in MANOVA) have been developed (Mantel 1967; Mantel & Valand 1970; Hubert & Schultz 1976; Mielke et al. 1976; Clarke 1988, 1993; Smith et al. 1990; Excofer et al. 1992; Edgington 1995; Pillar & Orlci 1996; Legendre & Anderson 1999). These methods generally have two things in common. First, they are based on measures of distance or dissimilarity between pairs of individual multivariate observations (which I will refer to generally as distances) or their ranks. A statistic is constructed to compare these distances among observations in the same group versus those in different groups, following the conceptual framework of ANOVA. Second, they use permutations of the observations to obtain a probability associated with the null hypothesis of no differences among groups. These non-parametric methods generally fall into two categories. First, there are those that can be based on any chosen distance measure. There are many such measures and these have different properties, which make them appropriate for different kinds of data (Legendre & Legendre 1998). For example, to express differences in community structure, the semimetric BrayCurtis measure of ecological distance (Bray & Curtis 1957) or Kulczynskis (1928) semimetric measure are generally preferred over metric measures, like Euclidean distance (Odum 1950; Hajdu 1981; Faith et al. 1987; Clarke 1993). The methods that are exible enough to be used with any such distance measure (e.g. Mantel 1967; Hubert & Schultz 1976; Smith et al. 1990; Clarke 1993) have much to recommend them for this reason. The drawback to using these methods is that they are not able to cope with multifactorial ANOVA. That is, they are not able to partition variation across the many factors that form part of the experimental design. Consequently, for most complex designs, one must analyse data as one-way analyses in multiple subsets within particular levels of factors. These multiple oneway analyses and qualitative interpretations of ordination plots are then used to infer something about interactions or variability at different spatial scales (e.g. Anderson & Underwood 1994; Kelaher et al. 1998). Some of the proposed non-parametric methods do allow partitioning for a complex design (e.g. Excofer et al. 1992; Edgington 1995; Pillar & Orlci 1996), but these are restricted for use with metric distance measures, which are not ideal for ecological applications.

Furthermore, even if these statistics were to be used, there has been disagreement concerning appropriate permutational strategies for complex ANOVA, particularly for tests of interactions (e.g. Edgington 1995; Manly 1997). There have been some recent examples of direct statistical analyses of BrayCurtis distances (Faith et al. 1995; Underwood & Chapman 1998). These are restricted, however, to very specic experimental designs or hypotheses and cannot be used for any multifactorial ANOVA design. Ecologists need a non-parametric multivariate method that can partition variation based on any distance measure in any ANOVA design. The method needs to be robust, interpretable by reference to the experimental design, and should lack formal assumptions concerning distributions of variables. The purpose of this paper is to outline just such a method and to give some ecological examples of its use. The more general mathematical theory underlying this method, along with simulations and a comparison with the related approach of Legendre and Anderson (1999), is described elsewhere (McArdle & Anderson, in press).

STRATEGY FOR NON-PARAMETRIC MULTIVARIATE ANALYSIS An outline for a general approach to the analysis of multivariate data in ecology was given by Clarke and Green (1988) and Clarke (1993). For experimental designs used to test hypotheses dened a priori, there are essentially four steps: (i) a choice is made concerning an appropriate transformation and/or standardization (if any) to apply to the data, given the hypothesis and the scales and nature of the species variables; (ii) a choice is made concerning the distance measure to be used as the basis of the analysis (e.g. BrayCurtis, Euclidean, 2 or other measure); (iii) ordination (and/or clustering) is performed in order to visualize patterns of resemblance among the observations based on their community composition; and (iv) a nonparametric multivariate test for differences among groups is done to obtain a rigorous probabilistic statement concerning multivariate effects of a priori groups. Note that (iii) is not essential in terms of the statistical test; ordination simply gives a visual representation by reducing the dimensionality of the data. In this paper, I focus on step (iv) of this procedure, which currently poses a problem for multifactorial designs.

DESCRIPTION OF THE TEST: ONE-WAY DESIGN The two essential considerations for the test are: (i) the construction of the test-statistic, and (ii) the calculation

34

M. J. ANDERSON

of a P-value using some method of permutation. I will describe the method, which I shall simply call nonparametric MANOVA, rst for the one-way design and then for more complex designs, followed by some ecological examples. I deal here only with the case of balanced ANOVA designs, but analogous statistics for any linear model, including multiple regression and/or unbalanced data, can be constructed, as described by McArdle and Anderson (in press).

This multivariate analogue can also be thought of geometrically (e.g. Cali ski & Harabasz 1974; Mielke n et al. 1976; Edgington 1995; Pillar & Orlci 1996), as shown in Fig. 1 for the case of two groups and two variables (dimensions). Here, SSW is the sum of the squared Euclidean distances between each individual replicate and its group centroid (the point corresponding to the averages for each variable, Fig. 1 and

The test statistic: an F-ratio

The essence of analysis of variance is to compare variability within groups versus variability among different groups, using the ratio of the F-statistic. The larger the value of F, the more likely it is that the null hypothesis (H0) of no differences among the group means (i.e. locations) is false. For univariate ANOVA, partitioning of the total sum of squares, SST, is achieved by calculating sums of squared differences (i) between individual replicates and their group mean (SSW, the within-group sum of squares; Table 1a), and (ii) between group means and the overall sample mean (SSA, the among-group sum of squares). Next, consider the multivariate case where p variables are measured simultaneously for each of n replicates in each of a groups, yielding a matrix of data where rows are observations and columns are variables. A natural multivariate analogue may be obtained by simply adding up the sums of squares across all variables (Table 1b). An F-ratio can then be constructed, as in the univariate case.

Fig. 1. A geometric representation of MANOVA for two groups in two dimensions where the groups differ in location. The within-group sum of squares is the sum of squared distances from individual replicates to their group centroid. The among-group sum of squares is the sum of squared distances from group centroids to the overall centroid. ( Distances ) from points to group centroids; () distances from group centroids to overall centroid; ( ), overall centroid; ( ), group centroid; ( ), individual observation.

Table 1. Calculations of within-group sums of squares for partitioning in (a) univariate ANOVA, (b) a multivariate analogue obtained by summing across variables, (c) a multivariate analogue equivalent to (b) obtained using sums of squared Euclidean distances, (d) the traditional MANOVA approach, which yields an entire matrix (W) of within-group sums of squares and cross products, and (e) the partitioning using inter-point distances advocated here, equivalent to (b) and (c) if Euclidean distances are used Univariate (a) One variable Multivariate (b) Summed across variables (c) Geometric approach (inner product, a scalar, based on Euclidean distances, correlations between variables ignored) (d) Traditional MANOVA (outer product, a matrix, based on Euclidean distances, correlations between variables matter) (e) Inter-point geometric approach (a scalar, based on any distance measure, correlations between variables ignored)
a i a i a i n j n j n j

SSW SSW SSW

( yij y i.)2 ( yijk y i.k)2 ( yij y i.)T( yij y i.)

1 p k

1 1

1 1

a i

n j

( yij y i.)( yij y i.)T

SS1

1 n

N1 i 1

N j i

d ij

ij

yij, univariate observation of the jth replicate (j 1,, n) in the ith group (i 1,, a); yijk, observation of yij for the kth variable (k 1,, p); yij, vector of length p, indicating a point in multivariate space according to p variables (dimensions) for observation j in group i. A superscript T indicates the transpose of the vector, bars over letters indicate averages and a dot subscript indicates averaging was done over that subscripted variable.

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

35

Table 1c). Note that this additive partitioning using a geometric approach yields one value for each of SSW, SSA and SST as sums of squared Euclidean distances. This geometric approach gives sums of squares equivalent to the sum of the univariate sums of squares (added across all variables) described in the previous paragraph. This differs from the traditional MANOVA approach, where partitioning is done for an entire matrix of sums of squares and cross-products (e.g. Mardia et al. 1979; Table 1d). The key to the non-parametric method described here is that the sum of squared distances between points and their centroid is equal to (and can be calculated directly from) the sum of squared interpoint distances divided by the number of points. This important relationship is illustrated in Fig. 2 for points in two dimensions. The relationship between distances to centroids and interpoint distances for the Euclidean measure has been known for a long time (e.g. Kendall & Stuart 1963; Gower 1966; Cali ski & n Harabasz 1974; Seber 1984; Pillar & Orlci 1996; Legendre & Legendre 1998; see also equation B.1 in Appendix B of Legendre & Anderson 1999). What is important is the implication this has for analyses based on non-Euclidean distances. Namely, an additive partitioning of sums of squares can be obtained for any distance measure directly from the distance matrix, without calculating the central locations of groups.

Why is this important? In the case of an analysis based on Euclidean distances, the average for each variable across the observations within a group constitutes the measure of central location for the group in Euclidean space, called a centroid. For many distance measures, however, the calculation of a central location may be problematic. For example, in the case of the semimetric BrayCurtis measure, a simple average across replicates does not correspond to the central location in multivariate BrayCurtis space. An appropriate measure of central location on the basis of BrayCurtis distances cannot be calculated easily directly from the data. This is why additive

Fig. 2. The sum of squared distances from individual points to their centroid is equal to the sum of squared interpoint distances divided by the number of points.

Fig. 3. Schematic diagram for the calculation of (a) a distance matrix from a raw data matrix and (b) a non-parametric MANOVA statistic for a one-way design (two groups) directly from the distance matrix. SST, sum of squared distances in the half matrix ( ) divided by N (total number of observations); SSW, sum of squared distances within groups ( ) divided by n (number of observations per group). SSA SST SSW and F = [SSA/(a 1)]/[SSW/ (N a)], where a the number of groups.

36

M. J. ANDERSON

partitioning (in terms of average differences among groups) has not been previously achieved using BrayCurtis (or other semimetric) distances. However, the relationship shown in Fig. 2 can be applied to achieve the partitioning directly from interpoint distances. Thus, consider a matrix of distances between every pair of observations (Fig. 3a). If we let N an, the total number of observations (points), and let dij be the distance between observation i 1,, N and observation j 1,, N, the total sum of squares is SST 1 N
N1 i l j N i 1

can be interpreted in the same way as they usually are for univariate ANOVA. The difference is that the hypothesis being tested for any particular term is a multivariate hypothesis.

OBTAINING A P-VALUE USING PERMUTATIONS The multivariate version of the F-statistic described here is not distributed like Fishers F-ratio under the null hypothesis. This is so because (i) we do not expect the individual variables to be normally distributed, and (ii) we do not expect that the Euclidean distance will necessarily be used for the analysis. Even if each of the variables were normally distributed and the Euclidean distance used, the mean squares calculated for the multivariate data would not each consist of sums of independent 2 variables, because, although individual observations are expected to be independent, individual species variables are not independent of one another. Thus, traditional tabled P-values cannot be used. A distribution of the statistic under the null hypothesis can be created, however, using permutations of the observations (e.g. Edgington 1995; Manly 1997). The only situation in which one could use the traditional tabled probabilities would be if one had a single variable that could reasonably be assumed to be normally distributed and one used Euclidean distances for the analysis. Suppose the null hypothesis is true and the groups are not really different (in terms of their composition and/or their relative abundances of species, as measured by the BrayCurtis distances). If this were the case, then the multivariate observations (rows) would be exchangeable among the different groups. Thus, the labels on the rows that identify them as belonging to a particular group could be randomly shufed (permuted) and a new value of F obtained (called, say, F ). This random shufing and re-calculation of F is then repeated for all possible re-orderings of the rows relative to the labels. This gives the entire distribution of the pseudo F-statistic under a true null hypothesis for our particular data. Comparing the value of F obtained with the original ordering of the rows to the distribution created for a true null by permuting the labels, a P-value is calculated as P (No. of F F) (Total no. of F ) (4)

d ij

(1)

That is, add up the squares of all of the distances in the subdiagonal (or upper-diagonal) half of the distance matrix (not including the diagonal) and divide by N (Fig. 3b). In a similar fashion, the within-group or residual sum of squares is SSW 1 n
N1 i l j N i 1

d ij

2 ij

(2)

where ij takes the value 1 if observation i and observation j are in the same group, otherwise it takes the value of zero. That is, add up the squares of all of the distances between observations that occur in the same group and divide by n, the number of observations per group (Fig. 3b). Then SSA SST SSW and a pseudo F-ratio to test the multivariate hypothesis is F SSA /(a 1) SSW /(N a) (3)

If the points from different groups have different central locations (centroids in the case of Euclidean distances) in multivariate space, then the among-group distances will be relatively large compared to the withingroup distances, and the resulting pseudo F-ratio will be relatively large. One can calculate the sums of squares in equations (1) and (2) and the statistic in equation (3) from a distance matrix obtained using any distance measure. The statistic in equation (3) corresponds exactly to the statistic in equation (4) of McArdle and Anderson (in press), who have shown more generally how partitioning for any linear model can be done directly from the distance matrix, regardless of the distance measure used. Another important aspect of the statistic described above is that, in the case of a Euclidean distance matrix calculated from only one variable, equation (3) gives the same value as the traditional parametric univariate F-statistic. This is proposed as a new non-parametric MANOVA statistic that is intuitively appealing, due to its analogy with univariate ANOVA, and that is extremely relevant for ecological applications. The results (in terms of sums of squares, mean squares and pseudo F-ratios) obtained for individual terms in a multivariate analysis

Note that we consider the original observed value of F to be a member of the distribution of F under permutation (i.e. it is one of the possible orderings of the labels on the rows). The usual scientic convention of an a priori signicance level of 0.05 is generally used for interpreting the signicance of the result, as in other statistical tests. It is also possible to view the

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

37

P-value as a measure of condence concerning the null hypothesis (Fisher 1955; Freedman & Lane 1983). With a groups and n replicates per group, the number of distinct possible outcomes for the F-statistic in a one-way test is (an)!/(a!(n!)a) (Clarke 1993). As it is usually not practical to calculate all possible permutations, because of the time involved, P can be calculated using a large random subset of all possible permutations (Hope 1968). However, the precision of the P-value will increase with increasing numbers of permutations. Generally, at least 1000 permutations should be done for tests with an -level of 0.05 and at least 5000 permutations should be done for tests with an -level of 0.01 (Manly 1997).

DISTINCTION FROM TRADITIONAL TEST STATISTICS Although the statistic described here is sensitive to differences in the relative dispersion of points among groups, it takes no account of the correlations among variables. In traditional MANOVA, the test-statistics (such as Wilks Lambda) use information contained in the between-group and/or within-group sample variance covariance matrices (e.g. Table 1d, see Olson 1974; Johnson & Field 1993). The traditional MANOVA tests assume not only that the variance for each variable remains constant across different groups (i.e. the points in different groups have similar scatter), they also assume that the relationships among the variables (their covariances or correlations) do not differ across groups. These differences in the sensitivities of different multivariate test statistics are shown diagrammatically in Fig. 4 for two variables (two dimensions). Figure 4(a) shows two groups that differ in their correlation structure, but not in their variances or location. Figure 4(b) shows two groups that differ only in their dispersions, but not in their correlation or location. Although all MANOVA statistics are designed to test for differences in location, the traditional statistics will also be sensitive to differences in correlations (Fig. 4a) as well as differences in dispersion (Fig. 4b). The method of nonparametric MANOVA described here will only be sensitive to differences in dispersion (Fig. 4b). The correlations among variables play no role in the analysis. If differences in the relationships amongst variables form a hypothesis of interest, then some other non-parametric techniques may be relevant (e.g. Biondini et al. 1991; Krzanowski 1993).

ASSUMPTIONS The only assumption of the test is that the observations (rows of the original data matrix) are exchangeable under a true null hypothesis. To assume exchangeability under the null hypothesis is generally to assume that the observations are independent and that they have similar distributions (e.g. Boik 1987; Hayes 1996). By similar distributions, I mean similar multivariate dispersions of points, not that the points are necessarily multivariate normal. The test described here is a test for differences in location (means or centroids) among groups of multivariate observations based on the chosen distance measure. Like its univariate counterpart, which is sensitive to heterogeneity of variances, this test and its predecessors that use permutations, like ANOSIM (Clarke 1993), will also be sensitive to differences in the dispersions of points, even if the locations do not differ. The sensitivity of ANOSIM to differences in dispersion has been suggested as an advantage by Clarke (1993). This is because it was introduced in the context of detecting environmental impacts, for which detection of differences of any kind between control and impacted locations is very important for environmental reasons. Here, I simply suggest that caution be exercised in interpreting the results of tests of signicance. Determining if signicant differences among groups may be due to differences in dispersion versus differences in location (or some combination of the two) is an important statistical and ecological issue. The use of permutation tests to obtain P-values does not avoid this issue. A useful comparative index of multivariate dispersion has been given by Warwick and Clarke (1993). Also, a separate permutation test for signicant differences in multivariate dispersions (after removing effects of differences in location), as an accompaniment to the nonparametric MANOVA approach given here, will be described elsewhere (Anderson, Dutilleul, Lapointe & Legendre, unpubl. data).

ONE-WAY EXAMPLE: EFFECTS OF GRAZERS Consider the following example, taken from an ecological study by Anderson and Underwood (1997). The study was designed to test the hypothesis that grazing by gastropods affects intertidal estuarine assemblages. The experiment was done at an intertidal oyster farm from January to July 1994 in Quibray Bay, south of Sydney, New South Wales, Australia. Experimental surfaces (10 cm 10 cm) were enclosed in cages to exclude gastropod grazers, while other surfaces were left open to grazing. A third treatment consisted of caged areas where natural densities of grazers were included. This was a control for the effect of the cage itself on assemblages. There were n 20 surfaces in each of the three treatments. The numbers of individuals of each of 21 taxa (invertebrates and algae) colonizing each surface were recorded. The rationale for increasing the severity of the transformation to increase the relative contribution of rare

38

M. J. ANDERSON

versus abundant species in the analysis, given by Clarke and Green (1988), is followed here. Note that the transformation is not done in an effort to make data conform to any assumptions of the analysis. In this example, the data contained some species that occurred on a very large relative scale of abundance (e.g. Spirorbid worms occurred in the thousands), so the data were transformed by taking double-square roots before the analysis. To visualize the multivariate patterns among observations, non-metric multidimensional scaling (MDS) was performed on the BrayCurtis distances (Kruskal & Wish 1978), using the PRIMER computer program. Non-parametric MANOVA was then done on BrayCurtis distances, as described in the previous section, using the computer program NPMANOVA, written by the author in FORTRAN.

The number of possible permutations for the oneway test in the case of the grazing experiment is 9.6 1025. A random subset of 4999 permutations was used (Fig. 5). In this case, the null hypothesis of no differences among groups was rejected, as the observed value was much larger than any of the values obtained under permutation (Fig. 5, Table 2).

A POSTERIORI TESTS As in univariate ANOVA where there is a signicant result in a comparison of 3 or more treatments, we may wish to ask for the multivariate case: wherein does the signicant difference lie? This can be done by using the same test, given above for the one-way comparison of groups, but where individual pair-wise comparisons between particular groups are done. To continue with the logic of the analogous univariate situation, we can use a t-statistic (which is simply the square root of the value of the F-statistic described above) for these
Table 2. Non-parametric MANOVA on BrayCurtis distances for assemblages of organisms colonizing intertidal surfaces in estuaries in three grazing treatments (grazers excluded, grazers inside cages, and surfaces open to grazers) Source Grazers Residual Total Comparison* Open versus caged Open versus cage control Caged versus cage control d.f. 2 57 59 SS 18 657.65 14 520.89 33 178.54 MS 9328.83 254.75 F P

36.61 0.0002

t 8.071 3.268 6.110

P 0.0002 0.0002 0.0002

*Pair-wise a posteriori tests among grazing treatments.

Fig. 4. Two variables in each of two groups of observations where (a) the groups differ in correlation between variables, but not in location or dispersion and (b) the groups differ in dispersion, but not in location or correlation between variables.

Fig. 5. Distribution of the non-parametric MANOVA F-statistic for 4999 permutations of the data on assemblages in different grazing treatments. The real value of F for these data is very extreme by reference to this distribution (F 36.62): thus there are strong differences among the assemblages in different grazing treatments.

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

39

pairwise comparisons. These have the same interpretation as univariate t-tests, but they test the general multivariate hypothesis of no difference between the groups on the basis of the BrayCurtis (or other chosen) distances. This is Students univariate t-statistic if Euclidean distances are chosen for the analysis of only one variable. P-values for each test are obtained using separate sets of permutations that are only done across the pair of groups being compared. In this example, a random subset of 4999 permutations was used (out of a possible 6.8 109) for each pairwise comparison. For the analysis of the experimental removal of grazers, there was a signicant difference among all pairs of treatments: assemblages colonizing surfaces in cages differed from those in the open or in cage controls (t 8.07, P 0.0002 and t 6.11, P 0.0002, respectively). Grazers had a signicant effect on the assemblages, which is consistent with the pattern of separation of points corresponding to different treatments in the non-metric MDS plot (Fig. 6). The fact that assemblages on open surfaces also differed signicantly from those in cage controls (t 3.27, P 0.0002) suggested that there was some additional artifact due to the presence of a cage in the experiment. An important point here is that the a posteriori comparisons just described did not make any correction for experiment-wise error rate (Day & Quinn 1989). Similarly, the multivariate pair-wise tests available in the computer program NPMANOVA are not corrected for experiment-wise error rate. This means that with an a priori signicance level of 0.05, one should expect to obtain a signicant result in one out of every 20

independent tests by chance alone. Nevertheless, the P-value obtained under permutation for any individual pair-wise test is exact. Many of the methods used for correcting error rates for multiple comparisons, such as the Bonferroni method, are very conservative but may be applied.

MORE COMPLEX DESIGNS For more complex designs, we can start by considering the situation with two factors in a factorial (or orthogonal) design. The principles used to partition the variation directly from the distance matrix and to obtain the statistics and permutation methods for individual terms in the model can be readily generalized to other multifactorial cases. The logic applied to multifactorial ANOVA of univariate data (e.g. see Underwood 1981, 1997) also applies to the analysis of multivariate data using this non-parametric procedure. For example, tests of main effects should be examined after tests for interactions.

Calculating the statistic

Let A designate factor 1 with a levels (treatments or groups) and B designate factor 2 with b levels, with n replicates in each of the ab combinations of the two factors. The total number of observations is N abn. The total sum of squares in the analysis is calculated as for the one-way case according to equation (1). To partition the variation, the within-group sum of squares for factor A, ignoring any inuence of B, is calculated as SSW(A) 1 bn
N1 i 1 j N i 1

d ij

2 (A) ij

(5)

where (A) takes the value 1 if observation i and obserij vation j are in the same group of factor A, otherwise it takes the value of zero. Similarly, the within-group sum of squares for factor B, ignoring any inuence of A, is SSW(B) 1 an
N1 i l j N i 1

d ij

2 (B) ij

(6)

Then, the corresponding sums of squares for each of the main effects in the analysis are SSA SST SSW(A) and SSB SST SSW(B). The residual sum of squares is calculated by considering the interpoint distances within each of the ab combinations of factor A and B, thus: SSR
Fig. 6. Non-metric MDS plot of assemblages colonizing intertidal surfaces in Quibray Bay in each of three different grazing treatments: ( ), cage control; ( ), open; ( ) caged.

1 n

N1 i l j

N i 1

d ij

2 (AB) ij

(7)

where (AB) takes the value 1 if observation i and obserij vation j are in the same combination of factors A and B, otherwise it takes the value of zero. We then can

40

M. J. ANDERSON

easily obtain the sum of squares corresponding to the interaction term: SSAB SST SSA SSB SSR. It may be easier to consider the squared distances being summed in equations (5) through (7) by reference to their physical location in the distance matrix itself, as illustrated in Fig. 7. In the case of a two-factor design where one factor is nested in the other, the same general approach is used. In this case, however, there is no interaction term in the analysis and we have instead SSB(A) SST SSA SSR, where B(A) denotes that factor B is nested in factor A. Having obtained appropriate sums of squares, the construction of the pseudo F-statistic for each term in the analysis for non-parametric MANOVA then follows the same rules and formulae as it would for the usual univariate ANOVA. The construction of the F-ratio depends on the experimental design, that is, whether factors are nested or factorial and whether they are xed or random, exactly as for univariate ANOVA (e.g. Underwood 1981, 1997; Winer et al. 1991).

Doing the permutations

The method of permutation required to obtain an exact test is not so simple if there is more than one factor in the design. The choice of an appropriate permutation method is not trivial and should be considered carefully for each term in the model. Indeed, the lack of exact tests or knowledge of how the available approximate permutation tests might behave for complex models has been a sticking point in the development of multivariate non-parametric methods (e.g. Crowley 1992; Clarke 1993). To construct exact tests, two important issues must be considered (Anderson & ter Braak, unpublished data). First, which units should be permuted (i.e. what are exchangeable under the null hypothesis) and second, should any restrictions be imposed on the permutations to account for other factors in the design? In many important situations, such as tests of interactions, no exact permutation test can be done. Also, there are times when the exact test imposes so many

Fig. 7. Schematic diagram of the interpoint distances used to partition the variability in the multivariate data set and to calculate the sum of squares for each term in a two-factor orthogonal design (each factor has two groups or levels). (a) SSW(A) sum of squared distances within groups of A ( ), divided by (bn). (b) SSW(B) sum of squared distances within groups of B ( ), divided by (an). (c) SSR sum of squared distances within combinations of AB ( ), divided by (n) (residual sum of squares). SST sum of squared distances in the total half matrix, divided by (abn), SSA SST SSW(A), SSB SST SSW(B), SSAB SST SSA SSB SSR.

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

41

restrictions as to render the test meaningless, due to there being too few possible permutations left. In these cases, approximate permutation tests should be used, of which there are several alternatives, including permutation of residuals and permutation of raw data across all terms in the analysis (e.g. Freedman & Lane 1983; ter Braak 1992; Manly 1997). Some empirical comparisons of these methods are provided by Gonzalez and Manly (1998) and Anderson and Legendre (1999).

ECOLOGICAL EXAMPLES
Two-way factorial design

The rst example is from an experiment in Middle Harbour (part of Sydney Harbour) to test for the effect of shade and proximity to the seaoor on assemblages of invertebrates and algae colonizing subtidal hard surfaces near marinas (Glasby 1999). The experiment was a two-way crossed (orthogonal) design with n 4 replicate settlement plates (15 cm 15 cm sandstone) either far from or near to the seaoor (the factor of position; all plates were at a similar depth of approximately 2 m below low water) and in one of three shading treatments: (i) shade (an opaque plexiglass roof), (ii) a procedural control (a clear plexiglass roof), and (iii) no shade. Organisms colonizing the plates after 33 weeks were counted and a total of 46 taxa were included in the analyses. Organisms that occurred only once across the entire data set were not included. Non-parametric MANOVA was done on BrayCurtis distances calculated from double-root transformed data using the FORTRAN program NPMANOVA. The sample size was reasonably small for this study (n < 5), so the
Table 3. Non-parametric MANOVA on BrayCurtis distances for assemblages of organisms colonizing subtidal sandstone settlement panels after 33 weeks in an estuary at different distances from the seaoor (positions either near or far) and in three different shading treatments Source Position Shade Position shade Residual Total Comparison* Shade versus control Shade versus no shade Control versus no shade d.f. 1 2 2 18 23 SS MS F 13.536 4.314 1.499 P 0.0002 0.0006 0.1394

test was done using unrestricted permutation of raw data (e.g. Manly 1997; Gonzalez & Manly 1998) with 4999 random permutations. Similar results were obtained using permutation of residuals under a reduced model (not shown). There was no signicant interaction of shade and position, but both main effects were signicant (Table 3, Fig. 8). It was then of interest to compare the groups corresponding to different shading treatments using a posteriori tests (Table 3). It was not necessary to do this for the effect of position, because this factor only had two groups. Assemblages of organisms on settlement plates near the bottom were extremely different from those far away from the bottom (Fig. 8). Also, assemblages on shaded plates were signicantly different from those on either the procedural control or on unshaded plates, which themselves did not differ (Table 3, Fig. 8). This analysis also shows how the effect of position relative to the bottom was much greater than the effect of shading on assemblages in this experiment (compare the values of their mean squares in Table 3). The non-parametric approach advocated here allows tests of signicance, but it also allows relative sizes of effects to be compared directly through the partitioning of the variation and examination of mean squares.

Three-way design, including nesting

The second example comes from an experiment to test the hypothesis that the size of a patch available for

5595.40 5595.40 3566.44 1783.22 1238.94 619.47 7440.66 17 841.43 413.37

t 1.783 1.987 0.866

P 0.0154 0.0018 0.5560 Fig. 8. Two-factor non-metric MDS plot of subtidal assemblages colonizing sandstone settlement plates after 33 weeks in Middle Harbour that were either near to or far from the seaoor and in one of three shading treatments. ( ), Shaded; ( ), control; ( ), no shade; 1, far from the sea oor; 2, close to the sea oor.

*Pair-wise a posteriori tests among shading treatments.

42

M. J. ANDERSON

colonization would affect the succession of assemblages in equal areas on those patches (Anderson 1998). This hypothesis was tested using wooden panels of three different sizes (10 cm 10 cm, 20 cm 20 cm and 40 cm 40 cm). Two panels of a given size were attached to sticks that were then strapped to the structure of an intertidal oyster farm in the Port Stephens estuary in New South Wales, Australia in January of 1995 (see Anderson 1998 for details). Six sticks (two for each of the three patch sizes) were then collected independently after periods of 3, 6, 9, 12 and 18 months of exposure to colonization. The experimental design thus consisted of three factors: time (5 periods of submersion), patch (3 sizes) and sticks (2 sticks per time patch combination, a random nested factor), with n 2 panels per stick. Organisms colonizing panels were quantied in a 10 cm 10 cm area from each panel (chosen randomly from the largersized panels). A total of 33 taxa were included in multivariate analyses. The analysis was done using NPMANOVA on Bray Curtis distances calculated on double-root transformed data, as for the previous examples. In this case, however, an exact permutation test for the nested factor (sticks) was carried out by permuting the observations randomly across sticks, but only within the 5 3 combinations of levels of time patch. Then, for the test of the upper-level terms (the main effects of time, patch size and their interaction), individual replicates on a stick were permuted together as a unit (i.e. whole sticks were permuted). This is done so that the upper-level effects can be tested against the variability across sticks, not across individual replicates, as is necessary under the null hypothesis for a nested hierarchy (e.g. Clarke 1993). For all tests, a subset of 4999 permutations was used. The nested factor of sticks was not signicant in the multivariate analysis, but the time patch size interaction was signicant (Table 4). Individual pair-wise comparisons of effects of patch size for each time were nearly identical to the one-way tests given in Anderson (1998) using analysis of similarities (ANOSIM, Clarke 1993). Assemblages were signicantly different on the smallest patches compared to the other sized patches after 3, 6, 9 or 12 months. Assemblages on the two larger sized patches did not differ signicantly from one another except after 12 months. After 18 months, assemblages were similar on all patch sizes. Non-metric MDS plots helped to interpret these results. Two separate ordinations were done on these data, as the stress value for the non-metric MDS plot that included all observations was too high for a reasonable interpretation. The effect of different patch sizes appeared to be fairly consistent (in terms of its magnitude and direction) after 3, 6 or 9 months (Fig. 9a). After 12 or 18 months, the observations become more scattered within and across the groups

and the effects of patch size become less clear (Fig. 9b). Increased dispersion (variability in assemblages) after these longer periods of time, compared to earlier periods, is seen clearly in the two-factor plot of stick centroids (Fig. 9c, which includes all data). As noted earlier, the tests are sensitive to such differences in dispersion. Overall, although the two factors did interact, the effect of time (i.e. succession) was relatively more important in distinguishing assemblages than the size of the patch (compare their mean squares in Table 4), and effects of patch size decreased through time for these assemblages (see Anderson 1998 for further details).

DISCUSSION Natural temporal and spatial variability is intrinsic to ecological systems. Indeed, variability might be considered the currency of ecological scientic work. It is for this reason that statistical analysis plays such an important role in the development of ecology as a science. In Design of Experiments, R. A. Fisher (1935, p. 4) wrote: We may at once admit that any inference from the particular to the general must be attended with some degree of uncertainty, but this is not the same as to admit that such inference cannot be absolutely rigorous, for the nature and degree of uncertainty may itself be capable of rigorous expression. Quantitative statistical inference is indeed what is needed for the rigorous interpretation of mensurative or manipulative ecological experiments. Although our conclusions may be uncertain, they are still rigorous in the sense that the degree of uncertainty can be expressed in terms of mathematical probability. In univariate analysis, W. S. Gosset (Student 1908) made this possible for comparisons of two treatments, while R. A. Fisher made this possible for many treatments and experimental factors. In a complex and intrinsically variable world, ANOVA allows us to identify simultaneous effects and interactions of more than one factor, and to identify the uncertainty of our inferences with rigour (e.g. Underwood 1981, 1997). An important advance in the analysis of multivariate data in ecology was the development of non-parametric methods for testing hypotheses concerning whole communities (e.g. Clarke 1988, 1993; Smith et al. 1990; Biondini et al. 1991). Some parallel advances were made in the context of tests for signicant clusters in cluster analysis (e.g. Good 1982; Gordon 1994). Before these applications, particularly ANOSIM (Clarke 1993), became widely available, most multivariate analyses in ecology focused on the reduction of dimensionality to produce and interpret patterns (ordination

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

43

methods) and the use of numerical strategies for placing observations into natural groups (clustering). These methods, although extremely useful towards their purpose, do not rigorously express the nature and degree of uncertainty concerning a priori hypotheses. Methods like Mantels test (Mantel 1967), ANOSIM (Clarke 1993) and multiresponse permutation procedures (Mielke et al. 1976) allow such rigorous probabilistic statements to be made for multivariate ecological data. The drawback to such non-parametric tests is that they cannot easily be extended to the multifactorial designs so common now in ecological studies. Two sticking points prevented this: (i) the lack of a generalized statistic for partitioning variation, and (ii) the lack of appropriate permutation methods (e.g. Clarke 1993; Legendre & Anderson 1999). Although traditional test-statistics used for MANOVA allow partitioning, their restrictive assumptions have prevented their effective use in ecology. The method of distancebased redundancy analysis (Legendre & Anderson
Table 4. Non-parametric MANOVA on BrayCurtis distances for assemblages of organisms colonizing wooden settlement panels of three different sizes after 3, 6, 9, 12 or 18 months on an intertidal oyster farm Source d.f. SS MS F P 0.0002 0.0002 0.0062 0.3384

Time 4 Patch Size 2 Time patch 8 Sticks (time 15 patch) Residual 30 Total 59

30 305.71 7576.43 20.50 6414.99 3207.49 8.68 6224.03 778.004 2.10 5544.66 369.64 1.28 8697.09 57 186.48 289.90

Comparison

3 6 9 12 18 months months months months months 1.86* 2.87* 1.47 1.48* 2.38* 1.69 2.00* 2.54* 1.75* 1.38 3.02* 1.43

Small versus 2.24* medium Small versus 2.30* large Medium versus 1.49 large

*P < 0.05; pair-wise a posteriori tests among patch sizes within each time using the t-statistic. Sticks were ignored in the pair-wise tests. There were 35 possible permutations for each.

Fig. 9. Two-factor non-metric MDS plots of assemblages colonizing intertidal wooden settlement panels of three different sizes for (a) periods of 3, 6 or 9 months and (b) periods of 12 or 18 months for raw data and (c) where centroids were plotted corresponding to each stick across all combinations of time patch size. The points corresponding to assemblages after particular times of submersion (numbers indicate the period in months) have been outlined for clarity. ( ), Small; ( ), medium; ( ), large.

44

M. J. ANDERSON

1999) largely solved these issues, but it has many rather complicated steps and involves the use of a correction constant to distances. Although this correction does not adversely affect the test (generally making it more conservative, if anything), accurate P-values are not given by this method in the case of anything other than a one-factor design (McArdle & Anderson, in press). The method presented here has, in some sense, combined the best of both worlds. Like the traditional test-statistics, it can partition variation according to any ANOVA design. Like the most exible nonparametric methods, it can be based on any symmetric dissimilarity or distance measure (or their ranks) and provides a P-value using appropriate permutation methods. That is, one can still choose a relevant transformation and an appropriate distance measure (or use ranks of distances), consistent with the method of ordination used to visualize patterns. By using permutations, the test requires no specic assumption concerning the number of variables or the nature of their individual distributions or correlations. The statistic used is analogous to Fishers F-ratio and is constructed from sums of squared distances (or dissimilarities) within and between groups. Another feature of this statistic is that it is equal to to Fishers original F-ratio in the case of one variable and when Euclidean distances are used. It is perhaps important to point out that the BrayCurtis measure of dissimilarity may or may not be the most appropriate measure to use in any given situation. A point still commonly ignored is that BrayCurtis and related measures, such as Kulczynskis coefcient, will tend to under-estimate true ecological distances when distances become large (i.e. when observations have very few species in common), as determined by simulations (Faith et al. 1987; Belbin 1991). The BrayCurtis measure may therefore only be useful insofar as it will produce reasonable ecological ordinations, through the ranks used for MDS. In light of this issue, Faith (1990) proposed doing MANOVA on ordination scores obtained from MDS. An analysis based only on ordination scores will omit some portion of the ecological information contained in the original distance matrix and will depend on the number of dimensions chosen for the MDS (Clarke 1993). The severity of this potential problem will obviously increase with increases in the stress value associated with an MDS plot. The actual amount and the kind of information lost in reducing dimensions using MDS are generally unknown and will depend on the particular data set. Also, any subsequent statistical inferences on ordination scores (using traditional MANOVA, as described in Faith 1990; or based on permutation tests) would be limited to points in the ordination plots, rather than being applicable to the original observations.

It is not possible to identify a single best distance measure for ecological data. The method described here may be used with any distance measure chosen (or on ranks of distances). It is useful to have the exibility to choose a distance measure appropriate for the data and hypothesis being tested. Although the BrayCurtis measure has now become commonplace in ecological studies (perhaps due to its availability in the PRIMER computer program, or due to its intuitive interpretation as percentage difference, or due to the results presented in Faith et al. 1987), there are still many rivals. Over 60 measures of similarity or dissimilarity have been described, with very few actual comparisons of their performance with different kinds of ecological data (see Lamont & Grant 1979; Legendre & Legendre 1998). This is undoubtedly an area needing further research. The approach advocated here is that multifactorial analysis of variance, as successfully applied to univariate data in ecology (e.g. Underwood 1981, 1997), can and should also be applied to multivariate data for testing hypotheses in a logical and rigorous way. It stands to reason that the developments in experimental design for ecology that require multifactorial ANOVA, in order to, for example, avoid pseudo-replication (Hurlbert 1984), test for generality (Beck 1997) or test for environmental impact (e.g. Green 1979, 1993; Underwood 1993; Glasby 1997), should be incorporated into multivariate analysis. The method described here allows that to happen, but within a framework that is general enough to suit our need for few assumptions and exibility in the multivariate analysis of ecological data.

ACKNOWLEDGEMENTS I am indebted to those who have worked on and developed multivariate methods and permutation tests that have led to the ideas in this manuscript, including K. R. Clarke, E. S. Edgington, R. A. Fisher, P. Legendre, B. F. J. Manly, N. Mantel, B. H. McArdle, L. Orlci, V. D. P. Pillar, E. P. Smith, and C. J. F. ter Braak. I also owe a great deal to A. J. Underwood for his work in the statistical analysis of ecological experiments using ANOVA, which inspired my pursuit of this topic for multivariate analysis. T. Glasby kindly provided data for the two-way factorial example. The PRIMER computer program was provided courtesy of M. R. Carr and K. R. Clarke, Plymouth Marine Laboratories, UK. My colleagues at the Special Research Centre for Ecological Impacts of Coastal Cities provided logistic support, tested out the computer program and commented on earlier versions of the manuscript. The computer program NPMANOVA is available from the author. This research was supported by a U2000 Post-doctoral Fellowship at the University of Sydney.

N ON -PA R A M ET R I C M A N OVA FOR ECOLOG Y

45

REFERENCES
Anderson M. J. (1998) Effects of patch size on colonisation in estuaries: revisiting the species-area relationship. Oecologia 118, 8798. Anderson M. J. & Clements A. (in press) Resolving environmental disputes: a statistical method for choosing among competing cluster models. Ecol. Applic. Anderson M. J. & Legendre P. (1999) An empirical comparison of permutation methods for tests of partial regression coefcients in a linear model. J. Stat. Comput. Simul. 62, 271303. Anderson M. J. & Underwood A. J. (1994) Effects of substratum on the recruitment and development of an intertidal estuarine fouling assemblage. J. Exp. Mar. Biol. Ecol. 184, 21736. Anderson M. J. & Underwood A. J. (1997) Effects of gastropod grazers on recruitment and succession of an estuarine assemblage: a multivariate and univariate approach. Oecologia 109, 44253. Bartlett M. S. (1939) A note on tests of signicance in multivariate analysis. Proc. Camb. Philos. Soc. 35, 1805. Beck M. W. (1997) Inference and generality in ecology: current problems and an experimental solution. Oikos 78, 26573. Belbin L. (1991) Semi-strong hybrid scaling, a new ordination algorithm. J. Veg. Sci. 2, 4916. Biondini M. E., Mielke P. W. & Redente E. F. (1991) Permutation techniques based on Euclidean analysis spaces: a new and powerful statistical method for ecological research. In: Computer Assisted Vegetation Analysis (eds E. Feoli & L. Orlci) pp. 22140. Kluwer Academic Publishers, Dordrecht. Boik R. J. (1987) The FisherPitman permutation test: a nonrobust alternative to the normal theory F test when variances are heterogeneous. Br. J. Math. Stat. Psychol. 40, 2642. Bray J. R. & Curtis J. T. (1957) An ordination of the upland forest communities of southern Wisconsin. Ecol. Monogr. 27, 32549. Caliski T. & Harabasz J. (1974) A dendrite method for cluster analysis. Commun. Stat. 3, 127. Chapman M. G., Underwood A. J. & Skilleter G. A. (1995) Variability at different spatial scales between a subtidal assemblage exposed to discharge of sewage and two control sites. J. Exp. Mar. Biol. Ecol. 189, 10322. Clarke K. R. (1988) Detecting change in benthic community structure. In: Proceedings XIVth International Biometric Conference, Namur: Invited Papers, pp. 13142. Socit Adolphe Quetelet, Gembloux. Clarke K. R. (1993) Non-parametric multivariate analyses of changes in community structure. Aust. J. Ecol. 18, 11743. Clarke K. R. & Green R. H. (1988) Statistical design and analysis for a biological effects study. Mar. Ecol. Prog. Ser. 46, 21326. Crowley P. H. (1992) Resampling methods for computationintensive data analysis in ecology and evolution. Ann. Rev. Ecol. Syst. 23, 40547. Day R. W. & Quinn G. P. (1989) Comparison of treatments after an analysis of variance. Ecol. Monogr. 59, 43363. Edgington E. S. (1995) Randomization Tests, 3rd edn. Marcel Dekker, New York. Excofer L., Smouse P. E. & Quattro J. M. (1992) Analysis of molecular variance inferred from metric distances among DNA haplotypes: application to human mitochondrial DNA restriction data. Genetics 131, 47991. Faith D. P. (1990) Multivariate methods for biological monitoring based on community structure. In: The Australian Society of

Limnology 29th Congress, p. 17 (Abstract). Alligator Rivers Region Research Institute. Faith D. P., Dostine P. L. & Humphrey C. L. (1995) Detection of mining impacts on aquatic macroinvertebrate communities: results of a disturbance experiment and the design of a multivariate BACIP monitoring program at Coronation Hill, N. T. Aust. J. Ecol. 20, 16780. Faith D. P., Minchin P. R. & Belbin L. (1987) Compositional dissimilarity as a robust measure of ecological distance. Vegetatio 69, 5768. Fisher R. A. (1935) Design of Experiments. Oliver & Boyd, Edinburgh. Fisher R. A. (1936) The use of multiple measurements in taxonomic problems. Ann. Eugen. 7, 17988. Fisher R. A. (1955) Statistical methods and scientic induction. J. Roy. Stat. Soc. 17, 6978. Freedman D. & Lane D. (1983) A nonstochastic interpretation of reported signicance levels. J. Bus. Econ. Stat. 1, 2928. Gaston K. J. & McArdle B. H. (1994) The temporal variability of animal abundances: measures, methods and patterns. Phil. Trans. Roy. Soc. Lond. 345, 33558. Glasby T. M. (1997) Analysing data from post-impact studies using asymmetrical analyses of variance: a case study of epibiota on marinas. Aust. J. Ecol. 22, 44859. Glasby T. M. (1999) Interactive effects of shading and proximity to the seaoor on the development of subtidal epibiotic assemblages. Mar. Ecol. Prog. Ser. 190, 11324. Gonzalez L. & Manly B. F. J. (1998) Analysis of variance by randomization with small data sets. Environmetrics 9, 5365. Good I. J. (1982) An index of separateness of clusters and a permutation test for its signicance. J. Stat. Comput. Simul. 15, 814. Gordon A. D. (1994) Identifying genuine clusters in a classication. Comput. Stat. Data Anal. 18, 56181. Gower J. C. (1966) Some distance properties of latent root and vector methods used in multivariate analysis. Biometrika 53, 32538. Gray J. S., Clarke K. R., Warwick R. M. & Hobbs G. (1990) Detection of initial effects of pollution on marine benthos: an example from the Ekosk and Eldsk oilelds, North Sea. Mar. Ecol. Prog. Ser. 66, 28599. Green R. H. (1979) Sampling Design and Statistical Methods for Environmental Biologists. Wiley, New York. Green R. H. (1993) Application of repeated measures designs in environmental impact and monitoring studies. Aust. J. Ecol. 18, 8198. Hajdu L. J. (1981) Graphical comparison of resemblance measures in phytosociology. Vegetatio 48, 4759. Hayes A. F. (1996) Permutation test is not distribution free. Psychol. Methods 1, 18498. Hope A. C. A. (1968) A simplied Monte Carlo signicance test procedure. J. Roy. Stat. Soc. 30, 58298. Hotelling H. (1931) The generalization of Students ratio. Ann. Math. Stat. 2, 36078. Hubert L. & Schultz J. (1976) Quadratic assignment as a general data analysis strategy. Br. J. Math. Stat. Psychol. 29, 190241. Hurlbert S. H. (1984) Pseudoreplication and the design of ecological eld experiments. Ecol. Monogr. 54, 187211. Johnson C. R. & Field C. A. (1993) Using xed-effects model multivariate analysis of variance in marine biology and ecology. Oceanogr. Mar. Biol. Ann. Rev. 31, 177221. Kelaher B. P., Chapman M. G. & Underwood A. J. (1998) Changes in benthic assemblages near boardwalks in

46

M. J. ANDERSON

temperate urban mangrove forests. J. Exp. Mar. Biol. Ecol. 228, 291307. Kendall M. G. & Stuart A. (1963) The Advanced Theory of Statistics, Vol. 1, 2nd edn. Charles Grifth, London. Kruskal J. B. & Wish M. (1978) Multidimensional Scaling. Sage Publications, Beverly Hills. Krzanowski W. J. (1993) Permutational tests for correlation matrices. Statistics and Computing 3, 3744. Kulczynski S. (1928) Die Panzenassoziationen der Pieninen. Bull. Int. Acad. Pol. Sci. Lett. Cl. Sci. Math. Nat. Ser. B, (Suppl. II) 1927, 57203. Lamont B. B. & Grant K. J. (1979) A comparison of twenty-one measures of site dissimilarity. In: Multivariate Methods in Ecological Work (eds L. Orloci, C. R. Rao & W. M. Stiteler) pp. 10126. International Co-operative Publishing House, Fairland. Lawley D. N. (1939) A generalization of Fishers IX test. Biometrika 30, 1807 (Corrections in Biometrika 30, 4679). Legendre P. & Anderson M. J. (1999) Distance-based redundancy analysis: testing multispecies responses in multifactorial ecological experiments. Ecol. Monogr. 69, 124. Legendre P. & Legendre L. (1998) Numerical Ecology, 2nd English edn. Elsevier Science, Amsterdam. McArdle B. H. & Anderson M. J. (in press) Fitting multivariate models to community data: a comment on distance-based redundancy analysis. Ecology. Manly B. F. J. (1997) Randomization, Bootstrap and Monte Carlo Methods in Biology, 2nd edn. Chapman & Hall, London. Mantel N. (1967) The detection of disease clustering and a generalized regression approach. Cancer Res. 27, 20920. Mantel N. & Valand R. S. (1970) A technique of nonparametric multivariate analysis. Biometrics 26, 54758. Mardia K. V. (1971) The effect of non-normality on some multivariate tests and robustness to nonnormality in the linear model. Biometrika 58, 10521. Mardia K. V., Kent J. T. & Bibby J. M. (1979) Multivariate Analysis. Academic Press, London. Mielke P. W., Berry K. J. & Johnson E. S. (1976) Multi-response permutation procedures for a priori classications. Commun. Stat. Theory Methods 5 (14), 140924. Odum E. P. (1950) Bird populations of the Highlands (North Carolina) Plateau in relation to plant succession and avian invasion. Ecology 31, 587605. Oliver I. & Beattie A. J. (1996) Designing a cost-effective invertebrate survey: a test of methods for rapid assessment of biodiversity. Ecol. App. 6, 594607. Olson C. L. (1974) Comparative robustness of six tests in multivariate analysis of variance. J. Am. Stat. Assoc. 69, 894908.

Pillai K. C. S. (1955) Some new test criteria in multivariate analysis. Ann. Math. Stat. 26, 11721. Pillar V. D. P. & Orlci L. (1996) On randomization testing in vegetation science: multifactor comparisons of relev groups. J. Veg. Sci. 7, 58592. Quinn G. P., Lake P. S. & Schreiber S. G. (1996) Littoral benthos of a Victorian lake and its outlet stream: spatial and temporal variation. Aust. J. Ecol. 21, 292301. Seber G. A. F. (1984) Multivariate Observations. John Wiley and Sons, New York. Skilleter G. A. (1996) Validation of rapid assessment of damage in urban mangrove forests and relationships with Molluscan assemblages. J. Mar. Biol. Ass. UK 76, 70116. Smith E. P., Pontasch K. W. & Cairns J. (1990) Community similarity and the analysis of multispecies environmental data: a unied statistical approach. Water Res. 24, 50714. Student. (1908) The probable error of a mean. Biometrika 6, 125. ter Braak C. J. F. (1992) Permutation versus bootstrap signicance tests in multiple regression and ANOVA. In: Bootstrapping and Related Techniques (eds K. H. Jckel, G. Rothe & W. Sendler) pp. 7986. Springer-Verlag, Berlin. Underwood A. J. (1981) Techniques of analysis of variance in experimental marine biology and ecology. Oceanogr. Mar. Biol. Ann. Rev. 19, 513605. Underwood A. J. (1993) The mechanics of spatially replicated sampling programmes to detect environmental impacts in a variable world. Aust. J. Ecol. 18, 99116. Underwood A. J. (1997) Experiments in Ecology: Their Logical Design and Interpretation Using Analysis of Variance. Cambridge University Press, Cambridge. Underwood A. J. & Chapman M. G. (1998) A method for analysing spatial scales of variation in composition of assemblages. Oecologia 107, 5708. Warwick R. M., Carr M. R., Clarke K. R., Gee J. M. & Green R. H. (1988) A mesocosm experiment on the effects of hydrocarbon and copper pollution on a sublittoral softsediment meiobenthic community. Mar. Ecol. Prog. Ser. 46, 18191. Warwick R. M. & Clarke K. R. (1993) Increased variability as a symptom of stress in marine communities. J. Exp. Mar. Biol. Ecol. 172, 21526. Wilks S. S. (1932) Certain generalizations in the analysis of variance. Biometrika 24, 47194. Winer B. J., Broan D. R. & Michels K. M. (1991) Statistical Principles in Experimental Design, 3rd edn. McGraw-Hill, Sydney.

FORUM

An Entomologist Guide to Demystify Pseudoreplication: Data Analysis of Field Studies With Design Constraints
LUIS FERNANDO CHAVES1,2

J. Med. Entomol. 47(3): 291298 (2010); DOI: 10.1603/ME09250

ABSTRACT Lack of independence, or pseudoreplication, in samples from ecological studies of insects reects the complexity of working with living organisms: the nite and limited input of individuals, their relatedness (ecological and/or genetic), and the need to group organisms into functional experimental units to estimate population parameters (e.g., cohort replicates). Several decades ago, when the issue of pseudoreplication was rst recognized, it was highlighted that mainstream statistical tools were unable to account for the lack of independence. For example, the variability as a result of differences across individuals would be confounded with that of the experimental units where they were observed (e.g., pans for mosquito larvae), whereas both sources of variability now can be separated using modern statistical techniques, such as the linear mixed effects model, that explicitly consider the different scales of variability in a dataset (e.g., mosquitoes and pans). However, the perception of pseudoreplication as a problem without solution remains. This study presents concepts to critically appraise pseudoreplication and the linear mixed effects model as a statistical solution for analyzing data with pseudoreplication, by separating the different sources of variability and thereby generating correct inferences from data gathered in studies with constraints in randomization. KEY WORDS linear mixed effects model, Culex quinquefasciatus, Anopheles nuneztovari, bootstrap, data analysis

Pseudoreplication is probably one of the most widely cited and misunderstood concepts in the statistical analysis of ecological studies on insects and other organisms. Pseudoreplication is dened as the use of inferential statistics to test for treatment effects with data from experiments where either treatments are not replicated . . . or replicates are not statistically different . . . (Hurlbert 1984). This concept has been very inuential and pervasive, to the extent that pseudoreplication is widely cited as a major aw of most eld studies (Heffner et al. 1996). Hurlberts major claim was correct, and he basically showed that mainstream statistical tools at that time (e.g., analysis of variance) were not suitable for the analysis of most experimental designs. However, the uncritical appraisal of his study has been a major barrier for the publication of results and, therefore, the advancement of ecology (Oksanen 2001). Hurlberts study did not prevent the unsuitable analysis of valuable datasets or the proliferation of unsound experimental practices, such as the movement of sampling units to control for spatial/temporal variability (Alto and Juliano 2001a, 2001b; Reiskind and Wilson 2004). As thoughtfully
1 Corresponding author: Department of Environmental Studies, Emory University, 400 Dowman Drive, Suite E510, Atlanta GA 30322 (e-mail: lfchave@emory.edu). 2 Laboratorio de Biologa Teorica, Instituto de Zoologa y Ecologa Tropical, Facultad de Ciencias, Universidad Central de Venezuela, Caracas, Venezuela.

presented by Oksanen (2001), the goal of ecological studies is not the application of statistical analysis to ecological data per se, but rather its application to the understanding of ongoing ecological phenomena from variation of individual phenotypic traits to the assemblages of organisms in populations, communities, and ecosystems. Unlike physics or chemistry, in which the supply of individual objects of study is practically unlimited, the objects of study for an entomologist (or more generally a naturalist) are nite and constrained. Thus, limitations in randomization will likely arise, and the science of statistics has developed new solutions to correctly analyze the lack of independence in eld data since Hurlberts study (Millar and Anderson 2004). The current forum article presents the following: 1) key concepts of experimental design to critically appraise and demystify the concept of pseudoreplication; 2) linear mixed effects models (LMEMs) as powerful tools to analyze data originated from constrained designs or to produce more general inferences from classical randomized designs (e.g., blocks); and 3) how these tools can be used to further gain insights from the data that can strengthen our understanding of insect ecology. In developing point 2, equations and a guide to interpret them as models used to analyze common entomological data are presented, as well as the implementation of this type of analysis in the open source software R.
2010 Entomological Society of America

0022-2585/10/02910298$04.00/0

292

JOURNAL OF MEDICAL ENTOMOLOGY

Vol. 47, no. 3

Field Studies, Experiments, and Statistical Data Analysis Experiments are one of the major tools for hypothesis testing (Fisher 1935). In general, the idea is to subject individual units of observation to varying degrees of independent and/or controllable factor(s), and to determine how the levels of variation explain a given pattern (Box 1980). Field studies focus on the impacts of natural (or controlled) variation in environmental factors on individual units of observation. Hypothesis testing has been central to the development of modern science, to the point that hypothesisdriven experiments or eld studies are one of the most prominent requirements for project support by funding agencies. One of the major reasons for the widespread appeal of hypothesis-driven experiments and eld studies has been their close association with tools for data analysis to determine the impact of different independent variables. The best example of a statistical tool guiding experimental design is the use of the linear model (LM). This model assumes that variability across a set of individual units of observation (yi) is explained by a series of n independent variables (x1, x2, . . . , xn) and by a unique source of unexplained variability, normally referred to as error ( ). These models are linear, because the parameters enter linearly into the equation that relates the independent variables to the outcome (Faraway 2006, Chaves and Pascual 2007). A major constraint of these models is that they assume total independence among the subjects of study, i.e., individual observation units are unrelated at least within strata (i.e., after accounting for the explanatory variables), which is the formal denition of replication. When there is a lack of independence across objects of study (i.e., pseudoreplication), the use of LMs with a unique source of variability is inappropriate, because the variability is modeled incorrectly and can lead to spurious inferences. For example, if mosquitoes are reared in pans (or kissing bugs in jars) to measure body size of emerging adults from different experimental conditions, the lack of independence that arises from the aggregation into pans (or jars) will inate the error value (a.k.a. residual variance) of the LM, in some cases leading to incorrect inferences when the LM is compared with a model that explicitly models the lack of independence because of the aggregation into a functional experimental unit (i.e., the pan or jar). More than 20 yr ago, because of the limited statistical toolbox in ecology, this issue was a major problem for the correct analysis of datasets from studies with design constraints (Hurlbert 1984). However, strategies to handle the problem of pseudoreplication were around at the time. For example, in evolutionary ecology, individuals have different degrees of common descent, and this variability by itself is often a subject of study. In the 1980s, it was common to use nested half-sibling designs to estimate the variance of families and individuals belonging to those families (Conner and Hartl 2004). Also, the use of dened designs such as Greco-Roman squares, Roman squares, and fractional factorials was

well established in the eld of engineering and process control (Montgomery 2005). Some of these balanced designs were even used in studies of medically important insects (Carpenter 1982, Chesson 1984). In fact, sophistication in randomization, when possible, can be very useful to evaluate the impact of strategies to control human-vector contact (Kirby et al. 2008). For example, randomized control trials have been used to demonstrate the importance of mosquito screening in reducing the risk of malaria transmission (Kirby et al. 2009). However, one of the major limitations of designs that handle pseudoreplication is the need for balanced designs, i.e., an equal number of replicates per treatment. The inability to analyze unbalanced designs with unequal number of replicates per treatment has been overcome with the development of maximum likelihood methods, especially the restricted maximum likelihood method and its application to estimate LMEM (Pinheiro and Bates 2000). LMEM has the same fundamental assumptions of the LM; it tries to explain the sources of variability across a set of individual units of observation (yi) as function of a series of n independent variables (x1, x2, . . . ., xn), referred to as xed factors, but it can incorporate additional sources of variability (the random factors), besides the error ( ). The random factors can accommodate the lack of independence among the individual units of observation as a result of spatial, temporal, genetic, or any exogenous environmental factor that is not fully randomized. For example, the variance of functional experimental units such as pans for mosquitoes or jars for kissing bugs can be explicitly modeled, thus allowing the proper estimation of the error variance, and thus limiting the chances of committing a type II error, i.e., rejecting the null hypothesis when true. Therefore, LMEMs allow the statistical analysis of pseudoreplicated data. The next section will provide a series of examples illustrating the use of LMEMs and how they compare with similar LMs. The data used in the examples and code to perform the analyses using the open source statistical software R are included as supplementary online material (http://www.envs.emory.edu/research/ Chaves_SOM_Pseudoreplication.html). LMs Versus LMEMs Factorial Designs. To illustrate the most basic differences between LMs and LMEMs, I reference a eld experiment designed to study oviposition by Culex quinquefasciatus Say in Atlanta, GA (Chaves et al. 2009). Cx. quinquefasciatus larvae are normally absent from lotic systems, such as rivers and creeks. However, several cities have relic sewage treatment systems where runoff water and sewage are combined in the same system, and after large rainfall events the combined sewage efuent can overow into urban water bodies (Chaves et al. 2009). In this experiment, the effects of combined sewage overow water and nutrient addition on oviposition site selection by this mosquito species were studied using 10 experimental pools (water containers) at four sites in a forest patch.

May 2010

CHAVES: DEMYSTIFYING PSEUDOREPLICATION

293

Table 1. Analysis of variance for the effects of water quality and nutrient addition on the ln of total number of egg rafts 1 oviposited over 5 d by Culex quinquefasciatus in Atlanta, GA (Chaves et al. 2009) Factor Water ( 1) Nutrient ( 2) Water nutrient ( 3) Block ( 4) Error ( ) df 1 1 1 3 9 Sum Mean square square 0.653 50.634 0.013 0.467 0.58 0.653 50.634 0.013 0.156 0.064 F Value 10.1401 786.0312 0.2055 2.4186 Pr( F) 0.01111* 4.54E-10* 0.66105 0.13339

ulation and models block variability as a random factor (Fig. 1). The equivalent to equation 1 for a LMEM is: y il
1 1

2 2

3 1 2

xx

il

[2]

This design was balanced. Original data and R code for analysis are in the supplementary online material (http://www.envs.emory.edu/ research/Chaves_SOM_Pseudoreplication.html). *Statistically signicant (P 0.05).

For this example, data will be used from the experiment when egg rafts were removed daily. Data use has been restricted to a randomly extracted subsample from the original data (only four pools per site) to have a dataset similar to that of a balanced design. In this experiment, oviposition (y) was measured by counting the total number of egg rafts oviposited over 5 d. The experiment has three independent variables (i.e., n 3): 1) x1 water quality (with two levels: combined sewage overow water and tap water as control); 2) x2 nutrient addition (added or absent as control); and 3) x3 sites (four in total). Only x1 and x2 are factors (each with two levels), because x3 is an independent variable considered to test the block effects of forest site on oviposition. Because all treatments were present at each site, this is a randomized block 2 2 factorial design, which is randomized because both factors were present in all four sites (blocks in the model), and 2 2 factorial because each factor has two levels. The goal of a factorial experiment is to test whether the factors interact, which can be expressed by the following LM: y il
1 1

2 2

3 1 2

xx

4 3l

il

[1]

where is the average value of the observations; 1, 2, and 4 quantify the impact of each independent variable; 3 the interaction of water quality and nutrient addition; and is the error, which is assumed to be normally distributed. The subscript l denotes block (site within the forest patch), and i is for individual pools within a block (containers in a forest site). Therefore, yil is the total number of rafts from a given pool and block (i.e., container in a site). Table 1 shows the results of the analysis of variance for the data using the model presented in (1). The natural logarithm transformation of ln (y 1) is done to normalize the data and fulll model assumptions. Table 1 shows that neither block nor the interaction between water quality and nutrient addition was signicant (P 0.05). The inuence of water quality and nutrients on Cx. quinquefasciatus oviposition can be reanalyzed using a LMEM. By contrast with the LM analysis, in which inferences are done over blocks, the LMEM assumes that blocks are random samples from a larger pop-

where , the s, and y and have the same interpretation as in equation 1, and quanties the variability across the blocks, which is assumed to be normally distributed. The signicance of factors can be tested using F tests, which work well when designs are balanced (equal number of samples per treatment and block) and parameters can be estimated using maximum likelihood. More generally, signicance may be tested using parametric bootstraps, for balanced or unbalanced designs where parameters are estimated by restricted maximum likelihood (Pinheiro and Bates 2000, Faraway 2006). Table 2 shows the results of an analysis of deviance, with parameters of equation 2 estimated using restricted maximum likelihood and inference based on 1000 replications of parametric bootstrap (an analysis in which datasets are simulated and the results of likelihood ratio tests for studied factors are compared with those of the true data to compute the signicance of factors). In this example, inference about the impact of water quality and nutrient addition is qualitatively similar using LM or LMEM. However, LMEM provides additional insight; it indicates oviposition is nely grained. The variance at the individual container level is larger than at the block level (Table 2: 0.059 and 0.024), a pattern also observed in the original study for the full dataset (Chaves et al. 2009). The data can be observed in Fig. 2. Constrained Designs. One of the major limitations of the LM is that it is not suited to analyze datasets with constraints in randomization. For example, all replicates from a treatment should be present in all blocks. This is a frequent limitation in eld studies in which features of a given landscape cannot be altered, an underlying motivation behind split-plot designs, in which some treatments do not vary across blocks. Split-plots are widely used in agriculture (Faraway 2006) and economic entomology (Blumberg et al. 1997, Haile et al. 2000, Oyediran et al. 2007). However, constraints in randomization can also arise as a product of other trade-offs in experimentation or by the nature of the questions asked. For example, the original design of Chaves et al. (2009) was unbalanced in the sense that each block had an unequal number of replicates for each one of the treatments. However, for each block the amount of total nutrients and water quality was constant; one of the questions was to determine the grain of mosquito perception for oviposition choices. The largest variance was among the individual oviposition containers, indicating a nely grained perception, in contrast to a scenario of coarsely grained mosquito perception, in which the largest variance would be expected for the blocks or sites. To illustrate the analysis of constrained designs, the original data of Chaves et al. (2009) are sampled in such a way that all experimental pools with nutrients added and combined sewage overow water

294

JOURNAL OF MEDICAL ENTOMOLOGY

Vol. 47, no. 3

Fig. 1. Blocks: xed or random? The left panel shows the case for blocks as a xed factor in LM, in which the assumption is that inferences are exclusive for the observed blocks (within squares). The right panel shows the case for blocks as a random factor in LMEMs, in which the observed blocks (within squares) come from a larger population (i.e., blocks inside and outside squares). (Online gure in color.)

came from the same block (Fig. 2), and therefore the design becomes unbalanced (Fig. 3). Thus, the LM from equation 1 cannot be employed to analyze the data, but the LMEM from equation 2 is suitable for such an analysis. Results are presented in Table 3 and are similar to those presented in Table 2, showing a decreased variability in the blocks and a larger error. In summary, LMEMs can uncover the same variance pattern in data under pseudoreplication, a major advantage over LMs.
Table 2. Analysis of deviance for the effects of water quality and nutrient addition on the ln of total number of egg rafts 1 oviposited over 5 d by Culex quinquefasciatus in Atlanta, GA (Chaves et al. 2009) Fixed Water ( 1) Nutrient ( 2) Water nutrient ( 3) Random Blocks ( ) Error ( ) df 1 1 1 Log likelihood 7.241 31.089 4.304 LRT 5.874 53.569 8.612 P 0.009* 0.000* 0.664

Spatial Variability Organisms can be clustered in space, for example, the larvae of mosquitoes can be associated with only certain habitats where eggs are oviposited, thus making their abundance autocorrelated in space (Pitcairn et al. 1994). Several statistical tools can accommodate the lack of spatial independence in data from eld studies (Fortin and Dale 2005), and they have been widely used with insects of medical importance (Koenraadt et al. 2007, 2008; Vazquez-Prokopec et al. 2008). However, their description is outside the scope of this article. LMEM can also be used to consider spatial variability. LMEM are especially suitable for cases when spatial scales are nested. For example, mosquito larval samples coming from containers in several houses that belong to the same neighborhood are hierarchically nested. Several studies have used this approach in recent studies on medically important insects (Harrington et al. 2008, Chaves et al. 2009, Gurtler et al. 2009). The LMEM tting procedure is similar to the one used next to consider the lack of temporal independence in longitudinal studies. Longitudinal Studies The fact that observations are repeated through time in the same place (or from the same organisms) can lead to data that are not independent and are autocorrelated in time. One approach to this problem

Mean square (variance) 0.024 0.059

This design was balanced. Original data and R code for analysis are in the supplementary online material (http://www.envs.emory.edu/ research/Chaves_SOM_Pseudoreplication.html). LRT, likelihood ratio test. Obtained with a parametric bootstrap. *Statistically signicant (P 0.05).

May 2010

CHAVES: DEMYSTIFYING PSEUDOREPLICATION

295

Table 3. Analysis of deviance for the effects of water quality and nutrient addition on the ln of total number of egg rafts 1 oviposited over 5 d by Culex quinquefasciatus in Atlanta, GA (Chaves et al. 2009) Fixed Water ( 1) Nutrient ( 2) Water nutrient ( 3) Random Blocks ( ) Error ( ) df 1 1 1 Log likelihood 10.45 29.03 7.711 LRT 5.477 42.638 0.0409 P 0.012* 0.000* 0.801

Mean square (variance) 5.0405e-14 0.125

This design was unbalanced (because of the sampling from the full dataset). Original data and R code for analysis are in the supplementary online material (http://www.envs.emory.edu/research/Chaves_ SOM_Pseudoreplication.html). LRT, likelihood ratio test. Obtained with a parametric bootstrap. *Statistically signicant (P 0.05).

Fig. 2. Boxplots (median and quartiles) for the natural logarithm number of Cx. quinquefasciatus egg rafts 1: (A) in the balanced (Tables 1 and 2) and (B) unbalanced (Table 3) block designs to study the effects of water quality and nutrient enrichment on oviposition. Tap indicates tap water and CSO indicates combined sewage overow water. Y stands for nutrient addition, and N for no additional nutrients. Data extracted from Chaves et al. (2009). Original data are available in the supplementary online material (http://www.envs.emory.edu/ research/Chaves_SOM_Pseudoreplication.html).

is to use repeated measurements analysis (Faraway 2006) and time series analysis techniques (Shumway and Stoffer 2000, Chaves and Pascual 2007). Time series techniques have been used for longitudinal studies of some vectors (Hayes and Downs 1980, Strickman 1988, Feliciangeli and Rabinovich 1998, Scott et al. 2000, Salomon et al. 2004). Alternatively, LMEM can model the lack of temporal independence as a random factor, which is one of the many methods

for repeated measurements analysis (Faraway 2006). Modeling the lack of temporal independence will be illustrated by examining data from a study on the biting and resting behavior of anophelines using experimental huts in three villages of western Venezuela (Rubio-Palis and Curtis 1992). Mosquitoes were collected during two nights per month and by catching the landing mosquitoes on the legs of two catchers between 1900 and 0700 hours, inside and outside experimental huts. Although several species were found, only data for Anopheles nuneztovari Gabaldon from Guaquitas collected between August 1988 and October 1989 will be analyzed in this study (Fig. 4). In this case, the response or dependent variable (y) is the total number of landings for all huts, as presented in the original study (Rubio-Palis and Curtis 1992). The xed factors are as follows: 1) x1, the site with two levels, inside and outside the hut; 2) x2, the landing time with 12 levels corresponding to the hours between 1900 and 0700 hours; and 3) x3, the rainfall season with two levels: dry (December-May) and wet (JuneNovember). The random factors consider the different scales of temporal variability: 1) l, the year l; 2) kl, the month k within a given year l; 3) jkl, the sampling day j within a given month k and year l; and 4) ijkl, which is the error i (error for an observation belonging to day j within a given month k and year l). All random factors are assumed to be independent and normally distributed. The model equation is as follows: y ijkl
1 1

2 2 l kl

3 3 jkl

4 2 3 ijkl

xx

[3]

Fig. 3. Unbalanced design. Note that one block contains all samples with added nutrients and combined sewage overow water, and other treatments have unequal number across other blocks. (Online gure in color.)

In this model, represents the mean value of all observations; 1, 2, and 3 quantify the impact of each independent variable on the number of landings; and 4 the interaction of season and landing time. Note that choice of factors is dictated by the study objective: quantication of the seasonal nocturnal biting pattern outside and inside the experimental huts. Such quantication requires landing time, site, and season to be treated as xed independent variables. The other temporal variables need to be random factors accounting

296

JOURNAL OF MEDICAL ENTOMOLOGY

Vol. 47, no. 3

Fig. 4. Boxplots (median and quartiles) for the hourly number of Anopheles nuneztovari landings: (A) outside the house, dry season; (B) inside the house, dry season; (C) outside the house, wet season; (D) inside the house, wet season. Data extracted from Rubio-Palis and Curtis (1992). Original data are available in the supplementary online material (http:// www.envs.emory.edu/research/Chaves_SOM_Pseudoreplication.html).

for the lack of independence that arises from the repeated measurements through time. There was no variability because of the year of the observation ( 0; see supplementary online material http://www.envs.emory. edu/research/Chaves_SOM_Pseudoreplication.html), and a simpler model, without a parameter for the annual variability, was t, as follows: y ijk
1 1

effects of site and season are statistically signicant (P 0.05). Daily observations for each month were more homogeneous, having a lower variance than those observations across months. For comparison purposes, a LM was also t, as follows: yi
1 1

2 2

3 3

4 2 3 i

xx

2 2

3 3 jk

4 2 3 ijk

xx

...

[5]

[4]

Results for this model are presented in Table 4. The interaction between season and time and the main
Table 4. Analysis of deviance for the effects of site, landing time, and season on Anopheles nuneztovari abundance in Guaquitas, Venezuela (Rubio-Palis and Curtis 1992) Factor Site ( 1) Landing time ( 2) Season ( 3) Landing time season ( 4) Random Month ( ) Day ( ) Error ( ) df 1 11 1 11 Log likelihood 3550 3619 3548 3542 LRT 15.1 154.2 11.2 124.4 P 0.000* 0.073 0.000* 0.000*

Table 5 shows the results for the analysis with equation 5. All factors are signicant with this model (P 0.05). However, when compared with the model with random effects, the main effect for landing time is not signicant when the lack of independence in the data is properly modeled with a LMEM (Table 4). These analyses illustrate one of the problems of incorrectly
Table 5. Analysis of variance for the effects of site, landing time, and season on Anopheles nuneztovari abundance in Guaquitas, Venezuela (Rubio-Palis and Curtis 1992) Factor df Sum square Mean square F Value Pr( F) 0.009704* 1.55E-08* 2.2e-16* 0.002432*

Mean square (variance) 751.46 182.68 1669.42

1 17,914 17,914 6.727 Site ( 1) 5.5483 Landing time ( 2) 11 162,525 14,775 1 286,446 286,446 107.5666 Season ( 3) 11 77,851 7077 2.6577 Landing time season ( 4) Error ( ) 671 178,6849 2663

Original data and R code for analysis are in the supplementary online material (http://www.envs.emory.edu/research/Chaves_ SOM_Pseudoreplication.html). LRT, likelihood ratio test. Obtained with a parametric bootstrap. *Statistically signicant (P 0.05).

Original data and R code for analysis are in the supplementary online material (http://www.envs.emory.edu/research/Chaves_ SOM_Pseudoreplication.html). *Statistically signicant (P 0.05).

May 2010

CHAVES: DEMYSTIFYING PSEUDOREPLICATION References Cited

297

modeling the lack of independence across observations: the LM rejects a null hypothesis that is true (type II error) by saying that landing time by itself is signicant (Table 5), when in reality it is only significant when considered in conjunction with the season (Table 4). Pseudoreplication: an Issue of the Past As shown in this forum, pseudoreplication no longer is an issue preventing the statistical analysis of experiments and eld studies. Current statistical tools such as LMEM can model the lack of independence in eld observations. However, pseudoreplication will most likely always be present in any ecological study, because of the complexity of working with living organisms that constrains full randomization or limits the number of replicates. Although other objects of study, like molecules or atoms, are numerous and widespread, samples of living organisms are comparatively few and organisms always are evolutionary and ecologically related at some scale. Although this forum has been focused on demystifying statistical concepts and presents how to use LMEM models to address the lack of independence in datasets, the ingenuity of statisticians is laudable because many other techniques outside the scope of this article have been developed over recent years. A best example includes the extension of LMEM to accommodate non-normal observations in generalized LMEMs (Bolker et al. 2009). Other tools that do not consider the individual variability of observations, but rather the average across all samples, like the generalized estimating equations (Faraway 2006), can address the lack of independence in observations, and have been used in the study of medically important insects (Lindblade et al. 2000, Gurevitz et al. 2009). A third line of new computer-based tools, including neural networks, trees (Olden et al. 2008), and random forests (Ruiz et al. 2010), does not have assumptions on data independence, and has been successfully used to study insects of public health importance (Hu et al. 2006, Ruiz et al. 2010). Thus, pseudoreplication should no longer be considered as a major aw that impairs the statistical analysis of experiments and eld studies. Independence constraints in the manipulation and observation of organisms are adequately handled by many available statistical tools, thus enabling valid inferences from valuable entomological data. Acknowledgments
I am thankful to Yasmin Rubio-Palis for sharing her original data on Anopheles nuneztovari from Guaquitas, Venezuela. This work was funded by a Gorgas Research Award from the American Society of Tropical Medicine and Hygiene and Emory University. This work also beneted from comments by the editor, anonymous reviewers, Jorge Rabinovich, Nicole Gottdenker, and Greg Decker, and helpful discussions from a National Institutes of Health-Research and Policy on Infectious Disease Dynamics (NIH-RAPIDD) study group on mosquito-borne diseases.

Alto, B. W., and S. A. Juliano. 2001a. Precipitation and temperature effects on populations of Aedes albopictus (Diptera: Culicidae): implications for range expansion. J. Med. Entomol. 38: 646 656. Alto, B. W., and S. A. Juliano. 2001b. Temperature effects on the dynamics of Aedes albopictus (Diptera: Culicidae) populations in the laboratory. J. Med. Entomol. 38: 548 556. Blumberg, A.J.Y., P. F. Hendrix, and D. A. Crossley. 1997. Effects of nitrogen source on arthropod biomass in notillage and conventional tillage grain sorghum agroecosystems. Environ. Entomol. 26: 3137. Bolker, B. M., M. E. Brooks, C. J. Clark, S. W. Geange, J. R. Poulsen, M. H. Stevens, and J. S. White. 2009. Generalized linear mixed models: a practical guide for ecology and evolution. Trends Ecol. Evol. 24: 127135. Box, J. F. 1980. R. A. Fisher and the design of experiments, 19221926. Am. Stat. 34: 17. Carpenter, S. R. 1982. Stemow chemistry: effects on population dynamics of detritivorous mosquitoes in tree-hole ecosystems. Oecologia 53: 1 6. Chaves, L. F., and M. Pascual. 2007. Comparing models for early warning systems of neglected tropical diseases. PLoS Negl. Trop. Dis. 1: e33. Chaves, L. F., C. L. Keogh, G. M. Vazquez-Prokopec, and U. D. Kitron. 2009. Combined sewage overow enhances oviposition of Culex quinquefasciatus (Diptera: Culicidae) in urban areas. J. Med. Entomol. 46: 220 226. Chesson, J. 1984. Effect of notonectids (Hemiptera, Notonectidae) on mosquitos (Diptera, Culicidae): predation or selective oviposition. Environ. Entomol. 13: 531 538. Conner, J. K., and D. L. Hartl. 2004. A primer of ecological genetics. Sinauer, Sunderland, MA. Faraway, J. J. 2006. Extending the linear model with R: generalized linear, mixed effects and nonparametric regression models. CRC, Boca Raton, FL. Feliciangeli, M. D., and J. Rabinovich. 1998. Abundance of Lutzomyia ovallesi but not Lu-gomezi (Diptera: Psychodidae) correlated with cutaneous leishmaniasis incidence in north-central Venezuela. Med. Vet. Entomol. 12: 121 131. Fisher, R. A. 1935. The design of experiments. Oliver & Boyd, Edinburgh, United Kingdom. Fortin, M. J., and M.R.T. Dale. 2005. Spatial analysis: a guide for ecologists. Cambridge University Press, Cambridge, United Kingdom. Gurevitz, J. M., U. Kitron, and R. E. Gurtler. 2009. Temporal dynamics of ight muscle development in Triatoma infestans (Hemiptera: Reduviidae). J. Med. Entomol. 46: 10211024. Gurtler, R. E., F. M. Garelli, and H. D. Coto. 2009. Effects of a ve-year citywide intervention program to control Aedes aegypti and prevent dengue outbreaks in northern Argentina. PLoS Negl. Trop. Dis. 3: e427. Haile, F. J., D. L. Kerns, J. M. Richardson, and L. G. Higley. 2000. Impact of insecticides and surfactant on lettuce physiology and yield. J. Econ. Entomol. 93: 788 794. Harrington, L. C., A. Ponlawat, J. D. Edman, T. W. Scott, and F. Vermeylen. 2008. Inuence of container size, location, and time of day on oviposition patterns of the dengue vector, Aedes aegypti, in Thailand. Vector Borne Zoonotic Dis. 8: 415 423. Hayes, J., and T. D. Downs. 1980. Seasonal changes in an isolated population of Culex pipiens quinquefasciatus (Dipter: Culicidae): a time series analysis. J. Med. Entomol. 17: 63 69.

298

JOURNAL OF MEDICAL ENTOMOLOGY

Vol. 47, no. 3

Heffner, R. A., M. J. Butler, and C. K. Reilly. 1996. Pseudoreplication revisited. Ecology 77: 2558 2562. Hu, W., S. Tong, K. Mengersen, B. Oldenburg, and P. Dale. 2006. Mosquito species (Diptera: Culicidae) and the transmission of Ross River virus in Brisbane, Australia. J. Med. Entomol. 43: 375381. Hurlbert, S. H. 1984. Pseudoreplication and the design of ecological eld experiments. Ecol. Monogr. 54: 187211. Kirby, M., P. Milligan, D. Conway, and S. Lindsay. 2008. Study protocol for a three-armed randomized controlled trial to assess whether house screening can reduce exposure to malaria vectors and reduce malaria transmission in The Gambia. Trials 9: 33. Kirby, M. J., D. Ameh, C. Bottomley, C. Green, M. Jawara, P. J. Milligan, P. C. Snell, D. J. Conway, and S. W. Lindsay. 2009. Effect of two different house screening interventions on exposure to malaria vectors and on anemia in children in The Gambia: a randomized controlled trial. Lancet 374: 998 1009. Koenraadt, C.J.M., J. Aldstadt, U. Kijchalao, A. Kengluecha, J. W. Jones, and T. W. Scott. 2007. Spatial and temporal patterns in the recovery of Aedes aegypti (Diptera: Culicidae) populations after insecticide treatment. J. Med. Entomol. 44: 6571. Koenraadt, C.J.M., J. Aldstadt, U. Kijchalao, R. Sithiprasasna, A. Getis, J. W. Jones, and T. W. Scott. 2008. Spatial and temporal patterns in pupal and adult production of the dengue vector Aedes aegypti in Kamphaeng Phet, Thailand. Am. J. Trop. Med. Hyg. 79: 230 238. Lindblade, K. A., E. D. Walker, A. W. Onapa, J. Katungu, and M. L. Wilson. 2000. Land use change alters malaria transmission parameters by modifying temperature in a highland area of Uganda. Trop. Med. Int. Health 5: 263 274. Millar, R. B., and M. J. Anderson. 2004. Remedies for pseudoreplication. Fisheries Res. 70: 397 407. Montgomery, D. C. 2005. Design and analysis of experiments. Wiley, New York, NY. Oksanen, L. 2001. Logic of experiments in ecology: is pseudoreplication a pseudoissue? Oikos 94: 2738. Olden, J. D., J. J. Lawler, and N. LeRoy-Poff. 2008. Machine learning methods without tears: a primer for ecologists. Q. Rev. Biol. 83: 171193. Oyediran, I. O., M. L. Higdon, T. L. Clark, and B. E. Hibbard. 2007. Interactions of alternate hosts, postemergence

grass control, and rootworm-resistant transgenic corn on western corn rootworm (Coleoptera: Chrysomelidae) damage and adult emergence. J. Econ. Entomol. 100: 557565. Pinheiro, J. C., and D. M. Bates. 2000. Mixed effects models in S and S-plus. Springer, New York, NY. Pitcairn, M. J., L. T. Wilson, R. K. Washino, and E. Rejmankova. 1994. Spatial patterns of Anopheles freeborni and Culex tarsalis (Diptera: Culicidae) larvae in California rice elds. J. Med. Entomol. 31: 54553. Reiskind, M. H., and M. L. Wilson. 2004. Culex restuans (Diptera: Culicidae) oviposition behavior determined by larval habitat quality and quantity in southeastern Michigan. J. Med. Entomol. 41: 179 86. Rubio-Palis, Y., and C. F. Curtis. 1992. Biting and resting behavior of anophelines in western Venezuela and implications for control of malaria transmission. Med. Vet. Entomol. 6: 325334. Ruiz, M. O., L. F. Chaves, G. L. Hamer, T. Sun, W. M. Brown, E. D. Walker, L. Haramis, T. L. Goldberg, and U. D. Kitron. 2010. Local impact of temperature and precipitation on West Nile virus in Culex species mosquitoes in northeast Illinois, U.S.A. Parasit. Vectors 3: 19. Salomon, O. D., M. L. Wilson, L. E. Munstermann, and B. L. Travi. 2004. Spatial and temporal patterns of phlebotomine sand ies (Diptera: Psychodidae) in a cutaneous leishmaniasis focus in northern Argentina. J. Med. Entomol. 41: 3339. Scott, T. W., A. C. Morrison, L. H. Lorenz, G. G. Clark, D. Strickman, P. Kittayapong, H. Zhou, and J. D. Edman. 2000. Longitudinal studies of Aedes aegypti (Diptera: Culicidae) in Thailand and Puerto Rico: population dynamics. J. Med. Entomol. 37: 77 88. Shumway, R. H., and D. S. Stoffer. 2000. Time series analysis and its applications. Springer, New York, NY. Strickman, D. 1988. Rate of oviposition by Culex quinquefasciatus in San Antonio, Texas, during three years. J. Am. Mosq. Control Assoc. 4: 339 344. Vazquez-Prokopec, G. M., M. C. Cecere, U. Kitron, and R. E. Gurtler. 2008. Environmental and demographic factors determining the spatial distribution of Triatoma guasayana in peridomestic and semi-sylvatic habitats of rural northwestern Argentina. Med. Vet. Entomol. 22: 273282. Received 12 October 2009; accepted 20 January 2010.