Ecology pp by the Ecological Society of America

Description

Niveau: Supérieur, Doctorat, Bac+8
Ecology, 87(10), 2006, pp. 2614–2625 2006 by the Ecological Society of America VARIATION PARTITIONING OF SPECIES DATA MATRICES: ESTIMATION AND COMPARISON OF FRACTIONS PEDRO R. PERES-NETO,1 PIERRE LEGENDRE, STEPHANE DRAY, AND DANIEL BORCARD Departement des sciences biologiques, Universite de Montreal, C.P. 6128, succursale Centreville, Montreal, Quebec H3C3J7 Canada Abstract. Establishing relationships between species distributions and environmental characteristics is a major goal in the search for forces driving species distributions. Canonical ordinations such as redundancy analysis and canonical correspondence analysis are invaluable tools for modeling communities through environmental predictors. They provide the means for conducting direct explanatory analysis in which the association among species can be studied according to their common and unique relationships with the environmental variables and other sets of predictors of interest, such as spatial variables. Variation partitioning can then be used to test and determine the likelihood of these sets of predictors in explaining patterns in community structure. Although variation partitioning in canonical analysis is routinely used in ecological analysis, no effort has been reported in the literature to consider appropriate estimators so that comparisons between fractions or, eventually, between different canonical models are meaningful. In this paper, we show that variation partitioning as currently applied in canonical analysis is biased. We present appropriate unbiased estimators. In addition, we outline a statistical test to compare fractions in canonical analysis.

• predictors


• canonical analysis


• response variable


• rda


• adjusted sample


• variables added


• canonical analysis involving


• multiple regression


• species

Sujets

• Species
• Centreville
• Analysis
• Canonical analysis
• Dependent and independent variables
• Linear regression

Informations

• Publié par: mijec
• Nombre de lectures: 27
• Langue: English

Extrait

Ecology , 87(10), 2006, pp. 2614–2625  2006 by the Ecological Society of America

VARIATION PARTITIONING OF SPECIES DATA MATRICES: ESTIMATION AND COMPARISON OF FRACTIONS P EDRO R. P ERES -N ETO , 1 P IERRE L EGENDRE , S TE ´ PHANE D RAY , AND D ANIEL B ORCARD De´partement des sciences biologiques, Universite´ de Montre´al, C.P. 6128, succursale Centreville, Montre´al, Que´bec H3C 3J7 Canada Abstract. Establishing relationships between species distributions and environmental characteristics is a major goal in the search for forces driving species distributions. Canonical ordinations such as redundancy analysis and canonical correspondence analysis are invaluable tools for modeling communities through environmental predictors. They provide the means for conducting direct explanatory analysis in which the association among species can be studied according to their common and unique relationships with the environmental variables and other sets of predictors of interest, such as spatial variables. Variation partitioning can then be used to test and determine the likelihood of these sets of predictors in explaining patterns in community structure. Although variation partitioning in canonical analysis is routinely used in ecological analysis, no effort has been reported in the literature to consider appropriate estimators so that comparisons between fractions or, eventually, between different canonical models are meaningful. In this paper, we show that variation partitioning as currently applied in canonical analysis is biased. We present appropriate unbiased estimators. In addition, we outline a statistical test to compare fractions in canonical analysis. The question addressed by the test is whether two fractions of variation are significantly different from each other. Such assessment provides an important step toward attaining an understanding of the factors patterning community structure. The test is shown to have correct Type I error rates and good power for both redundancy analysis and canonical correspondence analysis. Key words: adjusted coefficient of determination; bootstrap; canonical analysis; canonical correspond- ence analysis (CCA); ecological community; redundancy analysis (RDA); variation partitioning.

I NTRODUCTION mating habitat suitability, forecasting the effects of The search for causes dictating patterns in species  habitat change due to human interference, establishing distributionsinnaturalanddisturbedlandscapesisof d p i o c t t e in nt g ia h l o l w oc c a o ti m o m ns un f i o t r ys s t p r e u c c i t e u s re re m in a t y ro b d e u a c f t f i e o c n t , ed or by p t r h e e - primary importance in ecological science, and establish- ingrelationshipsbetweenspeciesdistributionsand i in m v p asi t o a n nt o f f or ex h o e t u ic rist s i p c ec is i s es u . es T s h u e ch se a c s o d n e d ter q m u i e n s i ti n o g n th is e environmentalcharacteristicsisawidelyusedapproach likel o i r hoodofcompetinghypothesestoexplainpartiul (e.g., Legendre and Fortin 1989, Jackson and Harvey  c ar 1993, Diniz-Filho and Bini 1996, Rodrı´  guez and Lewis  patterns in community structure (Peres-Neto et al. 1997,JenkinsandBuikema1998,BoyceandMcDonald 200 C 1 a ). nonicalanalysessuchasredundancyanalysis 1999, Peres-Neto and Jackson 2001). Habitat models  (RDA; Rao 1964), canonical correspondence analysis relating habitat characteristics and community structure (species occurrence or abundance) are expected to  (CCA; ter Braak 1986), and distance-based redundancy answeratleasttwoquestions.(1)Howwellisthe analysis(db-RDA;Legendre in andAnderson1999)are distributionofasetofspeciesexplainedbythegivenset i e n n v vi a r l o ua n b m le en t ta o l olsformodel he gc r o ov m id m e un t i h t e ies me t a h n r s ou f g o h r ofpredictivevariables?(2)Whichvariablesareirrele- conductingdire p c r t ed e i x ct p o la rs n . at T ory y a p nalysesinwhichthe vant or redundant in the sense of failing to strengthen  sociatio theexplanationofpatternsaftercertainothervariables t a h s eircom n m a o m n o a n n g d sp u e n c i i q es ue ca r n el b a e ti s o t n u s d h i i e p d s w w it it h h re e s n p v e i c r t o t n o - have been taken into account? The first question relates  redictors of to the predictive power of the model that can be used in  mental variables or any other set of p  ver conservationmanagement,forquestionssuchasesti- i 1 n 5 t 0 e 0 res s t t . ud A i s es a a d p e p m ly o i n n s g tra C ti C on A o o f r its R s D uc A ce i s n s,w m e o ll de o ling species–environment relationships have been published Manuscript received 26 August 2005; revised 9 February (see also Birks et al. [1996] for reviews on ecological 2006; accepted 16 February 2006; final version received 21 studies using these methods). RDA and CCA can be March 2006. Corresponding Editor: N. G. Yoccoz. 1 Present address: Department of Biology, University of r b e e g s r t es u si n o d n er t s h to at od ha a s s a m s e i t n h g o le ds re f s o p r on extending d m m u u l l t t i i p p l l e e Regina, Saskatchewan S4S 0A2 Canada. se y an E-mail: pedro.peres [e-mail] predictors X (e.g., several environmental predictors), to 2614

October 2006

VARIATION PARTITIONING

2615

F IG . 1. Variation partitioning scheme of a response variable Y between two sets of predictors X (e.g., environmental factors) and W (e.g., spatial predictors). The total variation in Y is partitioned into four fractions as follows: (1) fraction [a þ b þ c] based [ o b n þ b c o ] th ba s s et e s d o o f n pr m ed a i t c r t ix or W ma ([ tr b ic þ es c] [ X ¼ , R W 2 Y ] j ( W [ ) a ;( þ 4 c ) b ] t ¼ þ h [ e c a ] u þ ¼ ni b R q 2 Y u þ j½ e c X ] ; f W r – a  ) [ c ; a ti ( o þ 2) n b f ] o r ; f a ( c 6 v t ) i a o r t i h n a e t [ i a c o o n þ m e b x m ] p o b l n a a s i f e n r d e a d c o ti b n o y n m X o at , f r [ i v a x a ] ri X ¼ at ( [ i [ a o a n þ þ s b b h ] a þ r ¼ e c] d R 2 Y – b j X [b ); X þ (3 a c ) n ]; f d r ( a 5 W c ) t , i t [ o h b n e unique fraction of variation explained by W , [ y ] ¼ [a þ b þ c] – [a] – [c]; and (7) the residual fraction of variation not explained by X and W , [d] ¼ 1 – [a þ b þ c].

multiple regression involving multiple response variables Although canonical analysis and variation partition- Y (e.g., several species) and a common matrix of  ing may provide a robust approach for understanding predictors X . It follows that the percentage of variation  the relative influence of different ecological factors of the response matrix explained by the predictor matrix  driving community assembly, judging the importance (hereafter referred to as the redundancy statistic, or  of a factor solely on the basis of its proportional unique simply R 2 Y j X followingMillerandFarr1971)isthe co r n f t o ri r b m u e ti d o . n Th is es n t o a t tist a i s cal str b a i i a g s h r tf el o a r t w ed ar t d o a e s sti c m u a r t r i e n n g tly a canonical equivalent of the regression coefficient of  pe  on 2 based on a sample R 2 is a well-recogniz determination, R 2 . populati q ed Inmultipleregressionanalysis,wecanapplyvariation problem(Zar1999),as n s q a 2 m T pl h e eb estimates u t e e n n c d e , d o b n y partitioning (also known as commonality analysis;  average, to be larger tha . ias is infl Kerlinger and Pedhazur 1973) to identify common and  both the number of independent variables in the model uniquecontributionstomodelpredictionandhence a a s nd ad s j a u m st p m le e s n i t ze an ( d Kr ‘‘ o s m hr r i e n y k a a n ge d ’’ H r i e n f e e s r 1 to 99 t 5 h ). e T fa e c r t m t s h s a u t c a h better address the question of the relative influences of thegroupsofindependentvariablesconsideredinthe s p a ro m v p i l d e e -e a sti m m o a r t e ed acc R u 2 ra n t e e e e d s s ti t m o at b e e o r f ed q u 2 . ce B d y i t n ak o i rde i r nt t o o model(Mood1969).Whenpartitioningvariationisused accounttheappropriatedegreesoffreedom,thea n d g j inregressionanalysis,independentvariablesare mentprovidesawayofcomparingmodelswithdiffer u e s n t- t grouped into sets representing broad factors. In that context, variation partitioning is more suitable than  numbers of predictors (e.g., model selection) and sample analyzing the individual contributions of regressors via  sizes. Given that R 2 and R 2 Y j X are intrinsically related, the theirpartialcorrelationcoefficients.Inthisapproach, c b a ia n s oni o c b a s l er a v n e a d lys i i n s. m Al u t l h ti o p u le gh re v g a r r e i s a s t i i o o n np a a ls rt o iti e o x n i i s n ts gi i n n the total percentage of variation explained by the model  canonical analysis is routinely used in ecological ( R 2 ) is partitioned into unique and common contribu- tionsofthesetsofpredictors(Fig.1).Forexample, t a o na c l o ys n i s s i , de n r o a e p f p f r o o rt pri h a a t s ee b s e t e i n ma r t e o p r o s rt s e o d th in at t c h o e m li p t a e r r i a s t o u n r s e variation partitioning for RDA or CCA using two sets between fractions or eventually between different canon- of predictors ( X and W ) is straightforward as it is based on three canonical analysis (Fig. 1). The first one uses  ical models are meaningful. Specifically, our objective is both sets of predictors [ X , W ], the second only X ,andthe t sa wo m f p o l l e d: R ( 2 Y 1 j ) X , to and pro (2 vi ) de to ad o j u u t s li t n m e en a ts st f a o t r ist t ic h a e l b t i e a s s t i t n o last one only W . All remaining fractions of the  contrast partial effects in c anonical analysis (i.e., partitioning can be obtained by simple subtractions  compare fractions of variation). (Fig. 1). Note that the shared variation ([b], Fig. 1) may be negative due to suppressor variables (i.e., a regressor  R EDUNDANCY S TATISTIC IN C ANONICAL A NALYSIS having low, close to zero, correlation with the response variableandacorrelationwithanotherregressor,which use H d er i e n w c e a p n r o e n s i e c n a t l th a e na f l o y r si m s ul a at p i p o l n ie o d f t th o e s R p 2 Y e j c X ie s s tat d i a st t i a c i A n ze t n ur a n nd is B c u o d r e r s e c la u t [ e 2 d 00 w 3 i ] t f h or th m e or r e es d p e o t n ai s l e s) v o a r ri d a u b e le t s o ; t s w e o e matrices.InthecaseofRDA, R 2 Y j X is calculated as follows: ^ 2 trace s o L t p e r p g o e o n n s g i d l t y r e e c s o a i r n g r n d e s l L at ( e e o g d e n n e p d r r e p e d o i s 1 c i t 9 ti o 9 v r 8 e s : w S a i e n t c h d ti s o t tr n h o e 1 n 0 g o .3 t e . h f 5 f e ) e r . ct V n s a e o r g i n a a t t i y i v o o e n ; f R Y j X ¼ trace ð Y ð c 0 e Y nt 0 Y ^ Y c Þ ent Þ partitioning based on two sets of predictor matrices was ¼ 1  trace ½ð Y cent  Y ^ Þ 0 ð Y cent  Y ^ Þ i a n n tr d o B d o u r c c e a d r t d o a c n a d no L n e i g c e a n l d a r n e a ( l 1 ys 9 i 9 s 4) b , y la B t o er rc w ar a d se et xt a e l. n ( d 1 e 9 d 9 t 2 o ) trace ð Y 0 cent Y cent Þ ð 1 Þ three or more sets of predictor matrices (Anderson and  where Y ^ ¼ X ( X 0 X )  1 X 0 Y cent represents the matrix of Gribble 1998, Cushman and McGarigal 2002, Økland  predicted values. Note that this is identical to calculating 2003), and is now routinely used in direct gradient  predicted values for individual multiple regressions of analysis. each column of Y on X ; Y cent ¼ ( I – P ) Y is matrix Y

2616

PEDRO R. PERES-NETO ET AL. Ecology, Vol. 87, No. 10

centered by column means (i.e., column means ¼ 0). I is an ( n 3 n ) identity matrix and P is a ( n 3 n ) matrix with all elements ¼ 1/ n ; n refers to the number of sampling units. Matrix X can be either centered or standardized (column means ¼ 0 and column variances ¼ 1). The definition of R 2 Y j X presented here is the one used in ecological applications; it is called the RDA trace statistic in the Canoco program, Version 4.5 (ter Braak and Smilauer 2002) and the proportion of explained variation in Legendre and Legendre (1998). This definition is different from the one in redundancy analysis as used in behavioral research (Dawson-Sa- unders 1982, Lambert et al. 1988) where the response variables are standardized rather than centered prior to analysis. In that case, the R 2 Y j X is simply the mean of the R 2 statistics computed for each individual multiple regression of each column of Y on matrix X (Miller 1975). In ecological analysis the species are centered and not standardized, so R 2 Y j X is a weighted mean of the R 2 of individual models with weights proportional to the species variances divided by the total variance. The same definition based on a weighted mean applies to CCA, and for the sake of brevity we present the R 2 Y j X used in CCA in Appendix A. A N A DJUSTED R EDUNDANCY S TATISTIC FOR C ANONICAL A NALYSIS —T HE C ONTINUOUS C ASE Our first task was to determine whether adjustments for the multiple coefficient of determination ( R 2 adj ) developed for a single response variable could also be applied to the canonical R 2 Y j X . Dawson-Saunders (1982) has shown that Ezekiel’s adjustment (1930), commonly used in the case of multiple regressions (Legendre and Legendre 1998, Zar 1999), is appropriate for the case where response variables are standardized prior to analysis. Ezekiel’s formulation applied to canonical analysis based on centered values is as follows: R 2 ð Y j X Þ adj ¼ 1  n  n p   1 1 ð 1  R 2 Y j X Þ ^ ^ 1 trace ½ð Y cent  Y Þð Y cent  Y Þ = ð n  p  1 Þ ¼  trace ð Y 0 cent Y cent Þ = ð n  1 Þ ð 2 Þ where n is the sample size, p is the number of predictors, and R 2 Y j X is the sample estimation of the q 2 Y j X . Since fractions of variation represent redundancy statistics, they also need to be adjusted. Fractions [a þ b þ c], [b þ c], and [a þ b] can be adjusted directly, leading to [a þ b þ c] adj , [b þ c] adj , and [a þ b] adj . The individual fractions [a] adj , [b] adj , [c] adj , and [d] adj have to be calculated by appropriate subtractions based on [a þ b þ c] adj , [b þ c] adj , and [a þ b] adj . We conducted a Monte Carlo study equivalent to the one used by Kromrey and Hines (1995) who assessed the accuracy of different methods for adjusting sample R 2 in the univariate multiple regression case. The first step was to generate large population matrices (200 000 individ-

uals) with known q 2 Y j X and then draw a large number of samples with replacement from these populations and calculate R 2 Y j X and R ð 2 Y j X Þ adj for each sample. We decided to use large generated populations instead of standard protocols such as generating samples using established correlation matrices (see Peres-Neto et al. 2003 for an example) or by defining the q 2 as in the method introduced by Cramer (1987). The reason is that these previously used methods are capable of generating population values only for continuous variables; hence we cannot generate species-like data (e.g., abundance) where some sites are occupied (values . 0) and others are not (values ¼ 0). We started with a real data set comprised of stream fish communities of a watershed in eastern Brazil (Peres- Neto 2004). A total of 27 species and six environmental variables were considered. The first step was to calculate individual slopes between the species and the environ- mental variables (slopes are presented in Appendix B: Table B1). Slopes were calculated on the centered species and environmental matrices and used as the basis for our simulation study. The next step was to generate a matrix X containing six random normally distributed variables N (0,1) with 200 000 observations (rows). The columns of X were then standardized (i.e., mean ¼ 0 and variance ¼ 1). Then, a data matrix Y was generated as: Y ¼ XB mlt þ E , where B (Appendix B: Table B1) is a (6 3 27) matrix containing the slopes for each species on each environmental variable; mlt is a multiplication factor used to reduce the slopes so that we can manipulate them to attain the desirable R 2 Y j X values. The multiplication factor will be given for each simulated population. E represents a (200 000 3 27) matrix containing N (0,1) deviates. The last step was to calculate the q 2 Y j X based on the generated X (200 000 3 6) and Y (200 000 3 27) matrices. Since all slopes were different from zero, all predictors were active in the sense that they all contributed to the explanation of matrix Y . Assessing the accuracy of R 2 ð Y j X Þ adj using a single set of predictors (canonical analysis) The first set of simulations considered the simplest case of matrices X and Y made of continuous data; abundance-like data will be considered later. In the two sets of simulations, we considered the influence of random predictors by manipulating the number of random N (0,1) variables added to the set of true predictors X , as well as the sample size n . Two populations with q 2 Y j X ¼ 0.2007 (mlt ¼ 0.0004) and q 2 Y j X ¼ 0.6105 (mlt ¼ 0.001) were considered. In the first set of simulations, 1000 samples of 100 observations each were randomly drawn from the population [ Y , X ] and a certain number of random N (0,1) variables were added to the sample X . In the second experiment, 1000 samples with varying numbers of observations were randomly drawn and no random predictors were added to the model. Fig. 2 presents the results of the two