Partition the Variation of Community Matrix by 2, 3, or 4 Explanatory Matrices

Description

The function partitions the variation in community data or community dissimilarities with respect to two, three, or four explanatory tables, using adjusted R-squared in redundancy analysis ordination (RDA) or distance-based redundancy analysis. If response is a single vector, partitioning is by partial regression. Collinear variables in the explanatory tables do NOT have to be removed prior to partitioning.

Usage

varpart(Y, X, ..., data, transfo, scale = FALSE, add = FALSE,
    sqrt.dist = FALSE)
showvarparts(parts, labels, bg = NULL, alpha = 63, Xnames,
    id.size = 1.2,  ...)
## S3 method for class 'varpart234'
plot(x, cutoff = 0, digits = 1, ...)

Arguments

Details

The functions partition the variation in Y into components accounted for by two to four explanatory tables and their combined effects. If Y is a multicolumn data frame or matrix, the partitioning is based on redundancy analysis (RDA, see rda), and if Y is a single variable, the partitioning is based on linear regression. If Y are dissimilarities, the decomposition is based on distance-based redundancy analysis (db-RDA, see capscale) following McArdle & Anderson (2001). The input dissimilarities must be compatible to the results of dist. Vegan functions vegdist, designdist, raupcrick and betadiver produce such objects, as do many other dissimilarity functions in R packages. However, symmetric square matrices are not recognized as dissimilarities but must be transformed with as.dist. Partitioning will be made to squared dissimilarities analogously to using variance with rectangular data – unless sqrt.dist = TRUE was specified.

The function primarily uses adjusted R-squared to assess the partitions explained by the explanatory tables and their combinations, because this is the only unbiased method (Peres-Neto et al., 2006). The raw R-squared for basic fractions are also displayed, but these are biased estimates of variation explained by the explanatory table.

The identifiable fractions are designated by lower case alphabets. The meaning of the symbols can be found in the separate document (use browseVignettes("vegan")), or can be displayed graphically using function showvarparts.

A fraction is testable if it can be directly expressed as an RDA or db-RDA model. In these cases the printed output also displays the corresponding RDA model using notation where explanatory tables after | are conditions (partialled out; see rda for details). Although single fractions can be testable, this does not mean that all fractions simultaneously can be tested, since the number of testable fractions is higher than the number of estimated models.

An abridged explanation of the alphabetic symbols for the individual fractions follows, but computational details should be checked in the vignette (readable with browseVignettes("vegan")) or in the source code.

With two explanatory tables, the fractions explained uniquely by each of the two tables are [a] and [c], and their joint effect is [b] following Borcard et al. (1992).

With three explanatory tables, the fractions explained uniquely by each of the three tables are [a] to [c], joint fractions between two tables are [d] to [f], and the joint fraction between all three tables is [g].

With four explanatory tables, the fractions explained uniquely by each of the four tables are [a] to [d], joint fractions between two tables are [e] to [j], joint fractions between three variables are [k] to [n], and the joint fraction between all four tables is [o].

There is a plot function that displays the Venn diagram and labels each intersection (individual fraction) with the adjusted R squared if this is higher than cutoff. A helper function showvarpart displays the fraction labels. The circles and ellipses are labelled by short default names or by names defined by the user in argument Xnames. Longer explanatory file names can be written on the varpart output plot as follows: use option Xnames=NA, then add new names using the text function. A bit of fiddling with coordinates (see locator) and character size should allow users to place names of reasonably short lengths on the varpart plot.

Value

Function varpart returns an object of class "varpart" with items scale and transfo (can be missing) which hold information on standardizations, tables which contains names of explanatory tables, and call with the function call. The function varpart calls function varpart2, varpart3 or varpart4 which return an object of class "varpart234" and saves its result in the item part. The items in this object are:

Fraction Data Frames

Items fract, indfract, contr1 and contr2 are all data frames with items:

• Df: Degrees of freedom of numerator of the F-statistic for the fraction.

• R.square: Raw R-squared. This is calculated only for fract and this is NA in other items.

• Adj.R.square: Adjusted R-squared.

• Testable: If the fraction can be expressed as a (partial) RDA model, it is directly Testable, and this field is TRUE. In that case the fraction label also gives the specification of the testable RDA model.

Note

You can use command browseVignettes("vegan") to display document which presents Venn diagrams showing the fraction names in partitioning the variation of Y with respect to 2, 3, and 4 tables of explanatory variables, as well as the equations used in variation partitioning.

The functions frequently give negative estimates of variation. Adjusted R-squared can be negative for any fraction; unadjusted R-squared of testable fractions of variances will be non-negative. Non-testable fractions cannot be found directly, but by subtracting different models, and these subtraction results can be negative. The fractions are orthogonal, or linearly independent, but more complicated or nonlinear dependencies can cause negative non-testable fractions. Any fraction can be negative for non-Euclidean dissimilarities because the underlying db-RDA model can yield negative eigenvalues (see capscale, dbrda). These negative eigenvalues in the underlying analysis can be avoided with arguments sqrt.dist and add which have a similar effect as in capscale: the square roots of several dissimilarities do not have negative eigenvalues, and no negative eigenvalues are produced after Lingoes or Cailliez adjustment, which in effect add random variation to the dissimilarities.

The current function will only use RDA in multivariate partitioning. It is much more complicated to estimate the adjusted R-squares for CCA, and unbiased analysis of CCA is not currently implemented.

A simplified, fast version of RDA or dbRDA are used (functions simpleRDA2 and simpleDBRDA). The actual calculations are done in functions varpart2 to varpart4, but these are not intended to be called directly by the user.

Author(s)

Pierre Legendre, Departement de Sciences Biologiques, Universite de Montreal, Canada. Further developed by Jari Oksanen.

References

(a) References on variation partitioning

Borcard, D., P. Legendre & P. Drapeau. 1992. Partialling out the spatial component of ecological variation. Ecology 73: 1045–1055.

Legendre, P. & L. Legendre. 2012. Numerical ecology, 3rd English edition. Elsevier Science BV, Amsterdam.

(b) Reference on transformations for species data

Legendre, P. and E. D. Gallagher. 2001. Ecologically meaningful transformations for ordination of species data. Oecologia 129: 271–280.

(c) Reference on adjustment of the bimultivariate redundancy statistic

Peres-Neto, P., P. Legendre, S. Dray and D. Borcard. 2006. Variation partitioning of species data matrices: estimation and comparison of fractions. Ecology 87: 2614–2625.

(d) References on partitioning of dissimilarities

Legendre, P. & Anderson, M. J. (1999). Distance-based redundancy analysis: testing multispecies responses in multifactorial ecological experiments. Ecological Monographs 69, 1–24.

McArdle, B.H. & Anderson, M.J. (2001). Fitting multivariate models to community data: a comment on distance-based redundancy analysis. Ecology 82, 290-297.

See Also

For analysing testable fractions, see rda and anova.cca. For data transformation, see decostand. Function inertcomp gives (unadjusted) components of variation for each species or site separately. Function rda displays unadjusted components in its output, but RsquareAdj will give adjusted R-squared that are similar to the current function also for partial models.

Examples

data(mite)
data(mite.env)
data(mite.pcnm)

# Two explanatory matrices -- Hellinger-transform Y
# Formula shortcut "~ ." means: use all variables in 'data'.
mod <- varpart(mite, ~ ., mite.pcnm, data=mite.env, transfo="hel")
mod

## Use fill colours
showvarparts(2, bg = c("hotpink","skyblue"))
plot(mod, bg = c("hotpink","skyblue"))
# Alternative way of to conduct this partitioning
# Change the data frame with factors into numeric model matrix
mm <- model.matrix(~ SubsDens + WatrCont + Substrate + Shrub + Topo, mite.env)[,-1]
mod <- varpart(decostand(mite, "hel"), mm, mite.pcnm)
# Test fraction [a] using partial RDA:
aFrac <- rda(decostand(mite, "hel"), mm, mite.pcnm)
anova(aFrac, step=200, perm.max=200)
# RsquareAdj gives the same result as component [a] of varpart
RsquareAdj(aFrac)

# Partition Bray-Curtis dissimilarities
varpart(vegdist(mite), ~ ., mite.pcnm, data = mite.env)
# Three explanatory matrices 
mod <- varpart(mite, ~ SubsDens + WatrCont, ~ Substrate + Shrub + Topo,
   mite.pcnm, data=mite.env, transfo="hel")
mod
showvarparts(3, bg=2:4)
plot(mod, bg=2:4)
# An alternative formulation of the previous model using
# matrices mm1 amd mm2 and Hellinger transformed species data
mm1 <- model.matrix(~ SubsDens + WatrCont, mite.env)[,-1]
mm2 <- model.matrix(~ Substrate + Shrub + Topo, mite.env)[, -1]
mite.hel <- decostand(mite, "hel")
mod <- varpart(mite.hel, mm1, mm2, mite.pcnm)
# Use RDA to test fraction [a]
# Matrix can be an argument in formula
rda.result <- rda(mite.hel ~ mm1 + Condition(mm2) +
   Condition(as.matrix(mite.pcnm)))
anova(rda.result, step=200, perm.max=200)

# Four explanatory tables
mod <- varpart(mite, ~ SubsDens + WatrCont, ~Substrate + Shrub + Topo,
  mite.pcnm[,1:11], mite.pcnm[,12:22], data=mite.env, transfo="hel")
mod
plot(mod, bg=2:5)
# Show values for all partitions by putting 'cutoff' low enough:
plot(mod, cutoff = -Inf, cex = 0.7, bg=2:5)