Theory, Examples & Exercises
This is an old revision of the document!
Note: variation partitioning is sometimes also called commonality analysis in reference to the common (shared) fraction of variation (Kerlinger & Pedhazur 1973). It is also a synonym to variance partitioning1).
In case we have two or more explanatory variables, one may be interested in variation in species composition explained by each of them. If some of these explanatory variables are correlated, one must expect that variation explained by the first or the other variable cannot be separated - it will be shared.
The way how to approach this problem is variation partitioning, when variation explained by each variable (or set of variables) independently is partitioned into variation attributable purely to given environmental variable, and shared variation attributable to two or more variables.
Variation can be partitioned by individual variables (e.g. variation explained by soil pH vs variation explained by soil Ca) or by groups of variables (e.g. soil variables vs climatic variables).
Results can be visualized using Venn's diagram (see figure on the right). Meaning of the fractions in Venn's diagram is the following2):
If the variation is partitioned among groups with the same number of variables (e.g. two soil and two climatic variables), than variation explained by each group is comparable without adjustement. However, if groups contain different numbers of variables, variation explained by not adjusted R2 is not comparable, since R2 tends to increase with the number of explanatory variables. Here, use of adjusted R2 is recommended.
vegan offers function
varpart, which can partition variation among up to four variables (or groups of variables). Note that
varpart is based on redundancy analysis (
rda) and uses adjusted R2 to express explained variation. The reason for using only
rda is that in R, there is still no function available to calculate adjusted R2 for unimodal ordination methods (like
vegan) - variation partitioning (using linear constrained ordinatino -
rda) among up to four matrices of environmental variables. First argument (
Y) is dependent (usually species composition) matrix (but could be also only one variable - in that case
varpartis conductin linear regression). Next arguments (up to four) are (groups of) explanatory variables. Uses either formula interface (with ~) or matrices.
vegan) - draws Venn's diagram with fractions of explained variation. In default setting it doesn't show negative values of explained variation (argument
cutoff = 0). Consult
?plot.varpartfor more details.
vegan) - draws schema of Venn's diagram with codes of individual fractions.
Use data from Barley field weed community dataset, studying how does weed communities in barley fields changes after application of fertilizer. The problem is that fertilizer may influence the weed community directly, but also indirectly via increased cover of barley (indeed, it grows better if fertilized). The aim of this example is to separate variation in species composition of weed community caused by dose of fertilizer and by estimated cover of barley (both variables are correlated).
First, import the data:
fertil.spe <- read.delim ('http://www.davidzeleny.net/anadat-r/data-download/fertil.spe.txt', row.names = 1) fertil.env <- read.delim ('http://www.davidzeleny.net/anadat-r/data-download/fertil.env.txt', row.names = 1)
Species data are represented by estimated cover of weed community species in a plot and it is not necessary to transform them. One may first apply DCA ordination on species data to see the length of the first DCA axis and hence seek the recommendation whether linear or unimodal method should be used (the length is 3.8 SD, lying in grey zone). We will use RDA here, since
varpart method is based only on RDA (if the data are even more heterogeneous, Hellinger transformation would be advisable).
First, let's see how would the variation partitioning looks like if we do it step-by-step manually (without using
varpart function). We will partition variation explained by variables
fertil.env; for this, we need to define the global model (to see how much variation is explained by both of them), and also partial models with one variable as explanatory and the other as covariable:
# fractions [a+b+c]: rda.all <- rda (fertil.spe ~ dose + cover, data = fertil.env) # fraction [a]: rda.dose.cover <- rda (fertil.spe ~ dose + Condition (cover), data = fertil.env) # fraction [c]: rda.cover.dose <- rda (fertil.spe ~ cover + Condition (dose), data = fertil.env)
For completeness, let's define also models with simple (marginal) effect of each explanatory variable:
# fractions [a+b]: rda.dose <- rda (fertil.spe ~ dose, data = fertil.env) # fractions [b+c]: rda.cover <- rda (fertil.spe ~ cover, data = fertil.env)
Now, let's use the variation explained by individual fractions (using
# fraction [a+b+c] RsquareAdj (rda.all)
$r.squared  0.1430381 $adj.r.squared  0.1286354
# fraction [a]: RsquareAdj (rda.dose.cover)
$r.squared  0.04675401 $adj.r.squared  0.03988225
# fraction [c]: RsquareAdj (rda.cover.dose)
$r.squared  0.07414449 $adj.r.squared  0.06750099
For interpretation, we will use the adjusted R2 stored in
adj.r.squared element of the list returned by
Conditional effect of dose is slightly lower than conditional effect of cover (4.0% vs 6.9%), and shared variation is not that high (2.1%). Seems that species composition weeds is more affected by light conditions modified by increasing cover of barley, than by fertilizer itself.
Now, let's see the same, using function
varp <- varpart (fertil.spe, ~ dose, ~ cover, data = fertil.env) varp
Partition of variation in RDA Call: varpart(Y = fertil.spe, X = ~dose, ~cover, data = fertil.env) Explanatory tables: X1: ~dose X2: ~cover No. of explanatory tables: 2 Total variation (SS): 2581.2 Variance: 21.332 No. of observations: 122 Partition table: Df R.squared Adj.R.squared Testable [a+b] = X1 1 0.06889 0.06113 TRUE [b+c] = X2 1 0.09628 0.08875 TRUE [a+b+c] = X1+X2 2 0.14304 0.12864 TRUE Individual fractions [a] = X1|X2 1 0.03988 TRUE [b] 0 0.02125 FALSE [c] = X2|X1 1 0.06750 TRUE [d] = Residuals 0.87136 FALSE --- Use function 'rda' to test significance of fractions of interest
The results for individual fractions are identical with our results above.
varpart reports also simple (marginal) effect (both not-adjusted and adjusted variation) of individual predictors, i.e. fractions [a+b] and [b+c] - manually we would get the same numbers by applying
RsquareAdj function on
We may plot the results into Venn's diagram (argument
digits influences number of decimal digits shown in the diagram):
plot (varp, digits = 2)
Now, when we know both simple and conditional effect of each variables, we may want to know whether it is significant, and hence worth of interpreting. Results from
varpart contain the column
testable with logical values indicating whether given fraction is testable or not. To test each of them, we will need the models defined above, and the function
anova, which (if applied on single object resulting from
cca method, returns Monte Carlo permutation test of the predictor effect):
## fraction [a+b]: anova (rda.dose) # Permutation test for rda under reduced model # Permutation: free # Number of permutations: 999 # # Model: rda(formula = fertil.spe ~ dose, data = fertil.env) # Df Variance F Pr(>F) # Model 1 1.4696 8.8789 0.001 *** # Residual 120 19.8624 # --- # Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 ## fraction [b+c] anova (rda.cover) # Permutation test for rda under reduced model # Permutation: free # Number of permutations: 999 # # Model: rda(formula = fertil.spe ~ cover, data = fertil.env) # Df Variance F Pr(>F) # Model 1 2.0539 12.785 0.001 *** # Residual 120 19.2781 # --- # Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 ## the global model [a+b+c] anova (rda.all) # Permutation test for rda under reduced model # Permutation: free # Number of permutations: 999 # # Model: rda(formula = fertil.spe ~ dose + cover, data = fertil.env) # Df Variance F Pr(>F) # Model 2 3.0513 9.9313 0.001 *** # Residual 119 18.2808 # --- # Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 ## fraction [a] anova (rda.dose.cover) # Permutation test for rda under reduced model # Permutation: free # Number of permutations: 999 # # Model: rda(formula = fertil.spe ~ dose + Condition(cover), data = fertil.env) # Df Variance F Pr(>F) # Model 1 0.9974 6.4924 0.001 *** # Residual 119 18.2808 # --- # Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 ## fraction [b] anova (rda.cover.dose) # Permutation test for rda under reduced model # Permutation: free # Number of permutations: 999 # # Model: rda(formula = fertil.spe ~ cover + Condition(dose), data = fertil.env) # Df Variance F Pr(>F) # Model 1 1.5817 10.296 0.001 *** # Residual 119 18.2808 # --- # Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
From these results you may see that all simple (marginal) and conditional (partial) effects of both predictors are highly significant (P < 0.001 in all cases).