vegan (version 2.0-10)

ordisurf: Fit and Plot Smooth Surfaces of Variables on Ordination.

Description

Function ordisurf fits a smooth surface for given variable and plots the result on ordination diagram.

Usage

## S3 method for class 'default':
ordisurf(x, y, choices = c(1, 2), knots = 10,
         family = "gaussian", col = "red", isotropic = TRUE,
         thinplate = TRUE, bs = "tp", fx = FALSE, add = FALSE,
         display = "sites", w = weights(x), main, nlevels = 10,
         levels, npoints = 31, labcex = 0.6, bubble = FALSE,
         cex = 1, select = TRUE, method = "REML", gamma = 1,
         plot = TRUE, lwd.cl = par("lwd"), ...)
## S3 method for class 'formula':
ordisurf(formula, data, ...)
## S3 method for class 'ordisurf':
calibrate(object, newdata, ...)
## S3 method for class 'ordisurf':
plot(x, what = c("contour","persp","gam"),
     add = FALSE, bubble = FALSE, col = "red", cex = 1,
     nlevels = 10, levels, labcex = 0.6, lwd.cl = par("lwd"), ...)

Arguments

x

For ordisurf an ordination configuration, either a matrix or a result known by scores. For plot.ordisurf an object of class "ordisurf" as returned by

y

Variable to be plotted / modelled as a function of the ordination scores.

choices

Ordination axes.

knots

Number of initial knots in gam (one more than degrees of freedom). If knots = 0 or knots = 1 the function will fit a linear trend surface, and if knots = 2

family

Error distribution in gam.

col

Colour of contours.

isotropic, thinplate

Fit an isotropic smooth surface (i.e. same smoothness in both ordination dimensions) via gam. Use of thinplate is deprecated and will be removed in a future version of the package.

bs

a two letter character string indicating the smoothing basis to use. (eg "tp" for thin plate regression spline, "cr" for cubic regression spline). One of c("tp", "ts", "cr", "cs", "ds", "ps", "ad"). See

fx

indicates whether the smoothers are fixed degree of freedom regression splines (fx = FALSE) or penalised regression splines (fx = TRUE). Can be a vector of length 2 for anisotropic surfaces (isotropic = FALSE

add

Add contours to an existing diagram or draw a new plot?

display

Type of scores known by scores: typically "sites" for ordinary site scores or "lc" for linear combination scores.

w

Prior weights on the data. Concerns mainly cca and decorana results which have nonconstant weights.

main

The main title for the plot, or as default the name of plotted variable in a new plot.

nlevels, levels

Either a vector of levels for which contours are drawn, or suggested number of contours in nlevels if levels are not supplied.

npoints

numeric; the number of locations at which to evaluate the fitted surface. This represents the number of locations in each dimension.

labcex

Label size in contours. Setting this zero will suppress labels.

bubble

Use a bubble plot for points, or vary the point diameter by the value of the plotted variable. If bubble is numeric, its value is used for the maximum symbol size (as in cex), or if bubble = TRU

cex

Character expansion of plotting symbols.

select

Logical; specify gam argument "select". If this is TRUE then gam can add an extra penalty to each term so that it can be p

method

character; the smoothing parameter estimation method. Options allowed are: "GCV.Cp" uses GCV for models with unknown scale parameter and Mallows' Cp/UBRE/AIC for models with known scale; "GACV.Cp" as for "GCV.Cp

gamma

Multiplier to inflate model degrees of freedom in GCV or UBRE/AIC score by. This effectively places an extra penalty on complex models. An oft-used value is gamma = 1.4.

plot

logical; should any plotting be done by ordisurf? Useful if all you want is the fitted response surface model.

lwd.cl

numeric; the lwd (line width) parameter to use when drawing the contour lines.

formula, data

Alternative definition of the fitted model as x ~ y, where left-hand side is the ordination x and right-hand side the single fitted continuous variable y. The variable y must be in the workin

object

An ordisurf result object.

newdata

Coordinates in two-dimensional ordination for new points.

what

character; what type of plot to produce. "contour" produces a contour plot of the response surface, see contour for details. "persp" produces a perspective plot of the

...

Other parameters passed to scores, or to the graphical functions. See Note below for exceptions.

Value

• ordisurf is usually called for its side effect of drawing the contour plot. The function returns a result object of class "ordisurf" that inherits from gam used internally to fit the surface, but adds an item grid that contains the data for the grid surface. The item grid has elements x and y which are vectors of axis coordinates, and element z that is a matrix of fitted values for contour. The values outside the convex hull of observed points are indicated as NA in z. The gam component of the result can be used for further analysis like predicting new values (see predict.gam).

Warning

The fitted GAM is a regression model and has the usual assumptions of such models. Of particular note is the assumption of independence of residuals. If the observations are not independent (e.g. they are repeat measures on a set of objects, or from an experimental design, inter alia) do not trust the p-values from the GAM output.

If you need further control (i.e. to add additional fixed effects to the model, or use more complex smoothers), extract the ordination scores using the scores function and then generate your own gam call.

Details

Function ordisurf fits a smooth surface using penalised splines (Wood 2003) in gam, and uses predict.gam to find fitted values in a regular grid. The smooth surface can be fitted with an extra penalty that allows the entire smoother to be penalized back to 0 degrees of freedom, effectively removing the term from the model (see Marra & Wood, 2011). The addition of this extra penalty is invoked by setting argument select to TRUE. An alternative is to use a spline basis that includes shrinkage (bs = "ts" or bs = "cs").

ordisurf() exposes a large number of options from gam for specifying the basis functions used for the surface. If you stray from the defaults, do read the Notes section below and relevant documentation in s and smooth.terms.

The function plots the fitted contours with convex hull of data points either over an existing ordination diagram or draws a new plot. If select = TRUE and the smooth is effectively penalised out of the model, no contours will be plotted.

gam determines the degree of smoothness for the fitted response surface during model fitting, unless fx = TRUE. Argument method controls how gam performs this smoothness selection. See gam for details of the available options. Using "REML" or "ML" yields p-values for smooths with the best coverage properties if such things matter to you.

The function uses scores to extract ordination scores, and x can be any result object known by that function.

The user can supply a vector of prior weights w. If the ordination object has weights, these will be used. In practise this means that the row totals are used as weights with cca or decorana results. If you do not like this, but want to give equal weights to all sites, you should set w = NULL. The behaviour is consistent with envfit. For complete accordance with constrained cca, you should set display = "lc" (and possibly scaling = 2).

Function calibrate returns the fitted values of the response variable. The newdata must be coordinates of points for which the fitted values are desired. The function is based on predict.gam and will pass extra arguments to that function.

References

Marra, G.P & Wood, S.N. (2011) Practical variable selection for generalized additive models. Comput. Stat. Data Analysis 55, 2372--2387.

Wood, S.N. (2003) Thin plate regression splines. J. R. Statist. Soc. B 65, 95--114.

See Also

For basic routines gam, and scores. Function envfit provides a more traditional and compact alternative.

Examples

data(varespec)
data(varechem)
vare.dist <- vegdist(varespec)
vare.mds <- monoMDS(vare.dist)
with(varechem, ordisurf(vare.mds, Baresoil, bubble = 5))

## as above but without the extra penalties on smooth terms,
## and using GCV smoothness selection (old behaviour of `ordisurf()`):
with(varechem, ordisurf(vare.mds, Baresoil,col = "blue", add = TRUE,
                        select = FALSE, method = "GCV.Cp"))

## Cover of Cladina arbuscula
fit <- with(varespec, ordisurf(vare.mds, Cla.arb, family=quasipoisson)) 
## Get fitted values
calibrate(fit)

## Variable selection via additional shrinkage penalties
## This allows non-significant smooths to be selected out
## of the model not just to a linear surface. There are 2
## options available:
##  - option 1: `select = TRUE` --- the *default*
with(varechem,
     ordisurf(vare.mds, Baresoil, method = "REML", select = TRUE))
##  - option 2: use a basis with shrinkage
with(varechem,
     ordisurf(vare.mds, Baresoil, method = "REML", bs = "ts"))
## or bs = "cs" with `isotropic = FALSE`

## Plot method
plot(fit, what = "contour")

## Plotting the "gam" object
plot(fit, what = "gam") ## 'col' and 'cex' not passed on
## or via plot.gam directly
plot.gam(fit, cex = 2, pch = 1, col = "blue")
## 'col' effects all objects drawn...

### controlling the basis functions used
## Use Duchon splines
with(varechem, ordisurf(vare.mds, Baresoil, bs = "ds"))

## A fixed degrees of freedom smooth, must use 'select = FALSE'
with(varechem, ordisurf(vare.mds, Baresoil, knots = 4,
                        fx = TRUE, select = FALSE))

## An anisotropic smoother with cubic regression spline bases
with(varechem, ordisurf(vare.mds, Baresoil, isotropic = FALSE,
                        bs = "cr", knots = 4))

## An anisotropic smoother with cubic regression spline with
## shrinkage bases & different degrees of freedom in each dimension
with(varechem, ordisurf(vare.mds, Baresoil, isotropic = FALSE,
                        bs = "cs", knots = c(3,4), fx = TRUE,
                        select = FALSE))