monoMDS: Global and Local Non-metric Multidimensional Scaling and Linear and Hybrid Scaling

Description

Function implements Kruskal's (1964a,b) non-metric multidimensional scaling (NMDS) using monotone regression and primary (weak) treatment of ties. In addition to traditional global NMDS, the function implements local NMDS, linear and hybrid multidimensional scaling.

Usage

monoMDS(dist, y, k = 2, model = c("global", "local", "linear", "hybrid"), 
    threshold = 0.8, maxit = 200, weakties = TRUE, stress = 1,
    scaling = TRUE, pc = TRUE, smin = 1e-4, sfgrmin = 1e-7, 
    sratmax=0.99999, ...) 
## S3 method for class 'monoMDS':
scores(x, choices = NA, ...)
## S3 method for class 'monoMDS':
plot(x, choices = c(1,2), type = "t", ...)

Arguments

dist

Input dissimilarities.

Starting configuration. A random configuration will be generated if this is missing.

Number of dimensions. NB., the number of points $n$ should be $n > 2k + 1$, and preferably higher in non-metric MDS.

model

MDS model: "global" is normal non-metric MDS with a monotone regression, "local" is non-metric MDS with separate regressions for each point, "linear" uses linear regression, and "hybrid" uses

threshold

Dissimilarity below which linear regression is used alternately with monotone regression.

maxit

Maximum number of iterations.

weakties

Use primary or weak tie treatment, where equal observed dissimilarities are allowed to have different fitted values. if FALSE, then secondary (strong) tie treatment is used, and tied values are not broken.

stress

Use stress type 1 or 2 (see Details).

scaling

Scale final scores to unit root mean squares.

Rotate final scores to principal components.

smin, sfgrmin, sratmax

Convergence criteria: iterations stop when stress drops below smin, scale factor of the gradient drops below sfgrmin, or stress ratio goes over sratmax (but is still $< 1$).

A monoMDS result.

choices

Dimensions returned or plotted. The default NA returns all dimensions.

type

The type of the plot: "t" for text, "p" for points, and "n" for none.

...

Other parameters to the functions (ignored in monoMDS, passed to graphical functions in plot.).

Value

Returns an object of class "monoMDS". The final scores are returned in item points (function scores extracts these results), and the stress in item stress. In addition, there is a large number of other items (but these may change without notice in the future releases).

Details

There are several versions of non-metric multidimensional scaling in R, but monoMDS offers the following unique combination of features:

Weaktreatment of ties (Kruskal 1964a,b), where tied dissimilarities can be broken in monotone regression. This is especially important for cases where compared sites share no species and dissimilarities are tied to their maximum value of one. Breaking ties allows these points to be at different distances and can help in recovering very long coenoclines (gradients). FunctionsmacofSym(smacofpackage) also has adequate tie treatment.
Handles missing values in a meaningful way.
Offerslocalandhybridscaling in addition to usualglobalNMDS (see below).
Uses fast compiled code (isoMDSof theMASSpackage also uses compiled code).

Function monoMDS uses Kruskal's (1964b) original monotone regression to minimize the stress. There are two alternatives of stress: Kruskal's (1964a,b) original or stress 1 and an alternative version or stress 2 (Sibson 1972). Both of these stresses can be expressed with a general formula

$$s^2 = \frac{\sum (d - \hat d)^2}{\sum(d - d_0)^2}$$

where $d$ are distances among points in ordination configuration, $\hat d$ are the fitted ordination distances, and $d_0$ are the ordination distances under null model. For stress 1 $d_0 = 0$, and for stress 2 $d_0 = \bar{d}$ or mean distances. Stress 2 can be expressed as $s^2 = 1 - R^2$, where$R^2$ is squared correlation between fitted values and ordination distances, and so related to the linear fit of stressplot.

Function monoMDS can fit several alternative NMDS variants that can be selected with argument model. The default model = "global" fits global NMDS, or Kruskal's (1964a,b) original NMDS similar to isoMDS (MASS) or smacofSym (smacof). Alternative model = "local" fits local NMDS where independent monotone regression is used for each point (Sibson 1972). Alternative model = "linear" fits a linear MDS. This fits a linear regression instead of monotone, and is not identical to metric scaling or principal coordinates analysis (cmdscale) that performs an eigenvector decomposition of dissimilarities (Gower 1966). Alternative model = "hybrid" implements hybrid MDS that uses monotone regression for all points and linear regression for dissimilarities below or at a threshold dissimilarity in alternating steps (Faith et al. 1987). Function stressplot can be used to display the kind of regression in each model.

Scaling, orientation and direction of the axes is arbitrary. However, the function always centres the axes, and the default scaling is to scale the configuration ot unit root mean square and to rotate the axes (argument pc) to principal components so that the first dimension shows the major variation. It is possible to rotate the solution so that the first axis is parallel to a given environmental variable using function metaMDSrotate.

References

Faith, D.P., Minchin, P.R and Belbin, L. 1987. Compositional dissimilarity as a robust measure of ecological distance. Vegetatio 69, 57--68. Gower, J.C. (1966). Some distance properties of latent root and vector methods used in multivariate analysis. Biometrika 53, 325--328.

Kruskal, J.B. 1964a. Multidimensional scaling by optimizing goodness-of-fit to a nonmetric hypothesis. Psychometrika 29, 1--28.

Kruskal, J.B. 1964b. Nonmetric multidimensional scaling: a numerical method. Psychometrika 29, 115--129.

Minchin, P.R. 1987. An evaluation of relative robustness of techniques for ecological ordinations. Vegetatio 69, 89--107.

Sibson, R. 1972. Order invariant methods for data analysis. Journal of the Royal Statistical Society B 34, 311--349.

Examples

Run this code

data(dune)
dis <- vegdist(dune)
m <- monoMDS(dis, model = "loc")
m
plot(m)

Run the code above in your browser using DataLab