9
$\begingroup$

Using a biplot of values obtained through principal component analysis, it is possible to explore the explanatory variables that make up each principle component. Is this also possible with Linear Discriminant Analysis?

Examples provided use the The data is "Edgar Anderson's Iris Data" (http://en.wikipedia.org/wiki/Iris_flower_data_set). Here is the iris data:

  id  SLength   SWidth  PLength   PWidth species 

   1      5.1      3.5      1.4       .2 setosa 
   2      4.9      3.0      1.4       .2 setosa 
   3      4.7      3.2      1.3       .2 setosa 
   4      4.6      3.1      1.5       .2 setosa 
   5      5.0      3.6      1.4       .2 setosa 
   6      5.4      3.9      1.7       .4 setosa 
   7      4.6      3.4      1.4       .3 setosa 
   8      5.0      3.4      1.5       .2 setosa 
   9      4.4      2.9      1.4       .2 setosa 
  10      4.9      3.1      1.5       .1 setosa 
  11      5.4      3.7      1.5       .2 setosa 
  12      4.8      3.4      1.6       .2 setosa 
  13      4.8      3.0      1.4       .1 setosa 
  14      4.3      3.0      1.1       .1 setosa 
  15      5.8      4.0      1.2       .2 setosa 
  16      5.7      4.4      1.5       .4 setosa 
  17      5.4      3.9      1.3       .4 setosa 
  18      5.1      3.5      1.4       .3 setosa 
  19      5.7      3.8      1.7       .3 setosa 
  20      5.1      3.8      1.5       .3 setosa 
  21      5.4      3.4      1.7       .2 setosa 
  22      5.1      3.7      1.5       .4 setosa 
  23      4.6      3.6      1.0       .2 setosa 
  24      5.1      3.3      1.7       .5 setosa 
  25      4.8      3.4      1.9       .2 setosa 
  26      5.0      3.0      1.6       .2 setosa 
  27      5.0      3.4      1.6       .4 setosa 
  28      5.2      3.5      1.5       .2 setosa 
  29      5.2      3.4      1.4       .2 setosa 
  30      4.7      3.2      1.6       .2 setosa 
  31      4.8      3.1      1.6       .2 setosa 
  32      5.4      3.4      1.5       .4 setosa 
  33      5.2      4.1      1.5       .1 setosa 
  34      5.5      4.2      1.4       .2 setosa 
  35      4.9      3.1      1.5       .2 setosa 
  36      5.0      3.2      1.2       .2 setosa 
  37      5.5      3.5      1.3       .2 setosa 
  38      4.9      3.6      1.4       .1 setosa 
  39      4.4      3.0      1.3       .2 setosa 
  40      5.1      3.4      1.5       .2 setosa 
  41      5.0      3.5      1.3       .3 setosa 
  42      4.5      2.3      1.3       .3 setosa 
  43      4.4      3.2      1.3       .2 setosa 
  44      5.0      3.5      1.6       .6 setosa 
  45      5.1      3.8      1.9       .4 setosa 
  46      4.8      3.0      1.4       .3 setosa 
  47      5.1      3.8      1.6       .2 setosa 
  48      4.6      3.2      1.4       .2 setosa 
  49      5.3      3.7      1.5       .2 setosa 
  50      5.0      3.3      1.4       .2 setosa 
  51      7.0      3.2      4.7      1.4 versicolor 
  52      6.4      3.2      4.5      1.5 versicolor 
  53      6.9      3.1      4.9      1.5 versicolor 
  54      5.5      2.3      4.0      1.3 versicolor 
  55      6.5      2.8      4.6      1.5 versicolor 
  56      5.7      2.8      4.5      1.3 versicolor 
  57      6.3      3.3      4.7      1.6 versicolor 
  58      4.9      2.4      3.3      1.0 versicolor 
  59      6.6      2.9      4.6      1.3 versicolor 
  60      5.2      2.7      3.9      1.4 versicolor 
  61      5.0      2.0      3.5      1.0 versicolor 
  62      5.9      3.0      4.2      1.5 versicolor 
  63      6.0      2.2      4.0      1.0 versicolor 
  64      6.1      2.9      4.7      1.4 versicolor 
  65      5.6      2.9      3.6      1.3 versicolor 
  66      6.7      3.1      4.4      1.4 versicolor 
  67      5.6      3.0      4.5      1.5 versicolor 
  68      5.8      2.7      4.1      1.0 versicolor 
  69      6.2      2.2      4.5      1.5 versicolor 
  70      5.6      2.5      3.9      1.1 versicolor 
  71      5.9      3.2      4.8      1.8 versicolor 
  72      6.1      2.8      4.0      1.3 versicolor 
  73      6.3      2.5      4.9      1.5 versicolor 
  74      6.1      2.8      4.7      1.2 versicolor 
  75      6.4      2.9      4.3      1.3 versicolor 
  76      6.6      3.0      4.4      1.4 versicolor 
  77      6.8      2.8      4.8      1.4 versicolor 
  78      6.7      3.0      5.0      1.7 versicolor 
  79      6.0      2.9      4.5      1.5 versicolor 
  80      5.7      2.6      3.5      1.0 versicolor 
  81      5.5      2.4      3.8      1.1 versicolor 
  82      5.5      2.4      3.7      1.0 versicolor 
  83      5.8      2.7      3.9      1.2 versicolor 
  84      6.0      2.7      5.1      1.6 versicolor 
  85      5.4      3.0      4.5      1.5 versicolor 
  86      6.0      3.4      4.5      1.6 versicolor 
  87      6.7      3.1      4.7      1.5 versicolor 
  88      6.3      2.3      4.4      1.3 versicolor 
  89      5.6      3.0      4.1      1.3 versicolor 
  90      5.5      2.5      4.0      1.3 versicolor 
  91      5.5      2.6      4.4      1.2 versicolor 
  92      6.1      3.0      4.6      1.4 versicolor 
  93      5.8      2.6      4.0      1.2 versicolor 
  94      5.0      2.3      3.3      1.0 versicolor 
  95      5.6      2.7      4.2      1.3 versicolor 
  96      5.7      3.0      4.2      1.2 versicolor 
  97      5.7      2.9      4.2      1.3 versicolor 
  98      6.2      2.9      4.3      1.3 versicolor 
  99      5.1      2.5      3.0      1.1 versicolor 
 100      5.7      2.8      4.1      1.3 versicolor 
 101      6.3      3.3      6.0      2.5 virginica 
 102      5.8      2.7      5.1      1.9 virginica 
 103      7.1      3.0      5.9      2.1 virginica 
 104      6.3      2.9      5.6      1.8 virginica 
 105      6.5      3.0      5.8      2.2 virginica 
 106      7.6      3.0      6.6      2.1 virginica 
 107      4.9      2.5      4.5      1.7 virginica 
 108      7.3      2.9      6.3      1.8 virginica 
 109      6.7      2.5      5.8      1.8 virginica 
 110      7.2      3.6      6.1      2.5 virginica 
 111      6.5      3.2      5.1      2.0 virginica 
 112      6.4      2.7      5.3      1.9 virginica 
 113      6.8      3.0      5.5      2.1 virginica 
 114      5.7      2.5      5.0      2.0 virginica 
 115      5.8      2.8      5.1      2.4 virginica 
 116      6.4      3.2      5.3      2.3 virginica 
 117      6.5      3.0      5.5      1.8 virginica 
 118      7.7      3.8      6.7      2.2 virginica 
 119      7.7      2.6      6.9      2.3 virginica 
 120      6.0      2.2      5.0      1.5 virginica 
 121      6.9      3.2      5.7      2.3 virginica 
 122      5.6      2.8      4.9      2.0 virginica 
 123      7.7      2.8      6.7      2.0 virginica 
 124      6.3      2.7      4.9      1.8 virginica 
 125      6.7      3.3      5.7      2.1 virginica 
 126      7.2      3.2      6.0      1.8 virginica 
 127      6.2      2.8      4.8      1.8 virginica 
 128      6.1      3.0      4.9      1.8 virginica 
 129      6.4      2.8      5.6      2.1 virginica 
 130      7.2      3.0      5.8      1.6 virginica 
 131      7.4      2.8      6.1      1.9 virginica 
 132      7.9      3.8      6.4      2.0 virginica 
 133      6.4      2.8      5.6      2.2 virginica 
 134      6.3      2.8      5.1      1.5 virginica 
 135      6.1      2.6      5.6      1.4 virginica 
 136      7.7      3.0      6.1      2.3 virginica 
 137      6.3      3.4      5.6      2.4 virginica 
 138      6.4      3.1      5.5      1.8 virginica 
 139      6.0      3.0      4.8      1.8 virginica 
 140      6.9      3.1      5.4      2.1 virginica 
 141      6.7      3.1      5.6      2.4 virginica 
 142      6.9      3.1      5.1      2.3 virginica 
 143      5.8      2.7      5.1      1.9 virginica 
 144      6.8      3.2      5.9      2.3 virginica 
 145      6.7      3.3      5.7      2.5 virginica 
 146      6.7      3.0      5.2      2.3 virginica 
 147      6.3      2.5      5.0      1.9 virginica 
 148      6.5      3.0      5.2      2.0 virginica 
 149      6.2      3.4      5.4      2.3 virginica 
 150      5.9      3.0      5.1      1.8 virginica

Example PCA biplot using the iris data set in R (code below):

enter image description here

This figure indicates that Petal length and Petal width are important in determining PC1 score and in discriminating between Species groups. setosa has smaller petals and wider sepals.

Apparently, similar conclusions can be drawn from plotting linear discriminant analysis results, though I am not certain what the LDA plot presents, hence the question. The axis are the two first linear discriminants (LD1 99% and LD2 1% of trace). The coordinates of the red vectors are "Coefficients of linear discriminants" also described as "scaling" (lda.fit$scaling: a matrix which transforms observations to discriminant functions, normalized so that within groups covariance matrix is spherical). "scaling" is calculated as diag(1/f1, , p) and f1 is sqrt(diag(var(x - group.means[g, ]))). Data can be projected onto the linear discriminants (using predict.lda) (code below, as demonstrated https://stackoverflow.com/a/17240647/742447). The data and the predictor variables are plotted together so that which species are defined by an increase in which predictor variables can be seen (as is done for usual PCA biplots and the above PCA biplot).:

Example LDA biplot using the iris data set in R

From this plot, Sepal width, Petal Width and Petal Length all contribute to a similar level to LD1. As expected, setosa appears to smaller petals and wider sepals.

There is no built-in way to plot such biplots from LDA in R and few discussions of this online, which makes me wary of this approach.

Does this LDA plot (see code below) provide a statistically valid interpretation of predictor variable scaling scores ?

Code for PCA:

require(grid)

  iris.pca <- prcomp(iris[,-5])
  PC <- iris.pca
  x="PC1"
  y="PC2"
  PCdata <- data.frame(obsnames=iris[,5], PC$x)

  datapc <- data.frame(varnames=rownames(PC$rotation), PC$rotation)
  mult <- min(
    (max(PCdata[,y]) - min(PCdata[,y])/(max(datapc[,y])-min(datapc[,y]))),
    (max(PCdata[,x]) - min(PCdata[,x])/(max(datapc[,x])-min(datapc[,x])))
  )
  datapc <- transform(datapc,
                      v1 = 1.6 * mult * (get(x)),
                      v2 = 1.6 * mult * (get(y))
  )

  datapc$length <- with(datapc, sqrt(v1^2+v2^2))
  datapc <- datapc[order(-datapc$length),]

  p <- qplot(data=data.frame(iris.pca$x),
             main="PCA",
             x=PC1,
             y=PC2,
             shape=iris$Species)
  #p <- p + stat_ellipse(aes(group=iris$Species))
  p <- p + geom_hline(aes(0), size=.2) + geom_vline(aes(0), size=.2)
  p <- p + geom_text(data=datapc, 
                     aes(x=v1, y=v2,
                         label=varnames,
                         shape=NULL,
                         linetype=NULL,
                         alpha=length), 
                     size = 3, vjust=0.5,
                     hjust=0, color="red")
  p <- p + geom_segment(data=datapc, 
                        aes(x=0, y=0, xend=v1,
                            yend=v2, shape=NULL, 
                            linetype=NULL,
                            alpha=length),
                        arrow=arrow(length=unit(0.2,"cm")),
                        alpha=0.5, color="red")
  p <- p + coord_flip()


  print(p)

Code for LDA

#Perform LDA analysis
iris.lda <- lda(as.factor(Species)~.,
                 data=iris)

#Project data on linear discriminants
iris.lda.values <- predict(iris.lda, iris[,-5])

#Extract scaling for each predictor and
data.lda <- data.frame(varnames=rownames(coef(iris.lda)), coef(iris.lda))

#coef(iris.lda) is equivalent to iris.lda$scaling

data.lda$length <- with(data.lda, sqrt(LD1^2+LD2^2))
scale.para <- 0.75

#Plot the results
p <- qplot(data=data.frame(iris.lda.values$x),
           main="LDA",
           x=LD1,
           y=LD2,
           shape=iris$Species)#+stat_ellipse()
p <- p + geom_hline(aes(0), size=.2) + geom_vline(aes(0), size=.2)
p <- p + theme(legend.position="none")
p <- p + geom_text(data=data.lda,
                   aes(x=LD1*scale.para, y=LD2*scale.para,
                       label=varnames, 
                       shape=NULL, linetype=NULL,
                       alpha=length),
                   size = 3, vjust=0.5,
                   hjust=0, color="red")
p <- p + geom_segment(data=data.lda,
                      aes(x=0, y=0,
                          xend=LD1*scale.para, yend=LD2*scale.para,
                          shape=NULL, linetype=NULL,
                          alpha=length),
                      arrow=arrow(length=unit(0.2,"cm")),
                      color="red")
p <- p + coord_flip()

print(p)

The results of the LDA are as follows

lda(as.factor(Species) ~ ., data = iris)

Prior probabilities of groups:
    setosa versicolor  virginica 
 0.3333333  0.3333333  0.3333333 

Group means:
           Sepal.Length Sepal.Width Petal.Length Petal.Width
setosa            5.006       3.428        1.462       0.246
versicolor        5.936       2.770        4.260       1.326
virginica         6.588       2.974        5.552       2.026

Coefficients of linear discriminants:
                    LD1         LD2
Sepal.Length  0.8293776  0.02410215
Sepal.Width   1.5344731  2.16452123
Petal.Length -2.2012117 -0.93192121
Petal.Width  -2.8104603  2.83918785

Proportion of trace:
   LD1    LD2 
0.9912 0.0088
$\endgroup$
  • $\begingroup$ I can't follow your code (I'm not R user and I'd prefer to see actual data and results values rather than unexplained pictures and unexplained code), sorry. What do your plots plot? What are the coordinates of the red vectors - regressional weights of the latents or of the variables? What did you plot as well data poins for? What is discriminant predictor variable scaling scores? - the term seems to me not common and strange. $\endgroup$ – ttnphns Jan 21 '14 at 14:22
  • $\begingroup$ @ttnphns: thank you for suggesting question improvements which are now reflected in the question. $\endgroup$ – Etienne Low-Décarie Jan 22 '14 at 14:01
  • $\begingroup$ I still don't know what is predictor variable scaling scores. Maybe "discriminant scores"? Anyway, I added an answer which might be of your interest. $\endgroup$ – ttnphns Jan 23 '14 at 14:03
7
+100
$\begingroup$

Principal components analysis and Linear discriminant analysis outputs; iris data.

I will not be drawing biplots because biplots can drawn with various normalizations and therefore may look different. Since I'm not R user I have difficulty to track down how you produced your plots, to repeat them. Instead, I will do PCA and LDA and show the results, in a manner similar to this (you might want to read). Both analyses done in SPSS.

Principal components of iris data:

The analysis will be based on covariances (not correlations) between the 4 variables.

Eigenvalues (component variances) and the proportion of overall variance explained
PC1   4.228241706    .924618723 
PC2    .242670748    .053066483 
PC3    .078209500    .017102610 
PC4    .023835093    .005212184 
# @Etienne's comment: 
# Eigenvalues are obtained in R by
# (princomp(iris[,-5])$sdev)^2 or (prcomp(iris[,-5])$sdev)^2.
# Proportion of variance explained is obtained in R by
# summary(princomp(iris[,-5])) or summary(prcomp(iris[,-5]))

Eigenvectors (cosines of rotation of variables into components)
              PC1           PC2           PC3           PC4
SLength   .3613865918   .6565887713  -.5820298513   .3154871929 
SWidth   -.0845225141   .7301614348   .5979108301  -.3197231037 
PLength   .8566706060  -.1733726628   .0762360758  -.4798389870 
PWidth    .3582891972  -.0754810199   .5458314320   .7536574253    
# @Etienne's comment: 
# This is obtained in R by
# prcomp(iris[,-5])$rotation or princomp(iris[,-5])$loadings

Loadings (eigenvectors normalized to respective eigenvalues;
loadings are the covariances between variables and standardized components)
              PC1           PC2           PC3           PC4
SLength    .743108002    .323446284   -.162770244    .048706863 
SWidth    -.173801015    .359689372    .167211512   -.049360829 
PLength   1.761545107   -.085406187    .021320152   -.074080509 
PWidth     .736738926   -.037183175    .152647008    .116354292    
# @Etienne's comment: 
# Loadings can be obtained in R with
# t(t(princomp(iris[,-5])$loadings) * princomp(iris[,-5])$sdev) or
# t(t(prcomp(iris[,-5])$rotation) * prcomp(iris[,-5])$sdev)

Standardized (rescaled) loadings
(loadings divided by st. deviations of the respective variables)
              PC1           PC2           PC3           PC4
SLength    .897401762     .390604412   -.196566721    .058820016
SWidth    -.398748472     .825228709    .383630296   -.113247642
PLength    .997873942    -.048380599    .012077365   -.041964868
PWidth     .966547516   -.048781602    .200261695    .152648309  

Raw component scores (Centered 4-variable data multiplied by eigenvectors)
     PC1           PC2           PC3           PC4
-2.684125626    .319397247   -.027914828    .002262437 
-2.714141687   -.177001225   -.210464272    .099026550 
-2.888990569   -.144949426    .017900256    .019968390 
-2.745342856   -.318298979    .031559374   -.075575817 
-2.728716537    .326754513    .090079241   -.061258593 
-2.280859633    .741330449    .168677658   -.024200858 
-2.820537751   -.089461385    .257892158   -.048143106 
-2.626144973    .163384960   -.021879318   -.045297871 
-2.886382732   -.578311754    .020759570   -.026744736 
-2.672755798   -.113774246   -.197632725   -.056295401 
... etc.
# @Etienne's comment: 
# This is obtained in R with
# prcomp(iris[,-5])$x or princomp(iris[,-5])$scores.
# Can also be eigenvector normalized for plotting

Standardized (to unit variances) component scores, when multiplied
by loadings return original centered variables.

It is important to stress that it is loadings, not eigenvectors, by which we typically interpret principal components (or factors in factor analysis) - if we need to interpret. Loadings are the regressional coefficients of modeling variables by standardized components. At the same time, because components don't intercorrelate, they are the covariances between such components and the variables. Standardized (rescaled) loadings, like correlations, cannot exceed 1, and are more handy to interpret because the effect of unequal variances of variables is taken off.

It is loadings, not eigenvectors, that are typically displayed on a biplot side-by-side with component scores; the latter are often displayed column-normalized.


Linear discriminants of iris data:

There is 3 classes and 4 variables: min(3-1,4)=2 discriminants can be extracted.
Only the extraction (no classification of data points) will be done.

Eigenvalues and canonical correlations
(Canonical correlation squared is SSbetween/SStotal of ANOVA by that discriminant)
Dis1    32.19192920     .98482089 
Dis2      .28539104     .47119702
# @Etienne's comment:
# In R eigenvalues are expected from
# lda(as.factor(Species)~.,data=iris)$svd, but this produces
#   Dis1       Dis2
# 48.642644  4.579983
# @ttnphns' comment:
# The difference might be due to different computational approach
# (e.g. me used eigendecomposition and R used svd?) and is of no importance.
# Canonical correlations though should be the same.

Eigenvectors (here, column-normalized to SS=1: cosines of rotation of variables into discriminants)
              Dis1          Dis2
SLength  -.2087418215   .0065319640 
SWidth   -.3862036868   .5866105531 
PLength   .5540117156  -.2525615400 
PWidth    .7073503964   .7694530921

Unstandardized discriminant coefficients (proportionally related to eigenvectors)
              Dis1          Dis2
SLength   -.829377642    .024102149 
SWidth   -1.534473068   2.164521235 
PLength   2.201211656   -.931921210 
PWidth    2.810460309   2.839187853
# @Etienne's comment:
# This is obtained in R with
# lda(as.factor(Species)~.,data=iris)$scaling
# which is described as being standardized discriminant coefficients in the function definition.

Standardized discriminant coefficients
              Dis1          Dis2
SLength  -.4269548486   .0124075316 
SWidth   -.5212416758   .7352613085 
PLength   .9472572487  -.4010378190 
PWidth    .5751607719   .5810398645

Pooled within-groups correlations between variables and discriminants
              Dis1          Dis2
SLength   .2225959415   .3108117231 
SWidth   -.1190115149   .8636809224 
PLength   .7060653811   .1677013843 
PWidth    .6331779262   .7372420588 

Discriminant scores (Centered 4-variable data multiplied by unstandardized coefficients)
     Dis1           Dis2
-8.061799783    .300420621 
-7.128687721   -.786660426 
-7.489827971   -.265384488 
-6.813200569   -.670631068 
-8.132309326    .514462530 
-7.701946744   1.461720967 
-7.212617624    .355836209 
-7.605293546   -.011633838 
-6.560551593  -1.015163624 
-7.343059893   -.947319209
... etc.
# @Etienne's comment:
# This is obtained in R with
# predict(lda(as.factor(Species)~.,data=iris), iris[,-5])$x

About computations at extraction of discriminants in LDA please look here. We interpret discriminants usually by discriminant coefficients or standardized discriminant coefficients (the latter are more handy because differential variance in variables is taken off). This is like in PCA. But, note: the coefficients here are the regressional coefficients of modeling discriminants by variables, not vice versa, like it was in PCA. Because variables are not uncorrelated, the coefficients cannot be seen as covariances between variables and discriminants.

Yet we have another matrix instead which may serve as an alternative source of interpretation of discriminants - pooled within-group correlations between the discriminants and the variables. Because discriminants are uncorrelated, like PCs, this matrix is in a sense analogous to the standardized loadings of PCA.

In all, while in PCA we have the only matrix - loadings - to help interpret the latents, in LDA we have two alternative matrices for that. If you need to plot (biplot or whatever), you have to decide whether to plot coefficients or correlations.

And, of course, needless to remind that in PCA of iris data the components don't "know" that there are 3 classes; they can't be expected to discriminate classes. Discriminants do "know" there are classes and it is their natural job which is to discriminate.

$\endgroup$
  • $\begingroup$ So I can plot, after arbitrary scaling, either "Standardized discriminant coefficients" or "Pooled within-groups correlations between variables and discriminants" on the same axis as "Discriminant scores" to interpret the results in two different ways? In my question I had plotted "Unstandardized discriminant coefficients" on the same axis as the "Discriminant scores". $\endgroup$ – Etienne Low-Décarie Jan 24 '14 at 14:46
  • 1
    $\begingroup$ @Etienne I added details you asked for to the bottom of this answer stats.stackexchange.com/a/48859/3277. Thank you for your generosity. $\endgroup$ – ttnphns Jan 24 '14 at 16:56
  • 1
    $\begingroup$ @TLJ, should be: between variables and standardized components. I've inserted the word. See please here: Loadings are the coefficients to predict... as well as here: [Footnote: The components' values...]. Loadings are coefficients to compute variables from standardized and orthogonal components, by virtue of what loadings are the covariances between these and those. $\endgroup$ – ttnphns Aug 18 '14 at 17:46
  • 1
    $\begingroup$ @TLJ, "these and those" = variables and components. You said you computed raw component scores. Standardize each component to variance=1. Compute covariances between the variables and the components. That would be the loadings. "Standardized" or "rescaled" loading is the loading divided by the st. deviation of the respective variable. $\endgroup$ – ttnphns Aug 18 '14 at 20:33
  • 1
    $\begingroup$ Loading squared is the share of the variable's variance that is accounted for by the component. $\endgroup$ – ttnphns Aug 18 '14 at 22:41
4
$\begingroup$

My understanding is that biplots of linear discriminant analyses can be done, it is implemented in fact in R packages ggbiplot and ggord and another function to do it is posted in this StackOverflow thread.

Also the book "Biplots in practice" by M. Greenacre has one chapter (chapter 11, see pdf) on it and in Figure 11.5 it shows a biplot of a linear discriminant analysis of the iris dataset: enter image description here

$\endgroup$
2
$\begingroup$

I know this was asked over a year ago, and ttnphns gave an excellent and in-depth answer, but I thought I'd add a couple of comments for those (like me) that are interested in PCA and LDA for their usefulness in ecological sciences, but have limited statistical background (not statisticians).

PCs in PCA are linear combinations of original variables that sequentially maximally explain total variance in the multidimensional dataset. You will have as many PCs as you do original variables. The percent of the variance the PCs explain is given by the eigenvalues of the similarity matrix used, and the coefficient for each original variable on each new PC is given by the eigenvectors. PCA has no assumptions about groups. PCA is very good for seeing how multiple variables change in value across your data (in a biplot, for example). Interpreting a PCA relies heavily on the biplot.

LDA is different for a very important reason - it creates new variables (LDs) by maximizing variance between groups. These are still linear combinations of original variables, but rather than explain as much variance as possible with each sequential LD, instead they are drawn to maximize the DIFFERENCE between groups along that new variable. Rather than a similarity matrix, LDA (and MANOVA) use a comparison matrix of between and within groups sum of squares and cross-products. The eigenvectors of this matrix - the coefficients that the OP was originally concerned with - describe how much the original variables contribute to the formation of the new LDs.

For these reasons, the eigenvectors from the PCA will give you a better idea how a variable changes in value across your data cloud, and how important it is to total variance in your dataset, than the LDA. However, the LDA, particularly in combination with a MANOVA, will give you a statistical test of difference in multivariate centroids of your groups, and an estimate of error in allocation of points to their respective groups (in a sense, multivariate effect size). In an LDA, even if a variable changes linearly (and significantly) across groups, its coefficient on an LD may not indicate the "scale" of that effect, and depends entirely on the other variables included in the analysis.

I hope that was clear. Thanks for your time. See a picture below...

PCs and LDs are constructed differently, and coefficients for an LD may not give you a sense of how original variables vary in your dataset

$\endgroup$
  • $\begingroup$ This is all correct, and +1 from me, but I am not sure how your answer addresses the original question, which was very specifically about how to draw a LDA biplot. $\endgroup$ – amoeba says Reinstate Monica Feb 6 '15 at 17:51
  • $\begingroup$ I suppose you're right - I was responding to this, mostly "Using a biplot of values obtained through principal component analysis, it is possible to explore the explanatory variables that make up each principle component. Is this also possible with Linear Discriminant Analysis?" - and the answer is, yes, but the meaning is very different, as described above... Thanks for the comment and +1! $\endgroup$ – danno Feb 10 '15 at 15:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.