Glynn 2012 CA Correspondence Analysis

Correspondence Analysis
Exploring data and identifying patterns
Dylan Glynn
Lund University
1. A technique for visualising correlations in categorical data
Correspondence analysis is a multivariate exploratory space reduction tech-

nique for categorical data analysis.1 Although certainly true, such a descrip-
tion tells the linguist little. Equally true, but perhaps more helpful, is to de-
scribe correspondence analysis as an exploratory technique that reveals
frequency-based associations in corpus data. Most importantly, perhaps, the
technique visualises these associations to facilitate their identification. Lin-
guists often wish to find relations between given linguistic forms, between
their meanings and in what situations those forms and meanings are used.
Correspondence analysis is especially designed for identifying such usage
patterning. The visualisation of the relations takes the form of configuration
biplots, or maps, which depict degrees of correlation and variation through
the relative proximity of data points (which represent linguistic usage features
and / or the actual examples of use). This paper describes how to perform
correspondence analysis in R. It explains the R code needed to execute the
analyses and shows how to interpret the results.
1.1 Use – What does correspondence analysis do?
In their quotidian research, linguists, from all kinds of theoretical orientations,

analyse various usage-features of naturally occurring utterances. By way of
example, imagine that one obtains 600 examples of a given word, a grammat-
134 Dylan Glynn
ical case, or a syntactic pattern. These examples can then be analysed, using
traditional intuition-based analysis for a range of usage-features, such as
tense, aspect, argument structure, agent type, the ground or path type, and the
register or genre from which the example is taken. The results of analysing
the examples for these usage-features can be summarised as counts of how
often each of the features occurs. Significance tests can then be used to show
that the occurrence of certain features is substantially more common than
could be expected by chance. This statistically significant variation can then,
in turn, be interpreted as representing a distinct pattern of usage.
However, with more than a couple of different dimensions of analysis or
large numbers of features at play, interpreting the numbers of occurrences
becomes increasingly difficult, if not impossible. Quite simply, correspond-
ence analysis is an exploratory tool that helps one find which usage-features
co-occur with other usage-features, giving a map of their overall patterning.
Assuming that one is adopting a cognitive or functional approach to language,
these usage-patterns can be interpreted as grammatical description, operation-
alised in terms of relative frequency.
It must be stressed that this technique is designed solely for exploratory
purposes. In other words, it is a tool for finding things, not for establishing
their significance or discerning their relevance. Therefore, it offers you no
assurance that patterns found are anything more than a chance result, specific
to the sample under observation. Moreover, this tool does not tell you where
to look. Although exploratory, one must avoid ‘fishing’ for results by ran-
domly combining factors in the hope of finding correlations that could be
interpretable. Even if one finds correlations that ‘make sense’, such an ap-
proach increases the chance of finding co-incidental correlations or chance
patterns in the sample.2 A metaphor that might be helpful is that of the shovel
for the archaeologist: if one digs randomly, everywhere, it increases the
chances of finding irrelevant things. Correspondence analysis is a tool for
digging in the data for patterns and correlations. Yet the metaphor can still
serve us further: when an archaeologist finds an artefact, it is still up to the
archaeologist to interpret the finding as well as to verify its authenticity. Cor-
respondence analysis, assuming you have a reasonable hypothesis about
where to look, is a basic and useful tool for unearthing patterns in the data,
but it is no more than that.
Correspondence analysis 135
1.2 Concept – How does correspondence analysis work?
Basically, correspondence analysis takes the frequency of co-occurring fea-

tures and converts them to distances, which are then plotted, revealing how
things are related by how close to or far from each other they are in a two- or
three-dimensional visualisation. In the detail, there is much more to the tech-
nique, but this is the principle. Explaining a few key concepts will allow us to
better understand the functioning of the technique as well as to interpret its
results.
Distance Matrix
The distance matrix is sometimes also called a proximity matrix and even a
dissimilarity matrix. The concept is simple: the frequencies of co-occurrence
are converted to distances. The rrsulting distance matrix can then be visual-
ised in a two- or three-dimensional Euclidean space (‘normal’ perceptual
space). In fact, more precisely, it is the differences between the rows and col-
umns of frequencies that are converted to distances. Correspondence analysis
uses the chi-square distance measure to produce the distance matrix. This
measure is designed to compensate for different ‘amounts’ of a given catego-
ry. In other words, if one has only a few examples of a given feature, let us
say the ‘future tense’, it is highly likely that they will all, or mostly, co-occur
with some other feature, such as a given verb. However, due to the low num-
bers involved, this is much more likely to be chance than other correlations
identified. The chi-square distance measure attempts to compensate for this
kind of bias. Nevertheless, despite the use of the chi-square measure, with
experience, one will still observe (in the plots) correlations that are likely to
be due to small numbers of a given feature. It is always necessary to go back
to both the data, that is the actual language examples, and to the raw frequen-
cies, to see what the plots have ‘revealed’. Greenacre (2007: ch. 4) offers a
lucid explanation of the chi-square distance measure.
Euclidean cloud
The distance matrix takes the form of a Euclidean cloud. In other words, it is
a spread of points in a given space, like rice thrown onto a board or the holes
made by darts on a dartboard. Correspondence analysis computes the Eigen-
vectors of a correlation matrix and produces this Euclidean ‘map’, in one or
two dimensions of that correlation. It can be thought of as reducing a set of
Chi-square scores to Euclidean distances (natural perceptual distances), suita-
ble for two- or three-dimension visualisations. For the reader familiar with
136 Dylan Glynn
exploratory statistics, it is essentially the same as principal components analy-

sis, but modified for categorical data.
Profiles and mass

A profile is the behavioural characteristics of a given category in the analysis,
determined by the set of relative co-occurrence frequencies of that category.
In a frequency (contingency) table, it is a column or a row containing all the
relative frequencies of those co-occurrences. It is these profiles that the corre-
spondence analysis plots. To calculate the profile, you add the number of
occurrences for each feature in a row. You then divide each of those occur-
rences by the sum of them. This gives you a profile figure for each cell. The
same procedure is undertaken for the columns, giving you the column pro-
files.
However, as mentioned above, not all the co-occurrences are of equal im-
portance. Infrequent features would have a disproportionate effect if all were
taken equally. Correspondence analysis uses weighted averages of the pro-
files to compensate for this. In correspondence analysis, the term ‘mass’ is
used to mean ‘weight’. Weighting an average modifies the calculation to bias
certain scores. It is widely used in basic statistics, from calculating the aver-
age score in a class test to the average monthly profit of a franchised shop.
Inertia and Variation

The higher the explained inertia one obtains, the better. Inertia is the term
used in correspondence analysis to talk about the degree of variation. The
inertia is calculated on observed and expected frequencies of co-occurrence.
Inertia is high when column and row profiles have large deviations from their
averages. In multiple correspondence analysis (as opposed to binary corre-
spondence analysis), these scores are not interpretable, which is a major
drawback for this form of the technique. They are not interpretable because
the scores calculated seriously under-estimate the amount of accurately de-
scribed variation, giving unnecessarily ‘bad’ results. Two corrections to this
have been proposed, firstly by the original author of the technique, Benzécri
(1979 [reported in Greenacre 2006: 68]), and secondly by the current main
proponent of the technique, Greenacre (2006: 68). Greenacre argues that
Benzécri’s original correction was biased towards an overly optimistic result,
that is, explaining more variation than was actually the case. The {ca} pack-
age, described in section 2.3.2, includes an option to apply Greenacre’s inertia
adjustment. It application is explained in section 2.3.2.
Biplots
The concept behind the visualisation in a biplot is quite simple to understand.
The correspondence analysis has calculated proximity values for the combi-
nation of the cells across the rows and columns of a contingency table. These
can be plotted. Each dimension of the plot (there are two dimensions ina bip-
lot) will explain a certain percentage of the data variation, or ‘inertia’. Plot-
ting a single dimension, a simple line or the x-axis, will place the data points
on this line at varying distances from each other. However, in most situations,
this will poorly represent the relations between those features. If we add a
second dimension, the y-axis, we obtain a two-dimensional biplot, typical of
correspondence analysis and a range of other space reduction techniques. This
will, hopefully, explain a great deal of the variation in the data. The scores of
the explained inertia (or variation) are typically given for these first two axes.
Although, theoretically, it is possible to take any two dimensions and plot
these.
Normally, a combination of the first two dimensions captures a large per-
centage of the variation. Adding a third dimension, the z-axis, produces a
three-dimensional plot that will even more accurately represent the behaviour
of the data. Three-dimensional plots are also possible in R, but are not con-
sidered in this discussion. Sometimes, it is useful to examine combinations of
dimensions one and three or even two and three in biplots, especially when
the explained inertia is low. For most data sets, though, a combination of the
first two dimensions offers the most accurate and interpretable visualisation
of the variation and association in the data. The numerical summary of a cor-
respondence analysis will list all the dimensions, but above the third- or
fourth-dimension, it is rare that further dimensions represent anything more
than a small fraction of the variation. In order to completely represent a con-
tingency table, one would need all the dimensions. The number of possible
dimensions is equal to the number of rows or columns (which ever is smaller)
minus one. So, to visualise a table with five rows and eight columns, one
would need four dimensions.
Unfortunately, there is a range of terminology that varies from one book to
another and even from one R package to the next. A few basic terms that may
arise, especially in the numerical summaries of the analysis include: ‘Eigen-
values’, which indicates the inertia; the ‘percentages of explained variance’,
or simply the percentage of inertia; and ‘communalities’, which are the per-
centages of explained inertia for individual rows or columns. If one wishes to
work with the technique, there are three excellent books that explain its func-
tioning in a clear manner, accessible even to readers with no statistical train-
138 Dylan Glynn
ing. These books are Greenacre (2007) Correspondence Analysis in Practice,

Le Roux & Rouanet (2010) Multiple Correspondence Analysis and Hus-
son et al. (2011) Exploratory Multivariate Analysis by Example Using R.
1.3 Choice - Binary and Multiple Correspondence Analysis
Correspondence analysis is, in fact, a family of techniques and a family that is

growing. There exist at least three kinds of binary correspondence analysis
and three kinds of multiple correspondence analysis. Binary correspondence
analysis can be understood as the basis for the multiple correspondence anal-
ysis. It has two advantages over the latter. Firstly, it indicates the percentage
of explained inertia for each axis. This tells you how well your analysis fits
the data (how much of the variation the analysis captures). It is even possible
to add confidence ellipses that estimate statistical significance (see section
2.4.1). Secondly, the plots are simpler to interpret, which is an extremely
important advantage.
On the other hand, the advantage of multiple correspondence analysis is
that you can add more than two factors. The ability to capture the interaction
of more than two different factors should not be underrated. In linguistics,
lexical structure, syntactic structure, prosodic structure, argument structure, as
well as region, gender, register and so on, are all potentially and interde-
pendently relevant in language structure and its description. Although it is
possible to combine and / or concatenate factors as one normally does for
cluster analysis and binary correspondence analysis, doing so can lead to
overlooking important interactions in the data. This last point is important and
warrants explanation.
If we are looking at, for instance, the interaction of tense, aspect, mood
and a range of near-synonymous verbs, we may propose the hypothesis that
the grammatical semantics will indirectly reveal lexical semantic structure. In
other words, the grammatical semantic profile of each verb will be indicative
of the lexical semantic structure. To these ends, it is perfectly possible to
combine the different grammatical factors, giving us usage features such as:
feature 1 ‘present tense + indicative + perfective’, feature 2: ‘present tense +
indicative + imperfective’, feature 3: ‘present tense + conditional + perfec-
tive’, feature 4: ‘present tense + conditional + imperfective’ and so forth.
However, if we then want to add further semantic or sociolinguistic features,
we may miss potentially important correlations. Although combining factors
in this manner will permit us to perform cluster analysis and binary corre-
spondence analysis, we will not know if there are interactions between the
different factors. We may find that, for example, the conditional mood has an
important correlation with the imperfective aspect in a certain register. This
may be interesting in itself, but it may also severely bias the results if not
accounted for separately – for this given register, is the lexical-grammatical
correlation observed due to the conditional or the imperfect or a combination?
An answer to such a question is more difficult to discern in binary analysis.
Therefore, there is always a trade-off – binary correspondence analysis gives
more ‘reliable’ results and numerical indicators of explained variation, but it
can struggle to capture the interaction of more than two factors simultaneous-
ly.
Other than simple binary correspondence analysis, detrended correspond-
ence analysis and canonical correspondence analysis have been developed.
Detrended correspondence analysis includes a bias added to the distance ma-
trix calculation. It is designed to counter a well-known effect with ‘long’ gra-
dients, called the ‘horseshoe’ effect, whereby the data points tend to form an
arch. The effect is occasionally visible in plots, but any experience with cor-
respondence analysis should avoid misinterpreting results because of it.3 It
should be noted that Greenacre (1984: 232) is sceptical about detrended cor-
respondence analysis and it does not enjoy wide currency. It is, nevertheless,
straightforward to perform in R.4
Canonical correspondence analysis, also termed constrained correspond-
ence analysis, is popular in the life sciences, but is also directly relevant to
linguistics. For the reader familiar with mixed effects regression modelling,
the idea is similar. In a given study, it is perfectly common to be dealing with
two different kinds of variables. Some categories interact with each other, but
all relative to a different kind of category. For example, aspectual structure
interacts with Aktionsart and tense in complicated and close ways. The role
of, for example, register, in their interaction is of a different nature. We may
not want to have the correspondence analysis treating the register features of
conversation, news press and literature equally ‘mixed’ in with aspectual and
temporal features. We can, therefore, treat the register dimension as an ‘ex-
ternal’ factor and the grammatical semantics as the ‘internal’ factors. The
correspondence analysis then knows that we are actually interested in the
internal factors and it accordingly attempts to map that space relative to the
structure of the other. This results in less explained inertia overall, but (hope-
fully) more explained inertia for the factor that is the object of study.5
Multiple correspondence analysis techniques are an extension of binary
correspondence analysis for the treatment of multi-way tables (binary corre-
140 Dylan Glynn
spondence analysis is restricted to ‘normal’ two-way tables). In the R sessions

below, we consider three kinds – indicator matrix multiple correspondence
analysis, Burt matrix multiple correspondence analysis, and joint multiple
correspondence analysis.
Indicator multiple correspondence analysis is sometimes called homoge-
neity analysis (Gifi 1990, De Leeuw & Mair 2009a). It uses a ‘binary matrix
of indicators’ (dummy variables) to ‘combine’ the binary correspondence
analyses. Despite mathematical differences, its results are very similar to Burt
matrix multiple correspondence analysis. Indeed, Greenacre (2007: 141)
compares the two and concludes that there is no difference in the visualisation
of the results, but that the Burt matrix produces more ‘optimistic’ percentages
of inertia. However, for multiple correspondence analysis, it must be remem-
bered that the percentages of explained inertia cannot be interpreted because
they severely underestimate the representative quality of the biplot map. Nev-
ertheless, the Burt matrix multiple correspondence analysis is the most com-
monly implemented in R and what we will employ here.
A third type of multiple correspondence analysis is based on the Burt ma-
trix method and has been termed joint correspondence analysis. Greena-
cre (2006: 68; 2007: 145) argues that it is superior both in terms of the ex-
plained inertia and in the accuracy of the visualisation. It works by restricting
the analysis to the cross-tabulations that typically contain the correlations of
interest, those that explain the inertia. Greenacre (2007: ch. 19) explains the
technique in clear terms.
2. Performing and interpreting correspondence analysis in R
Before we begin with the application per se, we must cover a few general
questions that are relevant to every correspondence analysis. The first im-
portant question is – what to look for. There are four issues: ‘fishing’, over-
simplicity, over-complexity, and data sparseness. Let us briefly consider each
in turn.
By fishing, we mean the arbitrary (or near-arbitrary) selection of factors
in the hope that one will find correlations. Correspondence analysis is a tool
for identifying correlations, a tool that needs to be used in a reasoned fashion.
There is no point in establishing correlations between the use of language
features that bear no interpretable correlation in reality, or worse, bear an
interpretable correlation, but are just a result of few a chance occurrences. In
section 1.1, the metaphor of an archaeologist digging was used to explain this
point: by digging everywhere, it is sure that something will be found, but the
chances of finding irrelevant things increase exponentially.
Over-simplicity is less serious a problem, but still must be borne in mind.
There is no use in using correspondence analysis to identify a correlation that
a simple pie chart or histogram, combined with a test for significance, would
do even better. Similarly, obvious correlations can dominate results at the
expense of less obvious, and therefore, more interesting results. For example,
in the case of a correlation between first person singular uses of a mental
predicate and parenthetical uses of a mental predicate – since the vast majori-
ty all parenthetical uses will be in the first person, entering these factors
(grammatical person and parentheticality) will reveal an obvious correlation.
The problem is, if these two factors are amongst a more complex range of
factors, the obvious correlation could ‘override’, or ‘hide’, other correlations.
Although it is sometimes necessary to leave such obvious correlations in an
analysis because one is seeking structures in other parts of the data, if it is
possible to avoid doing so, then it should be avoided. Obvious correlations
run the risk of ‘hiding’ the more interesting results. In other words, the plot
will identify what is most strongly correlated instead of the subtler, yet ana-
lytically more important correlations.
Over-complexity occurs in binary correspondence analysis when using
concatenated tables (see section 2.2.3 below) and in a multiple correspond-
ence analysis when too many factors are examined simultaneously. For ex-
ample, there is obviously no point analysing, simultaneously, 22 factors, each
with 16 features, even if one has thousands of examples. Without even con-
sidering the impossibility of accounting for the variation (inertia), in such a
dataset, the results would not be interpretable for the simple reason that the
visualisation of so many factors becomes impossible to decipher. Moreover,
the chance of ‘false’ associations increases dramatically with the more varia-
bles and features that are considered simultaneously. There is no steadfast
rule, but thinking about how the analysis works and being realistic about its
limitations are the safest ways to avoid the problem of over-complexity.
One way to avoid such over-complexity is to work with subsets. Subsets
may be logical divisions within the data: for example, examining two dialects
independently from one another or examining two lexemes or grammatical
constructions separately. Similarly, certain features or factors can be com-
bined. As long as the choice is reasoned and reported, it can help to simplify
the interactions that the analysis is trying to explain.
This principle extends to data sparseness and ‘small cells’. As a rule of
thumb, one aims to have at least ten examples in each cell of the cross-
142 Dylan Glynn
tabulated matrix (see below, this section). Obviously, this is not always pos-
sible, but cells of less than eight tend to cause distortions in the analysis. One
may find that the analysis is ‘trying so hard’ to account for some relatively
infrequent use, that the important associations are not represented. A response
to this problem is to leave out the examples (the rows in a flat data-frame, see
below, this section) that contribute features only occurring a few times. First
performing the correspondence analysis on the full set of data and then grad-
ually taking out these small cells (rows of infrequent examples) is a good
heuristic. Not only will it result in a better final analysis, the exploratory na-
ture of correspondence analysis will help you to better understand the data
and the correlations within them. The numerical output of binary correspond-
ence analysis can be very helpful in identifying such problems. Using the
numerical output, one can quickly see which data points are being poorly
represented. This is explained in section 2.3.1.
Let us now turn to the computation and interpretation of correspondence
analysis in R. Some common packages for correspondence analysis include:
{MASS}, {ca}, {languageR}, {anacor}, {homals}, {FactoMineR}, {ve-
gan}, {ade4} and {pamctdp}. Unfortunately, for reasons of brevity, we re-
strict the demonstration to a small selection of functionalities in the first four
of these packages. However, references to further information and tutorials on
each are offered.
Each package is a suite of commands for performing correspondence
analysis. They have different options and possibilities. The program R, works
with functions, such as the function to read a table, to plot results of an analy-
sis and, of course, to perform a statistical analysis. In simple terms, the func-
tions are the commands that tell R what to do. Each function also has a set of
‘arguments’. These arguments are the ‘options’ that R should take into ac-
count in executing the command. They should be carefully typed – spaces
have no effect, but capitals commas, brackets, and so forth must be entered
exactly. Moreover, keeping a record of what you have done is vital in learn-
ing to use the program. There are lots of additions in R for keeping your
working history and also for storing the functions you use often. However,
when just beginning, it is perhaps simplest, to use a text file and to simply
‘copy’ and ‘paste’ to and from R. Also at the end of an R session, it is wise to
save the history (what you have done) either within R or in a separate text
file. This will help you to remember the steps you took the next time you
perform an analysis.
In the R sessions below, after each line of command another short line is
added, following the # sign. This sign indicates that the program R should
ignore what follows it and not try to interpret it as arguments belonging to the
function. It is standard practice to explain command lines after such hash (#)
signs.
For the purposes of explaining how to perform and interpret the analyses
in R, we will use artificial data. Let us take a range of near-synonymous verbs
in an imaginary language. In this language, let us say, there are three mental
predicates think, believe, and suppose, and three communication predicates,
say, speak, and talk, which can be used figuratively to also indicate epistemic
stance, just like the mental predicates. We take 575 occurrences of the verbs,
more or less equally distributed. Correspondence analysis does not require
equal distribution in such a situation, but we want to have as many examples
as possible of each form, so making a balanced selection fulfils this require-
ment. The imaginary language possesses a three-way distinction in the aspect-
mood system, distinguishing between ‘Perfective’, ‘Imperfective’ and ‘Mod-
al’ forms. Each of the examples is analysed for this grammatical category.
The examples are also analysed for the grammatical person of the verb and
the semantic type of the indirect object. Table 1 illustrates kind of results one
expects from such an analysis.
Table 1. Example of flat data-frame

Example Verb Gram. Category Person Ind. Obj. Semantics
example1 think Perfective 1st Human
example2 suppose Modal 3rd Concrete_Thing
example3 suppose Perfective 3rd Abstract_State_of_Affairs
example4 believe Imperfective 1st Concrete_Activity
example5 say Imperfective 3rd Abstract_State_of_Affairs
example6 talk Modal 1st Concrete_Thing
example7 suppose Imperfective 1st Concrete_Activity
example8 speak Perfective 1st Human
to 575 examples … … … …
Before we start the R session, an important aside must be made. There are
two different data formats that the R functions use. It is crucial that the data
be in the correct format. Details on loading the data can be found in van de
Weijer & Glynn (this volume), but this fact is essential enough that it is worth
repeating. We can call one format the flat ‘data-frame’ and the other numeri-
cal ‘cross-tabulation’ (or contingency table). The data-frame is typically what
one obtains after annotating (coding) linguistic examples in Excel, Filemaker
or some database application. The cross-tabulation is a result of calculating,
144 Dylan Glynn
or numerically summarising, the data-frame. This can be done in Excel with

the ‘pivot table’ option. The easiest way to understand the two different for-
mats is by way of example. Table 1 is an example of a data-frame; typically it
would be in a spreadsheet or in a text file with tab delimited columns. Table 2
is typical of a cross-tabulation. Note that Table 2 only cross-tabulates col-
umns two and three: “Gram. Category” by “Verb”. Again, van de Weijer &
Glynn (this volume), offers more information on the different formats and
how to convert data from one format to another. Once we have our data and
we have the two relevant formats of the data, we can begin the correspond-
ence analysis. Note also that the data, R sessions and commands (with more
detailed explanations) can be downloaded from XXXXXXX.
Table 2. Example of a numerical cross-tabulation contingency table

believe think suppose say speak talk
Perfective 32 28 22 16 20 14
Imperfective 24 24 34 42 49 44
Modal 44 52 48 29 26 27
2.1 R Package {MASS}
The package {MASS} (Venables & Ripley 2002) comes pre-installed and so it
only needs to be loaded. It has simple, but effective, functions for both binary
correspondence analysis and multiple correspondence analysis. The first step
is to load the package:
> library(MASS)
2.1.1 Binary correspondence analysis in MASS

Data must first be loaded in the numerical cross-tabulated format. We take the
cross-tabulation in Table 2. The first command line below loads data from a
text file containing a cross-tabulation. There are three ways of loading such
data into R, we will use the choose.file() function (cf. van de Weijer &
Glynn [this volume]).
> data.xtab <- read.table(file.choose(),

header= TRUE, sep="\t", row.names= 1)
# Loads data from text file containing cross-tabulation
# e.g.: table2.txt
# Specifies that the columns have labels, or ‘headers’.
This command line loads the data into R, calls it ‘data’ and specifies that the
first row in the table is the column labels, or headers (header= TRUE). It also
assumes that any blank space, tab or otherwise, is the sign of a new field. You
can use function sep= “ ” to specify how the columns are separated. This is
necessary if you have labels with blank spaces in them. If you are exporting
data from a spreadhseet or database, then the columns are most likley to be
tab delimited, in which case, add the argument sep= “\t”. When first be-
ginning, it is perhaps easiest to make sure there are no blank spaces in your
labels and let R guess the structure of the table. The argument row.names= 1
specifies that the first column is the labels for the rows. This last argument is
needed when loading a numerical cross-tabulation, not for the data-frame.
The function for performing a binary correspondence analysis in the
package {MASS} is corresp.
> ca_analysis <- corresp(data.xtab, nf= 2)

# performs correspondence analysis on ‘data’
> plot(ca_analysis)
# plots results of ‘ca_analysis’
Although the graphic options in R are excellent, the {MASS} package offers
only a simple, yet efficient, set of possibilities. To the last line of code above
(plot(ca_analysis)), we can add the graphic ‘arguments’ which determine
the appearance of the plot.6 For example, the argument cex changes the size
of the font, the argument col specifies the colours, and xlim and ylim
‘zoom’ the plot by delimiting the x and y axes. These arguments should be
added ‘inside’ the plot command, as shown in the example below:
> plot(ca_analysis, col= 1, cex= 1,

xlim= c(-.225,.375), ylim= c(-.4,.4))
# plots results of ca_analysis, in black, at font size 1,
‘zoomed’ on x and y-axes. See Figure 1.
The col argument can specify a range of colours by name “red”, black”,
“blue” or by numbers, “1”, “2”, “3” etc. The argument cex takes a number
that indicates the type size (1 = is default, 1.2 is larger, 1.3 larger still, and 0.8
smaller, and so forth).
Zooming can be tricky at first, but it is simple when mastered. The plots
present numbers on the x and y axes. These numbers show the distance from
the centre of the plot. The x and y axes can be limited both in the negative and
in the positive range with the following arguments xlim= c(-.05,.05),
ylim= c(-.05,.05). Adding this string to the plot function will make the
146 Dylan Glynn
cut-off points for the plot -0.05 and +0.05 on the x-axis and -0.05 and +0.05
on the y-axis. Change those numbers to delimit, and therefore ‘zoom’, the
plot. Experimenting with the zoom function will allow you to get a more leg-
ible plot. In Figure 1, data points were positioned to make maximum use of
the space, that is, to make sure that the entire box was used to display the
ordination of the data points.
The data submitted to this analysis are extremely simple, but the result
will allow us to understand the principles in interpreting correspondence bip-
lots. The dispersion of the data points represents the variation of the co-
occurrence of the different usage features – here six verbs (columns) and
three grammatical categories (rows). Proximity and distance represent de-
grees of correlation between the different features. The centre of the plot,
indicated by the numbers on the x and y axes and by the cross in the centre,
divides the plot into quadrants. This helps identify association.
-0.2 0.0 0.2 0.4
0.4
0.6
0.4
0.2
0.2
Modal suppose
think say talk
0.0
0.0
Imperfective
speak
believe
Perfective
-0.2
-0.2
-0.4
-0.6
-0.4
-0.2 -0.1 0.0 0.1 0.2 0.3
Figure 1. Binary correspondence analysis, function corresp, package{MASS}
The corresp function in {MASS} does not produce the most attractive plots,
but since it is the simplest to perform and comes pre-installed in R, we will
explain the principles of interpretation with it. In the following sections, you
will see how more sophisticated, yet still simple to use, packages produce
superior biplots.
In Figure 1, we see two distinct groupings of the verbs, distinguishing the
mental predicates and the communication predicates. The verb suppose lies
between the two groups, though clearly still on the same side of the plot as
the other mental predicates. From this, we know that although it behaves like
a mental predicate, its use is closer to the communication predicates. We also
see that although the mental predicates form a distinct group, they are also
distinguished along the horizontal axis. The items believe and think are dis-
tinguished by the grammatical categories of Modal and Perfective. Although
the distance between think and believe is relatively small, they are in different
quadrants of the plot, and most importantly, the grammatical categories are on
the ‘far side’ of the lexeme data points relative to each other. This shows the
association to be distinctive. If the data points Modal and think were inter-
changed, then believe would be distinctly associated with Perfective, but think
would only be associated with Modal, not distinctly so. It is for this same
reason that we know there is a distinct association between the communica-
tion predicates and the Imperfective. The data point for the Imperfective lies
on the ‘other side’ of the communication predicates data points relative to the
mental predicates. This shows the distinctiveness of the Imperfective use with
this group.
A note should be made about the scales printed on the axes. They are not
informative on their own, but help one to gauge relative distance. This is es-
pecially important when plots are not square, but are elongated or stretched to
permit the representation of all the data points. In Figure 1, we have a slightly
unusual situation where the plot is skewed. In this instance, it is not interfer-
ing with the results, but if the cloud of data points were more dense or the
array more complex, this would have to be taken into account.
We can summarise the interpretation of the plot as follows:
For the mental predicates

- Distinct usage in the aspect-mood system, relative to the communication
predicates
- Within the group: believe is distinctly associated with the Perfective;
think is distinctly associated with Modal; suppose is associated with the
Modal, but its use is also relatively close to that of the communication
predicates
148 Dylan Glynn
For the communication predicates

- Distinct in usage from the mental predicates due to their Imperfective
profiling
The numerical output for the {MASS} package is not helpful at this stage. We
will examine the numerical output below in section 2.3.1.
2.1.2 Multiple correspondence analysis in MASS

This function is delightfully simple and powerful, but therein lies its danger.
It is simple because the data from a flat data frame (such as the ‘raw’ data in
an excel file) do not need to be converted to a contingency table, and also
because two simple commands perform the analysis and plotting. It is power-
ful because (theoretically) one can add many variables to the analysis and
perform complex multivariate analysis. Although this technique and these
functions are excellent, care must be taken. It becomes increasingly difficult
to interpret the results of the analysis as one increases the number of variables
being treated. What is more, it becomes increasingly difficult to obtain relia-
ble results.
For demonstration purposes, let us add some more fictitious data. Column
5 in Table 1 shows the semantic type of the indirect object of the utterances.
We can add this factor and perform a multiple correspondence analysis. It
will show the interaction of three variables, the grammatical category of as-
pect – mood, the lexemes, and the semantic type of the indirect object.
The function mca performs a multiple correspondence analysis. For this
function, there is no need to produce a numerical contingency table, the func-
tion accepts the flat data-frame as the input format of its data:
> data.fr <- read.table(file.choose(), sep="\t", header= TRUE)

#loads data from text file containing data-frame
> mca_analysis <- mca(data.fr, abbrev= T)
# performs the multiple correspondence analysis
> plot(mca_analysis, rows= F, col= 1)
# plots results of mca_analysis, see Figure 2
We add abbrev= T (T for ‘true’) to the mca function. This tells the function
not to include the factor labels. Also, note row= F (F for ‘false’); this tells the
plot function not to add the row numbers (that is example numbers, in lin-
guistic analysis) to the plot. It is sometimes interesting to plot the row num-
bers in order to determine which examples are causing the interactions visual-
ised in the plot. As for the factor labels, if one has features whose labels are
the same for different factors, then one needs to add the factor labels to dis-
tinguish them on the plot. To activate these options, exchange F and T in the
command line.
Figure 2 presents the results of the multiple correspondence analysis of
the three factors. Firstly, you will notice that some of the data points overlap.
This can cause problems for interpretation. This is a natural result of visualis-
ing association through the proximity of data points, yet it means that often
one must enlarge plots (after having saved them as image files) or zoom in on
them in R, in order to discern what data points are overlapping. These, in
turn, must be explained and described in further detail when reporting results.
There exists a package, {FactoMineR}, which has an option for so called
dynamic graphing which allows one to move the labels (as opposed to small
data points) interactively so that they do not over lap. Details are given in
section 2.5. It is this packages that was used to make the plots in descriptive
studies presented in volume 2.
0.03
Abstr_SoA
believe
0.02
0.01
Perfect
Cncrt_thing
speak talk
Abstr_thing
0.00
suppose
Modal Imperfective
think
-0.01
Cncrt_activty
-0.02
Human
-0.03
say
-0.02 0.00 0.02 0.04
Figure 2. Multiple correspondence analysis, function mca, package{MASS}
Interpreting the plot of a multiple correspondence analysis can be complex. In

this plot, we have a multidimensional space that has been conflated to two-
150 Dylan Glynn
dimensions. This means that data points may appear close to each other but,
in fact, are placed far apart on the oblique dimension, back ‘into the page’, as
it were. Caution, and a little experience, makes interpreting such plots reason-
ably straightforward. However, as one increases the numbers of factors to
four or five, the plots can only be used as a rough guide and one must return
to the data to check every correlation visualised.
In Figure 2, above, imagine the plot divided diagonally, horizontally and
vertically. Locating the centre at the intersection of 0.00 and 0.00, we move
out to see the three grammatical features, Imperfective, Perfective, and Mod-
al, dividing up the dispersion of the plot. Once again, the Imperfective domi-
nates the right side, where two of the communication predicates are located,
talk and speak. With them is the indirect object semantic feature of Concrete
Thing (Cncrt_thing). The position of the Imperfective data point between the
centre and the two communication predicates suggests that having added the
semantic features makes this less distinctly associated with the communica-
tion predicates. This interpretation would still see the Imperfective as a char-
acteristic feature of the communication predicates, being located in the right
half of the plot, but would see its distinctiveness being lessened. Although at
first sight, this may seem reasonable, herein lies the trick of interpreting mul-
tiple correspondence analysis.
Another interpretation, and one more likely to be accurate, is that the Im-
perfective is still highly distinctive of the communication predicates, but it is
being drawn to the centre by the third communication verb, say, which is now
on the bottom left side of the plot. It is probable that there is a multiple inter-
action here along this dimension of use. Adding the indirect object semantics
has separated say from the group of communication predicates. Seeing that its
position on the plot almost overlaps with the indirect object semantic feature
of Human, but also that these two features cluster together a long way from
the centre of the plot, we can safely suppose that say is highly associated with
a Human indirect object. We also know, from the previous analysis, that the
Imperfective is highly associated with say. In this situation, a likely interpre-
tation would be that both the Imperfective and the Human indirect object are
associated with say, but that the Human Indirect object is ‘pulling’ the lexeme
away from its Imperfective – communication verb cluster, leaving the Imper-
fective data point stretched between say and the two other lexemes, speak and
talk. This happens because the Human indirect object must be highly associ-
ated with some other feature and / or highly disassociated with the other
communication predicates. If this were not the case, it too would group with
the Imperfective on the right of the plot. This interpretation is complex, but
rich. It is also reasonably clear, if one has some experience in interpreting

correspondence biplots. We will return to some of these complex associations
below and reconsider these relations using different packages and different
ways of handling the data.
Other findings include the distinct association of believe and the indirect
object semantics of Abstract State of Affairs (Abstr_SoA). Also, the position
of Concrete Activity (Cncrt_activity) as object semantics, lying between say
and think / suppose, shows that these three lexemes share this feature, yet
otherwise they are distinct.
Despite this complexity, if we step back and return to our research ques-
tion – the near-synonymy of the lexemes – we have a clear and coherent pic-
ture. The two sets of verbs, communication and mental, are still distinct. For
the mental predicates, believe is distinct relative to the communication predi-
cates. For the communication predicates, say is less so, relative to the mental
predicates. This gives us an uneven continuum from talk and speak to say,
which bridges the use with the mental verbs think and suppose. At the very
end of this continuum, believe is distinctly distant from the communication
predicates in its use.
Importantly, we know what characterises this continuum. At one end, the
communication predicates, speak and talk, are associated with Concrete
Things as indirect objects and with the Imperfective aspect. Although still
associated with Imperfectivity, midway on the continuum, say is more associ-
ated with the indirect objects of Concrete Activity and Human. At this point,
we meet the mental predicates with suppose and think sharing an association
with say of the object type of Concrete Activity. These lexemes, think and
suppose, are central to the mental predicate cluster determined by Abstract
Things as indirect objects and the Perfective aspect. Finally, believe, also
relatively close to the other mental predicates, is quite distinct due to its asso-
ciation with Abstract States of Affairs as object semantics.
In this description of near-synonymy, we have shown which verbs are
similar and which verbs are distinct. Most importantly, we have shown why
this is the case; what usage features are responsible for the similarities and
differences. With some experience, such an interpretation of the results is
reasonably clear. Of course, how these results inform the interpretation of
language structure would still be open to debate.
152 Dylan Glynn
2.2 R Package {languageR}
The {languageR} package, developed by Baayen (2011), has an impressive

range of statistical options. However, for correspondence analysis, it is re-
stricted to binary analysis. To these ends, it is simple to use and has an agree-
able graphical output. The application is explained in Baayen (2008: 128-
136). The package needs to be installed and then loaded separately:
> library (languageR)

Depending on your version of R and the operating system, it may require a
range of other packages to be installed and loaded before it can be loaded.
The R terminal will give you instructions to follow in case this happens. It is
a good practice to always select the option “install dependencies” when
downloading / installing R packages. This should make sure that other pack-
ages needed are also installed and will then load automatically.
2.2.1 Binary correspondence analysis in languageR

The function for binary correspondence analysis is corres.fnc. It expects
the data to be input as a numerical cross-tabulation.
> data.xtab <- read.table(file.choose(), header= TRUE,

sep="\t", row.names= 1)
> data.t <- t(data.xtab)
# transposes data, not necessary, added here to improve
plot legibility
> ca_analysis <- corres.fnc(data.t)
> plot(ca_analysis, ccex= 1.2, rcex= .85)
# plots ca_analysis results with different font sizes for
the two axes. See Figure 3.
> title("Lexical Variation of Grammatical Categories",
cex.main= .95, sub= "package - languageR, function -
corres.fnc", cex.sub= 0.85)
# Adds titles to the plot, specifying font size
Firstly, note that the cross-tabulation was transposed, or inverted (data.t <-
t(data)). This is simply because some of the data point labels were ‘hang-
ing’ off the edge of the plot, and transposing the table inverts the plot and
thus the direction of the labels, improving legibility. The title function
should be self-explanatory. It is entered after the plot line and adds the label-
ling to the plot. It works for most packages. There are a great many more
functionalities with plotting and labelling that we do not cover here. In R, if
one wishes to find what arguments (options) are available for a given func-
tion, one should type a question mark and then the function (e.g.: ?title), a
help page will appear with all the arguments available.
Interpreting Figure 3 should be straightforward, the only important differ-
ence to Figure 2 being the inversion. It is superior in its representation to the
plot produced in {MASS}. The four quadrants are clearly indicated and the
relationship between the different data points much more clearly depicted.
Lexical Variation of Grammatical Categories

0.15
Perfective
believe
0.10
speak
0.05
Factor 2 (7.4 %)
0.00
Imperfective
say
talk think
-0.05
Modal
suppose
-0.10
-0.4 -0.2 0.0 0.2 0.4
Factor 1 (92.6 %)
package - languageR, function - corres.fnc
Figure 3. Binary correspondence analysis,

function corres.fnc, package{languageR}
An important addition is the percentages indicated on each of the two axes.

These percentages indicate the amount of inertia (see section 1.2) that is ex-
plained by the first two dimensions, the plotted dimensions. It is an indication
of how well the analysis is able to account for the variation in the data and is
normally reported. Low explained inertia does not mean that the analysis is
154 Dylan Glynn
not valid, but it does mean that extra care should be taken in interpreting the
plot. It is difficult to indicate a figure that represents a ‘good’ level of ex-
plained inertia because it depends on the complexity of the data. Normally, in
simple binary correspondence analysis, the combination of the first two di-
mensions should be over 75%. This will often be lower for canonical corre-
spondence analysis and, as stressed above, for multiple correspondence anal-
ysis, the score is not interpretable.
2.2.2 Chi-square test for significant variation

Let us make a brief aside. In chapter 1, the Chi-square test was explained. We
can apply it here to make sure that there is, indeed, significant variation be-
tween the different lexemes. The function is chisq.test and it expects the
data to be a numerical cross-tabulation.
> x2.test <- chisq.test(data.xtab, correct= F)

> x2.test
data: data
X-squared = 40.6513, df = 10, p-value = 1.301e-05
This shows there is significant variation between the lexemes. We can then
call the Pearson residuals to see which categories are causing the most varia-
tion. Again, this is explained in chapter 1.
> x2.test$res
believe think suppose say speak talk
Perfect 1.887480 0.844261 -0.383690 -0.888823 -0.387302 -1.24804
Imperfect -2.236471 -2.433997 -0.837797 1.599814 2.195817 2.104913
Modal 0.748989 1.739817 1.114179 -0.888356 -1.855656 -1.108766
We see from the Pearson residuals that the Imperfective uses of speak and
talk and the lack of Imperfective uses of think and believe are the most im-
portant. We can bear this in mind when interpreting the plots. Looking at
Figure 3, we see this fact represented visually and clearly.
2.2.3 Concatenating tables and combining categories

We have, until now, worked with very simple data. Although correspondence
analysis can help us find correlations in such data, its true strength comes out
when applied to more complicated and multivariate data. In Table 1, we pre-
sented the imaginary feature analysis: there were four columns after the ex-
amples – lexeme, grammatical category, grammatical person, and indirect
object semantics. We saw, in section 2.1.2, that a multiple correspondence
analysis can help examine more than two factors simultaneously, but we also
saw the complexity in interpretation that arises in some data sets. Although
multiple correspondence analysis is necessary for finding correlations be-
tween the factors, there are two other ways of handling data that permit some
exploration of different dimensions of use.
First of all, using the pivot command in Excel, one can concatenate, or
‘stack’ tables. One literally generates two-dimensional tables and puts them
together in a row, creating one long table. For instance, we can take the data
for verb and grammatical category and, using the so-called ‘pivot’ function,
produce a cross-tabulation. Then we repeat the operation for the verb and
indirect object semantics. This gives us two tables we can join and submit to a
binary correspondence analysis. Further information on how to handle data is
presented in the Appendix, First Steps in R.
Table 3 is an example of such concatenated results. It is perfectly ac-
ceptable to tabulate data in this manner, but it must be remembered that when
used in multivariate statistics, it has conflated two conceptually different di-
mensions of language structure – the indirect object semantics and the gram-
matical categories of aspect and mood. Combining the factors in this way
means we will only be able to see the interaction of these categories relative
to the different verbs, not the interaction between them. If we are only inter-
ested in the verbs and know there are no important interactions between indi-
rect object semantics and mood-aspect, then this poses no problems.
Table 3. Example of concatenated cross-tabulation

Verb Imperfect Modal Perfect Abstr. Abstr. Cncrt Cncrt Human
SoA thing activity thing
believe 24 44 32 41 37 8 5 9
say 42 29 16 5 25 17 8 32
speak 49 26 20 5 22 8 51 9
suppose 34 48 22 5 62 15 16 6
talk 44 27 14 1 9 4 68 3
think 24 52 28 15 37 16 13 23
Table 4, below, shows another way of getting more information into a table.
Here we have combined two different factors. In Table 1, grammatical person
and verb were listed as separate columns. However, conceptually, it is per-
fectly reasonable to combine these two factors into one, a verb-person catego-
ry. To obtain the new combined factor, ‘sort’ the two columns in Excel and
add a new empty column, create the new combined factor using the ‘copy-
paste’ and ‘change all’ functions in Excel. Table 4 shows this combined fac-
tor of verb and person added to the cross-tabulation.
156 Dylan Glynn
A point of warning should be made about data sparseness. As we add

more complexity to the tables, we obtain smaller cells. The more different
things we consider simultaneously, the ‘thinner’ the data becomes. As a rule,
cells below a count of 8 should be avoided. In practice, at least when dealing
with manually coded data, this rule is broken. As long as care is taken when
considering the plots and it is remembered that this is only an exploratory
technique, some leeway on this front can be tolerated. However, small cells
should always be reported. The numerical output, which we consider below in
section 2.3.1, also helps one gauge the reliability or ‘accuracy’ of a data point
on a biplot. Let us now consider these tables submitted to binary correspond-
ence analysis.
Table 4. Example of concatenated cross-tabulation with combined factors

Verb + Imperf. Modal Perfect Abstr. Abstr. Cncrt Cncrt Human
Person SoA thing activity thing
believe 1st Prs 18 31 12 28 23 4 3 3
believe 3rd Prs 6 13 20 13 14 4 2 6
say 1st Prs 42 15 4 3 16 10 8 24
say 3rd Prs 0 14 12 2 9 7 0 8
speak 1st Prs 44 13 9 4 16 5 36 5
speak 3rd Prs 5 13 11 1 6 3 15 4
suppose 1st Prs 28 10 0 0 22 5 9 2
suppose 3rd Prs 6 38 22 5 40 10 7 4
talk 1st Prs 37 10 8 1 7 4 42 1
talk 3rd Prs 7 17 6 0 2 0 26 2
think 1st Prs 21 11 8 3 20 3 6 8
think 3rd Prs 3 41 20 12 17 13 7 15
Aplying the same command line as that used to produce Figure 3 to the data
presented in Table 3, produces the plot in Figure 4, below. Although the plot
is inverted, the dispersion of the data points has shifted a little, and the graph-
ical representation in this package is superior, the results are the same. In-
deed, the fact that the Imperfect is being ‘stretched’ between two quadrants of
the plot is clearer. Note also that the data point for Abstract States of Affairs
lies just off the plot, distinctly and highly associated with believe, just as in
Figure 2. It is also noteworthy that the explained inertia of the first two di-
mensions is 85% (dim 1: 63% + dim 2: 22.4%). This shows that the analysis
is stable and we can interpret the plot with some confidence.
0.6
Human
SAY
0.4
Cncrt_activty
0.2
Factor 2 (22.4 %)
Imperfect
THINK
SUPPOSE
Abstr_thing
0.0
Modal SPEAK
Perfect TALK
Cncrt_thing
-0.2
-0.4
BELIEVE
-0.6
Abstr_SoA -0.5 0.0 0.5 1.0
Factor 1 (63 %)
Figure 4. Binary correspondence analysis, from Table 3, function corres.fnc,

package{languageR}
Figure 5, below, visualises yet a more complex data set, presented in Table 4.
In these data, we have combined, or added, the grammatical person to the
verbs as well as concatenated the grammatical categories and the indirect
object semantics. The added complexity reduced the amount of explained
variation, which is now just over 72% for the first two dimensions. This is
still a relatively high figure and assures us that the plot remains stable, despite
the added complexity.
We see that speak and talk behave in a similar manner to what we saw
above in Figure 4. It seems that the addition of the variation in grammatical
person does not affect their interaction with the aspect-mood and object se-
mantics to any great extent.
158 Dylan Glynn
1.0
TALK_3rdPrs
0.5
Cncrt_thing
SPEAK_3rdPrs
Perfect
TALK_1stPrs
Factor 2 (21 %)
Modal
THINK_3rdPrs
SUPPOSE_3rdPrs
BELIEVE_3rdPrs
Abstr_SoA
SAY_3rdPrs
0.0
BELIEVE_1stPrs SPEAK_1stPrs
Cncrt_activty
Abstr_thing
THINK_1stPrs
Human Imperfect
SUPPOSE_1stPrs
-0.5
SAY_1stPrs
-0.5 0.0 0.5 1.0
Factor 1 (51.5 %)
Figure 5. Binary correspondence analysis, from Table 4, function corres.fnc,

package {languageR}
However, the behaviour of say, which we saw was distinct in Figure 4, has
been explained. The 1st person say is found on the Imperfective / communica-
tion verb side of the plot, but the 3rd Person usage of say is found right in the
centre of a mental predicate cluster, with the Modal and Perfective profilings.
We now know that it was a simplification to understand say as being between
the communication predicates and the mental predicates. It is, in fact, only the
3rd person uses that behave similarly to the mental predicates. Moreover, we
see now that believe has joined the mental predicate cluster, which suggests
that it was never so distinct from the mental predicate cluster as a whole. In-
stead, we see that it was just distinct from the 1st person uses of say, from
which it was being pushed away in the visualisation. Finally, suppose in the
1st person has shifted right across to the Imperfective, completing the picture
of a continuum between the two groups of verbs, where suppose in the 1st
person behaves like a communication verb and say in the 3rd person like a
mental predicate. Adding the extra dimensions of use has clarified the interac-
tion of the verbs. It is precisely this kind of multivariate complexity that cor-
respondence analysis is designed to capture.
If the reader wishes to perform these analyses, the command line present-
ed in section 2.2.1 will produce the plots.
2.3 R Package {ca}
The package {ca}, developed by Nenadić & Greenacre (2007) and Greenacre
& Nenadić (2010), is another commonly-used package for performing binary
and multiple correspondence analysis. The package does not come with the R
installation and must be downloaded separately using the package installer. It
is described in detail in Greenacre (2007: 232-240). Before we begin, load the
package:
> library (ca)
This package offers a wealth of possibilities, most of which go beyond our

discussion. We will focus on its numerical output and the various options for
multiple correspondence analysis that it includes.
2.3.1 Binary correspondence analysis in ca

As before, the data must be loaded in a cross-tabulated matrix.
> data.xtab <- read.table(file.choose(), header= T, sep="\t",

row.names= 1)
> ca_analysis <- ca(data.xtab)
> plot(ca_analysis, col= 1)
Due to the limitations of space, the plot of this analysis is not included. It
presents the same information as above. However, the numerical output in
{ca} is comprehensive and informative and so we will focus on this. Alt-
hough most of the output does not need reporting, as one becomes more expe-
rienced with correspondence analysis, the mass and explained inertia for the
individual rows and columns can help one interpret unusual patterns, espe-
cially with data sets more complex than those we are using here. There are
160 Dylan Glynn
two sets of numerical output. The first is obtained by simply typing the object
of the ca function. We called this object ca_analysis. This output is not
presented here because it is quite voluminous. The second numerical output is
obtained by asking for a summary of the results of the analysis. The summary
below is of the analysis presented, above, in Figure 5.
> summary(ca_analysis)
Principal inertias (eigenvalues):
dim value % cum% scree plot

1 0.252400 51.5 51.5 *************************
2 0.103030 21.0 72.5 **********
3 0.069058 14.1 86.6 *******
4 0.043973 9.0 95.6 ****
5 0.016743 3.4 99.0 **
6 0.004762 1.0 100.0
7 00000000 0.0 100.0
-------- -----
Total: 0.489967 100.0
Rows: DELETE
name mass qlt inr k=1 cor ctr k=2 cor ctr
1 | BELIEVE_1 | 106 383 133 | -484 382 98 | -19 1 0 |
2 | BELIEVE_3 | 68 659 71 | -575 642 89 | 94 17 6 |
3 | SAY_1 | 106 737 111 | 170 56 12 | -592 681 361 |
4 | SAY_3 | 45 604 62 | -636 601 72 | 42 3 1 |
5 | SPEAK_1 | 115 953 84 | 583 953 155 | -17 1 0 |
6 | SPEAK_3 | 50 760 27 | 106 42 2 | 438 718 94 |
7 | SUPPOSE_1 | 66 651 66 | 353 254 33 | -442 397 125 |
8 | SUPPOSE_3 | 115 524 88 | -428 485 83 | 121 39 16 |
9 | TALK_1 | 96 971 152 | 843 914 269 | 211 57 41 |
10 | TALK_3 | 52 910 99 | 574 354 68 | 719 556 262 |
11 | THINK_1 | 70 698 21 | 21 3 0 | -322 695 70 |
12 | THINK_3 | 111 775 85 | -517 717 118 | 147 58 23 |
Columns:
name mass qlt inr k=1 cor ctr k=2 cor ctr
1 | Impf | 189 958 199 | 574 638 246 | -407 320 303 |
2 | Modl | 197 764 73 | -311 534 76 | 204 230 80 |
3 | Prfc | 115 740 85 | -410 462 77 | 318 278 113 |
4 | A_SA | 63 427 165 | -740 425 136 | 43 1 1 |
5 | Abs_ | 167 408 77 | -237 248 37 | -191 160 59 |
6 | Cncrt_c | 59 373 33 | -279 289 18 | -151 84 13 |
7 | Cncrt_t | 140 998 271 | 845 753 396 | 481 244 315 |
8 | Humn | 71 326 97 | -223 74 14 | -411 252 117 |
The summary call begins with what it labels a ‘scree plot’. Scree plots are
used to help decide how many dimensions are needed to explain the variation
in the data. In principal components analysis, factor analysis and also in mul-
tidimensional scaling, such ‘plots’ are common. They offer a factor-by-factor
breakdown of how much variation the analysis has explained. One looks for
an ‘elbow’ in the plot, that is, a dimension where there is a marked drop in the
amount of variation explained. There are no mathematical rules to decide this
point, but typically it is clear – ‘most’ of the variation in the dispersion of
data is explained by only a few of the dimensions. The more gradual the ‘de-
scent’ of the scree plot, the more trouble the analysis is having in explaining
the dispersion of the data.
The table in the summary call, ‘Principal inertias’, contains the vital in-
formation for understanding the structure of a correspondence analysis. It
begins with the dimensions (dim), then lists the Eigenvalues (value), con-
verts these to percentages of explained variation (%), and then calculates the
cumulative explained variation with the addition of each dimension (cum%).
Since biplot visualisations of the results of a correspondence analysis typical-
ly depict the first two dimensions, the numerical output here tells us that the
first two dimensions explain 72.5% of the inertia. This means the plot that we
interpret does not account for just over one quarter of the variation in the data.
This information is a guide to how confident we can be about the accuracy of
the depiction. At 72.5%, less variation is explained than in the previous (sim-
pler) analyses, but this figure is still sufficiently high to interpret the plot,
though with some caution. The scree plot shows us that it might be informa-
tive to also visualise the third-dimension, either in a three-dimensional plot or
by producing two more biplots, with dimension 1 by dimension 3 and dimen-
sion 2 by dimension 3. Many of the R packages offer these possibilities, but
we do not consider them here.
Unlike {languageR}, the {ca} package does not automatically label the
plots with the amount of explained inertia. If one wishes to label the x and y
axes with the percentages of explained inertia, one simply uses the title func-
tion as above, for example: title(xlab= "Dim 1 (51.5%)", ylab= "Dim 2
(21%)").
In the second half of the summary we have two tables of information, one
for the rows and one for the columns of the contingency table that is the basis
of the analysis. Coordinates are only given for the first two dimensions (k=1
and k=2), the dimensions visualised in a biplot (note the plots produced in
languageR call these factors 1 and 2). Normally, in correspondence analysis,
interpretation is restricted to these first two dimensions. This table breaks
down the analysis for you. For each row and each column in the table, the
weight assigned to that column or row is indicated (mass). This was explained
in section 1.2. It is essentially a bias added to the calculation to stop small
numbers having a disproportionate effect.
162 Dylan Glynn
The next score listed is the quality (qlt) and, as the name would suggest,
this is a measure of the accuracy of the visualisation. This is a very useful
score to consider. A low quality score for any given ‘feature’, that is, row or
column, means the interpretation of its position on the plot should be treated
with extra care. The numbers are given in thousandths, so a figure of 375
would be 37.5% and it indicates the explained inertia for a given row or col-
umn (that is the labels on the plot). So, in the table above, the representation
in the plot of TALK_1 would be 97.1% accurate, where BELIEVE_1 would
be only 38.3% accurate. A quality score of less than 500 (50%) would suggest
that the position of data point in question does not necessarily accurately rep-
resent the relation of that feature to the others. This is often because a given
feature is common to a wide range of different situations, it correlates with
distinct phenomena. For example, it may be equally used in the past tense and
in the future tense, two tenses which are otherwise distinct in the analysis. In
such situations, the data point will lie close to the centre of the plot, the inter-
section of the two axes. The other situation is mathematically similar, but
analytically different. In situations where there are only few occurrences of a
given feature, and those few occurrences behave in different ways, the same
effect is obtained. In the latter situation, if it possible to do so with loosing too
much data, these examples can be omitted.
To understand why this is the case, we need to think about how the bi-
plots work. The plot is a representation of a complex n-dimensional set of
associations in just two dimensions. Therefore, the points of the labels are not
in their original, or mathematically true, positions, having been moved to
enable a two-dimensional representation. In other words, their true position is
better described by one of the axes not represented in the plot. By default, the
biplots present the first two dimensions, but recall that the actual number of
dimensions is the number or rows or columns (whichever is less), minus one.
So in the table above, we have seven dimensions (the columns, minus one).
Greenacre (2007: 87) explains this in greater detail, but the principle is that
the quality score here is the inertia score, explained above, broken down for
each row and column (or plot label / data point).
The inertia value (inr), to the right, is used to calculate the quality. But it
can also be directly interpreted. The figure listed is the contribution of that
row or column (feature) to explaining the total inertia. It is expressed in thou-
sandths, so that in the output above, BELIEVE_1 explains 13.3% of the iner-
tia in the analysis. So, given that the plot captures 72.5% of the inertia (distri-
bution / variation in the data), this particular feature accounts for nearly 20%
of the structure of the plot (13.3 / 72.5 x 100).
The next two sections of the table, to the right, give the correlation (cor)
and the contribution (ctr) for each of the two dimensions. These scores are,
perhaps, less useful in most circumstances, but warrant explanation. The ctr
is the contribution that a given row or column has made to explaining the
inertia along a single principal axis, that is, one of the first two dimensions.
For instance, the horizontal axis (k-1) in Figure 5 is largely determined by
three features, TALK_1, SPEAK_1, and THINK_3. The correlation scores
indicate the correlation between a principal axis and the row or column in
question. Typically, though not necessarily, a high correlation between a fea-
ture and an axis indicates high contribution to that axis.
2.3.2 Multiple correspondence analysis in ca

The {ca} package is one of the richest for multiple correspondence analysis.
Its main strength lies in the fact that one can choose which of the three kinds
of multiple correspondence analysis one wishes to perform. Moreover, one
can choose to automatically adjust the inertias (using Greenacre’s (2006)
calculation for adjustment). This option is only available for the Burt matrix
multiple correspondence analysis. The function for multiple correspondence
analysis is mjca. To this, one can ‘add’ the argument lambda where one spec-
ifies the kind of multiple correspondence analysis to perform: the indicator or
homogeneity analysis with lambda= “indicator”; the Burt matrix analysis
with lambda= “Burt”; the joint analysis with lambda= “JCA”, and the Burt
analysis with Greenacre adjusted inertia values with lambda= “adjusted”.
To perform and compare the different techniques, we use the data from
the previous analyses, but we do not use the concatenated table. Multiple
correspondence analysis may find important interactions between the aspect-
mood grammatical semantics and the indirect object semantics, and so we
must keep these two dimensions separate. We will, however, keep the com-
bined categories of verb and grammatical person.
For practical reasons, we cannot consider the plots of all four possibilities.
The joint multiple correspondence and the adjusted inertia Burt multiple cor-
respondence analysis produce the best results, both in terms of graphical out-
put and explained inertia. This is to be expected since, as mentioned above,
one cannot normally interpret the explained inertia scores in multiple corre-
spondence analysis.
> data.fr <- read.table(file.choose(), sep="\t", header= T)

> mca_indicator_analysis <- mjca(data.fr, lambda= "indicator")
> summary (mca_indicator_analysis)
164 Dylan Glynn

1 0.587336 10.4 10.4 *************************
2 0.479181 8.5 18.8 *******************
3 0.457792 8.1 26.9 ******************
4 0.442143 7.8 34.7 *****************
5 0.397579 7.0 41.7 ***************
6 0.358378 6.3 48.0 *************
7 0.333333 5.9 53.9 ************
8 0.333333 5.9 59.8 ************
…
The output here has been abbreviated. Importantly, we see only that 18.8% of
the inertia is explained. This is to be expected in multiple correspondence
analysis and need not be reported.
We will not consider the results of the Burt multiple correspondence
analysis. It suffices to point out that the explained inertia using the Burt ma-
trix is slightly better at 27.1%, but again this is unrealistically pessimistic.
The command for the Burt multiple correspondence analysis is:
> mca_Burt_analysis <- mjca(data.fr, lambda= "Burt")
The two most promising advances in multiple correspondence analysis are

certainly the joint analysis and the adjustment to inertias in the Burt matrix
analysis. Let us consider these. We begin with the joint correspondence anal-
ysis (Figure 7). By not including the diagonals, which offer little information
to the analysis, we greatly improve the explained inertia and also the projec-
tion onto the two-dimensional plane, the biplot. The command for the joint
multiple correspondence analysis is:
> mca_joint_analysis <- mjca(data.fr, lambda= "JCA")

> plot(mca_joint_analysis, labels= c(0, 2),
col= c("white", "black"))
# plots analysis with only data point labels, see Fig. 7.
The plot function includes two arguments that we have not seen before. The
command labels= c(0,2) hides the row numbers (which correspond to the
number of one’s actual language example in the raw data set). Obviously,
upon occasion, it is important to see which examples are causing the disper-
sion in a plot, especially when looking for exemplary occurrences in linguis-
tic discussion and result reporting. The second argument, col= c("white",
"black") hides the co-occurrence points. This can be used at times to show
how the co-occurrence of features is projected across the plot. For instance,
sometimes there is a single kind of co-occurrence that causes a given feature

to be pushed away from a group. Being able to add these points can help in
plot description. Finally, the function psch= changes the symbol representing
the point on the plot. Wide ranges of symbols exist and can be modified by
changing the number. We cannot consider the full extent of the graphic op-
tions here, but the reader is encouraged to experiment with the plot function.
Verb.TALK_3rdPrs
Verb.SPEAK_3rdPrs ObjSem.Cncrt_thing
0.5
Verb.SAY_3rdPrs
Verb.THINK_3rdPrs
Verb.SUPPOSE_3rdPrs
TM.Perfect
Verb.BELIEVE_3rdPrs
TM.Modal
Verb.BELIEVE_1stPrs
0.0
Verb.TALK_1stPrs
ObjSem.Abstr_thing
ObjSem.Cncrt_activty
ObjSem.Abstr_SoA
Verb.SPEAK_1stPrs
TM.Imperf
ObjSem.Human Verb.THINK_1stPrs
-0.5
Verb.SUPPOSE_1stPrs
Verb.SAY_1stPrs
-1.0 -0.5 0.0 0.5 1.0
Figure 7. Joint multiple correspondence analysis

function mjca, package {ca}, method=JCA
The call for the numeric summary is as above:
> summary(mca_joint_analysis)
…
Diagonal inertia discounted from eigenvalues: 0.2324778
Percentage explained by JCA in 2 dimensions: 70.4%
…
166 Dylan Glynn
We only consider two lines of the numerical summary of the joint corre-
spondence analysis. As mentioned, joint analysis functions by removing the
diagonals of the analysis. These ‘intersections’ of the tables contribute little to
the explanatory power of the analysis. The first line shows us that we im-
proved our explanation of inertia by 23% through their removal. The second
line tells us that the explained inertia for the first two dimensions is 70.4%.
Such a score should be reported, given the necessary caveat that estimating
the explained inertia in a multiple correspondence analysis is normally unre-
alistically pessimistic and that this score is produced through a joint analysis.
Plot interpretation is no different to the example interpretations presented
above. We will not, therefore, interpret Figure 7, but will move onto the ad-
justed Burt analysis. The command line follows what we saw above. We
begin with a numerical summary of the analysis:
> mca_adjusted_analysis <- mjca(data.fr, lambda= "adjusted")

> summary(mca_adjusted_analysis)

1 0.145165 42.0 42.0 *************************
2 0.047861 13.8 55.8 ********
3 0.034852 10.1 65.9 ******
4 0.026639 7.7 73.6 *****
5 0.009287 2.7 76.2 **
6 0.001411 0.4 76.6
…
The above numerical summary is truncated. We see that the explained inertia
in the first two dimensions is 55.8%. This is relatively low, but for a multiple
correspondence analysis, we can still confidently interpret the biplot. For the
plotting, once again we hide the data points and example numbers. The com-
mand for plotting Figure 8 follows:
> plot(mca_adjusted_analysis, labels= c(0, 2),

col= c("white", "black"))
# Figure 8.
The plots in Figure 7 and Figure 8 show the same results. However, if one
compares the dispersion of the data points carefully, the spread is a little
clearer in Figure 8, confirming Greenacre’s (2006) results in comparing the
two methods. Moreover, the results here mirror those of the binary corre-
spondence analysis presented in Figure 5. This shows that there are no im-
portant interactions between the factors aspect-mood and indirect object se-
mantics. If there existed interactions between these parameters of usage, this

extra complexity would be seen here.
Verb.TALK_3rdPrs
0.4
Verb.SPEAK_3rdPrs
TM.Perfect
ObjSem.Cncrt_thing
Verb.BELIEVE_3rdPrs
0.2
Verb.TALK_1stPrs
Verb.SUPPOSE_3rdPrs
ObjSem.Abstr_SoA
Verb.THINK_3rdPrs
Verb.BELIEVE_1stPrs
TM.Modal
Verb.SAY_3rdPrs Verb.SPEAK_1stPrs
0.0
ObjSem.Abstr_thing
ObjSem.Cncrt_activty
-0.2
TM.Imperf
Verb.THINK_1stPrs
ObjSem.Human
Verb.SUPPOSE_1stPrs
-0.4
Verb.SAY_1stPrs
-0.6
-0.5 0.0 0.5 1.0
Figure 8. Burt matrix multiple correspondence analysis (with adjusted inertias)

function mjca, package {ca}, method= adjusted
The plot offers a coherent picture of the clustering of the mental predicates
with perfective aspect and modal uses. The exceptions are the first person
uses of think and suppose, which are similar to the communication predicates
due to their association with the Imperfective. Relative to the mental predi-
cates, the communication predicates talk and speak also form a cluster in the
top right quadrant. This cluster is less homogenous, being based vaguely upon
Concrete Things as object semantics and the Imperfect. The position of the
Imperfect between the top and the bottom of the right side of the plot shows
how it is drawn between the 3rd person say and the rest of the communication
verb cluster.
168 Dylan Glynn
2.4 R-package {anacor}
De Leeuw & Mair (2009a) developed an excellent package for simple binary
and binary canonical correspondence analysis. Although canonical analysis is
a useful type of correspondence analysis, described in section 1.3, we cannot
cover the technique here. Beyond its ability to perform canonical analysis, the
package offers a rich variety of scaling and plotting options. We will consider
one of these. The package must be first downloaded. At the beginning of the
R session it must also be loaded:
> library(anacor)
2.4.1 Binary correspondence analysis with confidence intervals in anacor

The package {anacor} offers a range of scaling techniques and plot types –
two invaluable additions to correspondence analysis. The different scaling
methods can be applied to the rows and columns independently. This can be
useful for revealing or highlighting different associations that are hidden in
the blur of the plot. The centroid and standard scalings are typically the most
useful in this regard. For some data sets, using a combination of the two dra-
matically increases the legibility of the results by setting one of the dimen-
sions in the centre and ‘surrounding’ it with the other dimension. Unfortu-
nately, we cannot cover this option, but the reader is strongly encouraged to
experiment with combining different scaling methods with different datasets.
Moreover, the package offers no fewer than seven two-dimensional plotting
options. We will consider one plot type, what they term the joint plot. Both
the plotting and scaling options are explained in detail in De Leeuw & Mair
(2009a). The data are loaded as a numerical cross-tabulation:
> data.xtab <- read.table(file.choose(), header= T, sep="\t",

row.names= 1)
> ca_analysis <- anacor(data.xtab, scaling= c("standard",
"standard"))
> plot(ca_analysis, plot.type= "jointplot")
# see Figure 9
The anacor function takes the argument scaling=, which specifies the scal-
ing method for the x and y-axis. The plotting command takes the argument
plot.type=, which specifies the type of plot desired. We have used the joint
plot, which includes confidence ellipsoids. These ellipsoids are not based on a
test for statistical significance, but estimate it using what the authors call the
delta method (De Leeuw & Patrick 2009a). The ellipsoids are set at a ‘signifi-
cance’ level of 95%, matching the alpha level of p<0.05, standard in the so-
cial sciences. Although they do not represent p-values, but confidence inter-
vals, they can be thought of as a kind of estimated significance value. The
plot reveals the same associations as the binary plots above, but the addition
of the confidence ellipsoids is a welcome advance and will prove extremely
useful for some datasets. For example, it here reveals that the associations
between suppose in the 1st person and say in the 1st person and the Imperfect
are almost surely significant. Of course, we must return to the data for specif-
ic tests of association and / or move to configurational frequency analysis and
log-linear regression for confirmatory results. The confidence ellipsoids are
merely further guides to help understand
Jointrelations
plot visualised in a biplot.
1.5
TALK_3rdPrs
1.0
Cncrt_thing
SPEAK_3rdPrs
Perfect
0.5
Modal TALK_1stPrs
Dimension 2
THINK_3rdPrs
SUPPOSE_3rdPrs
BELIEVE_3rdPrs
Abstr_SoA
SAY_3rdPrs
0.0
BELIEVE_1stPrs SPEAK_1stPrs
Cncrt_activty
Abstr_thing
-0.5
THINK_1stPrs
Human Imperfect
SUPPOSE_1stPrs
-1.0
SAY_1stPrs
-1.0 -0.5 0.0 0.5 1.0 1.5
Figure 9. Binary Correspondence Analysis, scaling

Dimension 1 “standard” vs. “standard”
function res, package {anacor}, plot type “jointplot”
The numerical output is basic but clear. The percentage of explained inertia
for the first two-dimensions, using the standard scaling, comes out at 82% – a
stable result:
> summary(ca_analysis)
170 Dylan Glynn
z-test for singular values:

Singular Values Asymptotical SE p-value
D1 0.5024 0.0237 0
D2 0.3210 0.0282 0
2.5 Other R-packages for correspondence analysis
We have only treated some of the functionalities of the packages presented.

However, hopefully, enough detail in both the workings of R and the applica-
tion and interpretation of correspondence analysis has been covered to allow
the reader to delve further into the method and the packages presented. We
have omitted five packages that need to be mentioned. Armed with the expla-
nations above, these other packages should be approachable even for people
new to R. We will now briefly consider each of these packages.
R-package {homals}
De Leeuw & Mair (2009b), the developers of the {anacor} package, present-
ed above, also author {homals}. This package has even more powerful
graphic options than {anacor}. Not only does it offer joint plots and star
plots, but there is also the option of static and interactive three-dimensional
plotting. The interactive plotting allows one to turn the plot as an object in
space, in order to obtain the optimal viewing point or to see how the data
points are related when they are hidden behind each other. The R code is sur-
prisingly simple, though there are large numbers of dependent packages, so
make sure ‘install dependencies’ is selected when downloading / installing
{homals}.
R-package {vegan}
The {vegan} package, Oksanen (2006) and Oksanen et al. (2011), permits
the detrended correspondence analysis described in section 1.3. It also has
excellent ordination graphics. It offers the option to ‘build’ correspondence
plots, using a text function, to add labels for categories, rows, or columns,
one at a time. This option is excellent for complex datasets with large num-
bers of categories. The analysis is performed on the entire set, but only certain
data points are labelled, facilitating interpretation and reporting.
R-package {ade4}
The R package {ade4} (Dray & Dufour 2007) is, in fact, an impressive suite
of functions developed (and being developed) for the environmental sciences.
Many of the techniques available in the package are useful for linguists. The
suite includes a range of options for performing different kinds of corre-
spondence analysis as well as for plotting both the ordinate results and nu-
merical output.
R-package {FactoMineR}
A rich and powerful package, {FactoMineR} (Lê et al. 2008) performs prin-
cipal components analysis (a form of correspondence for continuous data),
binary correspondence analysis and multiple correspondence analysis. One of
its main advantages is an argument in the plot function invisible= which
allows one to hide certain rows or columns. Although it is possible in vegan
to build up a plot by adding rows and columns iteratively, this simple tool
makes it easy to quickly see data points that are hidden, but also to remove
complexity to aid in reporting. It also posseses interactive possibilities, ena-
bling the user to move the labels on the plots to improve legibility, but also to
hide certain clusters of data points and even to select certain clusters and
zoom in on them. For some of these functionalities, one needs to download
third party (yet free) software. The correspondence analyses presented in the
second volume use this package.
R-package {pamctdp}
The package {pamctdp} (Pardo 2010) offers a range of tools for dealing with
contingency tables and controlling the rows and columns in correspondence
analysis. One function that has wide application is the ability to produce
barplots of the profiles of the rows or the columns of a correspondence analy-
sis. This simple visualisation technique can help with reporting complex data
sets. Producing such plots ‘manually’ in R is straightforward, but time con-
suming.
3. Choice – correspondence or cluster
There are many statistical techniques and their number is growing. One of the
most confounding hurdles for anyone beginning to use quantitative tech-
niques is knowing which techniques are possible for a given data type and
which are most suitable for a given research question. This section outlines,
as briefly as possible, a technique comparable to correspondence analysis and
offers information on how to choose between them.
172 Dylan Glynn
3.1 Hierarchical cluster analysis
As an exploratory technique for categorical data, correspondence analysis

shares a great deal with cluster analysis (presented in the preceding chapter)
in that it visualises data that are interpretable in an intuitive way by ‘catego-
rising’ features relative to their occurrence with other features. Both tech-
niques are obviously the kind of tools useful to linguists, and indeed, all so-
cial scientists.
Despite their similarities, there are important differences. The most obvi-
ous difference lies in the visual representation of the results. The graphical
representation of a correspondence analysis can be more difficult to interpret
than the dendrograms typical of cluster analysis, but offers important ad-
vantages over the latter.
This visual representation and interpretation is at once correspondence
analysis’ strength and weakness. Firstly, the correspondence configuration
biplots do not give the false impression that any observed ‘clustering’ is dis-
crete, which is an unfortunate side effect of a poorly interpreted dendrogram.
Of course, the discrete visualisation in a dendrogram means that the algorithm
performs the ‘grouping’, not the interpreter (arguably more reliable). But
correspondence is not about groupings; it is about associations in the data.
This brings us to the second, more important, advantage – the configura-
tion biplot shows the interaction of the different features of the different fac-
tors, rather than merely sorting the features of a single factor. Correspondence
analysis reveals what is correlating with what, in other words, which features
of which variables are co-occurring with others. This is in contrast to cluster
analysis, where one only sees how the features of one factor are grouped, not
what features are responsible for that grouping. This representation of results
is both more complex and more ‘analogue’, and therefore, arguably, a more
socio-conceptually realistic representation of how different linguistic struc-
tures interact.
It should be obvious that added complexity / detail and an analogue / sub-
jective interpretation can pose serious problems in correspondence analysis.
The plots can be extremely difficult to read and interpret accurately. If the
goal of a study is to determine which features of a given dimension are simi-
lar and which are different, cluster analysis is a simple and powerful tech-
nique. However, if the goal is to understand the interactions of different di-
mensions of language use and structure, then correspondence analysis is
worth the extra effort
Cluster Dendrogram
0.15
20 Perfective
believe
0.10
speak
15
0.05
Factor 2 (7.4 %)
Height
0.00
Imperfective
10
suppose
say
talk think
-0.05
believe
think
speak
Modal
5
-0.10
suppose
say
talk
0
Figure 10. Hierarchical cluster analysis, -0.4

Figure 11 -0.2 0.0
Binary correspondence 0.2analysis 0.4
dist_mat
dendrogram, function
Cluster hclust , package
hclust (*, "single")
dendrogram with {MASS
AU/BP values (%) } biplot, function corres.fnc
Factor 1 , package
(92.6
Joint plot %) {languageR}
au bp
70
edge #
60
0.2
50
Perfective
believe
speak
40
Dimension 2
Imperfective
Height
0.0
say
30
talk think
Modal
suppose
20
90 82
4
100 96
-0.2
10
72 73
3 2
100 96
suppose
1
0
speak
believe
think
say
talk
-0.4
-0.4 -0.2 0.0 0.2
Figure 12. Hierarchical cluster

Distance: analysis,
euclidean Figure 13 Binary correspondence
Dimension 1 analysis
dendrogram, functionCluster method: ward
pvclust, package {pvclust} joint plot, function anacor, package {anacor}
3rdP.TALK
0.6
3rdP.think 3rdP.suppose
1stP.beleive CrtThing
3rdP.SPEAK
0.4
Perfect
3rdP.talk
3rdP.speak
Modal 1stP.TALK
0.2
3rdP.THINK
3rdP.say 3rdP.SUPPOSE
1stP.think 3rdP.BELIEVE
3rdP.beleive AbsSoA 3rdP.SAY
1stP.suppose
0.0
1stP.BELIEVE 1stP.SPEAK
CrtActivty
AbsThing
-0.2
1stP.say
1stP.THINK
1stP.SUPPOSE Imperfect
-0.4
1stP.speak Human
1stP.talk
1stP.SAY
-0.6
Figure 14. Hierarchical cluster analysis, Figure 15.

-0.5 Multiple0.0correspondence
0.5 analysis1.0
phylogenetic tree, function nj, package {ape} biplot, function mjca, package {ca}
.
174 Dylan Glynn
Let us move to a comparison of the visual outputs. Figures 10-15, above, set
out, side by side, the graphical representations resulting from the same data in
the two techniques. Figure 10 is a dendrogram of a hierarchical agglomerative
cluster analysis of the data presented in Figure 3, which is duplicated here as
Figure 11 to aid in comparison. The cluster analysis in Figure 10 was per-
formed using the package {MASS} and the function hclust. The distance
matrix used is Euclidean and the agglomeration method for clustering “aver-
age”. The functions expect the data to be in a numerical cross-tabulation (con-
tingency table) format. The command line is:
> data.xtab <- read.table(file.choose(), header= T, sep= "\t",

row.names= 1)
> data.dist <- dist(data.xtab, method= "euclidean")
# converts frequency table to distance matrix
> HCA <- hclust(data.dist, method= "average")
# performs the cluster analysis
> plot(HCA, frame.plot= T)
# Plots results and adds frame to plot, see Figure 10.
The comparison between Figure 10 and Figure 11 is self-explanatory. The

biplot and the dendrogram present similar information differently. The main
difference is that in the biplot, we see which features cause the ‘clustering’.
The dendrogram in Figure 12 is produced in the package {pvclust} with
the function pvclust. It also uses the Euclidean distance matrix, but this time
the agglomerating method is Ward. See Divjak & Fieller (this volume) for an
explanation of these concepts. Importantly, this function for cluster analysis
includes a set of bootstrapped p-values to offer confidence estimates for the
clusters.7 The confidence scores are labelled faintly next to the branches in
the clusters. The R-code for this technique follows. Again, the functions ex-
pect the data to be in a cross-tabulation format.
> library (pvclust)

row.names= 1)
> data.t <- t(data.xtab) # inverts the data
> PVClust <- pvclust(data.t, method.hclust= "ward", meth-
od.dist= "euclidean")
# produces distance matrix, performs cluster analysis and
bootstraps the results. Can take some time to process
> plot(PVClust, frame= T)
# Plots results, see Figure 12
The addition of the bootstrapped p-values is an important contribution to the

analysis. Compare the results with Figure 13, a binary correspondence analy-
sis with a similar estimation of significance included. The confidence ellip-
soids representing significant p-values are designed to capture similar infor-
mation. Both visualisations seem reasonably successful.
The third cluster analysis, presented in Figure 14, is not in the form of a
dendrogram but of an unrooted cluster or ‘phylogenetic’ tree. This representa-
tion has been used by Schmidtke-Bode (2009) and Divjak (2010). This form
functions poorly for only six features. Though it distinguishes them well, the
sparseness of the graphics leaves a lot to be desired. Therefore, the data from
Table 4 are used. The function for this kind of plot is nj and it is found in the
{ape} package.
> library (ape)

row.names= 1)
> data.dist <- dist(data.xtab)
# converts frequency table to distance matrix
> PhyloClust <- nj(data.dist)
# performs the cluster analysis
> plot(PhyloClust, type = "u", frame= T, cex= .9)
# Plots results, see Figure 14
The phylogenetic tree in Figure 14 is compared with a multiple correspond-

ence analysis in Figure 15. With this more complex data, we see clearly how
the two methods differ in their abilities to represent data structure. The phylo-
genetic tree is simple and intuitive in contrast to the correspondence analysis,
which needs detailed explanation. However, we know from the discussion
above that the correspondence analysis includes more information, infor-
mation missing in the cluster analysis.
4. Summary
There is a growing range of packages for performing correspondence analysis

as well as new and improved tools in R generally. Moreover, new lines of
statistical research are sure to bring interesting options for the technique in
the near future. Within statistics, there is work on implementing permutation,
or resampling, tests for correspondence analysis, as well as developing and
applying mathematical algorithms that better capture certain kinds of correla-
tions in the data – the development of grade correspondence analysis (Kow-
176 Dylan Glynn
alczyk et al. 2004) and so-called profile-based approaches (Speelman et al.

2003, Delaere et al. submitted.) are examples at hand. Within Cognitive Lin-
guistics, broadly speaking, the method of correspondence analysis has been
applied by Arppe (2006), Szelid & Geeraerts (2008), Plevoets et al. (2008),
Glynn (2009, 2010, in press), Krawczak & Glynn (2011), Glynn &
Sjölin (2011) and Kokorniak & Krawczak (in press).
There exists a good range of resources for learning more. J. P.
Benzécri originally developed the technique in the late 1960s. In English, his
work of reference is Benzécri (1992). More recently, Agresti (2002: 382-384)
and Greenacre (1984) have championed the method. These works are difficult
to approach for linguists, being concerned with the mathematics behind the
technique, rather than its application and interpretation. In recent years, a
range of volumes has appeared that can be used as manuals for performing
the analysis. As mentioned above, Greenacre (2007) and Husson et al. (2011)
are excellent manuals for both understanding the technique and performing it
in R. Along these lines, Baayen (2008: 128-136) also offers a brief descrip-
tion. Le Roux & Rouanet (2010) is an excellent and affordable book: essen-
tially it constitutes a detailed tutorial for the function-rich package {Fac-
toMineR}, though it offers no R commands. Lastly, although older, Weller &
Romney (1990) is another thorough and approachable alternative, but it also
offers no R code. More advanced, yet still reasonably practical publications
include Rencher (2002: ch. 15), Murtagh (2005), and Le Roux & Rouanet
(2005). Also consider Greenacre & Blasius (2006), which is a collection of
articles that seek to advance different facets of the method. Greenacre (2007:
appendix C) offers an annotated bibliography. For specifically information on
the plotting options, see Gower et al. (2010) and Greenacre (2010).
Correspondence analysis and the current range of packages for perform-
ing it in R offer a powerful and simple tool for identifying patterns in multi-
factorial data. The options for visualisation of its results can be difficult to
explain, but are extremely rich in the information that they display. It is an
exploratory method and an excellent heuristic for getting into complex data
and digging out what relates to what.
Notes
1. I would like to thank Koen Plevoets who first taught me to this technique and
Joost van de Weijer for his help polishing this paper. All shortcomings are my own.
2. Weller & Romney (1990: 57) suggest performing tests for independence, such
as those presented in chapter 1, to ascertain if there exists statistically significant
variation before one performs correspondence analysis. Although this is optional, it

may be a good guideline to avoid falling into the trap of thinking the correlations
identified are significant. It must be remembered, however, that obtaining statistical
significance is dependant upon the size of the sample. With a large enough sample,
very small variations will be significant, and with small samples (typical of semantic
research), significance is much less commonly obtained. This is, of course, how it
should be – it stops analysts making bold claims based on small samples.
3. See Greenacre (2007: 127-128) for an explanation of the horseshoe effect.
4. The package needed is {vegan} and the function is decorana. Oksanen (2006)
has written a clear tutorial for the {vegan} package that explains the needed R com-
mand and offers examples.
5. With the package {vegan} and the function cca, it is straightforward to perform.
Again, Oksanen (2006) offers an R tutorial on the possibilities. Canonical corre-
spondence analysis is also performed by the {anacor} package, explained in De
Leeuw & Mair (2008).
6. Saving plots is an important, yet often sidelined, element to using R. A full
explanation can be found in van de Weijer & Glynn (this volume). In Mac OSX, plots
are saved as .pdf, which gives perfect resolution magnified to infinity. Under Mac
OSX, in current versions of MSWord, importing with the Insert menu (not ‘cut and
paste’) will maintain this perfect resolution in an MSWord document. Under Win-
dows, depending on the version, MSWord does not accept .pdf or automatically con-
verts it with poor quality output. Under Windows, plots should be saved as .png.
Although the quality is not comparable to .pdf, it is more than acceptable.
7. Bootstrapping is a general statistical technique for obtaining an estimation of
significance. It is not a test for significance per se, but is rather a method that gener-
ates a large set of samples (from the sample under investigation) with which to com-
pare the results.
References
Agresti, Alan. 2002. Categorical Data Analysis (2nd ed.). Hoboken: John Wiley.
Arppe, Antti. 2006. Frequency Considerations in Morphology. Finnish Verbs Differ,
Too. SKY Journal of Linguistics 19: 175-189.
Baayen, R. Harald. 2008. Analyzing Linguistic Data: A practical introduction to
statistics using R. Cambridge: Cambridge University Press.
Baayen, R. Harald. 2011. languageR: Data sets and functions with "Analyzing Lin-
guistic Data: A practical introduction to statistics". R package version 1.1.
http://CRAN.R-project.org/package=languageR.
Benzécri, Jean-Paul. 1992. Correspondence Analysis Handbook. New York: Marcel
Dekker.
De Leeuw, Jan & Patrick Mair. 2009a. Simple and Canonical Correspondence Analy-
sis Using the R Package anacor. Journal of Statistical Software 31: 1-18.
http://www.jstatsoft.org/v31/i05/.
178 Dylan Glynn
De Leeuw, Jan de & Patrick Mair. 2009b. Gifi Methods for Optimal Scaling in R:
The Package homals. Journal of Statistical Software 31: 1-20.
http://www.jstatsoft.org/v31/i04/.
Delaere, Isabelle, Koen Plevoets, & Gert De Sutter. Submitted. Measuring text type
variation through profile-based correspondence analysis: How far apart are
translated and non-translated Dutch? Target. International Journal of Transla-
tion Studies.
Divjak, Dagmar. 2010. Structuring the Lexicon: A clustered model for near-
synonymy. Berlin: Mouton de Gruyter.
Dray, Stéphane & Anne-Beatrice Dufour. 2007. The ade4 package: Implementing the
duality diagram for ecologists. Journal of Statistical Software 22: 1-20.
Gifi, Albert. 1990. Nonlinear Multivariate Analysis. Chichester: Wiley.
Glynn, Dylan & Mette Sjölin. 2011. Cognitive Linguistic Methods for Literature. A
usage-based approach to metanarrative and metalepsis. In A. Kwiatkowska
(ed.), Texts and Minds. Papers in cognitive poetics and rhetoric, 85-102. Frank-
furt/Main: Peter Lang.
Glynn, Dylan. 2009. Polysemy, Syntax, and Variation. A usage-based method for
Cognitive Semantics. In V. Evans & S. Pourcel (eds.), New Directions in Cog-
nitive Linguistics, 77-106. Amsterdam: John Benjamins.
Glynn, Dylan. 2010. Synonymy, Lexical Fields, and Grammatical Constructions. A
study in usage-based Cognitive Semantics. In H.-J. Schmid & S. Handl (eds.),
Cognitive Foundations of Linguistic Usage-Patterns. Empirical studies, 89-118.
Berlin / New York: Mouton.
Glynn, Dylan. In press a. Cognitive socio-semantics. The theoretical and analytical
role of context in meaning. Review of Cognitive Linguistics.
Gower, John, Sugnet Gardner-Lubbe, Noël le Roux. 2010. Understanding Biplots.
Chichester: Wiley.
Greenacre, Michael & Jörg Blasius (eds.). 2006. Multiple Correspondence Analysis
and Related Methods. London: Chapman & Hall.
Greenacre, Michael & Oleg Nenadić. 2010. ca: Simple, Multiple and Joint Corre-
spondence Analysis. R package version 0.33.
http://CRAN.R-project.org/package=ca.
Greenacre, Michael. 1984. Theory and Applications of Correspondence Analysis.
London: Academic Press.
Greenacre, Michael. 2006. From Simple to Multiple Correspondence Analysis. In
M. Greenacre & J. Blasius (eds.), Multiple Correspondence Analysis and Relat-
ed Methods, 41-76. London: Chapman & Hall
Greenacre, Michael. 2007. Correspondence Analysis in Practice. London: Chapman
& Hall.
Greenacre, Michael. 2010. Biplots in Practice. Bilbao: Fundación BBVA.
Husson, François, Sébastien Lê, & Jérôme Pagès. 2011. Exploratory Multivariate
Analysis by Example Using R. London: Chapman & Hall.
Kokorniak, Ivona & Karolina Krawczak. In press. Subjective construal of THINK in
Polish. Pozna Studies in Contemporary Linguistics.
Kowalczyk, Teresa, Elżbieta Pleszczynska & Frederick Ruland (eds.). 2004. Grade
Models and Methods for Data Analysis. München: Springer.
Krawczak, Karolina & Dylan Glynn. 2011. Context and Cognition. A corpus-driven
approach to parenthetical uses of mental predicates. In K. Kosecki & J. Badio
(eds.), Cognitive Processes in Language, 87-99. Frankfurt/Main: Peter Lang.
Le Roux, Brigitte & Henry Rouanet. 2005. Geometric Data Analysis. From corre-
spondence analysis to structured data analysis. London: Kluwer.
Le Roux, Brigitte & Henry Rouanet. 2010. Multiple Correspondence Analysis. Lon-
don: Sage.
Lê, Sébastien, Julie Josse & François Husson. 2008. FactoMineR: an R package for
multivariate analysis. Journal of Statistical Software 25: 1-18.
Murtagh, Fionn. 2005. Correspondence Analysis and Data Coding with R and Java.
London: Chapman & Hall.
Nenadić, Oleg & Michael Greenacre. 2007. Correspondence analysis in R, with two-
and three-dimensional graphics: The ca package. Journal of Statistical Software
20(3). http://ww.jstatsoft.org/v20/i03.
Oksanen, Jari, Guillaume Blanchet, Roeland Kindt, Pierre Legendre, R. B. O'Hara,
Gavin L. Simpson, Peter Solymos, M. Henry H. Stevens & Helene Wagner.
2011. vegan: Community Ecology Package. R package version 1.17-11.
http://CRAN.R-project.org/package=vegan.
Oksanen. Jari. 2006. Multivariate Analysis of Ecological Communities in R: vegan
tutorial. http://cc.oulu.fi/jarioksa/softhelp/vegan.html.
Pardo, Campo. 2010. ‘pamctdp’. http://www.docentes.unal.edu.co/cepardot
Plevoets, Koen, Dirk Speelman & Dirk Geeraerts. 2008. The distribution of T/V
pronouns in Netherlandic and Belgian Dutch. In K. Schneider & A. Baron,
(eds.), Variational Pragmatics: Regional varieties in pluricentric languages,
181-209. Amsterdam: John Benjamins.
Rencher, Alvin. 2002. Methods of Multivariate Analysis (2nd ed.). Chichester: Wiley.
Schmidtke-Bode, Karsten. 2009. Going-to-V and gonna-V in child language: A
quantitative approach to constructional development. Cognitive Linguistics 20:
509-53.
Speelman, Dirk, Stefan Grondelaers & Dirk Geeraerts. 2003. Profile-based linguistic
uniformity as a generic method for comparing language varieties. Computers
and the Humanities 37: 317-337.
Szelid, Veronika & Dirk Geeraerts. 2008. Usage-based dialectology. Emotion con-
cepts in the Southern Csango dialect. Review of Cognitive Linguistics 6: 23-49.
Venables, William & Brian Ripley. 2002. Modern Applied Statistics with S (4th ed.).
London: Springer.
Weller, Susan & Kimball Romney. 1990. Metric Scaling Correspondence Analysis.
London: Sage.

Glynn 2012 CA Correspondence Analysis

Uploaded by

Copyright:

Available Formats

Glynn 2012 CA Correspondence Analysis

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Glynn 2012 CA Correspondence Analysis

Uploaded by

Copyright:

Available Formats

Correspondence Analysis

Exploring data and identifying patterns

1. A technique for visualising correlations in categorical data

Correspondence analysis is a multivariate exploratory space reduction tech-

1.1 Use – What does correspondence analysis do?

In their quotidian research, linguists, from all kinds of theoretical orientations,

1.2 Concept – How does correspondence analysis work?

Basically, correspondence analysis takes the frequency of co-occurring fea-

exploratory statistics, it is essentially the same as principal components analy-

Profiles and mass

Inertia and Variation

ing. These books are Greenacre (2007) Correspondence Analysis in Practice,

1.3 Choice - Binary and Multiple Correspondence Analysis

Correspondence analysis is, in fact, a family of techniques and a family that is

spondence analysis is restricted to ‘normal’ two-way tables). In the R sessions

2. Performing and interpreting correspondence analysis in R

Table 1. Example of flat data-frame

or numerically summarising, the data-frame. This can be done in Excel with

Table 2. Example of a numerical cross-tabulation contingency table

2.1 R Package {MASS}

2.1.1 Binary correspondence analysis in MASS

> data.xtab <- read.table(file.choose(),

> ca_analysis <- corresp(data.xtab, nf= 2)

> plot(ca_analysis, col= 1, cex= 1,

-0.2 -0.1 0.0 0.1 0.2 0.3

Figure 1. Binary correspondence analysis, function corresp, package{MASS}

For the mental predicates

For the communication predicates

2.1.2 Multiple correspondence analysis in MASS

> data.fr <- read.table(file.choose(), sep="\t", header= TRUE)

-0.02 0.00 0.02 0.04

Figure 2. Multiple correspondence analysis, function mca, package{MASS}

Interpreting the plot of a multiple correspondence analysis can be complex. In

rich. It is also reasonably clear, if one has some experience in interpreting

2.2 R Package {languageR}

The {languageR} package, developed by Baayen (2011), has an impressive

> library (languageR)

2.2.1 Binary correspondence analysis in languageR

> data.xtab <- read.table(file.choose(), header= TRUE,

Lexical Variation of Grammatical Categories

-0.4 -0.2 0.0 0.2 0.4

Figure 3. Binary correspondence analysis,

An important addition is the percentages indicated on each of the two axes.

2.2.2 Chi-square test for significant variation

> x2.test <- chisq.test(data.xtab, correct= F)

2.2.3 Concatenating tables and combining categories

Table 3. Example of concatenated cross-tabulation

A point of warning should be made about data sparseness. As we add

Table 4. Example of concatenated cross-tabulation with combined factors

Abstr_SoA -0.5 0.0 0.5 1.0

Figure 4. Binary correspondence analysis, from Table 3, function corres.fnc,

-0.5 0.0 0.5 1.0

Figure 5. Binary correspondence analysis, from Table 4, function corres.fnc,

2.3 R Package {ca}

> library (ca)

This package offers a wealth of possibilities, most of which go beyond our

2.3.1 Binary correspondence analysis in ca

> data.xtab <- read.table(file.choose(), header= T, sep="\t",