The Stata Journal (2005)

5, Number 2, pp. 208–223

Data inspection using biplots

Ulrich Kohler Magdalena Luniak
Wissenschaftszentrum Berlin Wissenschaftszentrum Berlin
[email protected] [email protected]

Abstract. Biplots display interunit distances, as well as variances and correla-

tions of variables of large datasets. They can be used as a tool to reveal clustering,
multicollinearity, and multivariate outliers, and to guide the interpretation of prin-
cipal component analyses (PCA). This article describes the uses of biplots and its
implementation in Stata.
Keywords: gr0011, biplot, biplot8, principal component analysis, exploratory data
analysis, multivariate statistics, euclidean distance, mahalanobis distance, relative
variation diagram, projection

[Editors note: This article was received before Stata 9 was announced. Stata 9 has
a biplot command, so the command documented here is named biplot8. biplot8
has some features not found in Stata 9 biplot (and vice versa). Additionally, the
exposition here acts as a helpful supplement to the Stata 9 biplot manual entry.]

1 Introduction
Biplots are projections of multivariate datasets that show the following quantities of a
data matrix:

• the variance–covariance structure of the variables

• the values of observations on variables

• the Euclidean distances between observations in the multidimensional space

They are helpful for revealing clustering, multicollinearity, and multivariate outliers
of a dataset, and they can be also used to guide the interpretation of principal component
analyses (PCA).
Biplots were first described thoroughly by Gabriel (1971) and were extended more
recently in a monograph by Gower and Hand (1996). They are heavily used in the
context of principal component analysis (Jolliffe 2002, 90–107) but also useful as a tool
for data inspection in the context of statistical modeling. As a projection technique,
they share similarities with many other projection techniques, such as multidimensional
scaling (Kruskal and Wish 1978), principal coordinate analysis (Fenty 2004), and cor-
respondence analysis (Blasius and Greenacre 1998).1
1 A discussion of the relative merits of several projection techniques can be found in

Schnell and Matschinger (1994), who recommend using biplots.

c 2005 StataCorp LP gr0011
U. Kohler and M. Luniak 209

In this article, we start with examples to explain the interpretation of biplots. We

then discuss the mathematical background and some computational issues. Finally, we
illustrate the uses of the Stata program biplot8.

2 Interpretation
Biplots consists of lines and dots. Lines are used to reflect the variables of the dataset,
and dots are used to show the observations. An example biplot is shown in figure 1,
which uses a dataset from Hamilton (1992, 268). The observations of this dataset are
planets, and the variables are their physical characteristics, for example the mass, the
number of moons, and the distance from the sun. With the exception of a dummy
variable for rings present, all variables are measured on a logarithmic scale.
1

Pluto
DIM 2 (16 % of Var)
.5

Neptune
logdist Uranus

logmoons
rings Saturn
lograd
0

Mars
logdens logmass
Jupiter

Earth
Mercury
Venus
−.5

−.5 0 .5 1
DIM 1 (82 % of Var)

Figure 1: Biplot of planets.dta

In a biplot, the length of the lines approximates the variances of the variables. The
longer the line, the higher is the variance. Inferring from figure 1 the logarithmic mass of
the planets (logmass) has by far the highest variance among the variables in the biplot,
while the dummy variable for rings present (rings) has the lowest.
The angle between the lines, or, to be more precise, the cosine of the angle between
the lines, approximates the correlation between the variables they represent. The closer
the angle is to 90, or 270 degrees, the smaller the correlation. An angle of 0 or 180
degrees reflects a correlation of 1 or −1, respectively. The biplot in figure 1 shows a
strong relationship between the ring dummy and the number of moons (logmoons),
and a weak relationship between the mass and distance from the sun (logdist). The
correlation between the density and each of the other variables is negative.
210 Data inspection using biplots

The cutpoint of a perpendicular from a specific point to a variable line approximates

the value of that observation on the variable that the line represents. If the cutpoint
falls on the origin, the value of the observation is approximately the average of the
respective variable. Cutpoints far off in the direction of the variable line indicate high
values, while cutpoints far off on the variable line, which has been extended through
the origin, represent low values. Therefore, Jupiter stands out with the highest mass,
followed by Saturn, and Neptune and Uranus, which have almost identical masses. Pluto
stands out as the planet with the lowest mass.
Finally, the distance between two points approximates the Euclidean distance be-
tween two observations in the multivariate space. Observations that are far away from
each other have a high Euclidean distance, and vice versa. In the example biplot, the
highest Euclidean distance is observed between Jupiter and Pluto, while Neptune and
Uranus are the other extremes.
Putting all these together, biplots reveal several characteristics of a dataset, which
are useful in the context of statistical modeling. First of all, you might be warned of
possible sources for multicollinearity, as for the variables rings and logmoons in the
biplot example in figure 1. Furthermore, biplots show multivariate outliers, such as the
planet Pluto. Finally, biplots can be used to detect clusters, such as the inner rocky
planets and the outer gas giants.
The latter two interpretations can be also found in a principal component score plot,
which is a common technique for plotting the results of a PCA (Hamilton 1992). In fact
(see section 4), for a certain type of the biplot, the scatter of observations is a principal
component score plot. In this special case, the positions of the observations approximate
the scores of the observations on the first two principal components, whereby the x- and
y-axes represent the first and second principal components, respectively.
Another useful application of biplots in the context of PCA is more obvious in the
biplot of the variables miles per gallon, price, weight, and displacement of auto.dta
in figure 2. As before, this plot reveals the correlation structure of the variables and
some clustering of observations. However, more important for here is the position of the
endpoints of the variable lines along the graph axes. The variables mpg, weight, and
displacement are relatively far from the origin along the x-axis but close to the origin
along the y-axis. For price, it is the other way around. These relative positions of the
variable lines represent the PCA coefficients (“loadings”) of the variables on the first two
principal components. Therefore, you might interpret the first principal component as
a consumption dimension and the second as a price dimension. In addition, looking at
the graph, you can conclude that a slight rotation of the axes of the PCA would improve
the ease of interpretation of both components.
U. Kohler and M. Luniak 211

4
price

DIM 2 (16 % of Var)

2
mpg

0
weight
displacement

−2
−4

−4 −2 0 2 4
DIM 1 (75 % of Var)
Domestic Foreign

Figure 2: Biplot of auto.dta

This interpretation of the biplot is similar to the interpretation of the plot of the PCA
coefficients, which is a common way to plot the results of a PCA (Tabachnik and Fidell
1989, 637–638). As for the principal component score plot, the plot of PCA coefficients
can be regarded as a special case of a biplot.

3 Mathematical background
Let Y be an n × k matrix holding the data. You can decompose Y with a singular value
decomposition (SVD) into
Y = ULV

where U is n × k, and both L and V are k × k. The elements of L, which is diagonal,

are the so called eigenvalues.
From the singular value decomposition, the coordinates of the observations are given
by
G = ULc (1)

and the coordinates for the variables are given by

H = L1−c V (2)

In (1) and (2), the scalar c can take any value between zero and one. Regardless of
the value of c, the equation
GH = ULc L1−c V = ULV = Y
212 Data inspection using biplots

always holds. However, as G is n × k and H is k × k, all the coordinates have k

dimensions. To plot these coordinates in a two-dimensional space, you must select two
of them. Usually this is done by choosing those columns of G and H that correspond to
the highest eigenvalues in L. This is the default setting in biplot8, but other settings
are possible (see section 5).
In any case, using fewer than k dimensions to plot the points will lead to a loss of
information, and the data matrix Y is only approximated by the multiplication of the
reduced forms of G and H. In effect, the interpretations shown in section 2 get less
valid if this approximation gets bad. To indicate the quality of the approximation, the
default axis titles mention the amount of explained variances by the selected dimensions.
Unless the sum of these explained variances is sufficiently large, “the interpretation of
the plot is suspect” (Jackson 1991, 199). However, there is no known boundary below
which the interpretation is erroneous. We have found explained variances of about 70%
enough to obtain good approximations of the key quantities for small datasets.
Choosing a value for c defines the coordinates for different types of biplots. Three
values for c are most commonly used and are therefore implemented in biplot8:

• c = 0, the GH, or column-metric preserving biplot

• c = 1, the JK, or row-metric preserving biplot
• c = .5, the SQ, or symmetric biplot

GH biplots are called column-metric preserving because the variance–covariance

structure of the variables is best approximated in the GH biplot. JK biplots, on the
other hand, are row-metric preserving, since the approximations of the Euclidean dis-
tances are optimal in this biplot. Finally, the SQ biplots represent the observational
values of Y better than the other types.

4 Computational issues
The Stata command to calculate a singular value decomposition is
. matrix svd U L V = Y
where Y is the name of the matrix that ought to be decomposed and U, L, and V are
arbitrary names for the resulting matrices of the SVD. To calculate the coordinates of
the biplot, this command requires that the complete data matrix be stored in Y. The
maximum dimension of a single matrix in Intercooled Stata is 800 × 800. In Intercooled
Stata, the SVD of a data matrix therefore can be only done for datasets with up to
800 observations. In Stata/SE, this limit is raised to 11,000 observations. Given that
there is no general maximum number of observations in Stata, the maximum number
of observations to be used in a biplot is restrictive2 .
2 In Stata 9, these limitations can be circumvented using Mata; see the Mata Reference Manual for

details.
U. Kohler and M. Luniak 213

In the case of the JK biplot (c = 1), the restriction can be circumvented. As Jolliffe
(2002, 94–95) shows, the elements in G are equal to the respective values of the ob-
servations on the principal components. Accordingly, the elements in H are equal to
the coefficients (loadings) of a PCA. Therefore, the coordinates of the JK biplot can be
easily calculated from a PCA, bypassing the calculation of the SVD. To this extent, the
biplot with c = 1 is nothing new since the component score plot and the plot of PCA
coefficients are widely used on their own. The superimposing of both plots, however,
gives additional information.
The possibility that you can calculate the plot coordinates by means of a PCA for
the JK biplot raises the question whether this is also possible for the other biplot types.
In fact, the coordinates of the JK biplot and the GH biplot are closely related. It follows
from the definition of both biplots and from (1) and (2) that

(GJK = UL ∧ GGH = U) ⇒ GJK = GGH L

(HJK = V ∧ HGH = LV ) ⇒ HGH = LHJK

Therefore, the coordinates of the JK biplot can be transformed into the coordinates
of the GH biplot with

GGH = GJK L−1 (3)

HGH = LHJK (4)

The SVD, however, is still needed to calculate L. At the same time, it is possible to
calculate the eigenvalues in L by transforming the eigenvalues of a PCA (LJK ) as shown
below3 :

L = U YS S−1 US LJK (5)

where S is the covariance matrix of the centered data matrix and US are the coefficients
of a PCA. Unfortunately, to get U, it is again necessary to calculate the SVD of Y, which
once more restricts the maximum number of observations to be used.
Right now, you cannot circumvent the restriction on the maximum number of obser-
vations for the GH or SQ biplot. In the future, it might be worthwhile for StataCorp to
program the calculation of the eigenvalues from the dataset without storing the dataset
in a matrix beforehand. In this case, at least the GH biplot could be easily derived from
a PCA with (3) and (4).
From a practical point of view, the described restriction is not as restrictive as it
sounds. It has been already stated that the interpretation of the biplot will be suspect
if the variance explained by the dimensions of the biplot are small. Small explained
variances, however, are quite common in working with datasets with many observations.
To this extent, the biplot has its strength mainly for datasets with small to moderate
number of observations. For huge datasets, the JK biplot can be calculated in any case.
3 The derivation of this formula can be found in the appendix.
214 Data inspection using biplots

5 The biplot8 command

5.1 Syntax





biplot8 varlist weight if exp in range ,


jk | sq | gh | mixed(jk | sq | gh jk | sq | gh) covariance mahalanobis rv



obsonly | varonly dimensions(##) generate(name1 name2 )


subpop(varname , scatter options ) stretch(#) flip(x | y | xy)

scatter options line options twoway options

aweights and fweights are allowed; see [U] 11.1.6 weight. However, no weights are
allowed with option rv, and aweights are not allowed with options sq and gh.

5.2 Options
jk | sq | gh specifies the biplot type. jk specifies the default, a JK biplot. gh and sq
specifies GH and SQ biplots, respectively (see section 5.4).
mixed(jk | sq | gh jk | sq | gh) can be used instead of the biplot types to combine the
relative advantages of the different biplot types. Inside the parentheses, you first
state the type for the observations and then a type for the variables (see section 5.4).
covariance is used to plot the unstandardized data matrix. The default is standard-
ization (see section 5.4).
mahalanobis can be used for GH biplots to rescale the graph in a way that the distances
between the observations approximate the Mahalnobis distances (see section 5.4).
rv is used to produce relative variation diagrams (see section 5.4).
obsonly | varonly are used to suppress the plotting of observations or variables, respec-
tively (see section 5.3).
dimensions(##) is used to specify the space in which the variables and observations
are drawn. The default is to use the dimension with the highest eigenvalues (i.e.,
the first two principal components for JK biplots) (see section 5.3).


generate(name1 name2 ) is used to store the coordinates for the observations and
the variables as variables in the dataset. The y-axis coordinates for the observations
are stored in name1 y, and the x-axis coordinates for the observations are stored in
name1 x. Accordingly, the coordinates for the variables are stored in name2 y and
name2 x.
subpop(varname) is used to highlight observations from different subpopulations with
different marker symbols (see section 5.5).
stretch(#) draws longer (or if needed shorter) lines for the variables. By default,
stretch() is set to a value that improves readability (see section 5.3).
U. Kohler and M. Luniak 215

flip(x | y | xy) exchanges the signs of the axes. flip(x) and flip(y) exchange signs
of the indicated axis, flip(xy) flips both axes. flip() is seldom used but might
be useful if you want to compare your results with the results of other software
packages.
scatter options are the following options allowed with twoway scatter.

jitter(relativesizelist) add spherical random noise to plot symbols

msymbol(symbolstylelist) shape of marker
mcolor(colorstylelist) color of marker, inside and out
msize(markersizestylelist) size of marker
mlabel(varlist) specify marker variables
mlabposition(clockposlist) where to locate label
mlabvposition(varname) where to locate label 2
mlabgap(relativesizelist) gap between marker and label
mlabsize(textsizestylelist) size of label
mlabcolor(colorstylelist) color of label

Up to two elements are allowed for each option. The first element refers to the
display of the observations, and the second element refers to the variables. Note
that the default plot symbol for the position of the variables is invisible; that is,
the default value for msymbol is msymbol(oh i). The lines for the variables are,
however, changed with the line options.
line options are the following set of the options allowed with line. Note that the
line options only refer to the display of the variable lines.
clpattern(linepatternstylelist) whether line is solid, dashed, etc.
clwidth(linewidthstylelist) thickness of line
clcolor(colorstylelist) color of line
twoway options are those options allowed with graph twoway; see [G] twoway options.

5.3 JK biplot and common PCA plots

Invoking the command biplot8 with a varlist and no other options brings up a JK
biplot (figure 3).4

(Continued on next page)

4 The examples in this section use the iris dataset. The data contains the sepal length, sepal width,

petal length, and petal width of 150 flowers from the iris species setosa, versicolor, and virginica. It was
collected by Anderson (1935) and was used by Fisher (1936) in his initiation of the linear-discriminant-
function technique.
216 Data inspection using biplots

. biplot8 sepallen-petalwid

4
sepalwid

2
DIM 2 (23 % of Var) sepallen

petalwid
petallen
0
−2
−4

−4 −2 0 2 4
DIM 1 (73 % of Var)

Figure 3: The standard JK biplot of iris.dta

As stated above, the JK biplot superimposes two of the most-often described plots
for principal component analysis: the component score plot and the plot of the PCA co-
efficients. However, in the default setting of the command biplot8, there is a difference
between the variable lines of the JK biplot and the plot of the PCA coefficients. The
biplot8 command stretches the variable lines to optimally fill the plot region given by
the observations (Digby and Kempton 1987, section 3.2). The positions of the variable
lines along the graph axis therefore represent the relative sizes of the PCA coefficients,
as opposed to the absolute ones, used in the plot of PCA coefficients. High values
still represent high “loadings”, but the square of the loadings cannot be interpreted as
communalities, as is the case for the plot of PCA coefficients.
It is, however, still possible to use biplot8 as a means to produce the plot of PCA
coefficients and the component score plot. The plot of PCA coefficients can be produced
with the options stretch(#) and varonly. In the former option, # stands for a
number by which the length of the variable lines are multiplied. By default, biplot8
automatically chooses this stretch factor to ensure optimal readability. Setting the
stretch factor to 1 forces Stata to use the original values, which are the PCA coefficients
in the case of the JK biplot. Using the option varonly, in addition, suppresses the
display of the observations entirely and thereby sets the graph scales according to the
coordinates of the variables. This brings up the plot of the PCA coefficients (figure 4).
U. Kohler and M. Luniak 217

. biplot8 sepallen-petalwid, st(1) varonly

1
sepalwid

DIM 2 (23 % of Var)

.5 sepallen

petalwid
petallen
0
−.5

−.5 0 .5 1
DIM 1 (73 % of Var)

Figure 4: Plot of PCA coefficients

Accordingly, the option obsonly as used in

. biplot8 sepallen-petalwid, obsonly

brings up the component score plot (figure 5).

4
2
DIM 2 (23 % of Var)
__000004
0 −2
−4

−4 −2 0 2 4
__000003
DIM 1 (73 % of Var)

Figure 5: Component score plot

218 Data inspection using biplots

As shown in section 3, the data coordinates of the biplot have k dimensions. To

plot these coordinates in a two-dimensional graph, you must select the dimensions to
be plotted. By default, this is done by selecting those coordinates that refer to the two
highest eigenvalues. The option dimensions(##) allows you to change this. Inside the
parentheses, you can state the ordinal rank of the eigenvalue for which the coordinates
ought to be selected. This is useful for JK biplots since you might be interested in
a display of the PCA coefficients for arbitrary principal components. Moreover, the
component score plot in the space of the two last principal components is said to show
a special kind of outlier (Gnanadesikan 1977, 261). Such a plot can be produced with
. biplot8 sepallen-petalwid, dim(3 4)

5.4 Biplot types and variations

The JK, GH, and SQ biplot can be displayed by using the options jk, gh, or sq, re-
spectively. It is possible in any case to calculate the coordinates from a standardized
or a nonstandardized data matrix. By default, biplot8 standardizes the data matrix,
which is why the variable lines tend to have the same length. To get lengths for the
variable lines according to variances of the variables, the option covariance must be
used. Figure 6 gives an example of the GH biplot for the nonstandardized data matrix,
which has been produced with the following command:
. biplot8 sepallen-petalwid, gh cov
.2
.1
DIM 2 (5 % of Var)

petalwid petallen
0

sepalwid sepallen
−.1
−.2

−.2 −.1 0 .1 .2
DIM 1 (92 % of Var)

Figure 6: GH biplot for unstandardized data

As mentioned in section 3, the biplot types differ in the quality of the approximations
of the key quantities shown in a biplot. While the approximation of the Euclidean
distance is best represented in the JK biplot, the variance–covariance structure is better
U. Kohler and M. Luniak 219

represented in the GH biplot. It seems, therefore, relatively straightforward to mix

the different biplot types. Gabriel (2002), for example, proposed a “correspondence
analysis” that uses the coordinates of a GH biplot for the variables and the coordinates
of a JK biplot for the observations. Such mixed biplots can be produced with the
option mixed(). The option allows you to list the names of two biplot types inside the
parentheses. The first name refers to the observational part, and the second refers to
the variable part. To obtain Gabriel’s correspondence analysis, you might type
. biplot8 sepallen-petalwid, mixed(jk gh)
4

sepalwid
2
DIM 2 (23 % of Var)

sepallen

petalwid
petallen
0
−2
−4

−4 −2 0 2 4
DIM 1 (73 % of Var)

Figure 7: Gabriel’s correspondence analysis

Note, however, that while it is possible to give optimal approximations to two of the
quantities shown in a biplot, this is not possible for all three of them (Gower and Hand
1996; Gabriel 2002). Mixing the GH and JK biplot as in the example above does not
optimally represent the observational values.
A further variant is biplots for compositional data. Compositional data are datasets
with constant row sums and only positive values, e.g., row percentages of contingency
tables. The standard data analysis techniques of compositional data usually tends to
be misleading, and therefore a set of specialized techniques are available for such data
(Aitchison 1986). The equivalent to biplots for compositional data is the “relative
variation diagram” (RV plot) (Aitchison 1990). A relative variation diagram refers to a
biplot of a transformed data matrix. The transformation is
∗
yik = ln yik − y i − y k

with yik being the untransformed value of Y in the ith row and kth column and y i and
y k being the row and column means of the data matrix. The option rv forces Stata to
make this transformation before producing the biplot.
220 Data inspection using biplots

Finally, the option mahalanobis can be used to rescale the coordinates in G and H
by
√
G∗ = G × n
H∗ = H× √1
n

before producing the biplot. According to Gabriel (1971) the resulting biplot reflects
the Mahalanobis distances between the observations instead of the Euclidean distances.

5.5 Options to control the graph appearance

Several options are available for controlling the appearance of the graph. Among them
are most of the options allowed for twoway scatter and twoway line. Here scat-
ter options allow up to two arguments, where the first argument refers to the obser-
vations (the dots) and the second refers to the points at the end of the variable lines
(which are invisible by default). line options refer to the variable lines.
The option subpop() is specific to biplot8 and is used to distinguish observations
from different subgroups with different markers. Therefore, the name of the variable that
identifies the subgroup is placed inside the parentheses. Note that the scatter options for
the observations are ignored if you specify subpop(). However, you can use the complete
set of scatter options as suboptions within subpop() to control the appearance of the
observations.
The subpop() option is especially useful for illustrating the substantial meaning of
data clusters. Figure 8, which has been produced with the command below, gives an
illustrative example.

(Continued on next page)

U. Kohler and M. Luniak 221

. biplot8 sepallen-petalwid, subpop(species, msymbol(Oh X Th))

> legend(ring(0) pos(4))

4
sepalwid

2
DIM 2 (23 % of Var)
sepallen

petalwid
petallen
0
−2

setosa versicolor
virginica
−4

−4 −2 0 2 4
DIM 1 (73 % of Var)

Figure 8: Illustrative example of representation options

Note that the default positioning of legends changes the aspect ratio of the biplot.
If you don’t like this, you can move the legend position to the inner ring, as shown in
the example. Alternatively, you can turn the legend off or refine the aspect ratio with
the options xsize() or ysize().

6 Appendix
Consider a PCA of the data matrix Y, which is a SVD of the variance–covariance matrix
S of Y
S = US LJK VS (6)

Also consider the coordinates of the observations for the JK biplot from (1):
GJK = UL (7)

From Jolliffe (2002, 94), it is known that GJK are equal to the scores of the obser-
vations on the principal components, which are given by
GJK = YUS (8)

From (7) and (8), we obtain

UL = YUS

U UL = U YUS
222 Data inspection using biplots

From the properties of the SVD, we know that U is an unitary matrix, so U U = I.

Hence

L = U YUS (9)

In order to find the relation between L and LJK , we look at US from (6). The matrix
S is symmetric, so

US = VS
S= US LJK US
SUS = US LJK US US

US is orthogonal, which means that US = U−1  

S and US US = US US = I. Hence

SUS = US LJK
−1
US = S US LJK (10)

Imposing (10) in (9) gives (5) on page 213:

L = U YS−1 US LJK

7 Acknowledgments
We like to point German readers to the book Graphisch gestützte Datenanalyse, written
by Rainer Schnell (1994, 176–186), upon which parts of this article are heavily based.
We also like to thank Frauke Kreuter for carefully reading an earlier draft of this article
and Vince Wiggins and Nick Cox for helping us write biplot8.

8 References
Aitchison, J. 1986. The Statistical Analysis of Compositional Data. London: Chapman
& Hall.
—. 1990. Relative variation diagrams for describing patterns of compositional variablity.
Mathematical Geology 22: 487–512.
Anderson, E. 1935. The irises of the Gaspé peninsula. Bulletin of the American Iris
Society 59: 2–5.
Blasius, J. and M. Greenacre, ed. 1998. Visualization of Categorical Data. San Diego:
Academic Press.
Digby, P. G. N. and R. A. Kempton. 1987. Multivariate Analysis of Ecological Com-
munities. London: Chapman and Hall.
U. Kohler and M. Luniak 223

Fenty, J. 2004. Analyzing distances. Stata Journal 4(1): 1–26.

Fisher, R. A. 1936. The use of multiple measurements in taxonomic problems. Annals

of Eugenics 7: 179–188.

Gabriel, K. 1971. The biplot graphic display of matrices with application to principal
component analysis. Biometrika 58(3): 453–467.

—. 2002. Goodness of fit of biplots and correspondence analysis. Biometrica 89(2):

423–436.

Gnanadesikan, R. 1977. Methods for Statistical Data Analysis of Multivariate Obser-

vations. New York: Wiley.

Gower, J. C. and D. J. Hand. 1996. Biplots. London: Chapman & Hall.

Hamilton, L. C. 1992. Regression with Graphics: A Second Course in Applied Statistics.

Pacific Grove, CA: Brooks/Cole.

Jackson, J. E. 1991. A User’s Guide to Principal Components. New York: Wiley.

Jolliffe, I. 2002. Principal Component Analysis. 2nd ed. New York: Springer.

Kruskal, J. B. and M. Wish. 1978. Multidimensional Scaling. Beverly Hills, CA: Sage.

Schnell, R. 1994. Graphisch gestützte Datenanalyse. München, Wien: Oldenbourg.

Schnell, R. and H. Matschinger. 1994. Multivariate graphics: Current use and imple-
mentations in the social sciences. In Computational Statistics. Papers Collected on
the Occasion of the 25th Conference on Statistical Computing at Schloß Reisensburg,
ed. P. Dirschedl and R. Ostermann, 275–294. Heidelberg: Physica.

Tabachnik, B. and L. S. Fidell. 1989. Using Multivariate Statistics. 2nd ed. New York:
Harper and Row.

About the Authors

Ulrich Kohler is a sociologist at the Wissenschaftszentrum Berlin (Social Science Research
Center) who has used Stata for several years. His research interests include social inequality
and political sociology. With Frauke Kreuter, he is author of the upcoming textbook Data
Analysis Using Stata.
Magdalena Luniak studies sociology at the Warsaw University and IT at the TU Berlin. The
main focus of her studies is the application of mathematics in sociology.