Better biplots
Introduction
Biplots can be viewed as the multivariate analogue of scatterplots, where samples/cases are plotted as points relative to two variables. With biplots, the multivariate distribution of a set of variables can be approximated in a low dimensional space (usually two dimensions) giving a useful visualization of the structure of the samples relative to the variables. Biplots were introduced by Gabriel (1971) in the field of principal component analysis; the first textbook on general biplot methods is that of Gower and Hand (1996).
Figure 2.5 of Gower and Hand (1996) illustrates many of the things that can be wrong with a biplot. Surprisingly, many of the deficiencies described in their artificial example continue to occur (compare also Fig. 8):
(i) Inelegant, and often needless, horizontal and vertical scales.
(ii) Unequal scaling in the and directions.
(iii) Vector representation of variables.
(iv) The inconvenience of evaluating inner-products.
(v) Variables, whether in vector form or not, being inextricably intermingled with sample points, so interfering with clear labelling.
(vi) Not a deficiency, but nevertheless a problem, is how to display very large numbers of samples.
With a little care, (i) and (ii) are easily handled. Items (iii) and (iv) may be circumvented, especially by the proposal that vector representations be replaced by using calibrated axes, so bringing biplots into line with the familiar everyday use of coordinate axes in all fields of study concerned with the relationships between two or more variables. Calibrated biplot axes are used in the same way as conventional calibrated coordinate axes, and differ only in having as many axes as there are variables. However, because few dimensions are used for the representation, these axes must be non-orthogonal. Thus, a sample point is projected orthogonally onto the axis, and the nearest calibrated point is taken, if necessary using a little visual interpolation between adjacent calibrations. The use of projection is a primary reason why it is essential to scale equally, i.e. a circle must look like a circle. The same point applies to calculating inner products which depend on accurate representations of angles. Although Gower and Hand (1996) had emphasised this point, it was particularly galling when a reviewer took the authors to task for not taking into account the excellent work of Cleveland (1993) on aspect ratios. Valuable as is the aspect ratio concept when used with differently scaled variables, it will destroy the metrics underlying biplots and should never be used with biplots or, indeed, with any form of MDS. A thing to watch is that however hard one tries to get scaling right; computer software, and some publishers, concerned with printing aesthetics may readjust diagrams to fit the page; the aspect ratio will suffer similarly.
What was not covered by Gower and Hand (1996) was (v) the placement of axes; something that is readily handled with manually produced plots, and (vi) what to do about large numbers of sample points. In this paper we address both of these matters, largely in the context of linear biplots. However, the simple ideas presented easily generalise to other situations and here, we provide an example, still in the linear context, with variables measured on ordinal category, rather than numerical, scales.
Section snippets
Calibrated axes
We shall not dwell further on the reasons for favouring calibrated axes. They are not an alternative to using inner products but merely a simple way for their evaluation. The details for calculating the positions of calibration markers are straightforward (see Gower and Hand (1996), pp 38–40) and are discussed in the Appendix. When there are points of special interest, such as the length of one standard deviation, there is no reason why this information should not be shown on a calibrated axis,
Density plots
There are several situations where scatterplots are very popular, for example, in the French Analyse des Données (cf. Le Roux and Rouanet (2004)), in microarray data as well as in the whole area of data mining. In microarray studies, scatterplots can contain tens of thousands of dots leading to perception problems. In this section we present the background to displaying (very) many data points as a smooth density image. This issue is important when the number of scattered data points is so
An example with ordered categorical variables
The International Social Survey Program (ISSP) is a continuing annual program of cross-national collaboration covering topics important for social science research, being one of the most recognized and best documented data sets world-wide (Scheuch, 2000). A self-administered questionnaire contains questions relevant to all participating countries. Sample size varies but should exceed 1000 cases in each country. 34 countries participated in the 2003 ISSP when 11 questions referring to attitudes
Discussion
The aim of the paper was to show useful ways of displaying biplots and to develop a program that aids automatic drawing. Although the developed program allows for iterative changes of axes and labels, some final manual improvement is possible and desirable, for example, using the Adobe Illustrator software. Further, we introduced some new ways of presenting these data, some already mentioned by Gower and Hand (1996). As examples we used a small data set on 21 aircraft described by four
References (14)
- et al.
Multivariate prediction with nonlinear principal components analysis: Application
Quality and Quantity
(2005) Visualizing Data
(1993)- et al.
Residuals and Influence in Regression
(1982) Nonlinear principal component analysis and related techniques
- et al.
Enhancing scatterplots with smoothed densities
Bioinformatics
(2004) The biplot-graphic display of matrices with applications to principal component analysis
Biometrika
(1971)Nonlinear Multivariate Analysis
(1990)
Cited by (26)
Categorical CVA biplots
2021, Computational Statistics and Data AnalysisCitation Excerpt :It is worthwhile noting that if the final z-quantifications of the ordinal variables do not preserve the natural order, ties will be introduced, which may hinder the biplots' overall interpretation. To further improve the interpretation of the biplot, Blasius et al. (2009) proposed that axes of the ordinal categorical variables are segmented according to different thickness levels across the various category levels. The segmentation is performed so that the lowest ordered level is given the thinnest segment while the highest ordered level is given the thickest segment.
Constructing bi-plots for random forest: Tutorial
2020, Analytica Chimica ActaCitation Excerpt :In the current study, a simple approach that allows for construction of bi-plots (i.e. spin plots) associated with RF models is demonstrated. Bi-plots are among the most popular and versatile visualization tool of multivariate models [10–13]. Bi-plots permit for the simultaneous display of information on the samples and the variables.
Visualising interactions in bi- and triadditive models for three-way tables
2017, Chemometrics and Intelligent Laboratory SystemsFLOOD: FLow cytometric Orthogonal Orientation for Diagnosis
2016, Chemometrics and Intelligent Laboratory SystemsThe impact of institutional and social context on corporate environmental, social and governance performance of companies committed to voluntary corporate social responsibility initiatives
2015, Journal of Cleaner ProductionCitation Excerpt :If two variables are closer, they are highly correlated (i.e. exhibit similar behaviour); if two other variables take different directions, they are inversely correlated. If they are perpendicular, they are independent (Blasius et al., 2009; Gardner et al., 2005). Finally, the smaller the angle between two vectors that join the centre of gravity with the points that represent the variables, the more concentrated the characters are.
Predictive nonlinear biplots: Maps and trajectories
2015, Journal of Multivariate AnalysisCitation Excerpt :In Section 4.3 it was noted that the origin for normal projection prediction trajectories constructed for biplots based on non-smooth dissimilarity functions of these prediction trajectories can be altered. Thus the origin can be placed away from the centre of the data points as suggested in [1]. However given that points lying on such a trajectory are, by construction, guaranteed to have correct predictions given by the trajectory there is an argument for keeping trajectories as close as possible to data points.