The American Statistician

ISSN: 0003-1305 (Print) 1537-2731 (Online) Journal homepage: https://www.tandfonline.com/loi/utas20

Violin Plots: A Box Plot-Density Trace Synergism

Jerry L. Hintze & Ray D. Nelson

To cite this article: Jerry L. Hintze & Ray D. Nelson (1998) Violin Plots: A Box Plot-Density Trace
Synergism, The American Statistician, 52:2, 181-184, DOI: 10.1080/00031305.1998.10480559
To link to this article: https://doi.org/10.1080/00031305.1998.10480559

Published online: 22 Mar 2012.

Submit your article to this journal

Article views: 2824

View related articles

Citing articles: 91 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=utas20
 Statistical Computing and Graphics
Violin Plots: A Box Plot-Density Trace Synergism
Jerry L. HINTZE and Ray D. NELSON

Hoaglin (1981); Chambers, Cleveland, Kleiner, and Tukey

(1983); Frigge, Hoaglin, and Iglewicz (1989), and others.
Many modifications build on Tukey's original box plot. A Box plots show four main features about a variable: cen-
proposed further adaptation, the violin plot, pools the best ter, spread, asymmetry, and outliers. As an example, con-
statistical features of alternative graphical representations sider the box plot in Figure 1 for the data published by
of batches of data. It adds the information available from Hamermesh (1994). The ASA Statistical Graphics Section's
local density estimates to the basic summary statistics in- 1995 Data Analysis Exposition analyzes these data, which
herent in box plots. This marriage of summary statistics and report compensation of professors from all academic ranks
density shape into a single plot provides a useful tool for in the United States. The labels in the diagram identify the
data analysis and exploration. principal lines and points which form the main structure of
the traditional box plot diagram. As shown, the violin plot
KEY WORDS: Density estimation; Exploratory data
includes a box plot with two slight modifications. First, a
analysis; Graphical techniques.
circle replaces the median line which facilitates quick com-
parisons when viewing multiple groups. Second, outside
1. INTRODUCTION points which are traditionally classified as mild and severe
outliers, are not identified by individual symbols.
Many different statistics and graphs summarize the charac- The density trace supplements traditional summary statis-
teristics of single batches of data. Descriptive statistics give tics by graphically showing the distributional characteristics
information about location, scale, symmetry, and tail thick- of batches of data. One simple density estimator, the his-
ness. Other statistics and graphs investigate extreme obser- togram, displays the distribution of data values along the
vations or study the distribution of data values. Diagrams real number line. Weaknesses of the histogram caused Tapia
such as stem-leaf plots, dot plots, box plots, histograms, and Thompson (1979), Parzen (1979), Silverman (1986),
density traces, and probability plots give information about Izenman (1991), and Scott (1992) to propose and summarize
the distribution of values assumed by all observations. numerous alternative density estimators. One of these alter-
The violin plot, introduced in this article, synergistically natives is the density trace described in Chambers, Cleve-
combines the box plot and the density trace (or smoothed land, Kleiner, and Tukey (1983). Defining the location den-
histogram) into a single display that reveals structure found sity d(xlh) at a point x as the fraction of the data values
within the data. The introduction of this new graphical tool per unit of measurement that fall in an interval centered at
begins with a quick overview of the combination of the box x gives
plot and density trace into the violin plot. Then, three illus-
trations and examples show the advantages and challenges ",n 8.
of violin plots in data summarization and exploration. d(xlh) = L..i=l t, (1)
nh
2. COMPONENT PARTS OF VIOLIN PLOTS
The violin plot, as depicted in Figure 1 and implemented
in NCSS (1997) statistical software, combines the box plot
and density trace into one diagram. The name violin plot
originated because one of the first analyses that used the
envisioned procedure resulted in a graphic with the ap-
pearance of a violin. Violin plots add information to the
simple structure of the box plot that Tukey (1977) initially
conceived. Although these original graphs are easily drawn
with pencil and paper, computers ease subsequent modifi-
cations, refinements, and computation of box plots as dis-
cussed by McGill, Tukey, and Larsen (1978); Velleman and

Jerry L. Hintze is President, NeSS, 329 North 1000 East, Kaysville, UT

84037 (E-mail: [email protected]). Ray D. Nelson is Associate Professor of Violin Plot
Business Management, Marriott School of Management, Brigham Young
University, Provo, UT 84602. Figure 1. Common Components of Box Plot and Violin Plot. Total
compensation for all academic ranks.

© 1998 American Statistical Association The American Statistician, May 1998 Vol. 52, No.2 181
 30 able density trace requires experience and judgment in de-

15
•• termining the appropriate amount of smoothing. As with
the selection of the bin width in the histogram, the inter-
val width h, which is usually specified as a percentage of
o the data range, must be selected. Experience suggests that
values near 15% of the data range often give good results.
-15
The choice of h, however, must be tempered by the size of
the sample. The density trace is subject to the same sample
size restrictions and challenges that apply to any density
-30 +-----r----'-'T---.......,...-----.., estimator. For small data sets, too small a value for h gives
Bimodal Uniform Normal
a wiggly density trace that suggest features that are simply
(a) artifacts of the individual data points. The oversmoothed
density estimate that results from too large h values gives
the illusion of knowing the shape of the distribution, while
30 in reality the data set is too small for any conclusions. As
a rule of thumb based on practice, the density trace tends
15 to do a reasonable job with samples of at least 30 observa-
tions. Even with sample sizes of several hundred, however,
o choosing too large a value for h causes the density trace to
oversmooth the data. In general, values of h greater than
-15 40% of the range usually result in oversmoothed densities,
while values less than 10 percent of the range result in un-
-30 +----....----.,----"""T'"---..., dersmoothed densities. Hence, percentages between 10 and
40 percent are recommended.
Bimodal Uniform Normal
(b) 4. ILLUSTRATIONS AND APPLICATIONS
Figure 2. Comparison of Box Plots and Violin Plots to Known Distri- With the addition of the density trace to the box plot,
butions. (a) Box plots; (b) violin plots. violin plots provide a better indication of the shape of the
distribution. This includes showing the existence of clusters
where n is the sample size, h is the interval width, in data. The density trace highlights the peaks, valleys, and
and Oi is one when the ith data value is in the interval bumps in the distribution. Three applications and examples
[x - h/2, x + h/2] and zero otherwise. In order to plot the of violin plots illustrate these advantages. The first example
density trace, first select a value for h and then compute demonstrates the ability of violin plots to distinguish among
d(xlh) on a dense grid of equally spaced x values. Connect the shapes of known distributions. The second highlights
the d(xlh) by lines. The shape of the d(xlh) curve is essen-
tially driven by the interval length, h. It is very smooth for 140 140
large values of h, and "wiggly" for small values.
Unfortunately, several density traces shown side by side
are difficult to compare. Contrasting the distributions of sev-
eral batches of data, however, is a common task. In order to
add information to the box plot and still make comparisons
possible, Benjarnini (1988) suggested "opening the box" of 204.--====:.... ------+ 20
the box plot. He makes the width of the box proportional
to the estimated density. The violin plot builds on the Ben- (a)
jarnini proposal by combining the advantages of box plots 5.0
with density traces.
The violin plot, as shown in Figure 1, combines the box 4.1
I
plot with density traces. The density trace is plotted sym- ?
metrically to the left and the right of the (vertical) box plot. ,5j
c 2 3.3
There is no difference in these density traces other than the ·gil
!~
direction in which they extend. Adding two density traces
2.4
gives a symmetric plot which makes it easier to see the
magnitude of the density. This hybrid of the density trace
I.S
I
and the box plot allows quick and insightful comparison of
several distributions. (b)

3. SPECIFICATION OF INTERVAL WIDTH Figure 3. Additional Information in Violin Plots. Two examples from
the density estimation literature: (a) annual snowfall for Buffalo, NY,
As with other density estimators, achieving an accept- 1910-1972; (b) Old Faithful eruption length.

182 Statistical Computing and Graphics

 120
These plots seem to indicate that since the mass of the
bimodal plot is less than the normal plot, the bimodal plot
95 is based on fewer observations. This is a weakness of this
·1 implementation of the violin plot, which adjusts the density
8.- 70 traces so that their maximum heights are equal. This allows
~~ a direct comparison of the shapes, but removes the visual
1'i 45
~ impact of sample size. A variation of this implementation
would keep a uniform scaling of the density traces.
20
Category I Category llA Category lIB 4.2 Density Estimation Examples
(a) Clusters of data appear as bumps in density estimators.
Box plots often do not alert analysts to their existence. Two
140 examples from the density estimation literature clearly il-
lustrate this ability. First, Parzen (1979) and Scott (1992)
used annual snowfall data for Buffalo, New York for 1910-
1972 to show the value of nonparametric density estima-
tion. The violin plot in Figure 3(a) illustrates the additional
insights available through density estimators that the basic
box plot does not reveal. The second example in Figure
3(b), which uses data previously considered by Silverman
20 -r-----.-----r-----r----.,
(1986) and Scott (1992), shows the bimodal nature of Old
Fun Associate Assistant
Professors Professors Professors Faithful eruption lengths. Once again, the violin plot clearly
(b) adds significant insight about the distribution of the process
generating the data.
Figure 4. Exploring Data with Violin Plots. Total compensation data
from ASA analysis competition. (a) Compensation of all professors by 4.3 Academic Compensation
university classification; (b) total compensation by academic rank.
The information that violin plots add to box plots in-
the ability to detect bumps or clusters of data. The third creases the potential of these tools when used in data ex-
shows their potential in exploring for structure and pattern ploration. As an example of the value of violin plots, con-
in the academic compensation data used previously in the sider the diagrams in Figure 4 for data published by Hamer-
illustration of the components of violin plots. The values for mesh (1994). The graphics in Figure 4(a) summarize total
the interval widths h are chosen using personal judgment compensation of all professors for three different classifica-
from values from within the recommended 10 to 40 percent tions of colleges and universities. The first category includes
interval. These examples establish the potential of violin institutions with a significant level of doctoral-level edu-
plots in data analysis and exploration. cation. The second encompasses institutions with diverse
post-baccalaureate programs but which do not have a sig-
nificant level of doctoral programs. The colleges and uni-
4.1 Comparison of Known Distributions versities in the third category focus their primary activity on
undergraduate baccalaureate-level education. The bumps in
Consider first the ability to detect general shapes for dis- the doctoral level and post undergraduate violin plots sug-
tributions of data. Figure 2 depicts box plots and violin gest that some universities in each of these categories might
plots for random samples of 10,000 simulated observations have compensation characteristics which distinguish them
drawn from three different known distributions. The three from other members of the general group.
distributions share identical location and scale characteris- Figure 4(b) shows the distribution of total compensation
tics as measured by the median and interquartile range. The for institutions in all three categories by academic rank.
first is a bimodal distribution with modes at -5 and 5 and All three of the categories appear to be somewhat posi-
range between -10 and 10. The second is a uniform distri- tively skewed with the skewness increasing with the aca-
bution on the interval [-10,10]. The third is a normal distri- demic rank. Comparison of the medians gives the expected
bution N(O, 54.95). The box plots in Figure 2(a) reflect the increase in compensation with the higher academic rank.
fact that all three have the same median and interquartile An interesting bulge grows in the upper tail of the distri-
range. butions as the academic rank increases. In an exploratory
As expected, the density trace accurately reveals the analysis, the violin plots point to the next question which
shape of the distribution from which the random samples might investigate the characteristics of the institutions in
are drawn. The violin plot for the bimodal distribution these clusters.
clearly shows the twin peaks of the known distribution.
Unfortunately, box plots cannot differentiate between the 5. SUMMARY AND CONCLUSIONS
shapes of the bimodal and uniform distributions. The box Individually, box plots provide succinct summaries of
plots do, however, show that the normal distribution differs data. By themselves, density traces reveal important infor-
from the others as it does have a larger range. mation about the distribution of data. The synergistic com-
The American Statistician, May 1998 Vol. 52. No.2 183
bination of the box plot and the density trace allows much Hamermesh, D. (1994), "Plus Ca Change: The Annual Report on the Eco-
of the information from each to be displayed in one plot. nomic Status of the Profession, 1993-1994," Academe, 5-89.
Hintze, J. (1997), User's Guide, NCSS 97, Statistical System for
This single plot structure makes comparisons of distribu-
Windows, Kaysville, UT: Number Cruncher Statistical Systems
tional factors of several variables much easier. Three dif- (http://www.ncss.com).
ferent illustrations show that violin plots retain much of Izenman, A. J. (1991), "Recent Developments in Nonparametric Density
the information of box plots and add information about the Estimation," Journal of the American Statistical Association, 86, 205-
shape of the distribution not obvious in box plots. Their 224.
McGill, R., Tukey, J. w., and Larsen, W. A. (1978), "Variations of Box
ability to detect clusters or bumps within a distribution is Plots," The American Statistician, 32, 12-16.
especially valuable. Parzen, E. (1979), "Nonparametric Statistical Data Modeling," Journal of
American Statistical Association, 74, 105-131.
[Received February 1997. Revised November 1997.J Scott, D. W. (1992), Multivariate Density Estimation: Theory, Practice,
and Visualization, New York: Wiley.
REFERENCES Silverman, B.W. (1986), Density Estimation for Statistics and Data Anal-
ysis, New York: Chapman and Hall.
Benjamini, Y. (1988), "Opening the Box of the Box Plot," The American Tapia, R. A., and Thompson, J. R. (1978), Nonparametric Probability Den-
Statistician, 42, 257-262. sity Estimation, Baltimore, MD: Johns Hopkins University Press.
Chambers, J. M., Cleveland, W. S., Kleiner, B., and Tukey, P. A. (1983), Tukey, J. W. (1977), Exploratory Data Analysis, Reading, MA: Addison-
Graphical Methods for Data Analysis, Belmont, CA: Wadsworth. Wesley.
Frigge, M., Hoaglin, D. C, and Iglewicz, B. (1989), "Some Implementa- Velleman, P. E, and Hoaglin, D. C. (1981), Applications, Basics and Com-
tions of the Box Plot," The American Statistician, 43, 50-54. puting of Exploratory Data Analysis, Boston: Duxbury Press.

184 Statistical Computing and Graphics