The Stata Journal (2004)

4, Number 1, pp. 66–88

Speaking Stata: Graphing distributions

Nicholas J. Cox
University of Durham, UK
[email protected]

Abstract. Graphing univariate distributions is central to both statistical graphics,

in general, and Stata’s graphics, in particular. Now that Stata 8 is out, a review of
official and user-written commands is timely. The emphasis here is on going beyond
what is obviously and readily available, with pointers to minor and major trickery
and various user-written commands. For plotting histogram-like displays, kernel-
density estimates and plots based on distribution functions or quantile functions,
a large variety of choices is now available to the researcher.
Keywords: gr0003, graphics, histogram, spikeplot, dotplot, onewayplot, kdensity,
distplot, qplot, skewplot, bin width, rug, density function, kernel estimation, trans-
formations, logarithmic scale, root scale, intensity function, distribution function,
quantile function, skewness

1 Introduction
The new graphics introduced in Stata 8 has been, by far, the most important step
forward in Stata’s graphical functionality since early releases in the mid-1980s. It is,
therefore, high time that this column turned to discuss graphics directly. I intend to
make 2004 a graphic year for Speaking Stata, starting with the basic and fundamen-
tal issue of graphing univariate distributions. Future columns are intended to discuss
graphing categorical and compositional data, comparisons, and model diagnostics. In
each case, the aim will be to provide an overview of Stata’s provision and to show ways
to go beyond what is obviously and readily available. The emphasis will be on graphics
commands of potential interest to the largest possible cross-section of Stata users. Thus
histograms clearly qualify, but justice cannot be done to details specific to analysis of
survival-time distributions.
The core commands for graphing distributions range from twoway kdensity and
its relative kdensity through twoway histogram and its relative histogram to graph
box and graph hbox. Related but perhaps less-often used commands include dotplot,
spikeplot, and those grouped as diagnostic plots.

2 Histograms, indigenous and exotic

2.1 Number of bins and bin width
With an eye to tradition, including Stata tradition, let us start the discussion with
histograms. Up until Stata 7, a histogram was the default graph type if graph was fed
just one variable. Before Stata 8, such histograms were relatively inflexible and could

c 2004 StataCorp LP

gr0003
N. J. Cox 67

not easily be combined with other graph types. Now we have both greater flexibility
and easier working with other types. Notable additions include the options to tune both
bin width and the start of binning, whereas previously only the number of bins could
be controlled directly. The start of binning could be controlled indirectly by tuning
xlabel() or xscale().
As every good introductory text explains, histogram construction is largely a trade-
off problem in which you seek a compromise between detail and generalization or be-
tween variance and bias. In doing this, you can tune either the number of bins or the
bin width. Theoretical discussions concentrate on the number of bins and its relation
to sample size and the kind of distribution being analyzed. However, my guess is that
people with their feet in application areas often find it natural to think in terms of a
sensible bin width for the variables they have, bearing in mind measurement issues and
the magnitude of important or interpretable differences. Whatever your preference, you
can now do it either way.

2.2 Varying bin widths

However, one feature that remains wired in histogram commands in Stata 8 is a restric-
tion to bins of equal width. No doubt this is often very sensible whenever the original
data are available, but there are occasions on which you might want to break this rule.
Let us drill down to some first principles here.
Recall that the idea behind histograms is that the area of each bar represents the
fraction of a frequency (probability) distribution within each bin (or class, or interval).
Among many books explaining histograms, Freedman, Pisani, and Purves (1998) is an
outstanding introductory text that strongly emphasizes the area principle. It is not part
of the definition that all bins have the same width, but rather that what is shown on the
vertical axis is, or is proportional to, probability density. Frequency density qualifies,
as does frequency if all bins have the same width.
In practice, the choice of bin width is often a little arbitrary. If the variable is discrete,
a width of 1 is clearly a natural choice. Even then, discrete variables may require some
grouping into bins wider than 1. If the variable is number of lifetime sexual partners,
the tail (apparently) stretches into very large numbers, and some grouping may be
desired. With continuous variables especially, there is always some arbitrariness. Many
researchers are most reluctant to compound that by varying the width of the intervals.
To do so would complicate the interpretation of the histogram, it might be argued, by
any variations in the way the bars were produced. Or, to put it another way, equal
widths are relatively simple, and any kind of complexity beyond them needs to be
justified.
Despite all that, sometimes the data come grouped into irregular intervals, and the
researcher has little or no choice because the raw data may be difficult or impossible to
access. Sometimes there is an underlying confidentiality issue. Nevertheless, researchers
may still want a histogram, which should be correctly drawn with density, not frequency,
on the vertical axis. For example, Altman (1991, 25) gives the ages of 815 road accident
casualties for the London Borough of Harrow in 1985:
68 Speaking Stata

age frequency
0–4 28
5–9 46
10–15 58
16 20
17 31
18–19 64
20–24 149
25–59 316
60+ 103

In this example and in other similar examples, density can only be calculated for
the open-ended class if we specify an upper limit; Altman suggests that 60+ be treated
as 60–80.
As usual in statistics, sampling variation is also an issue. If we regard the histogram
as a crude estimator of a density function, there is often a case for varying bin width to
match the structure of variation, in effect varying how we average probability density
locally.
But there is at least one other way to build a histogram in a simple, systematic way:
using as limits a set of quantiles equally spaced on a probability scale (e.g., Breiman
1973, 208–209; Scott 1992, 69–70). That way, each bar represents the same area. Un-
less our data come from something like a uniform distribution, the bin widths will
be markedly unequal, but they will reflect the character of the distribution. Breiman
points out that the associated error will be approximately a constant multiple of the
bar heights, so long as the bin frequencies are not too small.
A related problem is choice of class intervals for a chi-squared test of goodness of fit.
Mann and Wald (1942) and Gumbel (1943) urged the merits of choosing classes with
equal expected frequencies. That is a simple and definite procedure, which can reduce
difficulties arising from low expected frequencies, although data must arrive ungrouped
and there may be some loss of sensitivity in the tails of a distribution. Without getting
into a wider discussion of the merits of different tests of fit or of tests compared with
graphical analysis, it is clear that the equal probability idea is a natural one.
What can be done in Stata? Start with the messier problem in which the data arrive
grouped. Much can be done once you know about an undocumented feature of twoway
bar. We need to enter the lower bin limits and the bin frequencies and one final upper
limit as data. For Altman’s example, we need to enter data to get

(Continued on next page)

N. J. Cox 69

. list age freq

age freq

1. 0 28
2. 5 46
3. 10 58
4. 16 20
5. 17 31

6. 18 64
7. 20 149
8. 25 316
9. 60 103
10. 80 .

We then can calculate the densities:

. generate density = freq / (815 * (age[_n+1] - age))

If you want frequency density rather than probability density, you should omit scaling
by the sample size (here 815).
Finally, we can draw the graph, shown in figure 1:
. twoway bar density age, bartype(spanning) bstyle(histogram)
.04
.03
density
.02
.01
0

0 20 40 60 80
age

Figure 1: Example of histogram based on data supplied as frequency distribution with

varying bin widths.

The “spanning” extends bars to the right until they are curtailed; that is why it is
necessary to specify all lower limits and one upper limit for the graph. The data should
also be in the correct sort order, as in this example. The option bstyle(histogram) is
70 Speaking Stata

not compulsory, and you might like to check other possibilities. You might need to add
the option yscale(range(0)) if twoway bar does not automatically start bars at 0.
Turning to the more elegant problem, a user-written program for equal-probability
histograms can be described and, if desired, downloaded from the Statistical Software
Components (SSC) archive by using the ssc command; see [R] ssc:

. ssc describe eqprhistogram

. ssc install eqprhistogram

As an illustration, figure 2 is the result of

. use http://www.stata-press.com/data/r8/womenwage.dta
. eqprhistogram wage, bin(10) plot(kdensity wage, biweight width(5))
> legend(ring(0) position(1) column(1))
.1

Density
kdensity wage
.08
Density/kdensity wage
.04 .02
0 .06

0 20 40 60 80
wages in 1000s of dollars/x

Figure 2: Example of histogram in which bins represent intervals of equal probability.

In this case, bin boundaries are deciles, so that the area of each bin represents 1/10 of
the distribution. A kernel-density estimate is superimposed.

The bin limits are the deciles, so each bar represents 1/10 of the total probability in
the distribution. Note that you can superimpose a density estimate.
Although it may seem a curiosity, the equal-probability histogram has some ped-
agogic merit. First, it underlines the area principle on which histograms are based.
Wider bars are necessarily shorter and narrow bars necessarily taller. Second, it allows
a link to be made between histograms and quantile-based methods such as box plots.
Arguably, in some datasets it gives a better view of the tails than do the corresponding
box plots, especially if within those box plots no values are flagged beyond the quartiles,
and so no details are given on structure within the tails. (Box plots are especially poor
for U-shaped distributions. In some such cases, no values are identified beyond the
N. J. Cox 71

quartiles, and the box plot reduces to a long box and two short tails. Even experienced
people can misread this as indicating a unimodal distribution, forgetting that if half the
values lie inside the box, then, necessarily, half lie outside it.)
An equal-probability histogram is not suitable for all distributions. Given categor-
ical, discrete, or highly rounded data, quantiles may be tied, especially if the num-
ber of bins is large relative to the sample size. If the specified quantiles are tied,
eqprhistogram refuses to draw the graph. A technical aside: whenever it does this,
the exit code is 0. This is in part a diplomatic acknowledgment that inability to draw
the graph is either a feature of the data or a limitation of the method, rather than a
user error. In addition, it implies that a loop through equal-probability histograms of
different variables or groups will not fail merely because a particular graph is impossible.

2.3 Putting a rug underneath

One major merit of histograms is familiarity. All statistically minded people have looked
at many histograms, and nonstatistical people who use statistics have also usually come
across them. Nevertheless, the basis of histograms, a division of a range into bins, is
at best a means to an end, namely easy and effective visualization of a distribution,
and at worst a serious distraction. Both psychologically and numerically, densities or
frequencies calculated from a set of bins can convey a poor idea of the detailed shape
of the distribution of a variable.
One simple way to enhance a histogram by forging a closer link with the raw data is
to add a so-called rug, which as the name implies, is almost always placed underneath
the histogram. A rug is a very short, long display of point symbols, one for each distinct
value. Often a vertical pipe symbol | is used to minimize overlap. Rugs may also be
added to other kinds of plots. There are many varied examples in Davison (2003).
Before version 8, Stata had graph options to combine rugs, which in Stata were called
oneway plots, with box plots and with scatterplots. These options are still accessible
under graph7. However, they did not make the cut into the new graphics in that or
similar form.
Although rugs are not explicitly provided in Stata 8, the procedure for weaving your
own rug is straightforward. Starting with a basic histogram for the same wage data,
. histogram wage, start(0) width(5)

we see that with these choices density varies up to about 0.07 per 1,000 dollars. That
leads to a decision to put the rug at about −0.003 on that scale. We need a variable to
hold this value:
. gen where = -0.003
72 Speaking Stata

In practice, we can just choose a trial value and then use replace to improve upon
it. Next, there is no pipe symbol in the symbolstyle portfolio, so we must enlist the
pipe character as a marker label. Then, the rug is just a scatterplot of where against
wage, suppressing the default marker symbol and placing the marker label exactly on
target:
. gen pipe = "|"
. histogram wage, start(0) width(5)
> plot(scatter where wage, ms(none) mlabel(pipe) mlabpos(0)) legend(off)
.08
.06
Density
.04
.02
0

|||||||||||||||||||||||||||||||| ||||||||||| | | | | | | | |

0 20 40 60 80
wages in 1000s of dollars

Figure 3: Example of histogram with rug showing distinct values occurring in the data.

In this case, as shown in figure 3, the rug shows rounding of the data, and a tabulation
makes it explicit that all values are just multiples of $1,000. In general, a rug is a useful
but restrained way of showing some of the fine structure of a distribution.
A rug will take up a lot of bytes in a graph file if any point symbol stands for
many repeated values. Clearly, it is unnecessary to overwrite each symbol repeatedly.
A solution is to select each distinct value just once. There are two systematic ways
to do this, to select the first in each group after sorting or to select the last, and it is
immaterial here which you use, so you might as well go
. bysort wage: gen tag = _n == 1

A canned near-equivalent is
. egen tag = tag(wage)

The difference is that the egen call sorts your data while doing the calculation but then
returns it to its original sort order, which may differ. The first method may change your
sort order. Having done this, we select points for the rug as if tag.
N. J. Cox 73

2.4 Horizontal histogram bars

The histogram display, by default, has the frequency axis vertical, as is conventional.
The manual entry [G] graph combine shows how histograms may be placed vertically
and horizontally along the margins of a scatterplot. More generally, the horizontal
option may be used to reverse axes. This may sound merely cosmetic, but there are
occasions in which this layout appears more natural. In the environmental sciences,
among other fields, height above and depth below some surface are key natural variables.
The extra option yscale(reverse) would show depths the intuitive way up.
Here is a histogram of the mean elevations of 27,523 glaciers from Central Asia and
southern Siberia (figure 4). Data were extracted from the World Glacier Inventory. The
tendency to multiple modes is best interpreted as a consequence of lumping together
several distinct mountain ranges. The Stata command was

. histogram mean_elev, horizontal start(1600) width(100) frequency

> ylabel(, angle(horizontal))

7000

6000
mean elevation, m

5000

4000

3000

2000

0 500 1000 1500 2000 2500

Frequency

Figure 4: Example of histogram with horizontal bars. In this case, the response variable
is altitude.

2.5 Do-it-yourself density calculation

From a simple enhancement of histograms, we turn to the basic underlying calculations.
It may be useful to document how to calculate densities yourself, having first chosen a
start and a width. If we are counting so that each bin is defined precisely as lower limit
≤ value < upper limit, we could use floor() to generate lower limits as bin identifiers.
With the reverse convention, we could use ceil(). See Cox (2003c) for a note on these
functions. Then, the frequencies are the counts within each bin:
74 Speaking Stata

. gen double lower = width * floor((varname - start) / width)

. bysort lower: gen frequency = _N

To get fractions and percents, we must be careful to count each value just once:
. by lower: gen double sum = frequency * (_n == 1)
. replace sum = sum(sum)
. gen fraction = frequency / sum[_N]
. gen percent = 100 * fraction

The density is then

. gen density = frequency / (width * _N)

Real calculations get messier once you build in selections if or in, subdivision into
groups defined by other variables, or missing values. The messy details are coded up in
an egen function density() in the egenmore package on SSC.

3 Relatives of histogram: spikeplot, dotplot, and

onewayplot
One common reaction to histograms is to prefer more information in displaying distri-
butions, especially in the tails. The optimistic view is that more details will turn out
to be instructive fine structure. The corresponding pessimistic view is, naturally, that
such details will be best regarded as noise and, as such, an irreducible nuisance. Most
discussions stress the latter view over the former, but there can be real merit in playing
deterministic detective rather than stochastic skeptic.
The official commands spikeplot and dotplot and the user-written command
onewayplot offer different ways of showing more detail than do equivalent histograms.
spikeplot, by default, offers a spike for every distinct value—that is, no binning—
and the opportunity to control binning by a round() option, which in effect controls
bin width. Historically, spikeplot offered, before Stata 8, the most obvious official
alternative to graph, histogram for getting a histogram-like display with more than
50 bins. That role is now lost. However, its discrete representation of a frequency
distribution remains available for occasions when you want to emphasize the granularity
of data, either as defined in principle (counted variables, in particular) or as measured
in practice. The display of the age distribution of Ghana given at [R] spikeplot is a
good example of what spikeplot does best, revealing a fine structure of age preferences,
including multiples of 5, even rather than odd ages, and so forth. There is some scope for
controlling spike appearance if the default appearance (which is the default of twoway
spike under the prevailing scheme) appears too exiguous.
dotplot, in contrast, is based on the idea (or the ideal) of showing a point symbol for
each value; exactly the same description covers rugs and onewayplot. Similar plots un-
der a variety of names go back at least as far as van Langren (1644); see Tufte (1997, 15).
Wilkinson (1999) gives several further references of historical interest. Chambers et al.
(1983) used the term one-dimensional scatterplots. The term oneway plots appears to
N. J. Cox 75

have been introduced by StataCorp in its earlier guise as Computing Resource Center
(1985). Wild and Seber (2000) show many interesting examples of oneway plots.
dotplot, by default, offers, as far as possible, a point symbol for every value and
some binning. Binning can be controlled rather indirectly, although in practice, the
default is usually adequate, and when desired, the binning can be switched off with the
nogroup option. The main virtues of dotplots lie in their ability to show some features
that might otherwise be obscured by a series of touching bars, especially granularity and
details of outliers or other extreme values in the tails. You can also show, for example,
median and quartiles by horizontal marks and thus hybridize box plots and histograms.
The considerable flexibility of histogram, spikeplot, and dotplot might seem to
leave few important gaps in their territory. Nevertheless, onewayplot was written to
provide some extra possibilities in this area; it also may be downloaded using ssc. As
mentioned earlier, graph, oneway did not survive as such into Stata 8, although the
minor trickery needed to add rugs is just one illustration of how they can be emulated
fairly easily. onewayplot is essentially a convenience command that bundles together
various easy but tedious handles for making your own oneway plots. You can choose
between horizontal and vertical layouts, while stack and center options produce a vari-
ant on dotplot.1 There is, by default, no binning of data; binning may be accomplished
with the width() option.
In figure 5, we show the results of a onewayplot using the handle of a regional
classification to split the glacier elevations. Both histogram and dotplot struggle
given 18 regions, some with fairly long names.
. onewayplot mean_elev, by(region) ytitle("") stack ms(oh) msize(tiny) width(20)

(Continued on next page)

1 You can also type centre. An undocumented feature of dotplot is that centre is allowed as well

as center. This is a convenience for speakers of languages, such as English, which use that spelling,
and is emulated by onewayplot.
76 Speaking Stata

Qilian
Kodar
E Sayan
NE Altai
NW Altai
S Altai
Saur
Bogdo & Karlik
Iren Horu
Dzhungar
E Tien; Narat Tarliskay
C Tien; Pobedy
Zaili Kungey
Naryn Tien; Terskei Kokshaal
NW Tien; Chatkal Talas Kyrgyz
Alay Tien
SW Tien; Turkestan Hissar
Pamir

2000 3000 4000 5000 6000 7000

mean elevation, m

Figure 5: Example of multiple histograms produced by onewayplot.

4 Kernel-density estimation
4.1 Available commands
The histogram of a continuous variable is, from one point of view, an estimator of
the density function of that variable. Clearly the set of bins used to compute that
estimate imposes discontinuities on the estimate, which leads us directly to consider
smoother estimates, especially those based on convolution of the data and a sym-
metric kernel. twoway kdensity and kdensity are provided in official Stata as basic
commands. Recently, users have added variable kernel density estimation commands
(Salgado-Ugarte and Pérez-Hernández 2003; Van Kerm 2003).

4.2 Variations on the official theme

Transform before and after estimation

Some simple devices extend the range of applications of Stata’s official commands for
kernel-density estimation. First is the idea of estimating the density function on a trans-
formed scale and then back-transforming the estimate to one for the raw scale. Two of
the most natural transformations here, as elsewhere, are logarithms for positive variables
N. J. Cox 77

and logit-like transformations for proportions and other data measured on some interval
(a, b). The underlying general principle is that, for a continuous monotone transforma-
tion t(x), the densities f (x) and f {t(x)} are related by f (x) = f {t(x)}|dt/dx|. This
procedure is mentioned briefly by Silverman (1986, 27–30), although his worked exam-
ple (page 28) is not very encouraging. Good expositions are given by Wand and Jones
(1995, 43–45), Simonoff (1996, 61–64), and Bowman and Azzalini (1997, 14–16).
With a logarithmic transformation of x, we have

estimate of f (x) = estimate of f (log x) × (1/x)

given that d/dx(log x) = 1/x. Note in particular, if data are right skewed, that the
result of this transformation is more smoothing in the tail and less near the main part
of the distribution than in the default method. I have found this to be one of the
most valuable ways of going beyond the default. It fits very well both the common
finding that positive variables are right-skewed, suggesting a transformation, such as
the logarithm, and the common attitude that results on the original scale are of direct
scientific or practical interest. To put it another way, the transformation behaves more
like a link function than a classical transformation, given that end results are on the
scale of the original response. You can get the best of both worlds.
Returning to the wage data, here is an illustrative (and certainly not definitive)
example, in which we just use default kernel and width choice.
. gen logwage = log(wage)
. kdensity logwage, at(logwage) generate(densitylog)
. gen density = densitylog/wage
. levels wage, local(levels)
. line density wage, sort xtick(‘levels’, tposition(inside))
.08
.06
density
.04
.02
0

0 20 40 60 80
wages in 1000s of dollars

Figure 6: Example of a density function estimated on logarithmic scale and transformed

to the original scale of data.
78 Speaking Stata

The density function, shown in figure 6, is much smoother in the tails than the
equivalent default, which is not shown here. However, the step in the left-hand tail needs
investigation: is this some odd artifact or a genuine feature of the data? Incidentally,
another technique is used to show a rug by picking up a list of distinct values from
levels (added to Stata on 16 April 2003). However, this technique is not as general as
that previously illustrated, as it hinges on the variable concerned having only integer
values. levels is not designed to work with noninteger numeric values.
Similarly with a logit-like transformation,
(b − a)
estimate of f (x) = {estimate of f (logit x)}
(x − a)(b − x)
where logit x = log{(x − a)/(b − x)}, a slight generalization of the usual definition, for
which a = 0, b = 1. Note that d/dx(logit x) = (b − a)/{(x − a)(b − x)}.

Density on a log scale

It can be natural to calculate f (log x) as a way of getting a better estimate of f (x). It

can also be natural to calculate log{f (x)} as way of getting a better visualization of
f (x). This seems to be an old idea, periodically rediscovered. One venerable reference
is the work of the soldier, explorer, and scientist R. A. Bagnold (1937, 1941), who
worked on size distributions—especially those of sand—while at present the idea is
widely used in fields ranging from statistical physics (Bardou et al. 2002) to statistical
finance (Hazelton 2003). There is clearly no barrier also to looking at log{f (log x)} or
using some other transformation before density estimation if it seems appropriate.
The highly original contribution of Bagnold deserves some explanation, as it appears
to be little known within the statistical sciences. Born in England in 18962 , he joined
the British army from school and served in the First World War. He then took an
engineering degree at Cambridge. Remaining in the army, he used leaves to travel and
explore, particularly on pioneering long trips into the deserts of Egypt, Sudan, and
Libya using specially adapted cars. This provoked an interest in the physics of blown
sand, leading ultimately to a now-classic monograph (Bagnold 1941). Wind transport
of loose particles is highly size selective, as ordinary experience confirms: very coarse
material will not move, while very fine material may easily be lofted high into the
atmosphere and carried over vast distances. Thus, the particle size distribution of a
deposit (say from a sand dune) is of central interest. Bagnold found plots of log density
versus log grain diameter the most helpful way to show his data. It seems clear from his
very readable autobiography (1990), published just after his death, that the crucial first
step of plotting densities on a log scale owed most to an engineer’s feeling of a sensible
thing to do. By thinking for himself, he was not inhibited by ideas on what was or
was not standard statistical practice. Much later, Bagnold returned to the question and
contributed to the development of log-hyperbolic distributions by Ole Barndorff-Nielsen.
There is a full bibliography of his publications in Thorne et al. (1988).
2 His sister was Enid Bagnold, later a novelist, dramatist, and poet, and best remembered for the

children’s classic National Velvet.

N. J. Cox 79

Several properties are simple on a log-density scale. One exploited by Bagnold is

that a normal (Gaussian) density plots as a parabola
 √  (x − µ)2
log f (x) = − log σ 2π −
2σ 2
while exact or approximate exponential or power-law decay of density will show exact or
approximate linear patterns, the latter requiring also a logarithmic scale for the variable.
Let us illustrate with log of wage from the wage data considered above:

. gen logdensitylog = log(densitylog)

. qui summarize logwage
. local mean = r(mean)
. local sd = r(sd)
. scatter logdensitylog logwage
> || function normal =
> -log(‘sd’ * sqrt(2 * _pi)) - ((x - (‘mean’))^2 / (2 * ‘sd’^2)),
> ra(logwage) ytitle(log density) xtitle(log wage)
> legend(off) subtitle(log density plot)

log density plot

0
2
log density
4
6

1 2 3 4 5
log wage

Figure 7: Example of plot using log density. The parabola shows a normal density
function with the same mean and standard deviation as log wage.

The results in figure 7 suggest a good but not spectacularly good fit to a lognor-
mal. The slightly fat tails seem suggestive. At the same time, the density estimates,
especially in the tails, are, as always, subject to sampling variation and sensitive to
kernel bandwidth; note also that neighboring density estimates are necessarily highly
dependent.
80 Speaking Stata

What is implemented above is just a first stab. Hazelton (2003) suggests various
refinements, including robust estimation of the mean and standard deviation and, given
a density estimation bandwidth h, fitting a normal with variance sd2 + h2 to correct
for side-effects of using a kernel.

Density on a root scale

There would also be some advantages to a square-root scale, given that densities behave
a bit like counts, for which a root transformation is often the first to be tried. Also, the
square root of a Gaussian shape is another Gaussian shape. So we can have our cake
and eat it too: hunt for a Gaussian yet benefit from stabilized sampling fluctuation.
Check the assertion with
. twoway function sqrt(normden(x)), range(-4 4)

There is a root option in spikeplot for a similar reason. Tukey (1977, chapter 17)
worked through a bundle of related ideas, which seem to have been little explored since.

Intensities, too

Those interested in data on events, considered as the result of a point process in one
dimension (most obviously, time or space), should note that Stata’s kernel-density com-
mands can readily be used to estimate the intensity function (say, frequency per unit
of time or space). Suppose that a variable contains dates of earthquakes, eruptions,
strikes, honors for a sports team, or whatever else is of interest. To get results on an
intensity scale, just multiply ‘density’ by the number of observed data points. A key
detail is that intensities will be smoothed beyond the beginning and the end of the
interval in question; whether this is tolerable or further surgery is desired is a question
for the user.

5 Quantile plots and distribution plots

Another key approach eschews any kind of binning or smoothing and starts with the
idea of directly showing the pattern of the quantiles (the ordered values). Formally,
we order n data values for a variable x and label them such that x(1) ≤ · · · ≤ x(n) .
quantile has long been available as an official Stata command for quantile plots, in
which the x(i) are plotted against (i − 0.5)/n. See [R] diagnostic plots. The term
quantile plot appears in Chambers et al. (1983) and Cleveland (1993, 1994). Modern
use of quantile plots and their relatives stems largely from the path-breaking paper of
Wilk and Gnanadesikan (1968). Examples of antecedents from the nineteenth century
can be found in Quetelet (1827) and Galton (1875); see Stigler (1986, 167, 270).
Essentially the same information can be shown in a plot of cumulative distribution
functions or of survival (a.k.a., survivor, reliability, complementary, or reverse distribu-
tion) functions, in which we plot either probabilities or frequencies of values being ≤ x
N. J. Cox 81

or > x. In many biomedical or engineering applications, the survival function appears

closer to the practical problem, the name being suitably evocative if data are indeed
times to patient death, component failure, or something similar.
According to Hald (1990, 108), the first graph of (the complement of) a distribution
function appears in a 1669 letter from Christiaan Huygens (1629–1695) to his brother
Lodewijk (1631–1699). He plotted a survival function from data from the life table
of John Graunt (1620–1674). Huygens made numerous contributions to mathematics,
astronomy, and physics, studying, among many other matters, games of chance, the
collision of elastic bodies, the rings of Saturn, the pendulum clock, and the wave theory
of light.
In Stata 8, the graphics of quantile were revised to match the new graphics, but
the functionality was unchanged. A broader command is qplot (Cox 2004), which in
most respects is a generalization of quantile; just one detail is omitted, the reference
line. It supersedes the previous program quantil2 (Cox 1999b, 2001).
Stata already has an official graph command for survival functions, sts graph. If
your data really are survival times and you have any of the complications that are the
stuff of survival analysis, such as censoring or subjects entering at different times, you
should use sts graph. However, it is not and does not purport to be a general purpose
command for all kinds of distribution.
In addition, the official command cumul ([R] cumul) is available to calculate the
cumulative distribution function for a single variable, after which the function may be
plotted using twoway. The user with several variables to be compared or with an interest
in survival functions thus needs to repeat the cumul command or take the further step
of calculating survival functions from cumulative distribution functions. A command
distplot that bundles calculation and graphing steps together is, however, available
(Cox 1999a, 2003a,b).
qplot and distplot are, in effect, siblings. The choice between them is most obvi-
ously one of choice of axes and thus, in a sense, trivial, but different conventions may
seem natural for different problems and even different fields or traditions. In particular,
there seems to be a growth of interest in quantile functions as responses, which makes
qplot a possible choice (see, for example, Gilchrist 2000).
qplot and distplot have in common

1. Support for graphing several variables.

2. Support for graphing several groups, through a by() option.
3. Choice of twoway plottypes, from area, bar, connected, dot, dropline, line,
scatter, or spike. These are not in general equally useful or attractive, but there
is at least much choice, courtesy of twoway’s generous design.
4. Support for reversing the sort order so that values decrease from top left.
5. Support for alternative transformed scales.
82 Speaking Stata

In addition, qplot has support for choice of a in a general rule for plotting position
(i − a)/(n − 2a + 1) for i = 1, . . . , n. The default is a = 0.5, giving (i − 0.5)/n. Other
choices include a = 0, giving i/(n + 1), and a = 1/3, giving (i − 1/3)/(n + 1/3). The
choice is often immaterial, but some authorities have strong opinions on the best choice
on various grounds, some even statistical. For more discussion and references, see Cox
(1999b).

6 Skewness plots
The skewness of a variable is often of interest, perhaps especially as an indicator of
potential problems in subsequent analysis. Commonly a single measure is used, whether
the moment-based measure produced by summarize, detail or other measures (which,
in most cases, are readily calculated from the output of summarize). Graphically,
skewness may be assessed with varying degrees of ease and efficiency from the plots
mentioned so far, but there is also a case for a customized design.
Various possibilities are based on the quantiles (Gnanadesikan 1977, 1997). The
quantiles may be paired as lower and upper quantiles x(1) and x(n) , x(2) and x(n−1) ,
etc., and a median may be calculated in the usual way.
Stata supports symplot, a plot of (upper quantile − median) versus (median − lower
quantile), for which the reference situation of symmetry or lack of skew plots as a line of
equality. See [R] diagnostic plots. However, symplot will show only a single group of
data and thus cannot be used for comparisons, while a plot with a sloping reference line
is more difficult to deal with than the plot now to be described, which has horizontal
reference lines.
skewplot produces, by default, a plot of the midsummary versus the spread for the
variables supplied, also known as the mid-versus-spread plot. With the skew option,
it produces a plot of the skewness function versus the spread function. Such plots
convey both the general character and the fine structure of the symmetry or skewness of
datasets and can be used to compare distributions or to assess whether transformations
are necessary or effective.
There are some little-used terms here, so we need a few definitions. In a perfectly
symmetric set of data, the midsummaries (x(1) + x(n) )/2, (x(2) + x(n−1) )/2, etc., would
all be identical and equal to the median. A plot of each midsummary (or mean of lower
and upper quantiles) (x(i) + x(n−i+1) )/2 versus each difference or spread of lower and
upper quantiles x(n−i+1) − x(i) would, thus, yield a horizontal straight line. Conversely,
skewness in sets of data will be reflected by departures from horizontality. In particular,
right skewness would be shown by rising lines and left skewness by falling lines.
Apart from the divisor of 2, this plot was suggested by J. W. Tukey (Wilk and
Gnanadesikan 1968). See also Gnanadesikan (1977, 1997, chapter 6.2) or Fisher (1983).
The form used here and the name mid-versus-spread plot are found in Hoaglin (1985).
It is usual to plot only that half of the sample results for which spread is ≥ 0.
N. J. Cox 83

The skew option produces an alternative form promoted by Benjamini and Krieger
(1996, 1999). Consider the identity, which introduces their terminology,

x(n−i+1) = median + (x(n−i+1) − x(i) )/2 + (x(i) + x(n−i+1) − 2 × median)/2

= median + spread function + skewness function

for x(i) in the lower half of a sample. This leads to a plot of the skewness function
versus the spread function, known as the skewness versus spread plot. Note that the
skewness function is (midsummary − median) and so will be constant and zero for a
perfectly symmetric distribution and that the spread function is half the spread of the
mid versus spread plot. In short, the skew option does not change the configuration of
the plot but merely the labeling of the axes.
In addition, the ratio of the skewness and spread functions or
x(i) + x(n−i+1) − 2 × median
x(n−i+1) − x(i)

is a measure of skewness (in the traditional sense) originally suggested for quartiles by
Bowley (1902) and generalized to this form by David and Johnson (1956). Another
incarnation is as the p-skewness index (Gilchrist 2000, 54, 72).3 It varies between −1
and 1. A similar general measure was used by Parzen (1979). Graphically this measure
is the slope of the line connecting (0, 0) and each data point if the skew option is used.
See Benjamini and Krieger (1996, 1999) and Groeneveld (1998) for concise reviews
tracing such ideas from late 19th-century antecedents to recent work and further details
on the interpretation of the skewness-versus-spread plot.
Let us close with an example for data on 158 glacial cirques from the English Lake
District (Evans and Cox 1995). Glacial cirques are hollows excavated by glaciers that
are open downstream, bounded upstream by the crest of a steep slope (wall), and
arcuate in plan around a more gently sloping floor. More informally, they are sometimes
described as “armchair-shaped”. Glacial cirques are common in mountain areas that
have or have had glaciers present. Three among many possible measurements of their
size are length, width and wall height, and the distribution of all in the area studied is
shown by
. skewplot length width wall_height, legend(ring(0) position(5) column(1))

to be markedly right skew (figure 8). Logarithmic transformation seems an obvious

possibility, after which

3 Gilchrist calls the special case for quartiles Galton’s skewness (pages 8, 25, 53, and 72), but there

is no evidence that Galton used it.

84 Speaking Stata

1000
800
Midsummary
600400

length, m
width, m
200

wall height, m

0 500 1000 1500

Spread

Figure 8: Skewness plot for three variables. The systematic upward drift indicates
marked right skewness.

. skewplot log_length log_width log_wall_height, legend(ring(0) position(3)

> column(1))
2.8 2.6
Midsummary

log 10 length in m
log 10 width in m
log 10 wall height in m
2.4
2.2

0 .2 .4 .6 .8 1
Spread

Figure 9: Skewness plot for three log-transformed variables. Approximately horizontal

patterns indicate that transformations have yielded near symmetry of distributions.

shows approximate lack of skew (although a hint of mild overtransformation in wall

height, which is best left alone for simplicity) (figure 9). An important feature of such
plots is that the effects of outliers are localized.
N. J. Cox 85

7 Conclusions
With one command or another, users can now plot univariate distributions in many
different ways. You can choose between several depictions of the density function or
several depictions of the distribution function or its inverse, the quantile function. You
can choose discrete or continuous representations and vertical or horizontal alignments.
Less obviously, it is straightforward to add details (for example, rugs of distinct data
values) or exploit the inbuilt flexibility of graph (for example, by looking at density
estimates on a log scale or by constructing your own histogram with varying bin width).
The theme of distributions will continue into the next column but with a focus on
categorical data. Distributions of categorical variables may be shown in a variety of
displays: the survey will range from old staples to less well-known plots, with emphasis
on the important special cases of graded data and of three variables with constant sum.

8 Acknowledgments
Ian Evans provided the Lake District cirques data, pointed me to the World Glacier
Inventory data, and participated in many discussions on statistical graphics over more
than 30 years. Marcello Pagano persistently urged the merits of equal-probability his-
tograms and parenthetically underlined the connection with chi-squared tests. Vince
Wiggins specifically alerted me to spanning bars and generally advised on strategies and
tactics for using the new graphics. Elizabeth Allred, Ronán Conroy, Philip Ender, and
Roger Harbord made helpful comments during development of various predecessors or
versions of some programs discussed here. Richard Groeneveld kindly tracked down the
Bowley reference.

9 References
Altman, D. G. 1991. Practical Statistics for Medical Research. London: Chapman &
Hall.

Bagnold, R. A. 1937. The size-grading of sand by wind. Proceedings of the Royal

Society Series A 163: 250–264.

—. 1941. The Physics of Blown Sand and Desert Dunes. London: Methuen.

—. 1990. Sand, Wind, and War: Memoirs of a Desert Explorer. Tucson: University of
Arizona Press.

Bardou, F., J.-P. Bouchaud, A. Aspect, and C. Cohen-Tannoudji. 2002. Lévy Statistics
and Laser Cooling: How Rare Events Bring Atoms to Rest. Cambridge: Cambridge
University Press.

Benjamini, Y. and A. M. Krieger. 1996. Concepts and measures for skewness with
data-analytic implications. Canadian Journal of Statistics 24: 131–140.
86 Speaking Stata

—. 1999. Skewness—concepts and measures. In Encyclopedia of Statistical Sciences

Update, ed. S. Kotz, C. B. Read, and D. L. Banks, vol. 3, 663–670. New York: John
Wiley & Sons.

Bowley, A. L. 1902. Elements of Statistics. 2d ed. London: P. S. King.

Bowman, A. W. and A. Azzalini. 1997. Applied Smoothing Techniques for Data Anal-
ysis: The Kernel Approach with S-Plus Applications. Oxford: Oxford University
Press.

Breiman, L. 1973. Statistics: With a View towards Applications. Boston: Houghton

Mifflin.

Chambers, J. M., W. S. Cleveland, B. Kleiner, and P. A. Tukey. 1983. Graphical

Methods for Data Analysis. Belmont, CA: Wadsworth.

Cleveland, W. S. 1993. Visualizing Data. Summit, NJ: Hobart Press.

—. 1994. The Elements of Graphing Data. Summit, NJ: Hobart Press.

Computing Resource Center. 1985. STATA/Graphics User’s Guide. Los Angeles, CA:
Computing Resource Center.

Cox, N. J. 1999a. gr41: Distribution function plots. Stata Technical Bulletin 51: 12–
16. In Stata Technical Bulletin Reprints, vol. 9, 108–112. College Station, TX: Stata
Press.

—. 1999b. gr42: Quantile plots, generalized. Stata Technical Bulletin 51: 16–18. In
Stata Technical Bulletin Reprints, vol. 9, 113–116. College Station, TX: Stata Press.

—. 2001. gr42.1: Quantile plots, generalized: update to Stata 7.0. Stata Technical
Bulletin 61: 10–11. In Stata Technical Bulletin Reprints, vol. 10, 55–56. College
Station, TX: Stata Press.

—. 2003a. Software update: gr41 1: Distribution function plots. Stata Journal 3(2):
211.

—. 2003b. Software update: gr41 2: Distribution function plots. Stata Journal 3(4):
449.

—. 2003c. Stata tip 2: Building with floors and ceilings. Stata Journal 3(4): 446–447.

—. 2004. Software update: gr42 2: Quantile plots, generalized. Stata Journal 4(1): 97.

David, F. N. and N. L. Johnson. 1956. Some tests of significance with ordered variables.
Journal of the Royal Statistical Society, Series B 18: 1–20.

Davison, A. C. 2003. Statistical Models. Cambridge: Cambridge University Press.

Evans, I. S. and N. J. Cox. 1995. The form of glacial cirques in the English Lake District,
Cumbria. Zeitschrift für Geomorphologie 39: 175–202.
N. J. Cox 87

Fisher, N. I. 1983. Graphical methods in nonparametric statistics: a review and anno-

tated bibliography. International Statistical Review 51: 25–38.
Freedman, D., R. Pisani, and R. Purves. 1998. Statistics. New York: W. W. Norton.
Galton, F. 1875. Statistics by intercomparison, with remarks on the law of frequency
of error. Philosophical Magazine, 4th series, 49: 33–46.
Gilchrist, W. G. 2000. Statistical Modelling with Quantile Functions. Boca Raton, FL:
Chapman & Hall/CRC.
Gnanadesikan, R. 1977. Methods for Statistical Data Analysis of Multivariate Obser-
vations. New York: John Wiley & Sons.
—. 1997. Methods for Statistical Data Analysis of Multivariate Observations. 2d ed.
New York: John Wiley & Sons.
Groeneveld, R. 1998. Skewness, Bowley’s measure of. In Encyclopedia of Statistical
Scienes Update, ed. S. Kotz, C. B. Read, and D. L. Banks, vol. 2, 619–621. New
York: John Wiley & Sons.
Gumbel, E. J. 1943. On the reliability of the classical chi-square test. Annals of Math-
ematical Statistics 14: 253–263.
Hald, A. 1990. A History of Probability and Statistics and their Applications before
1750. New York: John Wiley & Sons.
Hazelton, M. L. 2003. A graphical tool for assessing normality. American Statistician
57: 285–288.
Hoaglin, D. C. 1985. Using quantiles to study shape. In Exploring Data Tables, Trends,
and Shapes, ed. D. C. Hoaglin, F. Mosteller, and J. W. Tukey, 417–460. New York:
John Wiley & Sons.
van Langren, M. F. 1644. La Verdadera Longitud po Mar y Tierra. Antwerp.
Mann, H. B. and A. Wald. 1942. On the choice of the number of class intervals in the
application of the chi-square test. Annals of Mathematical Statistics 13: 306–317.
Parzen, E. 1979. Nonparametric statistical data modeling. Journal of the American
Statistical Association 74: 105–131.
Quetelet, A. 1827. Recherches sur la population, les naissances, les décès, les prisons,
les dépôts de mendicité, etc., dans le Royaume des Pays-Bas. Nouveaux Mémoires de
l’Académie Royale des Sciences et Belles-lettres de Bruxelles 4: 117–192.
Salgado-Ugarte, I. H. and M. A. Pérez-Hernández. 2003. Exploring the use of variable
bandwidth kernel density estimators. Stata Journal 3(2): 133–147.
Scott, D. W. 1992. Multivariate Density Estimation: Theory, Practice, and Visualiza-
tion. New York: John Wiley & Sons.
88 Speaking Stata

Silverman, B. W. 1986. Density Estimation for Statistics and Data Analysis. Mono-
graphs on Statistics and Applied Probability, London: Chapman & Hall.

Simonoff, J. S. 1996. Smoothing Methods in Statistics. New York: Springer.

Stigler, S. M. 1986. The History of Statistics: The Measurement of Uncertainty before

1900. Cambridge, MA: Harvard University Press.

Thorne, C. R., R. C. MacArthur, and J. B. Bradley, ed. 1988. The Physics of Sediment
Transport by Wind and Water: A Collection of Hallmark Papers by R. A. Bagnold.
New York: American Society of Civil Engineers.

Tufte, E. R. 1997. Visual Explanations: Images and Quantities, Evidence and Narrative.
Cheshire, CT: Graphics Press.

Tukey, J. W. 1977. Exploratory Data Analysis. Reading, MA: Addison–Wesley.

Van Kerm, P. 2003. Adaptive kernel density estimation. Stata Journal 3(2): 148–156.

Wand, M. P. and M. C. Jones. 1995. Kernel Smoothing. London: Chapman & Hall.
Wild, C. J. and G. Seber. 2000. Chance Encounters: A First Course in Data Analysis
and Inference. New York: John Wiley & Sons.

Wilk, M. B. and R. Gnanadesikan. 1968. Probability plotting methods for the analysis
of data. Biometrika 55: 1–17.

Wilkinson, L. 1999. Dot plots. American Statistician 53: 276–281.

About the Author

Nicholas Cox is a statistically minded geographer at the University of Durham. He contributes
talks, postings, FAQs, and programs to the Stata user community. He has also co-authored
fourteen commands in official Stata. He was an author of several inserts in the Stata Technical
Bulletin and is Executive Editor of the Stata Journal.