UCD GEARY INSTITUTE FOR PUBLIC POLICY

DISCUSSION PAPER SERIES

Basic Stata graphics for economics students

Kevin J. Denny
School of Economics & Geary Institute for Public Policy
University College Dublin

Geary WP2018/13
July 11, 2018

UCD Geary Institute Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of
such a paper should account for its provisional character. A revised version may be available directly from the author.

Any opinions expressed here are those of the author(s) and not those of UCD Geary Institute. Research published in this series
may include views on policy, but the institute itself takes no institutional policy positions.

1
 Basic Stata graphics for economics students

Kevin Denny
School of Economics & Geary Institute for Public Policy
University College Dublin
11th July 20181

This paper provides an introduction to the main types of graph in Stata that economics
students might need. It covers univariate discrete and continuous variables, bivariate
distributions, some simple time plots and methods of visualising the output from estimating
models. It shows a small number of the many options available and includes references to
further resources.

Keywords: Stata, graphics, data visualization

JEL codes: A2, C87, Y10

1
This guide was written with my students in mind. I am releasing it into the wild as others may find it helpful but
mainly as a commitment device to prevent me extending it further. Comments to [Link]@[Link]
2
 1. Introduction

It has often been said that a picture paints a thousand words. This is certainly true when it
comes to data analysis. There are two good reasons to acquire some skills in graphing your
data: (1) Graphical methods are a powerful way for a researcher to explore data and (2)
graphs can be a very useful way of illustrating your data and results whether it is in a
presentation, a project or a thesis.

To motivate the first reason above, consider the following set of graphs:

These four graphs are collectively known as Anscombe’s Quartet. You may be surprised to
learn that the x variables in these four graphs have the same mean and the same variance.
This is also true of y and, moreover, the covariance between x and y is the same & hence the
regression line is the same. Clearly they are very different relationships. Without graphing
the data, you would probably never know.

In this document I show how to use Stata to generate some of the key graphs that economics
students should know about and should consider using in their projects, presentations and
theses. There are several good online treatments of Stata graphics (listed at the end). Stata’s
Youtube channel has videos on graphics which are excellent. The book by Michael Mitchell is

3
a fantastic resource which you could also draw on. Andrew Jones’ guide, though designed for
health econometrics, is of general interest if you are using Stata. Jesse Shapiro has a great set
of slides on preparing a good applied microeconomics presentation but is relevant for any
applied economics talk & probably other social sciences. Here I am going to outline the main
methods that I think economics students should know at a minimum. Along the way I show a
few of the many options available to whet your appetite. The definitive source of information
is the Stata Graphics Manual which is a mere 739 pages long. A classic text on data
visualization and graphics is Tufte (2001). For a shorter guide targeted at economists see the
paper by Schwabish (2014).

All of the datasets I use here are either available online & can be accessed in Stata using the
webuse command or they are provided with Stata and can be accessed using sysuse. To
switch from one dataset to another you need to use clear first. Stata commands will be in
bold. A basic knowledge of Stata is required. There are two ways to create graphs in Stata.
You can either (a) use a written command which can be done interactively in the command
line or written in a do-file or (b) you can use the dialogue boxes/pull-down menus at the top.

A nice feature is that if you use the dialogue box to create a graph Stata will show you the
equivalent syntax in the output window so you can learn how to generate the graph. You
could copy the syntax into a do-file so you can repeat the exercise. So I tend to use the pull-
down menus to experiment until I get the graph looking like I want. Then I copy the syntax
that generates it from the output window into my do-file so I can replicate it later.

When Stata produces a graph for you on the screen click on “file” at the top left: you can
either save it or you can open the editor to make further changes. Stata’s native format for
graphs is .gph. If you want to include it in a Word document or a Powerpoint file for example
you need to save it to a format like portable network graph (. png). , a tiff file (.tiff), a
Windows meta file (. wmf) or a postscript file (.ps). You may need to experiment saving to
different formats to get something that works with your document. As .wmf files only work
with Windows, if you are a Mac user or you are collaborating with one it might be best to
avoid them: .emf is better. If in doubt I recommend saving the graph as .png. Postscript files
can end up taking a lot of space if there are a large number of data points in your graph.

4
A feature I will not discuss here is that you can create two graphs separately and then
combine them into one graph. Koffman (2015) has a few slides on this or help graph
combine.

The Stata graphics editor has numerous options & you can customize the graph in many
many ways. It is beyond the scope of this document to describe how. Here I am mostly going
to use the graph commands that come with Stata. However there are some good user-written
commands for Stata graphics that are freely available online. You can find and download
them within Stata using the findit command. Here I will draw on just four of these:
binscatter, coefplot, fabplot and vioplot. To download the first of these say, just type:
findit binscatter in the command line or ssc install binscatter. Hit return and follow the
steps.

As these user created commands are occasionally revised, it is worth using the adoupdate
command periodically to ensure you have the latest version.

5
 2. Distributions

When you are analysing data it is essential that you carefully explore the data before you get
stuck into modelling using it using econometric methods. You really need to get to know your
data. There are a few reasons for this. One is that exploring the data will sometimes show up
anomalies, for example there might be crazy values like missing values code as -1 or 99. The
main reason is to get a sense of what the basic patterns are. This is particularly the case for
variables that you create from the raw data. It is very easy to make a mistake – even
experienced users do - so if you generate a new variable you want to check does it look
sensible.

2.1 Univariate
We will first consider looking at the distribution of a single variable. You should certainly
have a good look at your key variables before you do any modelling.

2.1.1 Discrete
For a discrete variable you should use a histogram:

webuse fullauto

ta rep78 generates a table of this discrete variable. This is fine as far as it goes and you may
want to include a table like this in your document particularly if this is your dependent
variable. Note that to keep the table nicely aligned as it is in Stata you need to use the Courier
font. However it may be hard to get a sense of the distribution simply by looking at the table.
If you are preparing a presentation, for example, you want the audience to easily grasp what
the data looks like. Let’s graph it next.

6
 Repair |
Record 1978 | Freq. Percent Cum.
------------+-----------------------------------
Poor | 2 2.90 2.90
Fair | 8 11.59 14.49
Average | 30 43.48 57.97
Good | 18 26.09 84.06
Excellent | 11 15.94 100.00
------------+-----------------------------------
Total | 69 100.00

histogram rep78, discrete fraction creates a histogram where the heights of the bars gives
the fractions in each category. If you want to show percentages instead simply replace
fraction with percent. If you want absolute frequencies use frequency.
.4
.3
Fraction

.2
.1
0

0 1 2 3 4 5
Repair Record 1978

You can generate a graph with separate histogram for different groups beside each other:
histogram rep78, discrete fraction by(, title(Figure 1) note(Data: auto)) by(foreign)

7
 Figure 1
Domestic Foreign

.6
.4
Fraction

.2
0

0 2 4 6 0 2 4 6
Repair Record 1978
Data: auto

See how I have added a title and a note at the bottom?

Another way of illustrating shares across categories is the pie chart. Pie charts are not to
everyone’s taste and some are less than helpful so use them wisely. Try graph pie,
over(rep78) as an alternative to a histogram. If you would like to see what % each slice
contains: graph pie, over(rep78) plabel(_all percent) .

A bar chart is a useful way of comparing some characteristic of a variable (like the mean)
across different categories of a variable. For example

graph bar (mean) price, over(rep78)

8
graph bar (mean) price, over(rep78) over(foreign) shows the mean across categories of
two variables

An alternative way of doing this where the two graphs are separate is:

graph bar (mean) price, over(rep78) by(foreign)

Note that bar charts aren’t just for means. You can use them to compare other statistics like
medians, standard deviations, maxima etc. When you use the pull-down menu there is a
“Statistic” option beside the variable. If you want the bars to be horizontal, replace bar with
hbar. The mean is the default. To generate a bar chart of standard deviations across the
categories of rep78 for example try:

graph bar (sd) price, over(rep78)

Stata has another type of graph which can be used to illustrate differences in means or
medians of a continuous variable across categories of one or more variables.

sysuse nlsw88, clear

grmeanby race married collgrad, summarize(wage) ytitle($) ytitle(, size(medlarge))

title(Mean wage differences) subtitle(hourly wage)

9
This example shows in one graph the differences in average earnings between the categories
of three variables. A glance at this graph suggests that education seems to be more important
than marital status. Note that I have added a title to the Y axis (the “$”) and changed the size
of the title. The command line I used continues onto a second line. If you were using this in a
do-file Stata has to know to read it as one line. Entering it like this in your do-file will work:

grmeanby race married collgrad, summarize(wage) ytitle($) ytitle(, size(medlarge)) /*

*/ title(Mean wage differences) subtitle(hourly wage)

The /* */ comments out the end of the line so Stata just reads it as one line. Adding median
after the “,” (the comma) in the command will show the medians instead e.g.:

grmeanby race married collgrad, summarize(wage) median

10
2.1.2 Continuous

There are a few ways to show the distribution of a continuous variable. You can use a
histogram as shown on page 5. My preferred method is to generate something called a kernel
density function.

kdensity price

The smoothness of the density is controlled by a bandwidth parameter. Stata calculates a

default parameter & reports it. You can see it is 605.6 in the above example. You can over-
ride this if necessary using the bandwidth option. For example by reducing the bandwidth to
400 you make it less smooth. Be careful not to over-smooth i.e. setting the bandwidth so high
that you remove key features of the data. There is a handy download akdensity which allows
the bandwidth to adapt optimally to how much data there is at a particular part of the
distribution. The “norm” option superimposes a normal distribution which can be useful if
you have reasons to believe that the variable should be normal:

kdensity price, bw(400) norm xtitle(US $) title(Distribution of car prices)

11
 Note how I have added a title and a label for the x axis. Sometimes you may wish to
superimpose two densities on top of each other. For example if you are looking at the
distribution of earnings, it might be useful to compare the earnings of men and women or
married and unmarried people.

Use the nlsw88 data

sysuse nlsw88, clear

kdensity wage if married==0 , addplot(kdensity wage if married==1)

Kernel density estimate

.15
.1
Density

.05
0

0 10 20 30 40
hourly wage

Kernel density estimate

kdensity wage
kernel = epanechnikov, bandwidth = 1.0162

The density shown in blue is for the unmarried (married==0) and red is for the married.
The legend below the graph is not very helpful unfortunately and you will need to edit it in
the graphics editor so you can end up with something like:

Kernel density estimate

.15
.1
Density

.05
0

0 10 20 30 40
hourly wage

Unmarried
Married
bandwidth = 1.0162

12
 An alternative way of comparing the distribution of a continuous variable across categories
of some other discrete variable is a boxplot (aka “box and whisker” plot). The middle line in
the box shows the median, the bottom and top of the box show the 25th & 75th quartiles,
respectively. So the height of the box is the IQR, the inter-quartile range.

The “whiskers” from the box extend vertically to the upper and lower adjacent values. Their
definition is somewhat tricky: think of a value U= the 75th percentile + (3/2)* IQR and L= the
25th percentile – (3/2)*IQR. The upper adjacent value is the value of x which is ≤ U. The lower
adjacent value is defined as the value of x which is ≥L.2 Essentially, the whiskers pick up the
extent to which the distribution is spread out outside of the IQR. Points outside this range are
shown as dots. In the example below we show the distribution over categories of two
variables, college graduate status and marital status:

graph box wage , over(collgrad) over(married):

40
30
hourly wage

20
10
0

not college grad college grad not college grad college grad
single married

The points at the top of each plot show that this variable is right (positively) skewed. These
points can sometimes distort the diagram so if you wish to omit them adding noout at the
end of the line will do. The command below will plot the bars horizontally and removes the
outliers:

graph hbox wage, over(married) noout

2
Noe that this is Stata’s implementation of box plots which goes back to the influential work of John Tukey (1977).
Other approaches are possible. For example some have the whiskers extend to the 10 th & 90th percentiles instead.
13
Violin plots are a useful way of combining box plots and densities invented by Hintze &
Nelson (1998). First download the vioplot package (“ssc install vioplot”). Then, using the
auto dataset:

vioplot mpg , over(rep78) title("Violin plot of mileage") subtitle("by repair record")

The white dot is a marker for the median, the thick line shows the interquartile range with
whiskers extending to the upper & lower adjacent values (as defined above). This is overlaid
with a density of the data. Violin plots contain a lot of information though you may need to
fiddle with the options to get it looking right.

14
2.2 Bivariate

To examine a bivariate distribution, start with a scatterplot

twoway (scatter price mpg , msymbol(Oh))

I used the msymbol option above to change the dots to an “O”. Scatterplots are not always
very illuminating and you may want to adjust them. It is simple to fit and plot a linear
regression to this data:
twoway (scatter price mpg) (lfit price mpg)

If you replace lfit with lfitci it shows the confidence interval around the line. Using qfit
instead fits a quadratic curve (and hence qfitci instead show the confidence intervals).

15
If you want to show a scatterplot for two different subsets of the data try:

scatter mpg weight if foreign || scatter mpg weight if !foreign , yline(20)

The blue dots refer to the first named subset (foreign cars). I have added a line
corresponding to y=20 with the yline(20) option. You can have more than one line: using say
xline(3000 4000) would create vertical lines corresponding to those values of x. You can
use this if there is a particular x or y value that is important (e.g. a particular year). This
syntax is another way of creating the same basic diagram:

twoway (scatter mpg weight if foreign) (scatter mpg weight if !foreign)

Note the variable foreign is either 0 or 1 so !foreign means “not foreign” i.e. foreign==0.
Sometimes a scatter plot has so many datapoints that you end up with a graph that’s not very
illuminating. Let’s switch to the nlsw88 data set to see this:

sysuse nlsw88, clear

scatter wage tenure (note this is the same as twoway (scatter wage tenure) )

16
 40
30
hourly wage

20
10
0

0 5 10 15 20 25
job tenure (years)

Not terribly clear is it? There is a handy Stata download called binscatter. It groups the x-
axis variable into equal-sized bins, computes the mean of the x-axis and y-axis variables
within each bin, then creates a scatterplot of these data points.

binscatter wage tenure

10
9
8
wage

7
6
5

0 5 10 15 20
tenure

What this brings out is there is a positive relationship between wages and job tenure.
binscatter has many nice feature that you can explore. It is possible to generate scatterplots
of 3 variables i.e. with 3 dimensional graphs using a download graph3d. These are trickier to
get in a form that is helpful. An alternative way of illustrating the bivariate relationship
between two variables is to fit a curve to the data. Stata has several ways of doing this. A
popular method is called lowess (for locally weighted scatterplot smoothing).

17
lowess wage tenure, by(married) lineopts(lwidth(medthick))

Here I have also made the line thicker than the default. An alternative curve-fitting technique
is local polynomial smoothing:

lpoly wage tenure, ci lineopts(lcolor(yellow) lwidth(thick)) title(Local polynomial

smoothing) note(nlsw88 data)

I have made the line even thicker & changed the colour to yellow. The confidence interval is
so tight for most of the range of tenure you can’t see it here except in the right tail where it
spreads out as there is so little data. Which technique you use is partly a matter of taste and
what works best with your data.

18
In the first graph on the previous page I used “by(married)” to generate separate graphs for
two subsets according to that variable. If I used twoway scatter wage tenure, by(married
union) I would have four separate graphs for each of the combinations e.g. single & non-
union, single & union.. etc:.

A download fabplot, created by Nick Cox, shows graphs in which each panel contains all the
points but the particular subsets are highlighted. For example:

fabplot scatter wage tenure , by(married union)

19
So in the top left graph, the blue dots show you where the single non-union members are in
the whole distribution. This may facilitate uncovering where in the joint distribution a
particular subset is concentrated.

To generate a matrix of scatterplots for several variables:

webuse auto, clear

graph matrix price mpg weight length, half

Price

40

30
Mileage
(mpg)
20

10
5,000

4,000
Weight
3,000 (lbs.)
2,000
250

200 Length
(in.)
150

5,000 10,000 15,000

10 20 30 402,000 3,000 4,000 5,000

Omitting the “half” option means that the upper triangle (symmetric to the lower one) is also
shown.

20
2.3 Time plots
If your data is time series it is best to use the dedicated line plot for time series command.

webuse klein, clear

tsset yr This tells Stata that this variable is the time variable
tsline consump invest govt This will graph the three variables over time:
80
60
40
20
0

1920 1925 1930 1935 1940

year

consumption investment
government spending

Use twoway (tsline consump, recast(scatter)) If you do not wish the points to be
connected.

To plot the correlogram i.e. autocorrelations between yt and yt-1, yt and yt-2 etc:
ac consump, lags(8)

Sometimes you have two variables and you want to illustrate the range between them over
time. For example they could be the upper and lower bounds for a given outcome, like a daily
price high and low.

sysuse tsline2
twoway rarea ucalories lcalories day

21
For a slightly different look replace rarea with rcap or rbar.

twoway area calories day shades the area under the curve

Some time series fluctuate have high variance and it may be preferable to create a smoother
version of the data that removes much of the short run volatility. The simplest way of doing
this is a moving average. To graph a moving average in Stata, create it using tssmooth ma
first.

tssmooth ma caloriessm = calories , window(3 1 2)

22
This generates a 6 year moving average with three lags, the current value and two leads. You
can apply weights to the different values if you wish. To show the original and the moving
average series:

tsline cal*

The widely used Hodrick-Prescott filter for macroeconomic data is available with the tsfilter
hp command.

Sometimes you might wish to plot two or more variables which have different dimensions
for example GNP and the unemployment rate. In that case you can use a separate y axis for
each of the two. Using the klein dataset:

twoway (scatter consump yr, c(l) yaxis(1)) (scatter taxnetx yr, c(l) yaxis(2))

.
23
To get a taste for some of the many options you can use in crafting an image, consider this:

twoway (scatter consump yr, c(l) msymbol(Dh) mcolor(blue) msize(large)

clwidth(thick) clcolor(maroon) )

2.4 Panel data

If you are using panel/longitudinal data it is possible to create line plots of a series for the
individuals:
sysuse xtline1
xtset person day
xtline calories, overlay

Dropping the overlay option generates a separate graph for each individual. If “N” is large
i.e. there are many individuals, then this form of graph may not be very useful.
24
 3. Graphs after estimation
After you have estimated your models there are several reasons to use graphs. One is that,
post-estimation, it may be wise to examine various characteristics of the residuals. A second
is that sometimes a plot of regression coefficients or marginal effects is an easier way of
showing the results.

sysuse nlsw88, clear

reg wage age married collgrad union [Link]
predict reshat , residuals
kdensity reshat , norm

This plots the residuals from the model and superimposes a normal distribution. You can see
the residuals don’t look normal. It is well known that the distribution of earnings is close to
being log-normal i.e. the log is normally distributed. If you change the dependent variable to
the log of wages & re-estimate the model you will find the residuals are remarkably close to a
normal distribution (2nd graph below)

Kernel density estimate

.15
.1
Density

.05
0

-10 0 10 20 30
Residuals

Kernel density estimate

Normal density
kernel = epanechnikov, bandwidth = 0.6171

25
 Kernel density estimate
log wage model

.8
.6
Density

.4
.2
0

-2 -1 0 1 2
Residuals

Kernel density estimate

Normal density
kernel = epanechnikov, bandwidth = 0.0932

If your data is time-series then you should be interested in whether the residuals are
autocorrelated. A graphical way of doing this is examine the correlogram which plots the
degree of autocorrelation for different lags i.e. the autocorrelation between the residuals in
period t and t-1, t and t-2 etc.

webuse klein, clear

tsset yr
reg consump totinc
predict reshat , residuals
ac reshat, lags(9) recast(connect)

So the autocorrelation between residuals in periods t and t-1 is about .65. The lags fade away
as we would expect so there is little correlation between residuals in period t and t-5, say.
The corrgram command displays a table of the autocorrelations and has a crude graph of
them.

26
To plot regression coefficients there is a user written command coefplot written by Ben Jann
(2013). By default it shows 95% confidence intervals but you can change that.

sysuse auto
regress price mpg headroom trunk length turn
coefplot, drop(_cons) xline(0)

Mileage (mpg)

Headroom (in.)

Trunk space (cu. ft.)

Length (in.)

Turn Circle (ft.)

-1500 -1000 -500 0 500

In the example above the coefficient on each variable is the marginal effect of that variable.
That is because the price variable is assumed to be a linear function of the x variables. If any
of the x variables enter non-linearly then the marginal effect of the variable will be different
at different values. Say for example if price depends on mpg (miles per gallon) and its square:

price = 0 + 1MPG +  2 MPG 2 + ...

price
 = 1 + 2 2 MPG
MPG

Stata’s margins command comes can be used to evaluate the marginal effect at different
values of mpg. First run the model, say: regress price [Link]##[Link] headroom trunk
length turn
margins, dydx(mpg) will give you the average marginal effect with a standard error but
doesn’t tell you how it varies with mpg. Note that if you had created the square of mpg as a
separate variable (say “mpgsq”) & then included in the model just like another variable then
margins would not provide the correct marginal effect of mpg as Stata, in using margins,
would not know what mpgsq is. That’s why it is necessary to use the [Link]##[Link]

27
formulation. Note that [Link]#[Link] will generate a quadratic function but without the
linear component i.e. the square only.

margins, dydx(mpg) at (mpg=(12(4)41)) evaluates the marginal effect at different values

of mpg starting at 12 and increasing by increments of 4 to 41. The results are presented in a
table but it is better at this stage to use the marginsplot command to get a nice graph of the
marginal effect of mpg as it varies with mpg. Since the model was quadratic in mpg, the
marginal effect is linear (see the 2nd equation above).

With coefplot and margins there are numerous options to customize the output. I have just
shown the basics. margins can also be used with interactions between variables. For
example

sysuse nhanes2
reg bpsystol agegrp##sex
margins agegrp#sex

<output omitted>
This regresses a continuous variable on a categorical age variable interacted with a dummy
variable and then calculates the marginal effects of the interactions. Typing marginsplot
produces a graph with the predicted value of the outcome for each category by sex with 95%
confidence intervals. Older people have higher (systolic) blood pressure and the age gradient
is steeper for females.

28
 4. Schemes

While you can tweak the look of graphs in many ways, one approach is to use different styles
of graphs (called schemes) that Stata has created. Taking the scatterplot we had on page 16
above, you could try:

scatter mpg weight if foreign || scatter mpg weight if !foreign , scheme(vg_teal)

Other schemes include s2color, s1mono, s2mono. Using scheme(economist) replicates the
look of the graphs in The Economist magazine. To see the different schemes available, when
you open the dialogue box for graphs, open the tab for “overall”. The “scheme” option is at
the top left. Alternatively, entering graph query, schemes in the Stata command line will
provide a full list.
29
5. The definitive pie chart:

ooOoo

30
 6. References & resources

Hintze, Jerry & Ray Nelson (1998). Violin Plots: A Box Plot-Density Trace Synergism. The
American Statistician 52(2):181-84.

Jann, Ben (2013). coefplot: Stata module to plot regression coefficients and other
results. [Link]

Jones, Andrew (2017) Data visualization and health econometrics

[Link]

Koffman, Dawn (2015) Introduction to Stata 14 graphics

[Link]

Mitchell, Michael N (2012) A visual guide to Stata graphics 3rd ed. Stata Press

Schwabish, Jonathan (2014) An economist’s guide to visualizing data. Journal of Economic

Perspectives, 28(1) 209-234.
[Link]

Shapiro, Jesse (nd) How to give an applied micro talk

[Link]

Tufte, Edward (2001) The visual display of quantitative information 2nd ed. Graphics
Press.

Tukey, John (1977) Exploratory Data Analysis, Addison-Wesley.

Van Kerm, Philippe (2012). Kernel-smoothed cumulative distribution function

estimation with akdensity. Stata Journal 12: 543-548.

Introduction to Stata Graphics: [Link]

Introduction to graphs in Stata

[Link]

Stata graphics tutorial [Link]

31