UCD GEARY INSTITUTE

DISCUSSION PAPER SERIES

A Practical Introduction to Stata

Mark E. McGovern

University College Dublin

Geary WP2012/01
January 2012

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.
 A Practical Introduction to Stata
Mark E. McGovern∗
January 2012

Abstract
This document provides an introduction to the use of Stata. It is designed to be an overview rather than
a comprehensive guide, aimed at covering the basic tools necessary for econometric analysis. Topics cov-
ered include data management, graphing, regression analysis, binary outcomes, ordered and multinomial
regression, time series and panel data. Stata commands are shown in the context of practical examples.

Contents
1 Introduction 4
1.1 Opening Stata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Audit Trails . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Getting Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Importing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.6 User Written Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.7 Menus and Command Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.8 Data browser and editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.9 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.10 Types of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Data Manipulation 9
2.1 Describing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Generating Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 if Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Summarising with tab and tabstat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Introduction to Labels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Joining Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.7 Tabout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.7.1 Tabout with Stata 9/10/11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.7.2 Tabout with Stata 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.8 Recoding and Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.9 Missing Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.10 Macros, Looping and Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.11 Counting, sorting and ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.12 Reshaping Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.13 Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
∗ Geary Institute & School of Economics, University College Dublin. PGDA Fellow, Harvard Center for Population and

Development Studies. I gratefully acknowledge funding from the Health Research Board. This document is based on notes
for the UCD MA econometrics module and a two day course in the UCD School of Politics. Preliminary, comments welcome.
Email: [email protected]

1
3 Regression Analysis 21
3.1 Dummy Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Outreg2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Post Regression Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.5 Interaction Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.6 Specification and Misspecification Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Binary Regression 26
4.1 The Problem With OLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Logit and Probit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3 Marginal Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5 Time Series 32
5.1 Initial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Testing For Unit Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.3 Dealing With Non-Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

6 Ordinal and Multinomial Regression 36

6.1 Ordinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.2 Multinomial Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

7 Panel Data 44
7.1 Panel Set Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.2 Panel Data Is Special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.3 Random and Fixed Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.4 The Hausman Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
7.5 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

8 Instrumental Variables 53
8.1 Endogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
8.2 Two Stage Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.3 Weak Instruments, Endogeneity and Overidentification . . . . . . . . . . . . . . . . . . . . . . 56

9 Recommended Reading and References 58

List of Tables
1 Logical Operators in Stata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Tabout Example 1 - Crosstabs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Tabout Example 2 - Variable Averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4 OLS Regression Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5 OLS Regression Output With Dummy Variables . . . . . . . . . . . . . . . . . . . . . . . . . 23
6 Outreg Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
7 Linear Probability Model Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
8 Logit and Probit Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
9 Marginal Effects Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
10 Alternative Binary Estimators for HighWage . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
11 Dickey Fuller Test Ouput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
12 A Comparison of Time Series Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
13 Ordered Probit Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
14 OLS and MFX for Ordinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2
 15 Multinomial Logit Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
16 MFX for Multinomial Logit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
17 xtdescribe Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
18 xtsum Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
19 xttrans Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
20 Test For A Common Intercept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
21 Random Effects Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
22 Fixed Effects Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
23 Comparison of Panel Data Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
24 Correlation Between Income, Openness and Area . . . . . . . . . . . . . . . . . . . . . . . . . 55
25 OLS and IV Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
26 Testing for Weak Instruments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

List of Figures
1 An example of a graph matrix chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Graph Example 1: Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Graph Example 2: Labelled Scatterplot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4 A Problem With OLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5 Problem Solved With Probit and Logit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6 Autocorrelation Functions For Infant Mortality and GDP . . . . . . . . . . . . . . . . . . . . 32
7 Partial Autocorrelation Functions For Infant Mortality and GDP . . . . . . . . . . . . . . . . 33
8 Using OLS To Detrend Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
9 Health Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
10 Height Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
11 Ordered Logit Predicted Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
12 Multinomial Logit Predicted Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
13 BHPS Income . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
14 BHPS Income by Job Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
15 Graph Matrix for Openess, Area and Income Per Capita . . . . . . . . . . . . . . . . . . . . . 53
16 Openness and Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

Objective
The aim of this document is to provide an introduction to Stata, and to describe the requirements necessary
to undertake the basics of data management and analysis. This document is designed to complement rather
than substitute for a comprehensive set of econometric notes, no advice on theory is intended. Although
originally intended to accompany an econometrics course in UCD, the following may be of interest to anyone
getting started with Stata. Topics covered fall under the following areas: data management, graphing,
regression analysis, binary regression, ordered and multinomial regression, time series and panel data. Stata
commands are shown in red. It is assumed the reader is using version 11, although this is generally not
necessary to follow the commands.

3
1 Introduction
1.1 Opening Stata
Stata 11 is available on UCD computers by clicking on the “Networked Applications”. Select the “Mathe-
matics and Statistics” folder and Stata v11. It is also possible to run Stata from your own computer. Log into
UCD connect and click “Software for U” on the main page. You will first need to download and install the
client software, then you will be able to access Stata 11, again in the “Mathematics and Statistics” folder. For
further details see http://www.ucd.ie/itservices/teachinglearningit/applications/softwareforu/d.en.21241
Stata 11 is recommended, however Stata 8.0 may also be available on the NAL (Novell Application
Launcher). Click Start and open the NAL. Open the Specialist Applications folder and click into Economics.
Open wsestata.exe, or right-click and add as a shortcut to your desktop. Alternatively, click Start > Run,
paste in Y:\nalapps\W95\STATASE\v8.0 and click enter.

1.2 Preliminaries
Before starting, we need to cover a very important principle of data analysis. It is vital that you keep track
of any changes you make to data. There is nothing worse than not knowing how you arrived at a particular
result, or accidentally making a silly mistake and then saving your data. This can lead to completely incorrect
conclusions. For example you might confuse your values for male and female and conclude that men are
more at risk of certain outcomes, etc. These mistakes are embarrassing at best, and career threatening at
worst. There are three simple tips to avoid these problems. Firstly keep a log of everything. Secondly, to
ensure you don’t embed any mistakes you’ve made in future work, most econometricians never save their
datasets. Generally people initially react badly to this suggestion. However you don’t need to saves changes
to the dataset itself if you implement all manipulations using do files. The final tip therefore, is to use do
files. We will cover each of these in what follows.
The first thing we need to do is open our data. If we have a file saved somewhere on our hard disk
we could use the menus to load it. FILE, OPEN. Or we could write out the full path for the file, e.g.
“h:\Desktop\”. The path for your desktop will differ depending on the computer your are using, however,
if you are on a UCD machine this should be it. This is awkward, and we will also need somewhere to store
results, and analysis. So we will create a new folder on our desktop called “Stata”. Right click on your
desktop, and select NEW, FOLDER. Rename this to “Stata”. We will also create a new folder within this
called “Ado” which we will use to install new commands. Save the files for this class into the “Stata” folder.
Stata starts with a default working directory, but it is well hidden and not very convenient, so we want to
change the working directory to our new folder. First we check the current working directory with pwd. Now
we can change it cd ‘‘h:\Desktop\Stata’’. If you are unsure where your new “Stata” folder is, right click
on it and go to PROPERTIES. You will see the path under LOCATION. Add “\Stata ” to this. Now we
can load our data files. One final piece of housekeeping, because we can only write to the personal drive
(“h:\”) on UCD computers we need to be able to install user written commands here. So we set this folder
with sysdir set PLUS ‘‘h:\Desktop\Stata\Ado’’. This is only necessary if you are running Stata from
a UCD computer.
Now we have this set up, accessing files saved in Stata format (.dta) is straightforward. use icecream2.
If you make changes to the data, you will not be allowed to open another dataset without clearing Stata’s
memory first. gen year=2010. We will encounter the gen command later. Now if we try and load the data
again use icecream2 we get the error message “no; data in memory would be lost”. We need to use the
command clear first, then we can reload the dataset use icecream2. Alternatively, using the clear option
automatically drops the dataset in current use use icecream2, clear. This raises a very important point,
we need to keep track of our analysis and our changes to the data. Never ever save changes to a dataset.
If you have no record of what you have done not only will you get lost and not be able to reproduce your
results, neither will anyone else. And you won’t be able to prove that you’re not just making things up.
This is where do files come in. A do file (not to be confused with an ado file)1 is simply a list of commands
1 This is a do file which contains a programme. Stata uses these to run most of its commands. This is also how we are able

4
that you wish to perform on your data. Instead of saving changes to the dataset, you will run the do file
on the original data. You can add new commands to the do file as you progress in your analysis. This way
you will always have a copy of the original data, you will always be able to reproduce your results exactly,
as will anyone else who has the do file. You will also only need to make the same mistake once. The top
journals require copies of both data and do files so that your analysis is available to all. It is not uncommon
for people to find mistakes in the analysis of published papers. We will look at simple example. Do files have
the suffix “.do”. You can execute a do file like this do intro. 2 do tutorial1 would run all of the analysis
for this particular tutorial. There are several ways to open, view and edit do files. The first is through Stata.
Using the menus go to WINDOW DO-FILE EDITOR, NEW DO-FILE. Or click on the notepad icon below
the menus. Or type doedit in the command window. Or press CTRL F8. Each of these will open the dofile
editor. Alternatively you can write do files in notepad or word. They must be saved as .do files however.
You don’t have to execute a whole do file, you can also copy and paste commands into the command window.
Here we will create our own do file using the commands in this document.
As well as using do-files to keep track of your analysis, it is important to keep a log (a record of all
commands and output) in case Stata or your computer crashes during a session. Therefore you should open
a log at the start of every session. log using newlog, replace. To examine the contents of a log using the
menus go to FILE, VIEW. Alternatively type view logexample. Also useful is set more off, which allows
Stata to give as much output as it wants. This setting is optional but otherwise Stata will give only one
page of output at a time. Finally, you must have enough memory to use your data. You can set the amount
of memory Stata uses. By default, it starts out with 10 megabytes which is not always enough. If you run
out of memory you will get the error message “no room to add more observations”. For most data files 30
megabytes will be enough, so we will start by setting this as the memory allocation. set mem 30m. To check
the current memory usage type memory. You could set memory to several hundred megabytes to ensure that
Stata will never run out, but this makes your computer slow (especially if you have a slow computer) and so
is not recommend. None of the files we will be examining require more than this. Note that if you run out of
memory you will have to clear your data, set the memory higher and re-run your analysis before proceeding.
In general all of these items are things you will want to place at the start of every do file.
clear
set mem 30m
cd "h:\Desktop\Stata"
sysdir set PLUS "h:\Desktop\Stata\Ado"
set more off
capture log close
local x=c(current_date)
log using "h:\Desktop\Stata\‘x’", append
Lines 7 and 8 require some explanation. The outcome of this is that Stata will record all analysis you
conduct on a particular day in a log file, the name of which will be that day’s date. We will explain how this
works when we discuss macros. Note that Stata ignores lines with begin with “*”, so we will use this to write
comments. The command “capture” is also important. If you are running a do file and it encouters and
error, the analysis will stop. The “capture” command tells Stata to proceed even if it encounters a mistake.
If you are running Stata on your own computer, there is a way to alter the default settings that Stata
starts with. When it launches, Stata looks for a do file called “profile.do” and runs any commands it contains.
You can create this file so that these changes are made automatically every time you launch Stata. (i.e.
memory is set, directory is set and a log is started). As well as a working directory, Stata also has other
directories where programmes are stored. We need to put our “profile.do” into the “personal” folder. To
find it, type sysdir. We now paste the following into a text file (either using notepad or Stata), and save it
as “profile.do” into that directory.
to install new user written commands. Usually we will be able to install these automatically, however sometimes we need to do
this manually. All that is involved here is saving the appropriate ado file into the appropriate directory which you can locate
with sysdir.
2 run intro executes the do file but suppresses any output.

5
1.3 Audit Trails
1. Remember to keep a record of everything
2. Never alter the original dataset

(a) Place the original dataset in a separate folder

(b) Make a backup of the dataset
(c) Use .dta files for Stata use
3. Before completion, do a test run on a backup

1.4 Getting Help

Stata has an inbuilt help function. In fact you can easily access the help file for any command. Suppose
we are interested in the “tabstat” command, we can type help tabstat. However these are often aimed at
experienced users and you may have difficulty understanding them. The syntax for this command is given as
“tabstat varlist [if] [in] [weight] [, options]”. Items in square brackets are optional. So all we require to run this
command is the command itself followed by at least one variable. The various options are explained, followed
by some examples. Looking at the examples is often the best way of getting to grips with how a command
works. If you cannot solve the problem using the help file, Stata has an extensive online support system. It is
more than likely that someone else has encountered the same problem. A google search usually throws up sev-
eral items of interest. There are also several excellent websites which detail how to deal with various aspects
of data analysis. The best of these are the UCLA (http://www.ats.ucla.edu/stat/stata/sk/default.htm) and
Princeton
(http://dss.princeton.edu/online help/stats packages/stata/stata.htm) Stata websites. We are now ready
to begin analysing our data.

1.5 Importing Data

We have already seen that importing data in Stata format (.dta files) is simple. The command is “use
filename”. We need to load the icecream dataset for this analysis. use icecream.

Note: We’re opening a dataset. Remember, do not save any

changes you make.

Often you will have to make do with Microsoft Excel (.xls) files. Fortunately Stata can import these quite
easily. If you have an Excel file named myfile.xls, you can import it using Stata’s insheet command.
First, in Microsoft Excel, click File > Save As. Now instead of saving as a Microsoft Office Excel file, save
the file as a CSV (Comma Delimited) file using the dropdown menu. You can then load the data using the
command insheet using myfile.csv. The “infile” command is an alternative for loading other types of
data. If a direct import with this command fails, try opening it in excel and following the instructions above.
We will discuss how to import SPSS files when discussing an example of a user written command.

1.6 User Written Commands

One of the major advantages of Stata is the manner in which users can write their own commands. In the
unlikely event that you are trying to do something that doesn’t have an official command, you are practically
guaranteed that someone else has had the same problem, and can be reasonably confident that someone has
written their own code to deal with the issue. Finding the answer to your particular problem is not always
straightforward, but it can be as easy as a google search. Knowing exactly what to look for can be the main
problem. If you can find the name of your programme it can be relatively straightforward to install. Later
we will encounter two other user written commands (“tabout” and “catplot”), now we will consider the

6
programme “usespss” which is used to import data saved in SPSS format.3 . If we try to import the SPSS
file in our directory, use spssfile.sav we get the error message “file SPSSfile.sav not Stata format”. You
would imagine that it should be relatively easy to transfer files between different statistics packages but this
is not the case. Without this command you might need to use the expensive StatTransfer programme. As we
know what we’re looking for, the process of installing the programme is easy. We use the “findit” command.
findit usespss. In fact if you’re sure of the name you can simply type ssc install usespss. If you are
not exactly sure of the name, but have a general idea of what it’s called you can use the command search
usespss, all. A new window will appear, and clicking on the blue link will take you to a new page. Click
on CLICK HERE to install. This programme is now ready to run. You can also access the help file, help
usespss. We can now open the SPSS file in our folder. usespss using spssfile.sav. We could now save
this in Stata format for future use. save, replace. It is saved as “SPSSfile.dta”. However as you can see,
this is just the same as our icecream dataset.4 We will re-load the Stata version. use icecream2, clear.

1.7 Menus and Command Window

Across the top of the Stata programme are a number of menus, and you may be tempted to use these to
carry out your analysis. For example, going to ‘FILE’, ‘OPEN’, and selecting your file will open the dataset.
As we will explain later on, this is not ideal. Not only is it slow, but it makes reproducing your results very
tricky. The alternative is to type commands directly into the command window, which requires becoming
familiar with the language of Stata, but is ultimately much more efficient and reliable. One benefit of using
the menus is that when you run a command this way, it appears in the main window, which is where results
of analysis are displayed. So this is a useful way to learn commands if you are unsure of them. Copying this
from the main window into the command window and pressing enter will reproduce the command. This is
in fact how to keep track of your analysis, by recording every change you make to the data. The Review
window provides you with a list of all entered commands. Clicking on a command in the review window will
cause it to appear in the command window. Failed commands are shown in red. The variable window gives
a list of all the variables in the dataset, their labels, and their attributes.

1.8 Data browser and editor

Stata holds the data in memory like an excel file. To see the actual data select ‘DATA’ then ‘DATA
BROWSER’ from the menus. You will see the variables ordered horizontally with the observations ordered
vertically. Selecting ‘DATA EDITOR’ instead provides you with the same view, except you are now able
to edit the data, like you would a spreadsheet. In fact this is an alternative way of entering data, you can
simply paste it in. As before this is not recommended for reasons of reproducibility. Clicking on a particular
case gives you the exact entry for that particular variable and observation. The data editor and browser
windows must be closed before you can enter any new commands into Stata.

1.9 Syntax
Getting to grips with how to communicate with Stata is perhaps the most daunting aspect of starting out.
Generally programmes and commands take the form of “command name” “variable name(s)” “, options.”
We will shortly see examples with tab and later regress. The exact syntax for a particular command is
detailed in the help file. For example, help tab. Here the aim is to introduce you to some of the most
important commands. As you become more familiar with them you will be able to use the various options
available, depending on the particular task you wish to perform.
Stata understands abbreviations, once the abbreviation can only be interpreted one way. For example,
the full command to run a regression is “regress”, however Stata understands what you mean if you only type
“reg”. The same principle applies to variable names. Using the icecream dataset, typing tab ti is equivalent
3 SPSS is a statistics package popular in the other social sciences
4 An alternative for transferring SPSS files into Stata is to download SPSS which is available from UCD Connect, open the
SPSS file and save in Stata format.

7
 Table 1: Logical Operators in Stata
And &
Or |
Not ! or ˜
Multiplication *
Division \
Addition +
Subtraction -
Less Than <
Greater Than >
Less Than or Equal =<
More Than or Equal =>
To The Power Of ˆ
Wildcard *

to tab time. However, the command tab t will return the error message “t ambiguous abbreviation.” In
the help file for a command, the shortest acceptable abbreviated version is underlined.

1.10 Types of Variables

There are essentially two types of variables in Stata, words (referred to as strings) and numbers. Each is
handled slightly differently when you are manipulating your data. Within numerical variables, there are two
further types; continuous data, such as income or height, and categorical data (such as level of education
or gender). In the second case a value will take on a particular meaning, e.g. 1=male and 2=female. As
we will see these variables are often labelled in the data. In the Icecream dataset, the first 5 are obviously
continuous numerical variables (these appear in black in the data browser), however county, var7 and weekend
are different. County is a string variable, with the entries appearing as words in the dataset. Strings such
as this appear in red. If you click on the county variable for observation 1, you will see “Antrim” appear as
the entry for that case. On the other hand, for var7 the entry for observation 1 appears as “Ulster” in the
data browser, but “1” when you click on it. This means the data is in numerical form (in this case 1-4), but
has been labelled so that each number refers to a different province. We will discuss manipulating each of
these types of variables in the next section.

8
2 Data Manipulation
Here we introduce the basic commands for manipulating your data. The most important logical operators in
Stata are outlined in table 1. The most frequently used are & (and), | (or) and ! (not). These are essential
for manipulating the data correctly. We can illustrate some of these using the “display” command, which we
can shorten to “di”: di 10*20 and di 6/(5-2) +18. Notice that strings require double quotes di Welcome
does not work but di ‘‘Welcome’’ does. You can also access system values and programme results with this
command, for example today’s date di ‘‘‘c(current date)’’’. Note again that di ‘c(current date)’
returns an error message. We need single quotes because “c(current date)” is a macro. This is how we were
able to name our log file, see section 2.10 for more details on macros. We will explain the use of the wildcard
in section 2.9. Stata will know whether you mean multiplication or the wildcard depending on the situation.

2.1 Describing Data

Stata provides several ways of investigating and describing data. It is generally a good idea to browse the
data once you have it loaded. This allows you to view the data in spreadsheet format. For a broader look
at the variables in the dataset, use the sum command, which summarises all the variables. Alternatively,
specifying a variable after the command (e.g. sum price) will summarize the price variable on its own. If
you would like more precise information (e.g. percentiles) then you can add the detail option to the end of
that command, i.e. summ price, detail. A very useful command is inspect. This summarizes a variable’s
missing values (if any) and provides a simple plot of the variable’s distribution.
Sometimes you may wish to break down summaries by a particular variable. For example, you might
like to see how some variable changes over time. The bysort command is very useful here. It is best
explained with an example. Suppose you want to see how consumption changes as temperature changes.
The command here would be: bysort temp: summ cons. This means, in English, “summarize consumption
for every distinct value of temperature.” Other variables that you may use again include describe and list.
To investigate basic correlations, use the corr command followed by the variables you want to correlate.
For example, corr price temp cons will provide a matrix of Pearson correlation coefficients between price,
temperature and consumption. You may want to use the pwcorr command instead. It is essentially the
same as corr but it allows for a little bit more detail. If you type pwcorr temp cons, sig it will provide
p-values of the test with the null of the correlation being zero.

2.2 Generating Variables

You will regularly have to generate variables to aid econometric analysis. For example, you may want to
create dummy variables or run log-log regressions. To create a variable named loggedprice equal to the
natural log of price, the command is gen loggedprice = ln(price). Similarly, to generate a variable equal
to twice the square root of temperature, use the command gen twice root temp = 2 ∗sqrt(temp). Note
that variable names cannot contain spaces.
The egen (“extended gen”) command works just like gen but with extra options. For example, egen
avg price = mean(price). With egen we can also break commands down into percentiles very eas-
ily. For example, to create a variable equal to the 99th percentile of price, enter egen high price =
pctile(price), p(99). Changing the 99 to 50 in that command would produce a variable equal to the
median price. “egen” is often used to obtain a breakdown of a particular statistic by another variable. For
example, we could obtain a variable containing the minimum income value for that particular province egen
minincome=min(income), by(var7).

2.3 if Commands
Oftentimes we want Stata to run a command conditional on some requirement. For example, the correlation
between price and consumption if the temperature is greater than 25◦ C. This is easily achieved: corr
price cons if temp>25. To add more conditions to a command, for example to examine the correlation

9
if temperature is both greater than 25◦ C and less than 35◦ C, we use the & operator: corr price cons if
temp>25 & temp<35.
We can also use if to investigate data more closely: summ cons if price >.275. We can create
dummy variables with if commands. Typically two steps are needed. First we create a variable set equal
to zeros: gen expensive = 0. Now we replace it: replace expensive = 1 if price >avg price. See
the tutorial on regression analysis for more details on dummy variables. Similarly we can control for outliers
using if commands. For example if you want to eliminate the most expensive 5% of observations, the
following would work:
egen top_fivepercent_prices = pctile(price), p(95)
drop if price > top_fivepercent_prices

We remove these variables from the data with the “drop” command5 as we do not need them in this
analysis. drop loggedprice twice root temp avg price high price expensive.

2.4 Summarising with tab and tabstat

One of the first things you will want to do with your data is to summarise its main features. Crosstabs are a
useful pre-regression tool, and are also useful for presenting the main points of your data succinctly. The two
most important commands for this are “tab” and “tabstat”. “Tab” tells you how many times each answer is
given in response to a particular variable. This is only suitable for variables with relatively few entries, such
as categorical data. If you try to use “tab” on a variable which has hundreds of different entries you will get
an error message. Typing tab var7 will show the how many entries there are for each province. As with all
commands, it can also be accessed through the menus via: STATISTICS, SUMMARIES, TABLES. It is also
easy to obtain crosstabs which give a breakdown of a variable by another. For example typing tab var7
weekend will show how many weekend and weekday entries there are for each province. It is often useful
to know the percentages as well as the actual numbers. We need to add an option to our “tab” command.
tab var7 weekend, col will give us the percentage with each column. Typing tab var7 weekend, row
will give us the percentage within each row.
The second command which is useful here is “tabstat”. This is used for continuous variables, and its
main use is to provide mean values. For example tabstat price will give the average price in the dataset.
Using the command options we can also access other useful statistics such as the median tabstat price,
stats(med), or the variance tabstat price, stats(var). For a full list of the available statistics, type
help tabstat. As before we can obtain these statistics according to different levels of a second variable.
tabstat price, by(var7) gives the average price for each province.

2.5 Introduction to Labels

Labels are designed to help the user of a dataset understand and present their findings. These are often
essential, for example if you have a categorical variable gender with two values 1 and 2, and no label you are
in trouble as you will not know which refers to male and which to female. Generally these will be provided
in your data, but not necessarily. Often the variable name itself will be self explanatory, for example in the
icecream dataset the meaning of the variable “income” is obvious, although we do not know the unit it is
measured in. It is not so obvious for var7, but a look at descriptives or the data browser makes it clear that
it refers to Irish provinces. It is easy to rename a variable. Type rename var7 province into the command
window.
As well as their actual names, you can also label a variable. These can be used to provide additional
information that is not apparent from the variable name. So far none of the variables in the dataset have these.
Adding a variable label to a variable is also straightforward. Type label variable temp ‘‘Temperature
Degrees C’’ and label variable cons ‘‘The number of ice-creams purchased.’’. You can see the
5 We can also select the variables we wish to remain in the data, with the “keep” command.

10
result of this in the variable window where this label will appear. The label will also be used in tables and
other output generated by Stata.
We have already encountered the other type of labels, value labels. These are labels attached to particular
values of a variable and are mainly used with categorical data. In the case of the variable province, we have
already seen how this works. For a quick way of seeing exactly which labels are attached to each value, type
codebook province to obtain the name of the value label. Then lablebook province1 will show all the
values and their labels. From this we can see that in the variable province, the value “1” is labelled with
the name “Ulster”, 2 “Leinster” etc. If there are many value labels you may need to use the option all.
codebook province, all.
You will often need to either create your own value labels, or else modify existing ones. The procedure
is almost the same in both cases. If you are starting from scratch, you will need to pick a name for your
label. In this case we will create value labels for the weekend variable, and call the label weekend1. Suppose
we know that “1” refers to “weekend” and “2” refers to “weekday”. We use the “label define” command,
followed by the values and their labels in double quotation marks. label define weekend1 1’’Weekend’’
2’’Weekday’’. We then need to attach our value label to the existing variable using label values weekend
weekend1. We tab weekend to confirm the change. The only difference in the case of modifying an existing
set of values labels is that you need to obtain the name of the value label (“codebook” will supply you with
this). Then you use the “label define” command with the “modify” option to change one or more of the
labels. You do not need to reattach the modified value label to the variable, this is done automatically. You
also do not need to write out the full set of labels, only the one you want to change. For example, if we
wanted to change the label on the province variable for the value “2” from “Leinster” to “Dublin”, we would
use label define prov 2’’Dublin’’, modify.

2.6 Joining Datasets

There are two different situations when you will need to join datasets, each requires a different command.
Broadly these involve adding more observations, or adding more variables. We will demonstrate each using
our icecream dataset. For the first situation we will add two extra observations. In the second case we will
add an extra variable (rainfall).
The first is easier. Suppose you want to join two rounds of a survey into a single dataset. This often
happens with the likes of the European Social Survey, the QNHS or other surveys which involve repeated
cross sections. If you think about what you would do in excel, all you need to do in this case is stack the
two datasets on top of each other. But do not do it in excel because there will be no record of what you
have done or how you have done it. We will use the “append” command to add our new observations which
are in the dataset icecream3. It is as straightforward as append using icecream3. Now we have a dataset
with extra observations. We confirm this with tab county and the data browser.
Adding new variables is slightly more tricky. We can’t just add the information at random, we need
to make sure that each observation is matched in each dataset. This often arises in the context of surveys
such as SHARE which have several modules (and datafiles) dealing with different domains such as health,
income, demographics etc. If we want to join modules we have to make sure that the extra variables are really
the responses which that particular individual gave. So you need a variable in each dataset that uniquely
identifies each individual (or firm, country etc. whatever your level of observation is).6 In this case we will
assume we are interested in time periods, and add an extra variable (rainfall) for each time period. In this
case our unique identifier is time, and we will need this variable to be the same in the new dataset. The first
thing we need to do is make sure the data is ordered correctly, so we need to sort by our identifier. sort
time. The new dataset will also need to be sorted by this variable. Now we can use the merge command
with the dataset containing the new variable icecream4. merge time using icecream4. We can now check
that we have an additional rainfall variable for each time period. tabstat rainfall, by(time). And in
the data browser. A new variable, “ merge” is created, which tells us whether the variables were present in
6 In certain situations you may need to merge across more than one variable, for example merge mergeid year if you have a

panel dataset.

11
the master dataset, the dataset we merged with , or both. See help merge for details. For no we will drop
this variable with drop merge.

2.7 Tabout
We have already discussed tabulating variables and looking at crosstabs. Here we will examine how to
extract these results in a way that can be easily used in presentations or papers. To do this we need the user
written command “tabout”. To find and install, we first search for it. search tabout, all. The option all
allows us to search the internet. We find an entry in blue for tabout, with the description:
‘TABOUT’: module to export publication quality cross-tabulations / tabout is a table building program
for oneway and twoway / tables of frequencies and percentages, and for summary tables. It / produces
publication quality tables for export to a text file.

2.7.1 Tabout with Stata 9/10/11

You can install tabout from this link, or else by typing ssc install tabout.
As before, there are essentially two cases involved, one analogous to “tab”, and the other analogous
to “tabstat”. tabout county using filename.xls, replace gives us an excel table with the numbers
in each county. The replace option is important (and an option which is available in most programmes
which involve exporting results) and tells Stata to overwrite the file if it already exists. The alternative is
append, which tells Stata to add new results to the same file. If we want percentages as well as numbers in
each category we can use the following option: tabout county using filename.xls, append cells(freq
co). We may also want to display crosstabs which we can do by adding a second variable, for example tabout
county province using filename.xls, append. If we want percentages as well as numbers we can use
the following: tabout county province using filename.xls, append cells(freq co). “Tabout” only
works with variables which have labels, as these kind of tables do not really make sense for continuous
variables such as income. However, we can make tables of means (or other statistics) like with “tabstat”.
tabout province using filename.xls, append cells(mean income) sum gives us the average income
for each province.

2.7.2 Tabout with Stata 8

Clicking on the link after running search tabout, all will open a page which allows you to install the pro-
gramme. You will notice that Stata version 9 is required. We need a version which is compatible with
the verision of Stata we are running. Here we need “tabout8”, but a search is unsuccessful. search
tabout8, all. We will need to install it ourselves. If we google “tabout8” we will find the website
“http://www.ianwatson.com.au/stata.html” as the first entry. Here we find a link to the .ado file, which we
download and save in our personal directory which we can locate with the command sysdir. We also save
the corresponding help file in this directory. This may seem like a lot of effort, however this only needs to
be done once, and will save you a lot of time if you are handling datasets with large numbers of variables.
It is possible to copy and paste from the Stata window into excel after using the “tab command”, but this
is a painstaking process, especially if you are constantly updating your analysis. “Tabout” automates this
process.
If we want percentages as well as numbers in each category we can use the following option: tabout8
county using filename.xls, append cells(fcount fper). We may also want to display crosstabs which
we can do by adding a second variable, for example tabout8 county province using filename.xls,
append. If we want percentages as well as numbers we can use the following: tabout8 county province
using filename.xls, append cells(double). “Tabout8” only works with variables which have labels,
as these kind of tables do not really make sense for continuous variables such as income. However, we can
make tables of means (or other statistics) like with “tabstat”. tabout8 province using filename.xls,
append cells(mean income) gives us the average income for each province.

12
 Table 2: Tabout Example 1 - Crosstabs
Province
Ulster Dublin Munster Connacht Total
Weekend Num % Num % Num % Num % Num %
Weekend 5 55.6 7 58.3 3 60.0 3 50.0 18 56.3
Weekday 4 44.4 5 41.7 2 40.0 3 50.0 14 43.8
Total 9 100.0 12 100.0 5 100.0 6 100.0 32 100.0

Table 3: Tabout Example 2 - Variable Averages

Mean Income
By Weekend
Weekend 84
Weekday 85
Total 85
By Province
Ulster 83
Dublin 85
Munster 85
Connacht 86
Total 85

“Tabout” is at its most powerful when used in conjunction with LATEX, which is the software which was
used to create this document. Note that the more recent version has a slightly different syntax, but the
general idea is the same. You can also get confidence intervals using the latest version.

2.8 Recoding and Strings

Sometimes we may need to recode particular values in order to carry out our analysis. For example we have
information on all four provinces, but we may only be interested in comparing Ulster to the other provinces.
We will generate a new variable province2, which at first is exactly the same as our province variable. gen
province2=province. Now we will recode the Ulster value to be equal to one, and the other provinces
to be equal to zero. In case we’ve forgotten, codebook province and then lablebook province1 will tell
us the value labels. Ulster is already coded as 1, so we leave that as it is. There’s more than one way
to change the value for the other countries. We could use the replace command. Either of the following
would do the job: replace province2=0 if province>1 or replace province2=0 if province2!=1 or
replace province2=0 if province2==2 | province2==3 | province2==4. The recode command often
accomplishes the same task with less effort. recode province2 (2/4=0). The “/” tells Stata to recode
values 2 through 4 to be equal to 0.7 “Recode” is also important if you want to change values within a
variable simultaneously. For example suppose we want to reverse the coding on the variable “province”, we
could do so with recode province (1=4)(2=3)(3=2)(4=1). This would not be possible using the “replace”
command. If you do this, don’t forget to change the value labels. Now if we tab province2 we see only
two values as required. To be sure that we’re not confused when we come back to the data later, or
someone else is using it, we should label this new variable properly. We define a new label, lab def prov2
0’’Leinster Munster or Connacht’’ 1’’Ulster’’. Then attach this to the variable province2. label
values province2 prov2. Now tab province2 is self explanatory.
7 An even quicker command would be gen province2=(province==1).

13
 Several issues arise with trying to manipulate string variables. For example, if we try replace prov2=15
if county==Armagh, we get an error message. This is because in general Stata requires strings to be
surrounded with double quotations marks. replace province2=15 if county==‘‘Armagh’’ works. But
we should undo that change. replace province2=1 if county==‘‘Armagh’’. As we will see later on,
single quotation marks have other uses. In general, dealing with strings can be tricky, for example, we cannot
replace a string value with a numerical value. We also want to avoid having to type out a string every time
we want to manipulate the data. Expressions involving >and <clearly don’t work with strings. Fortunately,
there is a command which converts string variables into categorical variables with the appropriate value labels
(e.g. the province variable). Here we use this on the county variable. encode county, gen(county2). Now
we have a categorical numerical variable county2 which is appropriately labelled. We can check this in
the data browser. encode county, replace would have given us almost the same result, except with the
original county variable replaced rather than generating a new variable. There is a command, “decode”
which reverses the process.

2.9 Missing Values

Missing values are practically unavoidable, particularly in micro surveys. Individuals may not know how to
respond to a question, or may simply refuse. In well established surveys such as Share or Living in Ireland,
these will be coded as some value such as “99”, and be appropriately labelled. However this is not always
the case. Often there will just be a blank entry. This appears in Stata as “.”. Stata actually equates this
with infinity, so if you try something like replace var9=100 if var9>100 then your missing values will
be (unintentionally) included in your recode. So let’s reverse this replace var9=. if var9==100. For
most commands, observations which have missing values (for any of the variables which are involved in
running that command) are excluded. This is particularly important to remember when we come to look
at regressions. For example we know there are now 32 observations in our dataset. If we tab income we
find a total of 32 as expected. If we look at var9 in the data browser we notice it contains some missing
values. So if we tab var9, we will only see a total of 29. We notice that there are some entries labelled as
“Refusal”. Depending on the circumstances we may or may not want to exclude these from our analysis. To
do this we need to find the value for “Refusal” with codebook var9. We find it is “99”. To recode this as
missing we could do as above replace var9=. if var9==99 or recode var9 (99=.). An alternative is
the “mvdecode” command. mvdeocde var9, mv(99). An advantage of using this command is that we can
simultaneously recode all missing values for several variables at the same time, e.g. mvdecode time income
weekend, mv(99). But of course in this case the variables time and income do not contain any observation
with the value “99”. It is also possible to reverse this process with the command “mvencode”. You may
also want to recode existing non-missing values as missing, for example to deal with outliers. One way to do
this is replace income=. if income>100. Again, an alternative is “mvencode”. But here there are no
observations with income greater than 100. Note that dealing with missing values is a very important topic
in applied econometrics, and can have a major impact on your results. Earlier we mentioned the wildcard
“*”. Suppose we wanted to drop all variables with beginning with “var”, we could type drop var*. Or if we
wanted to recode missing values for all variables and the value “99”, we could type mvdecode *, mv(99).

2.10 Macros, Looping and Programming

An important thing to remember about Stata is that there is nearly always an easier and quicker way to do
things, especially if you find yourself having to repeat the same task over and over. This could be recoding,
generating variables etc. Stata has features designed to automate this processes, and as you become more
familiar with the programme you will literally be able to save yourself hours (by using loops, for example).
We will start with macros. These are simply shortcuts which stand in for something else, and can be used
to store everything from strings (words) to values, variable lists and results from programmes. We define a

14
macro with the “local” command, and access it using single quotation marks ‘’.89 First we define a macro
x which takes the numerical value 10. local x 10. Now every time we call the macro ‘x’ we have the value
10. To check this we type di ‘x’. We can now use this macro in expressions, for example di 100-‘x’.
We can also use it to manipulate variables. gen income2=income*‘x’. We can also store words in macros.
Suppose you wanted to add the word “icecream” to each variable name. You could type rename price
icecreamprice and rename cons icecreamscons etc. To save time you could store the word “icecream” in
a macro. local y icecream. Then rename price ‘y’price would give the same result. You may wonder
as to how useful this is, and in these cases it is probably not particularly helpful. A better example is when
we want to store a list of variables. Rather than typing out the whole list every time, we can save the
variables in a macro. local z price income temp. Then suppose we wanted to recode all missing values
in for all of these, instead of typing mvdecode price income temp, mv(100) we can type mvdecode ‘z’,
mv(100). This is a small dataset so it’s not a particularly big deal here, but it’s a different matter when you
have 100s of variables.
Macros are also important for accessing results stored by programmes.10 The macros saved by a pro-
gramme are listed in the help file. For example, if we look at help summarize, we see the list for this
command.11 Suppose we are interested in constructing new versions of our variables which are in the form
of a z score (standardised deviation from a variable’s mean: zi = (µ−x i)
sd(x) . We can see from the help file that
the command “summarize” stores the two results we need, the mean and standard deviation in the macros
“r(mean)” and “r(sd)” respectively.12 We can use these to form our new variables. To access the stored
results we need an “=” that we didn’t when we were defining our own macros. First we run the command
sum time. Then we define our macros local a=r(mean) and local b=r(sd). To check we have the correct
results di ‘a’ and di ‘b’. Now we can generate our new variable. gen ztime=(‘a’-time)/‘b’. If we now
sum ztime, we see that the mean is effectively zero (it’s actually a very small number due to rounding), as
it should be seeing as the average variation around a mean is zero by definition. The standard deviation is
also as expected. We will later write our own programme which will allow us to transform all our variables
in this way in a single line of code. Think how long it would take to do this in excel by hand.
Loops are another time saving device which employ the use of macros. For example in our icecream
dataset, we notice from the data browser that the variable hour has been badly inputted. These should all
be in the format of the 24 hour clock, however you can see that the final two zeros are missing from some
of the entries. This makes analysis difficult, for example sum hour will give misleading results. In order to
correct this we make use of a loop. This involves the forvalues command. Essentially we want to add two
zeros to every entry less than 100.

forvalues i=1(1)24 {
replace hour=hour*100 if hour==‘i’
}

Notice the syntax, we are creating a macro “i” which will start at the value 1, execute the command for
that value, move on to the next value (“2”), execute the command for that value etc until the loop ends.
The first number refers to the starting value, the number in brackets is the increment, and the final number
is the end value. We need a curly brace at the end of this line, the command on a separate line, and another
curly brace again on a separate line. Executing this command will give us a well behaved variable with every
entry in the same format. sum hour.
The other type of loop (the “foreach” command) is generally used when you want to perform the same
task on a number of different variables. We will write our own programme to transform every variable into a
z score. We will call it “zscore”. We then use the foreach loop, with “i” being the macro that corresponds to
8 Note that ‘ is the inward pointing single quotation mark, and is usually the button to the left of the number 1 on your

keyboard. You may need to press it twice.

9 There are also “global” macros which are rarely used.
10 This is what we are doing at the beginning of our do file. Stata is told to access the date, and open a log file under that

name.
11 Type mac list to see which macros are currently in use.
12 For more on accessing macros see help extended fcn.

15
every individual variable we wish to transform. The macro “0” refers to the list of variables we are interested
in, essentially everything we type after our new command.13 Like forvalues, we need a curly brace at the
end of this line, and each of our commands also on a separate line. For every variable in “0” (our variable
list) we are running the sum command, obtaining the macros for the mean and standard deviation, and then
generating a new variable which will have the prefix “zscore”. We also summarise this new variable. After
the final curly brace we need “end” to tell Stata the programme is finished.

capture program drop zscore

program define zscore
foreach i in ‘0’ {
sum ‘i’
local m=r(mean)
local n=r(sd)
gen zscore‘i’=(‘m’-‘i’)/‘n’
sum zscore‘i’
}
end

We can now run the programme on whichever variables we are interested in. For example, zscore time
cons price. The summary statistics confirm we have what we wanted. Note that if you try to call a new
programme by the same name as an existing programme you get an error message. program define zscore.
So if you want to modify an existing programme you will first need to drop it. program drop zscore. But
if there is no program called zscore this will produce an error message. Hence the use of “capture” just like
when we were opening our log file. Also, you will need to define your programme again each time you start
a new Stata session, unless you save it as an ado file.14

2.11 Counting, sorting and ordering

It is possible to access individual observations with your data using subscripts. This takes the form of square
brackets containing the observation number after the variable of interest. For example, di time[2] displays
the value of the time variable for the second observation. This can be very important if you want to access
information in the responses of other observations. For example you may have data on households and may
want to use information provided by parents to analysis the outcomes of children. We will illustrate this
with our province variable. Two special cases of subscripts are n and N. The latter is used to count the
total number in each case. For example, gen totalno= N gives us a new variable which is the total number
of observations in the dataset. It is obviously the same for each observation. tab totalno. On the other
hand, gen totalno2= n tells us the ranking of each observation in the dataset, and runs from 1 to 32. tab
totalno2. These are most useful when used with the by command.
We will generate two variables which gives us the total number in each province, and also the rank of
each observation in each province. First we need to sort our data sort province. Then by province: gen
provinceno= N15 and bysort province: gen provinceno2= n.
We can use the data browser to confirm that this generated the variables we expected. In this case we
have some province level data in var10, however it is only present for one observation in each province, and
we need it in all observations.
This is a case where you may be tempted to use excel, but apart from the replication issue, this will
simply not be possible if you have a datatset with 1000s of observations and hundreds of variables. Instead
we can use subscripts and a loop to make the change:

forvalues i=1(1)10 {
by province: replace var10=var10[‘i’] if var10==. & var10[‘i’]!=.}
13 Within the macro “0”, the macro “1” refers to the first variable, “2” the second etc.
14 Itso happens that Stata already has a way of creating standardised variables with egen newvar=std(var).
15 These two steps could be combined with “bysort”: bysort province: gen provinceno= N

16
 Within each province we are simply replace var10 with the value of the first observation in each province,
provided that variable is missing. We are then looping over ten values as there are 10 at most 10 observations.
The variable is only replaced if there is a missing value for that observation. Another useful command in
this context is gsort, which orders the variables according to the values of some other variable.

2.12 Reshaping Datasets

Sometimes data will be in the wrong “shape”. This is difficult to explain without an example. If we open the
file cyear use cyear2, and look in our data browser we will see that we have several variables that refer to
repeated measures over a number of years. In fact there are three outcomes: u5m (under 5 mortality), gdppc
(GDP Per Capita) and hivp (Proportion of Population infected with HIV/AIDS), and 9 years. Depending on
the analysis you wish to conduct, this may be awkward with the data in this “shape”. For example, tracking
an indicator across time is difficult with this format. We can use the reshape command to transform the
data into something more useable. The syntax is a little tricky, but the most important part is to identify
our outcome variables, our time variable, and our country (or firm or individual) identifier. The command
then takes the form “reshape” “outcome variables”, i(“identifier”) j(“time variable”). So in this case we
have reshape long u5m gdppc hivp, i(country) j(year). Now we can track an indicator across time,
e.g. tabstat gdppc, by(year). Of course we now essentially have a panel dataset, which is a whole other
topic. See section 7 for more details.

2.13 Graphs
Like the tables we discussed above, graphs are a powerful tool for exploring, summarising and presenting
your data. The basic graph commands for Stata are straightforward, however getting to grips with all the
available options is tricky. It would be impossible to discuss all the different types of graph, however we will
discuss the most common types. Like with tables we will divide graphs into two types, those that deal with
continuous variables and those that deal with categorical data. We will load our icecream data again. use
icecream2, clear.
For continuous data, the easiest way to visualise the relationship between two variables is to produce a
scatterplot of them, e.g. scatter cons temp. If, instead, we want a graph of the line of best fit between
the two variables, the relevant command is graph twoway lfit cons temp. We can also combine multiple
plots in one graph using the twoway command. For example, try using twoway (scatter cons temp) (lfit
cons temp). We can add some complexities very easily. For example you may wish to add confidence interval
“bands” around your line of best fit. To achieve this, use lfitci instead of lfit.16 Stata also can produce
several graphs in the one chart. For example we can create a 3x3 matrix of scatterplots with by inputting
graph matrix cons temp price, scheme(s1mono). To investigate the distribution of a single variable, we
can create a histogram of it using the histogram command. 17 For example, histogram temp.
We can produce a bar chart showing the mean of our variables using graph bar time cons price
income temp. We can display this breakdown for values of a particular variable in the same graph graph
bar time cons price income temp, by(province) or in different graphs graph bar time cons price
income temp, over(province). We may also want to graph categorical variables like province, with the
aim of showing the per cent in each category. The best way to do this is with the user written command
“catplot”. As before we use the search command search catplot, all or ssc install catplot, and click
on the blue link to install. Now we can use this to graph our province variable by itself catplot province,
or by another variable catplot province weekend.
We will discuss this further in the time series section, but if you have time series data then line graphs
can be important. First we need to sort our data by the time variable, in this case time. sort time. Then
we use the “line” command to graph the variables. The last variable needs to be our time variable. So in
this case we could have line cons income time.
16 Note that the easiest way to change the overall look of a graph is with schemes. See help schemes. We will use
“scheme(s1mono)” to generate the graphs in this document as we want them in black and white.
17 A similar chart is produced with graph7.

17
 Figure 1: An example of a graph matrix chart

20 40 60 80
.6

cons .4

.2
80

60
temp
40

20
.29

.28
price
.27

.26
.2 .4 .6 .26 .27 .28 .29

We can save any graph we produce in Stata using the graph export command. After drawing our graph,
Stata will open it in a new window. It is possible to save it using the menus FILE, SAVE AS. We can also
type graph export filename.png, replace. Stata can save graphs in various different formats, but .png
is the most straightforward. 18 . We can then use the graphs in other documents and presentations.19
It would take too long to go through all of the available options, but some of the most important ones
refer to the title and axis labels. For example:

line income temp time, title(Time Series Graph of Income and Tempreature Over Time) ///
xtitle(Time Period) ytitle(Euro(Income) and Degrees(Temp)) caption(Source: icecream.dta)

18 .wmf is best for word documents, .eps is best for LATEX.

19 There is a useful graph editor available in Stata version 10 onwards.

18
 The “///” tells Stata to read the next line as part of the same command. Two examples of the kind
of graphs that are possible are provided below. The first graph was made using the user written command
“spmap”. The second graph was generated using the cyear dataset with the following code:

twoway(scatter gdppc u5m if gdppc<1000 & year==2005, mlabel(country2) mlabsize(tiny)) ///

(lfit gdppc u5m if gdppc<1000 & year==2005), ///
caption(’’Source: WHO and Penn World Tables’’, size(tiny) span) legend(order(2 ’’Linear Fit’’)) ///
title(’’National Income Per Head and Under 5 Mortality’’, span) ///
ytitle(’’GDP Per Capita in US Dollars’’, size(small)) note(’’Note: Correlation=-0.1737’’) ///
plotregion(style(none)) legend(pos(4) col(1)) ///
xtitle(’’Under 5 Mortality, Deaths Per 1,000 Births’’) ///
graphr(lwidth(thin) ilwidth(vvthin) ilcolor(black) ilpattern(solid)) ///
subtitle(’’Developing Countries in 2005’’, span)

For more examples see:

http://www.survey-design.com.au/Usergraphs.html

And:
http://www.ats.ucla.edu/stat/stata/library/GraphExamples/default.htm

19
 Figure 2: Graph Example 1: Map
Proportion In Cluster 4
Family And Community Integrated

Sweden

Denmark

Percent
Ireland 33 − 40
Netherlands Poland 25.5 − 33
Germany 16 − 25.5
Belgium
Czech Rep 9 − 16

Austria
France Switzerland

Italy

Spain

Greece

Source: SHARE 2004 & 2006

Figure 3: Graph Example 2: Labelled Scatterplot

National Income Per Head and Under 5 Mortality

Developing Countries in 2005
Senegal
500
GDP Per Capita in US Dollars

Zimbabwe Lesotho
Nigeria
Kenya
400

Benin Zambia
Tanzania
300

Mozambique
Ghana Uganda
Mali

Burkino Faso
Madagascar
200

Malawi Liberia
100

80 100 120 140 160 180

u5m

Linear Fit
Note: Correlation=−0.1737
Source: WHO and Penn World Tables

20
3 Regression Analysis
The purpose of this lab is to provide an introduction to regression analysis using Stata, particularly the
interpretation of the output produced by this and other statistical packages.
If we type help regress we see that the basic command takes the form regress y x1 x2 x3. Load the
icecream data. use icecream2, clear. If we want to regress consumption on income temperature and price
from our icecream dataset we type reg cons price temp income.20 The output from this is presented in
table 4. Stata will present the output almost immediately as this is a small dataset. We see that there are
30 observations used in the regression. Also note that Stata automatically includes a constant term, the
variable cons, not to be confused with your y variable, cons. You do not need to create a constant term
yourself. 21 Interpreting output like this needs to become second nature if you are interested in pursuing
research in economics. The most important of these numbers are the coefficients and their p values so we
start with these. The coefficients (second column) tell us the effect of each X on Y (dy/dx). In this case
none of the variables are in log form so the interpretation is straightforward. The coefficients tell us how Y
changes if X changes by one unit. So in this case if price goes up by one unit, then consumption changes by
-1.044, i.e. consumption falls by 1.044. Likewise, if income goes up by one unit, consumption will go up by
.003. This is not the whole story however, some of these coefficients may be indistinguishable from zero (in a
statistical sense). It is important to remember the difference between a coefficient and a t-stat. Coefficients
are the “real” effects of variables but aren’t very useful without knowledge of the variance. Each coefficient is
assigned a t statistic in the 4th column. t = coefficient . The standard error of the coefficient is displayed
se(coefficient)
in the third column.
As we know the distribution of the t statistic, so we are able to assess the probability that the population
coefficient is equal to zero. This probability is the number displayed in the 5th column. So in this case we
cannot reject the possibility that price has no effect on consumption, however we can reject the possibility
that temperature and income have no effect. Another way of saying this is that the 95% confidence interval
for price includes 0. Also included in the output are the Residual Sum of Squares, the Model (or Explained)
Sum of Squares, and the Total Sum of Squares. We also have the R2 , which describes the proportion of
variation in the outcome explained by the model. Remember that R2 = M SS/T SS. We also have the F
statistic, which is a test of whether the coefficients are all jointly equal to zero. An F test is analogous to
a t test, except it examines several hypotheses at the same time. In this case F=22. As we know the F
distribution we can assign it a p value in the same way we did with the t values. In this case we reject the
possibility that the model coefficients are all zero. The F stat can also be seen as a test of whether the R2
value has arisen at random. The R2 term can only increase as more variables are added to the regression
and can therefore be misleading, as the model may contain several redundant variables which have little
explanatory power. The adjusted R2 takes account of the number of variables used.
Interpretation of coefficients when some of the variables are in log form is a little different. If both X
and Y variables are in logs, then the X coefficient tells you the percentage change in Y when X goes up
by 1%. If Y is in log form and X is not, then the X coefficient tells you the percentage change in Y when
X goes up by one unit. If Y is not in log form but X is, then the coefficient tells you the unit change in
Y if X goes up by 1%. For example, suppose we generate a new variable which is the log of income gen
logincome=ln(income), and include this in our regression instead of our original variable. reg cons price
temp logincome. Now the coefficient on logincome is .3 and we interpret this as meaning that a 1% rise in
income results in consumption rising by .3 units.

3.1 Dummy Variables

Suppose we also want to include the province variable in our regression. As before we rename our province
variable rename var7 province. It would be incorrect to run the model reg cons price temp income
province as province is not a continuous variable (we can remind ourselves that it is a categorical variable
20 Remember that Stata understands abbreviations.
21 The constant term can be omitted with the option noconstant, but you need good reasons to justify this.

21
 Table 4: OLS Regression Output

Source | SS df MS Number of obs = 30

-------------+------------------------------ F( 3, 26) = 22.17
Model | .090250523 3 .030083508 Prob > F = 0.0000
Residual | .035272835 26 .001356647 R-squared = 0.7190
-------------+------------------------------ Adj R-squared = 0.6866
Total | .125523358 29 .004328392 Root MSE = .03683

------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
price | -1.044413 .834357 -1.25 0.222 -2.759458 .6706322
temp | .0034584 .0004455 7.76 0.000 .0025426 .0043743
income | .0033078 .0011714 2.82 0.009 .0008999 .0057156
_cons | .1973149 .2702161 0.73 0.472 -.3581223 .752752
------------------------------------------------------------------------------

which runs from 1-4 with tab province). Instead we need to include a separate dummy variable for each
category.22 We could create 3 new variables, taking the value one for each category and zero otherwise.23 gen
leinster=(province==2) and gen munster=(province==3) and gen Connacht=(province==4).24 Now
we can include these 3 variables in our regression. reg cons price temp income leinster munster
connacht. There is an easier way of doing this however. We do not actually need to generate the sep-
arate dummy variables ourselves if we use the xi command, which expands categorical variables for us. xi:
reg cons price temp income i.province. Remember to include “i.” before the categorical variable you
wish to expand. Note that these dummies will appear in your variable window.
It is important to be able to interpret the coefficients on these dummy variables correctly. Notice again
that there are only three dummies, despite the fact that there are four categories in the province variable.
Each of these refers to the effect of each category relative to the omitted category. We can clearly see that
the omitted category is province 1 (Ulster). In any case, Stata tells us which one is omitted. If we want to
change the omitted category, for example if we wanted to compare the effect of the other provinces to being
in Leinster, we can set which group Stata omits. char province[omit] 2. If we run the regression again
we see this is indeed the case xi: reg cons price temp income i.province. The output is presented in
table 5.

3.2 Outreg2
You will encounter dozens if not hundreds of regressions, and if you need to present these results copying
and pasting from Stata is not really an option. Fortunately there is a user written command, similar to
tabout, which automates the export of regression results. First we need to install the programme. findit
outreg2 or search outreg2, all, or ssc install outreg2. Having done this, we can check the help file
help outreg2. We run the outreg command after our regression to save our results in excel format (this is
probably best, unless you are using LaTeX). We need to specify the filename. outreg2 using test, excel
replace. This exports all coefficients, if required you can specify the variables for export, e.g. outreg2
temp income using test, excel append. The replace and append options operate as with tabout. If you
choose append, new regression results will appear in different columns in the file. The excel option is also
22 Including binary variables on the right hand side of a regression does not violate any of the assumptions which make OLS

BLUE.
23 We don’t need 4 variables, this would be falling into the Dummy Variable Trap.
24 A quicker way to do this is tab province, gen(province).

22
 Table 5: OLS Regression Output With Dummy Variables

i.province _Iprovince_1-4 (naturally coded; _Iprovince_1 omitted)

Source | SS df MS Number of obs = 30
-------------+------------------------------ F( 6, 23) = 10.27
Model | .091413007 6 .015235501 Prob > F = 0.0000
Residual | .034110351 23 .001483059 R-squared = 0.7283
-------------+------------------------------ Adj R-squared = 0.6574
Total | .125523358 29 .004328392 Root MSE = .03851

------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
price | -1.272991 .9760936 -1.30 0.205 -3.292195 .7462123
temp | .0034345 .0004815 7.13 0.000 .0024385 .0044306
income | .0030756 .0012691 2.42 0.024 .0004503 .005701
_Iprovince_2 | .0093664 .0190921 0.49 0.628 -.0301285 .0488614
_Iprovince_3 | .003358 .0225584 0.15 0.883 -.0433076 .0500236
_Iprovince_4 | .0181066 .0210197 0.86 0.398 -.025376 .0615892
_cons | .2737497 .3113172 0.88 0.388 -.370259 .9177583
------------------------------------------------------------------------------

important as this ensures that excel will recognise the way the standard errors are formatted. An example is
shown in table 7. It is also possible to export the results directly to a word document with the “word” option
outreg2 using test2, word replace. In this case you will get a document with the extension “.rtf”.

3.3 Hypothesis Testing

We have already established the meaning of the t statistics and p values, these refer to a test of the hypothesis
that the coefficient on that variable in the population is equal to zero. However, we can also test any linear
hypothesis using the test command. For example, theory may tell us that the coefficient on the log of
income should be equal to one. test b[income]=1. We use b and square brackets to refer to the regression
coefficients on our variable(s) of interest. In this case we find a very large F statistic, and as the p value is <.05
or even .01 we reject the hypothesis that the coefficient on income is equal to one. We can also test several
hypotheses jointly test ( b[income]=1) ( b[temp]=.01). If we wish to test a non-linear hypothesis, such
as price*temperature=income, we use the testnl command testnl b[price]* b[temp]= b[income].
Suppose you run a regression of wages that includes experience and experience squared as independent
variables and find that neither have significant t-tests. You should do a joint-test of the significance of the
two terms measuring experience. The distribution of multiple variables differs from that of a single variable,
so you cannot use a t-test; you need to run an F -test and use the appropriate critical values. To test several
parameters, use the test or testparm commands. For example, testparm exper expersq outputs the F -
and p-values of the test that these coefficients are jointly-insignificant.

3.4 Post Regression Commands

There are a number of useful postregression commands involving the “predict” command. To obtain the pre-
dicted values after our regression xi: reg cons price temp income i.province, we use predict xbhat.
We can them compare these with the actual outcome with a scatter diagram. scatter cons xbhat. If the
model was perfect the correlation between the two variables would be 1. We can also obtain the residuals
with the command predict residuals, res, and examine the residual corresponding to each value of our

23
 Table 6: Outreg Example
(1)
VARIABLES cons

price -1.273
(0.976)
temp 0.00343***
(0.000482)

income 0.00308**
(0.00127)
Iprovince 1 -0.00937
(0.0191)
Iprovince 3 -0.00601
(0.0233)
Iprovince 4 0.00874
(0.0202)
Constant 0.283
(0.318)

Observations 30
R2 0.728
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

x variables scatter res temp and scatter res income. Both of these plots show the residuals centred
around zero which is not indicative of any misspecification, however we can test this more formally.

3.5 Interaction Effects

Interaction effects are an important tool in regression analysis. Suppose we have reason to believe that
the effect of income differs depending on the province. We can examine this by interacting income with
our province. xi: reg cons price temp i.province*income. Now we have three extra variables, in-
come*province2, income*province3 and income*province4. These variables tell us the differential effect of
income in each province relative to income in our base province. Note that the provincial dummy variables
and income variables (which gives us the average effect of income) are still included. As before we compare
the effects of the income interaction to the base category which is income*province1. In this case there is no
statistically different effect of income in each province. We can test whether all of these coefficients are equal
test b[ IproXincom 2]= b[ IproXincom 3]= b[ IproXincom 4]. We fail to reject the hypothesis that all
these coefficients are equal.

3.6 Specification and Misspecification Testing

It is important to test whether the errors have constant variance, i.e.the assumption of homoskedasticity,
otherwise statistical inference becomes unreliable, even if the coefficients are unbiased. The Breusch-Pagan
tests whether the residuals depend on the level of the X-variables (e.g. if the variance of income increases as
the income increases.) This is activated in Stata with the hettest command. (If installed, you can conduct
a White test with the command whitetst.)

24
 If you have a problem with heteroskedasticity, you can use White (1980) heteroskedastic-consistent stan-
dard errors by adding the “robust” option, e.g. regress y x1 x2 x3, robust. Robust errors are not
always appropriate. You should also consider whether you should cluster your standard errors with the
“cluster(variable)” option. See chapter 8 in Mostly Harmless Econometrics for more details on both these
issues. With regards to autocorrelation you can produce a correlogram with the ac command. The Durbin-
Watson test is called with dwstat. If you’d like to run a Cochrane-Orcutt regression instead of OLS, you
can use prais regression method. The standard test for model specification is the Ramsey RESET test.
The relevant command is ovtest. A high F -test suggests your model is improperly specified. To test for
multicollinearity, enter the vif command after a regression. This tests the variable inflation factor, which is
basically the effect of collinearity. One suggestion is that values greater than ten warrant further investiga-
tion. You can use Stata’s sktest to test normality. You could plot the residuals with predict myresids,
resid which produces the residuals in a variable named myresids. Then you can produce a histogram of
this variable, or use a qnorm plot. For example try kdensity myresids, normal to compare your graph to
the normal bell-curve.

25
 Table 7: Linear Probability Model Output

Linear regression Number of obs = 3296

F( 3, 3292) = 135.69
Prob > F = 0.0000
R-squared = 0.1010
Root MSE = .42809
------------------------------------------------------------------------------
| Robust
highwage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
exper | .0101499 .003119 3.25 0.001 .0040345 .0162652
male | .1605543 .0150249 10.69 0.000 .1310953 .1900133
school | .0781009 .0044197 17.67 0.000 .0694352 .0867666
_cons | -.7894874 .0600989 -13.14 0.000 -.9073225 -.6716523
------------------------------------------------------------------------------

4 Binary Regression
Including categorical or binary measures on the right hand side of a regression does not pose any problems
for estimation, however it is a different matter when the Y variable is categorical. We first examine the
case of a binary dependent variable. We will use the wages data. use wages1. We have information on
an individual’s years of schooling, their current wage(in logs), their experience (in years), and their gender.
We first summarise the data describe and summarize. So we have 4 variables and 3296 observations. We
examine the wage variable with histogram wage. As there are clearly a number of outliers we restrict this
graph to values less than 20, and add a kernel density plot. histogram wage if wage <20, kdensity. The
variable appears roughly lognormal. Suppose we are interested in the determinants of wages. As an initial
step, we could examine the correlation between our variables. pwcorr wage exper male school, sig.
Wages are significantly positively correlated with being male and years of schooling. We could also examine
the average wage by years of schooling tabstat wage, by(school). Again we see evidence of positive
correlation, however the relationship appears to be non-linear. The lowest average wage is for those in the
8 and 9 years of schooling. This is easier to see graphically graph bar wage, over(school) title(Wages
by Years of Schooling). We could run a standard regression with this data, but instead suppose we are
interested in the determinants of earning a wage with a value greater than 7. We generate a variable that
takes the value 1 if the individual earns this value or greater, and zero otherwise. gen highwage=(wage>7).
28% of the sample are in this category tab highwage. Males are more likely to be in this category tab
highwage male, row.

4.1 The Problem With OLS

We could run our analysis using OLS reg highwage exper male school, r, and outreg the results outreg2
using test, replace excel ctitle(OLS).25 We need robust standard errors as the binary nature of our
outcome induces heteroskedasticty. We now interpret our coefficients in terms of the probability of earning
a high wage (wage greater than 7).26
So for example, being male increases the probability of being in this category by 16%. Likewise an extra
year of schooling raises the probability of earning a high wage by 7%. Although OLS is still unbiased in
these situations, there is a difficulty as some groups may be predicted to have a negative probability of the
outcome. To see this graphically, we examine a plot of wages and schooling, and fit an OLS regression line.
25 The ctitle option places a title on the regerssion column.
26 Using OLS on a binary outcome is often referred to as the linear probability model (LPM).

26
As our outcome is binary, all observations will lie either on 1 or 0. twoway(scatter highwage school)(lfit
highwage school), scheme(s1mono) title(Scatterplot of High Wage and Schooling) ytitle(High
Wage) xtitle(Years of Schooling). Because there are a limited number of outcomes, each point could
represent several hundred observations or 10, it is not possible to tell. To get a better idea of the distribution
of our variables, we add some random noise to our plot using the jitter option twoway(scatter highwage
school, jitter(10))(lfit highwage school), title(Scatterplot of High Wage and Schooling) ytitle(High
Wage) xtitle(Years of Schooling). This issue is clear from this graph, for individuals with less than 9
years of schooling, the predicted probability of earning a high wage is negative, which clearly does not make
sense. To confirm this we obtain the predicted probabilities from our model. predict xbhat and check
them against years of schooling tabstat xbhat, by(school).

Figure 4: A Problem With OLS

Scatterplot of High Wage and Schooling

1
.5
High Wage
0
−.5

0 5 10 15
Years of Schooling

highwage Fitted values

4.2 Logit and Probit

Two of the most popular alternatives are the probit and logit estimators.27 Both are maximum likelihood es-
timators which involve slightly different distributional assumptions, but should produce roughly the same re-
sults.28 We run regressions with each of these probit highwage exper male school and logit highwage
exper male school. We also verify that the predicted probabilities are with the 0-1 range. predict xbhat2
and tabstat xbhat2, by(school). We can also graph these graph bar xbhat2, over(school). Figure
5 compares the predicted probabilities from each model. In this case logit and probit both give very similar
results.
27 A problem with the logit estimator is the independence of irrelevant alternatives assumption. The nested logit, “nlogit”

command, is a designed to overcome this difficulty.

28 See any standard econometrics text for further details, and how these relate to the concept of an underlying latent variable.

27
 Table 8: Logit and Probit Output

Iteration 0: log likelihood = -1968.4829

Iteration 1: log likelihood = -1794.7776
Iteration 2: log likelihood = -1792.4022
Iteration 3: log likelihood = -1792.4006

Probit regression Number of obs = 3296

LR chi2(3) = 352.16
Prob > chi2 = 0.0000
Log likelihood = -1792.4006 Pseudo R2 = 0.0895

------------------------------------------------------------------------------
highwage | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
exper | .0320678 .0113016 2.84 0.005 .009917 .0542186
male | .5130279 .0496018 10.34 0.000 .4158102 .6102456
school | .2565379 .0161262 15.91 0.000 .2249311 .2881447
_cons | -4.138153 .2306573 -17.94 0.000 -4.590233 -3.686073
------------------------------------------------------------------------------

Iteration 0: log likelihood = -1968.4829

Iteration 1: log likelihood = -1796.4414
Iteration 2: log likelihood = -1791.3207
Iteration 3: log likelihood = -1791.3022

Logistic regression Number of obs = 3296

LR chi2(3) = 354.36
Prob > chi2 = 0.0000
Log likelihood = -1791.3022 Pseudo R2 = 0.0900

------------------------------------------------------------------------------
highwage | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
exper | .0530249 .0194082 2.73 0.006 .0149855 .0910644
male | .8678092 .0847772 10.24 0.000 .7016489 1.03397
school | .4386614 .0283132 15.49 0.000 .3831685 .4941544
_cons | -7.028232 .4058646 -17.32 0.000 -7.823712 -6.232752
------------------------------------------------------------------------------

28
 Figure 5: Problem Solved With Probit and Logit

Comparison of Predicted Probabilities

By Years of Schooling
1 .5
High Wage
0 −.5

0 5 10 15
Years of Schooling

High Wage OLS/LPM Predicted Pr

Probit Predicted Pr Logit Predicted Pr

4.3 Marginal Effects

It is important to recognise that these coefficients are not the same as the output generated by an OLS
regression as they refer to the unobserved latent variable. They are not marginal effects, i.e. they do
not tell us the average effect of a change in the X variable on Y (dy/dx). This is the case with OLS
as it is a linear estimator. All we can interpret from probit or logit coefficients is the direction of the
average effect, and the significance, but not the magnitude. Marginal effects are rarely reported in other
disciplines in these models 29 , however economists are usually interested in the magnitude of an effect, not
just its statistical significance. We can calculate the marginal effects using the mfx compute command.
Whereas the coefficients from logit and probit will differ due to scaling, the marginal effects should be
almost identical. Outreg2 also works for exporting marginal effects, we will use this to compare different
ways of calculating marginal effects. logit highwage exper male school then mfx compute and outreg2
using test, mfx excel append ctitle(mfx logit). probit highwage exper male school then mfx
compute and outreg2 using test, mfx excel replace ctitle(mfx probit).
We see that this is indeed the case. Not only that, but these are also almost identical to the OLS results.
For example, for the probit model, an increase in a year of schooling increases the probability of earning a
high wage by 8%, and likewise for the logit model. As the marginal effects differ depending on the value of
the x variables, there are a number of different ways of calculating these.

29 Often odds ratios from logit models are reported in psychology, sociology and epidemiology.

29
 Table 9: Marginal Effects Output

Marginal effects after probit

y = Pr(highwage) (predict)
= .26504098
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
exper | .0105044 .0037 2.84 0.005 .003251 .017758 8.04187
male*| .1656721 .01561 10.61 0.000 .13508 .196264 .523968
school | .0840339 .00521 16.14 0.000 .073827 .09424 11.6302
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

Marginal effects after logit

y = Pr(highwage) (predict)
= .26010828
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
exper | .0102047 .00373 2.74 0.006 .002892 .017517 8.04187
male*| .1645799 .01554 10.59 0.000 .134131 .195029 .523968
school | .0844213 .00525 16.09 0.000 .074139 .094704 11.6302
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

By default, Stata calculates these marginal effects at the mean of the independent variables, however it is
also possible to evaluate them at other values. For example, suppose you suspect that the effect of experience
and schooling on wages differs for men and women. Then you could evaluate the marginal effects for women.
mfx compute, at(male=0) and outreg2 using test, mfx excel append ctitle(mfx probit women).
For men mfx compute, at(male=1) and outreg2 using test, mfx excel append ctitle(mfx probit
men). To reiterate, because OLS is a linear estimator the estimated marginal effect is the same at every set
of X values. You can think about this in terms of the slope of the regression line in the bivariate case. OLS
is a straight line, so the slope of the line is the same at every point, unlike with logit and probit.
The marginal effects of experience and schooling appear larger for men than women. There other ways
of approaching how to estimate marginal effects. An alternative is to calculate the marginal effects for every
value and then take the average to obtain the average partial effect. For this we use the user written command
margeff. ssc install margeff to install. Then we run the command with the replace option as we wish to
export the results. margeff, replace. We use outreg2 without the mfx option this time. outreg2 using
test, excel append ctitle(amfx probit). We can see that in this case both approaches give similar
results. Mfx2 is another user written command which produces marginal effects. If you are using a probit
model, you can obtain the marginal effects directly (without the coefficients) with the command dprobit
highwage exper male school.

30
 Table 10: Alternative Binary Estimators for HighWage
(1) (2) (3) (4) (5)
VARIABLES mfx probit mfx probit women mfx probit men amfx probit mfx hetprob

exper 0.0105*** 0.00856*** 0.0119*** 0.00987*** 0.0101***

(0.00370) (0.00304) (0.00418) (0.00347) (0.00368)
male 0.166*** 0.166*** 0.166*** 0.158*** 0.161***
(0.0156) (0.0156) (0.0156) (0.0171) (0.0155)
school 0.0840*** 0.0685*** 0.0951*** 0.0785*** 0.0812***
(0.00521) (0.00441) (0.00606) (0.00437) (0.00546)

Observations 3296 3296 3296 3296 3296

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

If you are concerned about potential misspecification, the heteroskedastic probit model allows you to
specify the variance as a function of one or more of the independent variables. Suppose we suspect that
the variance of our error term depends on your level of schooling, we could estimate the following hetprob
highwage exper male school, het(school) and compute the marginal effects mfx compute and export
them outreg2 using test, excel mfx append ctitle(mfx hetprob). The LR test of lnsigma is a test
of our assumption that the variance depends on schooling. We reject at the 5% significance level but fail to
reject at the 10% level. Finally, note that if you are including interaction effects in these kinds of binary
models you must be careful about interpreting them. The user written command “inteff” is recommended.

31
 1.00

1.00
Autocorrelations of im_rate_per_annum

Autocorrelations of gdp_maddison
0.50

0.50
0.00

0.00
−0.50

−0.50
−1.00

−1.00
0 5 10 15 20 0 5 10 15 20
Lag Lag
Bartlett’s formula for MA(q) 95% confidence bands Bartlett’s formula for MA(q) 95% confidence bands

Figure 6: Autocorrelation Functions For Infant Mortality and GDP

5 Time Series
We will examine the basic fundamentals of time series analysis using data on Irish infant mortality. First
we load the data use im, clear and apply some labels. lab var netm ’’Net Migration’’ and lab var
at ’’Average Annual Tempreature’’ lab var ar ’’Average Annual Rainfall’’. We also generate a
time squared variable gen year2=year^2. We will first summarise the data describe and sum, and plot
our variables. As we have time series data we need to tell Stata. tsset year. First we sort the data. sort
year.

5.1 Initial Analysis

We could naively run a regression of infant mortality on gdp and other factors, but this is potentially a
very serious mistake. reg im gdp at ar netm and export the results outreg2 using timeseries, excel
replace ctitle(Naı̈ve Regression). These results suggest that GDP has an impact on infant mortality,
however time series analysis involves dealing with the specifics of how variables behave over time. Ignoring
the fact that variables are trending over time could return spurious results.30 The high R2 is a warning
sign that this is a problem here. We could check the Durbin Watson test statistic estat dwatson and
estat bgodfrey. Both of these indicate the potential for the residuals to be non-stationary, as we reject
the assumption of no serial correlation. We should first examine our variables. line im year and line gdp
year, title(’’Time Series Plot of Irish GDP’’), remembering that Stata understands unambiguous
abbreviations. GDP is clearly non-stationary, as it infant mortality, at least for the period after 1947. We
can also check the autocorrelation and partial autocorrelation plots. ac im pac im ac gdp pac gdp. In each
case we have linearly declining autocorrelation functions and a high value of the first partial autocorrelation
function, indicating non-stationarity.

5.2 Testing For Unit Roots

We now confirm this with formal dickey fuller tests. Two issues arise when testing for stationarity, choice
of lag length and whether constants and trends should be included in the test. For the second issue, it is
recommended that the Dolado procedure31 is followed, however here we will assume that it is correct to
include both. We deal with the first issue by choosing lag lengths that reflect the best fit to the variables.
varsoc im. In each case of infant mortality we choose two lags with one for GDP. dfuller im, trend
lags(2). dfuller gdp, trend lags(1).
30 See
any standard econometrics text for details.
31 See
Dolado, J., Jenkinson, T., & Sosvilla-Rivero, S. (1990). Cointegration and Unit Roots. Journal of Economic Surveys,
4(3), 249-27

32
 1.00
1.00
Partial autocorrelations of im_rate_per_annum

Partial autocorrelations of gdp_maddison

0.50
0.50

0.00
−0.50 0.00

−0.50
0 5 10 15 20 0 5 10 15 20
Lag Lag
95% Confidence bands [se = 1/sqrt(n)] 95% Confidence bands [se = 1/sqrt(n)]

Figure 7: Partial Autocorrelation Functions For Infant Mortality and GDP

Table 11: Dickey Fuller Test Ouput

Augmented Dickey-Fuller test for unit root Number of obs = 38

---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
------------------------------------------------------------------------------
Z(t) -1.624 -4.260 -3.548 -3.209
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 0.7830

It is important to be able to interpret these tests properly. The null hypothesis is the presence of a unit
root. If the test statistic does not exceed the 5% critical value in absolute terms (i.e. is not more negative
then the 5% critical value), we cannot reject non-stationarity. In each case here we fail to reject the null
hypothesis of a unit root, and conclude that both variables are integrated of order 1. There are several
potential ways of addressing this. Remember that under normal conditions non-stationarity violates the
regularity assumption on which OLS is based. To deal with this problem, we could detrend the variables,
we could check for the presence of a cointegrating relationship between gdp and infant mortality, and finally
we could run a model in first differences.

5.3 Dealing With Non-Stationarity

To detrend the variables we regress each variable on time and time squared to obtain the residuals. We
include time squared as we suspect the trend to be quadratic. The residuals from this regression will be that
particular variable with the time trend removed. reg im year year2 and predict imres, r. Note that
the predicted values will correspondingly give us the infant mortality time trend. predict imtime. We
can graph this with line imtime imres year, title(’’Residuals and Fitted Values From Infant
Mortality’’) legend(lab(1 ’’Fiitted Values(IM Time Trend)’’) lab(2 ’’Residuals(IM Detrended)’’))
scheme(s1mono).
The detrended series looks like it could be stationary here, but this is could be due to the scaling on
the graph. For gdp reg gdp year year2 and predict gdpres, r. We could run the regression on our
detrended variables reg imres gdpres at ar netm and export the results outreg2 using timeseries,
append excel ctitle(Detrended Regression). However we should make sure that this is correct. varsoc
imres and varsoc gdpres. We can now use formal dickey fuller tests to check whether these detrended
variables are stationary. dfuller imres, trend lags(1) and dfuller gdpres, trend lags(2). In each

33
 Figure 8: Using OLS To Detrend Variables

80
60
40
20
0
−20 Residuals and Fitted Values From Infant Mortality

1920 1930 1940 1950 1960

year

Fiitted Values(IM Time Trend) Residuals(IM Detrended)

case, detrending the variables is not enough to induce stationarity. So we cannot simply include these
detrended variables in our regression. Note that in fact this kind of detrending procedure is hardly ever
necessary, as it is equivalent to simply including the time trend on the right hand side of our regression.32 .
Compare the coefficients on gdp from these two regression reg im gdp year year233 and reg imres
gdpres. There is a second possibility for running a valid regression, it may be that despite the fact that
infant mortality and gdp are non-stationary, they may be cointegrated, i.e. share a common trend. If we
were running Stata version 10 or above we could run a Johansen test for cointegrating vectors between the
two variables, however there are alternatives. The principle in testing for cointegration is that if such a
relationship exists between two variables, the residuals from their regression should be stationary. So we
run our basic model, this time including year and year squared. We know from the above analysis that
we cannot rely on this detrending to induce stationarity, however we will examine the residuals to test for
a cointegrating relationship.34 reg im gdp at ar netm year year2. We first graph the residuals line
cointres year and ac coint and pac coint. varsoc cointres and then dfuller cointres, trend
lags(1). We therefore cannot reject the null that the residuals are non-stationary, and therefore reject
the presence of a cointegrating relationship. We finally return to the option of running a regression in
first differences. Fortunately Stata provides a number of time series operators which simplify the process.
L.var is the lag of that variable, i.e. l.im is imt−1 . D.im is the first difference (imt − imt−1 ), and f.im
is the first lead (imt+1 ). We can easily generate a new variable that is the first difference of gdp. gen
dgdp=d.gdp and for infant mortality gen dim=d.im . Both now look roughly stationary, although gdp
32 This is the Frisch-Waugh-Lovell theorem.
33 Note that we need to add underscores in order not to confuse Stata as we created two new variables imres and gdpres.
34 This form of unit root test is known as an Engle Granger cointegration test.

34
 Table 12: A Comparison of Time Series Models
(1) (2) (3)
VARIABLES Naive Regression Detrended Regression First Differences

gdp maddison -0.0117***

(0.00115)
at 1.330 -0.846 -3.418
(3.139) (2.881) (2.704)

ar -0.00902 -0.00642 -0.0134

(0.00975) (0.00839) (0.00851)
netm 0.0531 -0.0256 0.00644
(0.0889) (0.0586) (0.0612)
gdpres -0.0130***
(0.00331)
dgdp -0.00154
(0.00446)
Constant 167.4*** 14.37 47.12
(35.22) (32.72) (30.56)

Observations 35 35 35
R2 0.860 0.379 0.119
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

still causes some concerns. line dim year line dgdp year. We reject non-stationarity in both cases in
formal tests. varsoc dim varsoc dgdp. dfuller dim, lags(1) trend. dfuller dgdp, trend lags(0).
We now proceed with a regression in first differences reg dim dgdp at ar netm and export the results
outreg2 using timeseries, excel append ctitle(First Differences).35 We conclude that the only
valid model is the first differences regression, and that gdp has no effect on infant mortality.

35 We could include year as a variable here, however it makes little difference.

35
6 Ordinal and Multinomial Regression
6.1 Ordinal Data
We have seen an example of how to analyse data where the outcome variable is binary. We will now examine
the case where the Y variable is categorical. There are two situations; where the variable has a clear ranking
(e.g. health or education) we will consider the ordered logit and ordered probit models. When the Y variable
has no clear ranking the multinomial logit or multinomial probit is appropriate. We will load data taken
from the Health Survey for England. use ordinal, clear. As always we examine our data. describe
and sum. We will also examine our outcome variable in more detail. rename genhelf health renames
the variable. tab genhelf. We see that this is the standard health question asked in surveys. Individuals
are asked to assess their general health on a scale of 1 to 5, with 1 being very bad and 5 excellent. If we
examine the labelling on this variable, we see that the coding is reversed. codebook health and labelbook
genhelf. We will fix this, and also tidy the value labels. recode health (1=5)(2=4)(4=2)(5=1). We
can now change the labels lab def genhelf 1’’Very Bad’’ 2’’Bad’’ 3’’Fair’’ 4’’Good’’ 5’’Very
Good’’, modify. Now this looks better tab health. Suppose we are interested in comparing health
across gender. We could take the average health score, tabstat health, by(gender), but this is not
ideal and in fact there is little difference here. Instead we should compare across categories. We could
graph this relationship catplot health sex, percent scheme(s1mono) title(Self Assessed Health
by Gender) blabel(bar, format(%8.1f)).

Figure 9: Health Distribution

Self Assessed Health by Gender

Very Bad 0.5

Bad 2.5

men Fair 7.7

Good 18.8

Very Good 15.5

Very Bad 0.9

Bad 2.7

women Fair 10.2

Good 23.3

Very Good 17.9

0 5 10 15 20 25
percent

There are more women in the very good and good categories, but also more in the bottom two categories.
Note how we added labels onto the end of the bars with the “blabel” option. We also examine height.
rename htval height and codebook height and labelbook htval. We replace missing values mvdecode
height, mv(-1).

36
 Figure 10: Height Distribution

Height Distribution
.04
.03
Density
.02.01
0

140 160 180 200

(d) valid height (cm)

We see that height is roughly normally distributed hist height, normal scheme(s1mono) title(Height
Distribution). Suppose we wish to examine the determinants of self reported health status. We generate
an age squared term gen age2=age2 . We could start with a simple OLS model xi:reg health age age2
i.sex i.topqual3 height. We save the results outreg2 using ordinal, excel replace.
However, the difficulty with OLS is potentially more important in this situation than in the previous
case where we had a binary variable. The assumption behind these kinds of models is that the 5 point
scale is really a manifestation of an underlying continuous health index. OLS assumes that movement
from very bad to bad is equivalent to movement from good to very good. A simple examination of the
distribution shows that it is highly skewed towards good health. The problem is that we do not know
where these thresholds lie in this unobserved continuous index. The ordered logit and probit models take
account of these thresholds. They are similar to the standard probit and logit models we encountered in
the last section. xi:oprobit health age age2 i.sex i.topqual3 height and xi:ologit health age
age2 i.sex i.topqual3 height. The output from the ordered probit model is shown in table 13. The
“cut” values refer to estimates of the thresholds.
The same issue as in the binary case arises here. These are coefficients and not marginal effects, so we
can only identify the significance of the variable and not the magnitude. We can obtain marginal effects, but
in this case it will be the marginal effect of the variable for being in that particular category compared to all
others. In the ordered logit case, if we were interested in the marginal effect for being in very good health
mfx compute, predict(outcome(5)). Note that these are computationally intensive and can take a long
time to estimate. In this case we see that a unit increase in height increases the probability of being in the
very good category by .3%. Being in education category 8 reduces the probability of being in this category
by nearly 20%. mfx2 will estimate marginal effects for all categories. This raises a difficult question about
interpretation. Here there are 5 categories, and with 11 variables, which means 55 coefficients.

37
 Table 13: Ordered Probit Output

i.sex _Isex_1-2 (naturally coded; _Isex_1 omitted)

i.topqual3 _Itopqual3_1-8 (_Itopqual3_1 for topqual3==-1 omitted)
Iteration 0: log likelihood = -8403.7684
Iteration 1: log likelihood = -8109.4856
Iteration 2: log likelihood = -8109.3117
Iteration 3: log likelihood = -8109.3117
Ordered probit regression Number of obs = 6859
LR chi2(11) = 588.91
Prob > chi2 = 0.0000
Log likelihood = -8109.3117 Pseudo R2 = 0.0350

------------------------------------------------------------------------------
health | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | -.0056614 .003921 -1.44 0.149 -.0133464 .0020236
age2 | -.0000358 .0000398 -0.90 0.369 -.0001139 .0000423
_Isex_2 | .1133844 .0373295 3.04 0.002 .0402199 .1865488
_Itopqual3_2 | .0691763 .0826651 0.84 0.403 -.0928444 .2311969
_Itopqual3_3 | -.0504849 .08567 -0.59 0.556 -.218395 .1174252
_Itopqual3_4 | -.1730846 .080075 -2.16 0.031 -.3300286 -.0161406
_Itopqual3_5 | -.1893723 .0780889 -2.43 0.015 -.3424238 -.0363208
_Itopqual3_6 | -.3217053 .0921638 -3.49 0.000 -.5023429 -.1410677
_Itopqual3_7 | -.2502062 .097808 -2.56 0.011 -.4419063 -.0585061
_Itopqual3_8 | -.5067183 .0812331 -6.24 0.000 -.6659322 -.3475045
height | .0085651 .0020132 4.25 0.000 .0046193 .012511
-------------+----------------------------------------------------------------
/cut1 | -1.47487 .3508896 -2.162601 -.7871385
/cut2 | -.7184513 .3490263 -1.40253 -.0343723
/cut3 | .1835116 .3488053 -.5001342 .8671575
/cut4 | 1.393712 .3490541 .7095789 2.077846
------------------------------------------------------------------------------

Also, suppose you find that a variable increases the probability of being in the top category and the
bottom category. It is difficult to know what to conclude from this kind of result. Often researchers will
divide the variable into two categories (e.g. very goood and good health vs all other categories). OLS is a
potential alternative if there are a large amount of categories.
A useful written command is “spost” which is aimed at the analysis of this type of categorical data. It
provides a number of diagnostic tests, as well as tools for interpretation. We will make use of it to graph
the predicted probabilities for our outcome. ssc install spost. We obtain the predicted probabilities for
each of our outcome categories by height prgen height, gen(height1), and graph these with the following
code. This is useful as it is apparent from this that height increases the chances of being in the top category
and reduces the chances of being in the “fair” category. Interestingly, it appears to have little effect on the
probability of being in the bottom two groups.

38
twoway connected heighp1 heighp2 heighp3 heighp4 heighp5 heighx, scheme(s1mono) ///
title(’’Predicted Probabilities - Health Condition by Height’’, span) ///
subtitle(’’Ordered Logit Model’’, span) ytitle(Probability) ///
legend(lab(1 ’’Pr(Very Bad)=1’’) lab(2 ’’Pr(Bad)=1)’’) lab(3 ’’Pr(Fair)=1’’) ///
lab(4 ’’Pr(Good)=1’’) lab(5 ’’Pr(Very Good)=1’’)) xtitle(Height)

Figure 11: Ordered Logit Predicted Probabilities

Predicted Probabilities − Health Condition by Height

Ordered Logit Model
.5
.4
Probability
.2 .3
.1
0

140 160 180 200

Height

Pr(Very Bad)=1 Pr(Bad)=1)

Pr(Fair)=1 Pr(Good)=1
Pr(Very Good)=1

6.2 Multinomial Regression

Now suppose we are interested in the determinants of housing conditions. We first examine this variable tab
tenureb. We remove the missing values mvdecode tenure, mv(-9 -8). It would not make sense to run an
OLS model here as there is no clear way to rank these alternatives. For this we use the multinomial logit
or multinomial probit model. We first tell Stata to use married as the base outcome for marital status char
marstatb[omit]2 and run our model xi:mlogit tenureb hhsize i.sex age age2 i.marstatb and for
probit xi:mprobit tenureb hhsize i.sex age age2 i.marstatb. We will not run this command as it
generally takes some time.36 The interpretation of this output is a little different. Again remember that we
can only interpret the significance and direction of these variables. The coefficients are like pairwise logit
comparisons, they refer to the effect of a variable of being in that category relative to the base category.
Note that we have 3 categories, but output for only 2. So for example, women are more likely to own and
have a mortgage relative to the base (renting). Household size has a positive effect on renting relative to
having a mortgage, but not for owning. As with the ordinal case, we can compute the marginal effects for
any of the three categories. mfx2, replace and export the results outreg2 using multinomial, excel
replace. As before, we will also examine the results graphically with the spost command.
36 The MNP should be considered if you are concerned about the IIA assumptions with MNL.

39
 Table 14: OLS and MFX for Ordinal Data
(1) (2)
VARIABLES OLS Logit MFX For Very Good Health

age -0.00472 -0.000611

(0.00307) (0.00147)
age2 -2.92e-05 -2.89e-05*
(3.13e-05) (1.50e-05)

Isex 2 0.0917*** 0.0428***

(0.0290) (0.0139)
Itopqual3 2 0.0670 0.0167
(0.0628) (0.0310)
Itopqual3 3 -0.00878 -0.0284
(0.0654) (0.0304)
Itopqual3 4 -0.114* -0.0652**
(0.0610) (0.0268)
Itopqual3 5 -0.117** -0.0761***
(0.0594) (0.0265)

Itopqual3 6 -0.223*** -0.115***

(0.0709) (0.0269)
Itopqual3 7 -0.171** -0.0873***
(0.0755) (0.0308)
Itopqual3 8 -0.389*** -0.180***
(0.0621) (0.0236)
height 0.00679*** 0.00320***
(0.00156) (0.000758)
Constant 3.273***
(0.271)

Observations 6859 6859

R2 0.087
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

40
 This time we will look at the predicted probabilities for men by age prgen age, x( Isex 2=0) f(20)
t(80) gen(male). The probability of having a mortgage unsurprisingly decreases with age, and increases
for the other two options.

twoway connected malep1 malep4 malep2 malex, scheme(s1mono) ///

title(’’Predicted Probabilities - Home Ownership For Men’’, span) ///
subtitle(’’Multinomial Logit Model’’, span) ytitle(Probability) ///
legend(lab(1 ’’Pr(Own)=1’’) lab(2 ’’Pr(Mortgage)=1)’’) lab(3 ’’Pr(Rent)=1’’)) ///
xtitle(Age)

Figure 12: Multinomial Logit Predicted Probabilities

Predicted Probabilities − Home Ownership For Men

Multinomial Logit Model
.8 .6
Probability
.4 .2
0

20 40 60 80
Age

Pr(Own)=1 Pr(Mortgage)=1)
Pr(Rent)=1

41
 Table 15: Multinomial Logit Output

i.sex _Isex_1-2 (naturally coded; _Isex_1 omitted)

i.marstatb _Imarstatb_1-7 (_Imarstatb_3 for marstatb==2 omitted)
Iteration 0: log likelihood = -7729.6211
Iteration 1: log likelihood = -6337.5541
Iteration 2: log likelihood = -6233.3063
Iteration 3: log likelihood = -6227.7351
Iteration 4: log likelihood = -6227.7068
Iteration 5: log likelihood = -6227.7068
Multinomial logistic regression Number of obs = 7271
LR chi2(20) = 3003.83
Prob > chi2 = 0.0000
Log likelihood = -6227.7068 Pseudo R2 = 0.1943
------------------------------------------------------------------------------
tenureb | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Own |
hhsize | -.0236389 .032346 -0.73 0.465 -.0870359 .0397581
_Isex_2 | .199169 .0679491 2.93 0.003 .0659912 .3323468
age | -.0012009 .0148197 -0.08 0.935 -.030247 .0278451
age2 | .0010277 .0001468 7.00 0.000 .00074 .0013155
_Imarstatb_1 | 1.337345 .2719465 4.92 0.000 .8043398 1.87035
_Imarstatb_2 | 1.326905 .1286428 10.31 0.000 1.07477 1.57904
_Imarstatb_4 | -.0765315 .284506 -0.27 0.788 -.6341531 .4810901
_Imarstatb_5 | -.0434973 .1650001 -0.26 0.792 -.3668915 .279897
_Imarstatb_6 | .3076225 .1870105 1.64 0.100 -.0589114 .6741564
_Imarstatb_7 | -.146967 .1565473 -0.94 0.348 -.453794 .15986
_cons | -3.364784 .4225025 -7.96 0.000 -4.192874 -2.536694
-------------+----------------------------------------------------------------
rent it |
hhsize | -.0507241 .0267455 -1.90 0.058 -.1031444 .0016962
_Isex_2 | .1605255 .0628003 2.56 0.011 .0374392 .2836117
age | -.1082971 .0123948 -8.74 0.000 -.1325904 -.0840037
age2 | .0015481 .0001318 11.74 0.000 .0012897 .0018065
_Imarstatb_1 | .5218535 .1950537 2.68 0.007 .1395553 .9041517
_Imarstatb_2 | 1.565435 .1134457 13.80 0.000 1.343085 1.787784
_Imarstatb_4 | 1.862286 .190297 9.79 0.000 1.489311 2.235262
_Imarstatb_5 | 1.58454 .1345246 11.78 0.000 1.320877 1.848204
_Imarstatb_6 | 1.2927 .1937198 6.67 0.000 .9130164 1.672384
_Imarstatb_7 | 1.373128 .1063721 12.91 0.000 1.164643 1.581613
_cons | .0742578 .3289097 0.23 0.821 -.5703933 .718909
------------------------------------------------------------------------------
(tenureb==Mortgage is the base outcome)

42
 Table 16: MFX for Multinomial Logit
(1) (2) (3)
VARIABLES Own Mortgage Rent

hhsize -0.000749 0.00940 -0.00865

(0.00596) (0.00582) (0.00535)

Isex 2 0.0262** -0.0443*** 0.0181

(0.0120) (0.0134) (0.0119)
age 0.00784*** 0.0143*** -0.0221***
(0.00239) (0.00291) (0.00220)
age2 8.15e-05*** -0.000323*** 0.000241***
(2.32e-05) (3.07e-05) (2.23e-05)
Imarstatb 1 0.251*** -0.219*** -0.0317
(0.0634) (0.0394) (0.0417)
Imarstatb 2 0.108*** -0.320*** 0.212***
(0.0253) (0.0185) (0.0250)

Imarstatb 4 -0.165*** -0.275*** 0.440***

(0.0235) (0.0279) (0.0350)
Imarstatb 5 -0.141*** -0.236*** 0.377***
(0.0168) (0.0228) (0.0267)
Imarstatb 6 -0.0647*** -0.209*** 0.274***
(0.0199) (0.0355) (0.0351)

Imarstatb 7 -0.136*** -0.198*** 0.334***

(0.0188) (0.0198) (0.0245)

Observations 7271 7271 7271

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

43
7 Panel Data
7.1 Panel Set Up
This tutorial presents the basics of panel data analysis, using a sample of employed workers in the British
Household Panel dataset.37 It is available for download from the UK Data Archive http://www.data-archive.ac.uk/.
First we load the data use bhps1, clear. Note that this is quite a large dataset. If we browse, we see that
the data is in the incorrect “long” format so we need to use the reshape command to fix this. reshape long
mlstat jbstat hlstat smoker jbsat fisit caruse age qfedhi hhsize hhtype ncars fiyr fihhyr jbhrs,
i(pid) j(year). Now we have a unique identifier (“pid”), a time variable (“year”), and our outcomes.
As always, we first label and tidy our data.
lab var mlstat ‘‘Marital Status’’
lab var jbstat ‘‘Job Status’’
rename hlstat health
lab var health ‘‘General Health’’
lab var smoker ‘‘Smoker’’
lab var jbsat ‘‘Job Satisfaction’’
lab var fisit ‘‘Financial Situation’’
lab var caruse ‘‘Use A Car’’
lab var age ‘‘Age’’
lab var qfedhi ‘‘Highest Educational Qualification’’
lab var hhsize ‘‘Household Size’’
lab var hhtype ‘‘Type of Household’’
lab var ncars ‘‘Number of Cars’’
lab var fiyr ‘‘Annual Income’’
lab var fihhyr ‘‘Household Annual Income’’
lab var year ‘‘Wave’’
lab var jbhrs ‘‘Hours Worked Per Week’’
gen age2=age^2
lab var age2 ‘‘Age Squared’’
Suppose we are interested in the relationship between income and job satisfaction. We first examine
our job satisfaction variable tab jbsat. There are over 40,000 observations on this variable in the data;
remember that this will involve information on the same individuals over time. We will ignore missing
values and individuals who didn’t answer. First we check these values codebook jbsat and labelbook
kjbsat. Alternatively, the user written command lablist is useful, which you can install with ssc install
lablist. Then we can recode these to missing values replace jbsat=. if jbsat<1. And check tab
jbsat. We also do this for missing values in our health and education variables mvdecode health, mv(-9
-1) and mvdecode qfedhi, mv(-9 -8). Let’s examine our income variable hist fiyr. This suggests
the presence of outliers which we may wish to exclude from our analysis. We first remove missing neg-
ative values replace fiyr=. if fiyr<0. One way to do this is to exclude the bottom and top 5%
of observations. First we find these values using the “centile” command. centile fiyr, centile(5
95) and recode these values as missing replace fiyr=. if fihhyr<2500 | fihhyr>32000. Now we
have a much more sensible distribution hist fiyr. We take the log of income gen logincome=ln(fiyr)
and graph the distribution hist logincome, kdensity scheme(s1mono) title(Income Distribution)
xtitle(Log Income) caption(’’Source: BHPS’’, span).
We can also examine our job satisfaction variable tab jbsat and graph it catplot jbsat sex, percent
blabel(bar, format(%8.1f)) title(Job Satisfaction) subtitle(By Gender). We can obtain the av-
erage income for each category tabstat logincome, by(jbsat). We notice that those in the highest cate-
gory for job satisfaction have the lowest income. We can use a scatterplot to confirm that the relationship
is non-linear.
37 This is a fantastic resource covering 18 waves, see http://www.iser.essex.ac.uk/survey/bhps for more details.

44
 Figure 13: BHPS Income

Income Distribution
.8
.6
Density
.4.2
0

8 8.5 9 9.5 10 10.5

Log Income
Source: BHPS

graph twoway (scatter logincome jbsat if logincome>6, jitter(6)) ///

(qfit logincome jbsat if logincome>6), scheme(s1mono) legend(lab(2 ’’Quadratic Fit’’)) ///
ytitle(Log Income) title(Income and Job Satisfaction) caption(’’Source: BHPS’’, span)

If we look at the health variable, tab health and codebook health, we notice that the highest level of
health corresponds to “1” and the lowest to “5”. This may be confusing so we reverse this recode health
(1=5)(2=4)(4=2)(5=1). Don’t forget to change the labels labelbook khlstat and label define khlstat
1’’Very Bad’’ 2’’Bad’’ 3’’Fair’’ 4’’Good’’ 5’’Very Good’’, modify.
We will now start our regression analysis. We could estimate a simple OLS model xi:reg logincome
sex age age2 i.jbsat i.health i.year i.qfedhi with income as a function of gender, age age squared,
job satisfaction health and education, with controls for survey year. This involves ignoring the fact that we
have information on the same individuals over time, which may not be ideal as you can imagine.
If we try to run a panel model xi:xtreg logincome sex age age2 i.jbsat i.health i.year i.qfedhi,
re we get an error message. Like with the time series case, we must inform Stata that we have a panel, and
identify our time and identifier. xtset pid year. However, before running the model, we should examine
the panel aspects of the data. The xtdescribe command we obtain the characteristics of our data.
We have a strongly balanced panel (that is all individuals are present in all waves), with 6429 people
tracked over 11 years. We can check that this gives us over 70,000 observations in total bysort pid: gen
no= N and tab no. “xtsum” is another useful command. Remember that there is variation within individuals
as well as across individuals in a panel. We examine this for our job satisfaction variable xtsum jbsat.
Time invariant variables have no variation within individuals over time xtsum sex. This will be impor-
tant when we examine the fixed effects model. Another command is “xttrans” which gives the transition
probabilities for a variable. xttrans jbsat. See table 19. So if you are in category 7, you have a 50/50
chance of remaining there next period.

45
 Figure 14: BHPS Income by Job Satisfaction

Income and Job Satisfaction

10 9
Log Income
8 7
6

0 2 4 6 8
Job Satisfaction

logincome Quadratic Fit

Source: BHPS

Table 17: xtdescribe Output

pid: 10007857, 10014608, ..., 19140509 n = 6429

year: 1990, 1991, ..., 2000 T = 11
Delta(year) = 1 unit
Span(year) = 11 periods
(pid*year uniquely identifies each observation)

Distribution of T_i: min 5% 25% 50% 75% 95% max

11 11 11 11 11 11 11
Freq. Percent Cum. | Pattern
---------------------------+-------------
6429 100.00 100.00 | 11111111111
---------------------------+-------------
6429 100.00 | XXXXXXXXXXX

Table 18: xtsum Output

Variable | Mean Std. Dev. Min Max | Observations

-----------------+--------------------------------------------+----------------
jbsat overall | 5.413611 1.348332 1 7 | N = 40011
between | 1.05996 1 7 | n = 6274
within | .9841251 .0136113 10.14088 | T-bar = 6.37727

46
 Table 19: xttrans Output

Job |
Satisfacti | Job Satisfaction
on | 1 2 3 4 5 | Total
-----------+-------------------------------------------------------+----------
1 | 20.69 15.71 12.45 11.88 12.26 | 100.00
2 | 8.15 15.54 20.76 9.94 18.73 | 100.00
3 | 3.57 8.31 24.72 14.58 24.92 | 100.00
4 | 2.73 4.29 14.00 23.48 28.49 | 100.00
5 | 1.10 2.59 8.26 10.64 35.05 | 100.00
6 | 0.57 1.25 3.67 3.87 19.40 | 100.00
7 | 0.74 0.56 1.25 1.62 6.97 | 100.00
-----------+-------------------------------------------------------+----------
Total | 1.60 2.71 6.99 7.45 21.52 | 100.00
Job |
Satisfacti | Job Satisfaction
on | 6 7 | Total
-----------+----------------------+----------
1 | 19.92 7.09 | 100.00
2 | 22.93 3.95 | 100.00
3 | 20.91 2.99 | 100.00
4 | 22.43 4.57 | 100.00
5 | 37.85 4.51 | 100.00
6 | 60.99 10.25 | 100.00
7 | 40.04 48.82 | 100.00
-----------+----------------------+----------
Total | 45.24 14.48 | 100.00

47
 It is also likely that the Y variable is correlated over time. We can confirm this by looking at the
correlation within income over time pwcorr logincome l.logincome l2.logincome l3.logincome, sig.
Note that we are using a time series operator, all of which also work in a panel context. As expected, income
is serially correlated. Not that the pwcorr command uses parwise deletion of missing values, whereas corr
uses listwise.

7.2 Panel Data Is Special

It is important to realise that panel data differs from cross sectional and time series data for a number
of reasons. In our standard cross sectional model we have our outcome as a function of other variables
across individuals (or firms, countries etc). yi = α + βXi + i . In the context of panel data we have our
outcome as a function of other variables across individuals and time yit = αi + βXit + it . There are
others, but the most important advantage of panel data is that it can be used to deal with the problem
of unobserved heterogeneity. Individuals differ due to omitted variables which can often not be measured.
This can bias estimation in other contexts. The error term for an individual i in time t now consists of
two parts uit = αi + it . The first is something that makes that individual different, but does not change
over time (αi ). The second is the usual random error term in any particular period that we are famil-
iar with from our cross sectional models (it ). The main econometric choices in panel analysis relate to
how you model the αi component. These can be thought of in terms of intercepts. For an illustration,
see Kennedy (2003). Consider the problems with OLS in this context. Firstly, the standard errors in
this model are incorrect as they are almost certainly correlated within individuals over time. If you had
a large positive error in the first period, you are likely to have a large positive error in the second. We
can correct this problem by clustering the standard errors at the individual level. xi:reg logincome sex
age age2 i.jbsat i.health i.year i.qfedhi, vce(cluster pid) and save the results for the coeffi-
cients we’re interested in outreg2 age age2 sex Ijbsat 2 Ijbsat 3 Ijbsat 4 Ijbsat 5 Ijbsat 6
Ijbsat 7 using panel, excel replace ctitle(OLS(Clustered)). We include a dummy variable for
each year to capture anything unique to that time period. We can test whether the cofficients are jointly 0
with the testparm command. We see that being in the highest category for job satisfaction appears to be
associated with a reduction in income. However, in this case (referred to as pooled OLS), you are imposing
everyone to have the same intercept αi = α. As we have stated, individuals are likely to differ on some
unobserved characteristic(s), so this assumption is unlikely to hold. We can easily test the assumption of a
common slope.
An alternative is to assume that the unobserved characteristics are random, and are not correlated with
any of our X variables and consequently the period specific error term. I.e. αi is not correlated with it .

7.3 Random and Fixed Effects

We will now implement our random effects estimator with the “xtreg” command. xi:xtreg logincome
sex age age2 i.jbsat i.health i.year i.qfedhi, re and export the results outreg2 age age2 sex
Ijbsat 2 Ijbsat 3 Ijbsat 4 Ijbsat 5 Ijbsat 6 Ijbsat 7 using panel, excel append ctitle(RE).
The output is shown in the table below (the coefficients have been omitted). After an RE regression we can
test whether POLS is an appropriate model with the command xttest0. The null hypothesis is that there
is a common intercept, so in this case we strongly reject and proceed with the RE model.
Note that the model explains 34% of the variation in income overall, but 27% of the variation within
individuals over time, and 36% of the variation across individuals. σu is the standard deviation of the
individual time invariant error term (αi ), while σe is the standard deviation of it . Rho essentially refers to
the correlation of the error term within individuals over time. Notice that Stata kindly reminds you about
the assumption of X and ui (or αi ) being uncorrelated.
If we believe that the individual error term is correlated with the time varying term, then the ran-
dom effects model is not appropriate and will produce biased estimates. This can be thought of as a
missing variable problem. For example, in this model we may believe that personality determines both
income and job satisfaction. A way to deal with this is to use changes within individuals over time.

48
 Table 20: Test For A Common Intercept

Breusch and Pagan Lagrangian multiplier test for random effects

logincome[pid,t] = Xb + u[pid] + e[pid,t]

Estimated results:
| Var sd = sqrt(Var)
---------+-----------------------------
logincome | .6037161 .7769917
e | .1510079 .3885974
u | .2929869 .5412827

Test: Var(u) = 0
chi2(1) = 33273.67
Prob > chi2 = 0.0000

Table 21: Random Effects Output

Random-effects GLS regression Number of obs = 33021

Group variable: pid Number of groups = 5990

R-sq: within = 0.2797 Obs per group: min = 1

between = 0.3613 avg = 5.5
overall = 0.3478 max = 10

Random effects u_i ~ Gaussian Wald chi2(33) = 13946.18

corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000

sigma_u | .38981447
sigma_e | .25416102
rho | .70169989 (fraction of variance due to u_i)

49
 Table 22: Fixed Effects Output

Fixed-effects (within) regression Number of obs = 33021

Group variable: pid Number of groups = 5990

R-sq: within = 0.2867 Obs per group: min = 1

between = 0.0432 avg = 5.5
overall = 0.0968 max = 10

F(32,26999) = 339.13
corr(u_i, Xb) = -0.0551 Prob > F = 0.0000
sigma_u | .54079082
sigma_e | .25416102
rho | .81908033 (fraction of variance due to u_i)

This eliminates the influence of all time invariant characteristics of the individual, including unobserved
heterogeneity. The intuition is as follows. As personality doesn’t change over time (let’s just assume it
doesn’t for now!), it cannot explain the change in the outcome over time. Of course it may explain the
level, but this doesn’t matter when we’re looking at first differences or deviations from a mean. This is
commonly referred to at the fixed effects estimator, as it is equivalent to adding a fixed effect (dummy
variable) for each individual. xi:xtreg logincome sex age age2 i.jbsat i.health i.year i.qfedhi,
fe and export the results outreg2 age age2 sex Ijbsat 2 Ijbsat 3 Ijbsat 4 Ijbsat 5 Ijbsat 6
Ijbsat 7 using panel, excel append ctitle(FE). Notice that the overall explained variance is much
lower.
Table 22 compares these estimates across the different approaches. We see that the coefficient on job
satisfaction category 7 (completely satisfied) is -.04 (i.e. being in this category reduces income by 4%,
remember it is measured in logs) for POLS, -.03 for RE, and -.025 for FE, and is also only marginally
significant. So taking account of unobserved heterogeneity reduces the effect substantially. It is important
to understand how the FE coefficient is generated. Only movement from one category of job satisfaction to
another (changes over time within individuals), is used to estimate the effect on income. People who move
categories are generally referred to as “switchers”.

50
7.4 The Hausman Test
The drawback of the FE estimator is that it is less efficient than the RE estimator (i.e. higher variance), and
also the fact that you cannot recover the coefficients on time invariant characteristics. So if the assumption
of no correlation between the individual error and explanatory variables holds, then you should use Random
Effects. Fortunately there is a way of testing which estimator is more appropriate in any given situation.
This is based on the fact that under the null hypothesis of random individual effects the estimators should
give similar coefficients. The Hausman test can be implemented by comparing the estimates from the two
models. Suppose we run a simple case xi:xtreg logincome i.jbsat, re. We need to save these estimates
so we can use them for comparison estimates save RE. Now we run our FE model xi:xtreg logincome
i.jbsat, fe and save estimates save FE. Now we test hausman FE RE, sigmamore.

---- Coefficients ----

| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| FE RE Difference S.E.
-------------+----------------------------------------------------------------
_Ijbsat_2 | .0603609 .0606981 -.0003372 .0027453
_Ijbsat_3 | .0712668 .0786652 -.0073983 .0027158
_Ijbsat_4 | .0326557 .0330731 -.0004174 .0028292
_Ijbsat_5 | .0671614 .0700455 -.002884 .0027512
_Ijbsat_6 | .0518018 .0504758 .001326 .0028042
_Ijbsat_7 | -.0425275 -.0607399 .0182125 .0031464
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from xtreg
B = inconsistent under Ha, efficient under Ho; obtained from xtreg

Test: Ho: difference in coefficients not systematic

chi2(6) = (b-B)’[(V_b-V_B)^(-1)](b-B)
= 193.25
Prob>chi2 = 0.0000

In this case, we can reject the null hypothesis that the coefficients are equal, therefore we conclude that
the FE model is correct.

7.5 Dynamics
We have already seen that there is a high degree of correlation in our income variable, and we may wish to
account for this with a more dynamic model, for example by including income in the previous period (yt−1 )
as a regressor. Unfortunately the FE estimator is biased (RE is worse here as αi is incorporated into it ) in
the case where T is small (T>30 is recommended). We run this specification for comparison only xi:xtreg
logincome l.logincome sex age age2 i.jbsat i.health i.year i.qfedhi, fe. Note again the use of
the lag operator (l.logincome). A potential solution is to use an instrumental variables approach. The
details are not straightforward, and require further assumptions about the structure of the error term. We
implement the Arellano and Bond estimator for illustration xi:xtabond logincome age age2 i.jbsat
i.health i.year i.qfedhi, which uses the second lag as an instrument for the first difference.

51
 Table 23: Comparison of Panel Data Estimators
(1) (2) (3) (7) (8)
VARIABLES OLS(Clustered) RE FE FE Lag FE ABond

L.logincome 0.365*** 0.346***

(0.00576) (0.0158)
sex -0.464*** -0.511***
(0.0113) (0.0111)
age 0.0569*** 0.0642*** 0.0604*** 0.0414*** 0.0350***
(0.00259) (0.00179) (0.00767) (0.00688) (0.00867)
age2 -0.000618*** -0.000694*** -0.000724*** -0.000373*** -0.000253***
(3.09e-05) (2.06e-05) (2.53e-05) (2.51e-05) (5.76e-05)
Ijbsat 2 0.0247 -0.00271 -0.00615 0.00882 0.00542
(0.0238) (0.0152) (0.0153) (0.0140) (0.0155)
Ijbsat 3 0.0772*** 0.00947 0.000456 0.0163 -0.00175
(0.0220) (0.0135) (0.0136) (0.0126) (0.0142)
Ijbsat 4 0.0359* 0.00368 0.00129 0.0166 -0.00257
(0.0217) (0.0133) (0.0135) (0.0127) (0.0142)
Ijbsat 5 0.0792*** 0.0113 0.00234 0.0210* -0.00471
(0.0210) (0.0127) (0.0129) (0.0121) (0.0137)
Ijbsat 6 0.0479** -0.00313 -0.0101 0.0139 -0.0108
(0.0208) (0.0126) (0.0127) (0.0120) (0.0136)
Ijbsat 7 -0.0434* -0.0301** -0.0258* 0.00844 -0.0201
(0.0225) (0.0131) (0.0133) (0.0127) (0.0145)
Constant 8.470*** 8.541*** 8.129*** 4.896*** 5.152***
(0.0915) (0.0577) (0.273) (0.321) (0.366)

Observations 33021 33021 33021 25314 17458

R-squared 0.366 0.287 0.413
Number of pid 5990 5990 4916 4145
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

52
8 Instrumental Variables
8.1 Endogeneity
This tutorial will provide an introduction to Instrumental Variables regression. One of the basic assumptions
required for the unbiasedness of OLS is exogeneity, that there is no correlation between the X variables and
the error term in the model. This can arise under a number of scenarios, including omitted variables,
simultaneity and measurement error. Common examples include the effect of schooling on wages and the
effect of aid on growth. In these circumstances IV can be used to produce consistent estimates of the effect of
the endogenous variable on the outcome. The idea is to find some variable (Z) that predicts the X variable,
but not the Y. This isolates the exogenous part of X, which we can use to examine the true causal effect
of X on Y. In the case of the schooling example, this could be a change in compulsory schooling leaving
age which exogenously (i.e. is outside the decision making capacity or characteristics of the individual)
increases the amount of schooling received. Here we will illustrate the technique using a simple example of
how openness affects growth. First we load the data use openness, clear and have a preliminary look
at it sum and describe. sum pcinc open land. If we graph scatterplots of these three variables we see
that there are a number of outliers, so we remove these from our analysis graph matrix pcinc open land,
scheme(s1mono) title(’’Income, Openness and Area’’).

Figure 15: Graph Matrix for Openess, Area and Income Per Capita

20 40 60 80
.6

cons .4

.2
80

60
temp
40

20
.29

.28
price
.27

.26
.2 .4 .6 .26 .27 .28 .29

replace pcinc=. if pcinc>15000 replace open=. if open>100 replace land=. if land>1000000.

We are primarily interested in the relationship between income per head and openness, so we graph that
relationship twoway(lfit pcinc open)(scatter pcinc open). We see that there may be some suggestion
of a week relationship in the descriptive data. We could run an OLS model with robust standard errors reg
pcinc open inf oil good, r and save the results outreg2 using iv, excel replace ctitle(OLS). On
the basis of this model, 1% increase in the proportion of imports as a % of GDP raises income per head by $20.

53
Oil producers have higher income, as do countries with “good” data. Higher inflation has a negative effect.
However it is important to recognise that the relationship between GDP and openness may be determined
by another unobserved factor, or that higher income may cause higher levels of openness. In other words,
openness is potentially endogenous, and therefore the model estimated above has no causal interpretation.
To deal with this issue we need to instrument for this variable. The two criteria for an instrument are that
the variable must be correlated with the X (openness) but not correlated with the Y (income). One possi-
bility here is land area. If you believe that smaller countries are more likely to engage in international trade
(because of economies of scale, specialisation etc.) but that area does not necessarily affect GDP per capita
otherwise, then area is a valid instrument. The first assumption is testable, if we graph the relationship we can
clearly see a significant negative correlation twoway (scatter open land if land<600000) (lfit open
land if land<600000), ytitle(Imports % GDP) title(Openness and Land Area) scheme(s1mono).

Figure 16: Openness and Area

Openness and Land Area

100
80
Imports % GDP
40 20
0 60

0 100000 200000 300000 400000 500000

land area, square miles

imports as % GDP, ’73− Fitted values

If you do not believe the second assumption in this case, that is probably wise! We will just use this as
an example of how to implement the procedure. IV papers ultimately depend on how much you believe this
assumption, and you will encounter many you do not. Unfortunately there is no econometric test which will
settle the matter entirely. We could look at the correlation between these three variables pwcorr pcinc open
land, sig. First, openness is significantly negatively correlated with area. Secondly income is positively
correlated with openness, although not significantly. The assumption of no correlation between Z and U
refers to the relationship in the population, here we only have a sample, and a relatively small sample at
that. However, if we have more instruments than exogenous variables, it is possible to test how close the
corresponding sample moment is to zero. We will illustrate this later.

54
 Table 24: Correlation Between Income, Openness and Area

| pcinc open land

-------------+---------------------------
pcinc | 1.0000
open | 0.0669 1.0000
| 0.4796
land | 0.1507 -0.3083 1.0000
| 0.1095 0.0008

Table 25: OLS and IV Comparison

(1) (2)
VARIABLES OLS IV

open 23.76* 49.29*

(13.45) (29.28)
inf -13.82** -9.959
(6.684) (7.702)
oil 5022*** 4883***
(1390) (1358)
good 3679*** 3482***
(489.6) (493.9)
Constant 563.0 -466.3
(545.7) (1174)

Observations 110 104

R2 0.330 0.339
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

8.2 Two Stage Least Squares

IV is often referred to as two stage least squares (2SLS). This is because the procedure can be conducted
with OLS by regressing the endogenous variable on the instrument and obtaining the predicted values (the
first stage). The second stage consists of regressing the Y variable on these predicted values and other
exogenous variables. To illustrate the procedure, consider the simple case where we instrument for openness
with land area. reg open land and predict openhat and reg pcinc openhat. It is recommended that
you instead use the “ivregress” command or the user written “ivreg2”. For one thing, the standard errors
from the manual procedure are incorrect. ivregress 2sls pcinc (open = land). You can see that the
coefficients are almost the same here. We will implement the procedure for our full model with robust
standard errors ivregress 2sls pcinc inf oil good (open = land), vce(robust) and save the results
outreg2 using iv, excel append ctitle(IV). So compared to our OLS model, the effect of openness has
doubled, although it is still only significant at the 10% level. This is not surprising, we have a small sample,
and the price we pay for using IV is less precision in our estimation. The lower correlation between the X
and the Z, the higher these standard errors will be.

55
 Table 26: Testing for Weak Instruments

First-stage regression summary statistics

--------------------------------------------------------------------------
| Adjusted Partial Robust
Variable | R-sq. R-sq. R-sq. F(1,99) Prob > F
-------------+------------------------------------------------------------
open | 0.1801 0.1470 0.1415 17.6524 0.0001
--------------------------------------------------------------------------
Minimum eigenvalue statistic = 16.3221

Critical Values # of endogenous regressors: 1

Ho: Instruments are weak # of excluded instruments: 1
---------------------------------------------------------------------
| 5% 10% 20% 30%
2SLS relative bias | (not available)
-----------------------------------+---------------------------------
| 10% 15% 20% 25%
2SLS Size of nominal 5% Wald test | 16.38 8.96 6.66 5.53
LIML Size of nominal 5% Wald test | 16.38 8.96 6.66 5.53
---------------------------------------------------------------------

8.3 Weak Instruments, Endogeneity and Overidentification

It is also important to recognise that the validity of IV rests on asymptotic properties. IV estimates can be
badly biased in situations where the instrument is weakly correlated with the X so it is important to test for
weak instruments. This is generally done by examining the F statistic from the first stage regression. Low
values are an indication of potentially weak instruments. We can examine this by using the “firststage” test
estat firststage, forcenonrobust . We use the “forcenonrobust” option as we used robust standard
errors and the theory for this test relies on the usual standard errors (so again caution is advised in this
case). There is a lot of information in this output, but the important things to focus on are the F test
for land “17.65” (a general rule of thumb is that if this is less than 10 there is cause for concern) and the
minimum eigenvalue statistic “16.32”, which is the Stock and Yogo test for weak instruments. We compare
both of these to the critical value (at the 5% level) of “16.38”. The test passes on the first statistic, but
fails for the second. As we are using robust errors we should only reject the presence of weak instruments if
our statistics greatly exceeded the critical values. As this is not the case, we should be sceptical about our
results.
If you are confident about the validity of your instrument, it is possible to test whether the variable is
in fact endogenous. This is based on the fact that if the variable is indeed exogenous there should be little
difference between the OLS and IV estimates. This test is referred to as the Durbin-Wu-Hausman test, and
is implemented with the following command estat endogenous. In this case we fail to reject the hypothesis
that OLS and IV are equivalent, and therefore fail to reject the hypothesis that openness is exogenous. This
may seem strange, as the IV estimate is twice that of OLS. The reason for this is again our small sample
size and large IV standard errors. If we had more data, this could reduce the standard errors. Remember
that this test is only valid with a valid instrument.

Tests of endogeneity
Ho: variables are exogenous
Robust score chi2(1) = .467088 (p = 0.4943)
Robust regression F(1,98) = .423783 (p = 0.5166)

56
 Earlier we stated that it was possible to examine the validity of an instrument if the model was over-
identified (i.e. more instruments than endogenous variables). This is based on the fact that the sample
correlation between the instrument and error term should be close to zero if the instruments are valid. We
will illustrate this using a second instrument, whether the country is a major oil producer. Clearly this
will affect GDP as well as openness, so this is not a good instrument. Again we just using this example
for illustration of how to implement the test. First we run our model with our two instruments, and the
gmm option. The generalised method of moments is similar to the 2sls estimator, but is suitable for use in
the overidentified case. ivregress 2sls pcinc inf good (open = land oil), wmatrix(robust). We
can now implement the test, which is often referred to as the “overidentifying restrictions test” or the
“Hansen-Sargan” test. estat overid.

Test of overidentifying restriction:

Hansen’s J chi2(1) = 4.05487 (p = 0.0440)

In this case we reject the null hypothesis that all instruments are valid, which is hardly surprising.
Finally, there are a number of other important commands relating to IV that you may require in your
analysis. The user written command “ivreg2” provides extensions to the Stata “ivregress” command. When
the instrument is binary, the “treatreg” command is applicable, and when your Y variable is binary, the
“ivprobit” command. With weak instruments, the LIML (limited information maximum likelihood) and
JIVE (Jackknife IV) estimators can be used to provide better small sample estimates. The command for IV
in a panel context is “xtivreg”. There are also a number of different estimators for dealing with dynamic
panel models, such as “xtabond” which implements the Arellano and Bond estimator.

57
9 Recommended Reading and References
Dublin Microeconomics Blog http://dublinmicroblog.blogspot.com/

UCLA Stata Website http://www.ats.ucla.edu/stat/stata/sk/default.htm

Princeton Stata Website:http://dss.princeton.edu/online help/stats packages/stata/stata.htm

Graphs 1 http://www.survey-design.com.au/Usergraphs.html

Graphs 2 http://www.ats.ucla.edu/stat/stata/library/GraphExamples/default.htm

A Visual Guide to Stata Graphics, Michael N. Mitchell http://www.stata.com/bookstore/vgsg.html

UK Data Archive http://www.data-archive.ac.uk/

Irish Social Science Data Archive http://issda.ucd.ie/

Econometric Analysis 6th Edition, William H. Greene http://pages.stern.nyu.edu/ wgreene/Text/econometricanalysis.htm

Econometric Analysis of Cross Section and Panel Data, Jeffrey Wooldridge

https://www.msu.edu/ ec/faculty/wooldridge/book2.htm

A Guide To Econometrics, Peter Kennedy http://mitpress.mit.edu/catalog/item/default.asp?ttype=2tid=9576

Microeconometrics Using Stata, A. Colin Cameron and Pravin K. Trivedi http://www.stata.com/bookstore/mus.html

Regression Models for Categorical Dependent Variables Using Stata, J. Scott Long and Jeremy Freese
http://www.stata.com/bookstore/regmodcdvs.html

Mostly harmless econometrics: an empiricist’s companion, JD Angrist and JS Pischke

http://www.mostlyharmlesseconometrics.com/

University of Essex. Institute for Social and Economic Research. ESDS Longitudinal and University of Essex. UK
Data Archive. ESDS Longitudinal, British Household Panel Survey: Waves 1-11, 1991-2002: Teaching Dataset
(Work, Family and Health) [computer file]. 2nd Edition. University of Essex. Institute for Social and Economic
Research, [original data producer(s)]. Colchester, Essex: UK Data Archive [distributor], November 2011. SN: 4901,
http://dx.doi.org/10.5255/UKDA-SN-4901-2

58