A guide to Stata

Florian Chavez Juarez

Version 1.7 - May 18, 2015

Contents
1 Introduction 5
1.1 Why to use Stata? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Alternative tutorials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Interface, components and data structure 6

2.1 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Command based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Help files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 File types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Data types: database and variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.6 Data type and format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.6.1 Data type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.6.2 compress - Optimize the memory use of your database . . . . . . . . . . . . . . 10
2.6.3 Further compressing by using a ZIP-file . . . . . . . . . . . . . . . . . . . . . . 10
2.6.4 Display format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6.5 format - Change the display format of a variable . . . . . . . . . . . . . . . . . 11
2.7 Missing values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Basic commands 12
3.1 do-files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1 Line breaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.2 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Prepare Stata for your analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Loading and saving data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 The generic Stata command . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Types of commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.7 Relational and logical operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.8 Placeholders and selection of multiple variables . . . . . . . . . . . . . . . . . . . . . . 16
3.9 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.9.1 matrix define - Generic command for matrices . . . . . . . . . . . . . . . . . 17

Centro de Investigracion y Docencia Economicas (CIDE), Mexico City, Mexico. [email protected]

1
CONTENTS CONTENTS

3.9.2 Some basic matrix functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.9.3 matrix list - Displaying matrices . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.9.4 matrix rownames/colnames - Renaming matrix columns and rows . . . . . . . 18
3.9.5 matrix get - Getting system matrices . . . . . . . . . . . . . . . . . . . . . . . 18
3.9.6 mkmat - Data to Matrix conversion . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.9.7 svmat - Matrix to Data conversion . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.10 Factor variables and time series operators . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.10.1 i. - Categorical variables in the regression . . . . . . . . . . . . . . . . . . . . 19
3.10.2 l. - Lagged variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.10.3 f. - Forwarded variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.10.4 d. - Difference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.10.5 c. - Continuous variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.10.6 # - Interaction terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.11 Loops and programming conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.11.1 if/else - if/else/elseif clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.11.2 while - while loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.11.3 foreach - Looping through elements . . . . . . . . . . . . . . . . . . . . . . . . 21
3.11.4 forvalues - Looping through numerical values . . . . . . . . . . . . . . . . . . 22

4 System commands 23
4.1 System commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.1 exit - Leaving the do-file or Stata . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.2 query - Displaying Stata options . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.3 set - Change settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.4 about - Getting information of the version and license . . . . . . . . . . . . . . 23
4.1.5 update - Update your Stata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.6 findit - Find ado-files online . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.7 search - Search ado-files online . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.8 help - Display help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

5 Data handling 25
5.1 Variable manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.1.1 generate - Generate a new variable . . . . . . . . . . . . . . . . . . . . . . . . 25
5.1.2 replace - Replacing a variable . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.1.3 egenerate - Generate a new variable with summary statistics . . . . . . . . . . 25
5.1.4 recode - Recoding a variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1.5 label - Label your variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1.6 rename - Rename a variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.1.7 drop - Deleting variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 Describing and sorting the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.1 describe - Describe the dataset . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.2 codebook - Display the codebook . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.3 sort - Sorting the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.4 gsort - Sorting the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.5 order - Sorting the variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2
CONTENTS CONTENTS

5.2.6 aorder - Sorting the variables alphabetically . . . . . . . . . . . . . . . . . . . 29

5.3 Joining several databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.3.1 merge - Merging/appending two databases (horizontally) . . . . . . . . . . . . 29
5.3.2 append - Appending/merging two databases (vertically) . . . . . . . . . . . . . 30
5.4 Changing structure of a database (panel data) . . . . . . . . . . . . . . . . . . . . . . 31
5.4.1 reshape - Reshape your panel data . . . . . . . . . . . . . . . . . . . . . . . . 31

6 Summary statistics and graphics 32

6.1 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.1.1 summarize - Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.1.2 tabstat - More flexible summary statistics . . . . . . . . . . . . . . . . . . . . 32
6.1.3 tabulate - One- and twoway tables of frequencies . . . . . . . . . . . . . . . . 32
6.1.4 correlate - Correlation and Covariance . . . . . . . . . . . . . . . . . . . . . . 33
6.2 Graphs and plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2.1 plot - Easy scatter plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2.2 graphics - General graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.2.3 spmap - Build vectorial maps with shapefiles (shp) . . . . . . . . . . . . . . . . 36
6.2.4 Export graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

7 Econometric analysis 39
7.1 Continuous outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.1.1 regress - OLS estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.1.2 Other estimators of continuous outcome . . . . . . . . . . . . . . . . . . . . . . 40
7.2 Categorical outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.2.1 probit - Probit estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.2.2 dprobit - Probit estimation with marginal effects . . . . . . . . . . . . . . . . 41
7.2.3 logit - Logit regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.2.4 mlogit - Multinominal logit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.2.5 oprobit - Ordered probit model . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.2.6 ologit - Ordered logit model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.3 Count data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.3.1 poisson - Poisson regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.3.2 nbreg - Negative binomial regression . . . . . . . . . . . . . . . . . . . . . . . . 42
7.4 Panel data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
7.4.1 xtset - Set-up the panel data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
7.4.2 xtdescribe - Describe the pattern of the panel data . . . . . . . . . . . . . . . 42
7.4.3 xtreg - Panel regression: fixed and random effects . . . . . . . . . . . . . . . . 43
7.5 Time series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7.6 Extracting estimation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7.7 Post estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.7.1 hettest - Breusch-Pagan / Cook-Weisberg test for heteroskedasticity . . . . . 44
7.7.2 test - Test linear hypotheses after estimation . . . . . . . . . . . . . . . . . . . 44
7.8 Saving and reusing estimations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.8.1 est store - Save an estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.8.2 est restore - Restore an estimation . . . . . . . . . . . . . . . . . . . . . . . 45

3
CONTENTS CONTENTS

7.8.3 est replay - Replay an estimation . . . . . . . . . . . . . . . . . . . . . . . . . 45

7.8.4 est table - Display an estimation table of several regressions . . . . . . . . . 45
7.9 Marginal effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
7.9.1 margins - Compute marginal effects (post estimation) . . . . . . . . . . . . . . 46

8 Stata meets LaTeX 50

8.1 Exporting high quality graphics to LaTeX . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.1.1 graph export - Graph export to eps . . . . . . . . . . . . . . . . . . . . . . . . 50
8.2 Exporting estimation results to LaTeX . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.2.1 estout - Creating estimation tables . . . . . . . . . . . . . . . . . . . . . . . . 50
8.2.2 tabout - Creating descriptive statistics tables . . . . . . . . . . . . . . . . . . . 52

9 Importing data in other formats 54

9.1 usespss - Read SPSS data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
9.2 fdause - Read SAS XPORT data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

10 Stata is not enough? Add user-written modules to enhance Stata 55

11 Programming your own command 56

11.1 Where and how to save your routine? . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
11.2 Useful commands for programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
11.2.1 tokenize - Tokenize a string . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
11.2.2 macro shift - Shifting the content of a local . . . . . . . . . . . . . . . . . . . 58
11.2.3 marksample - Selecting the sample considering if and in . . . . . . . . . . . . . 58
11.2.4 preserve - Making an image of your current data . . . . . . . . . . . . . . . . 58
11.2.5 restore - Restoring an image of your data . . . . . . . . . . . . . . . . . . . . 58
11.2.6 local:extended - Extract labels from variables and value labels . . . . . . . . 58

12 Simulation 60
12.1 Uniform distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
12.2 Normal distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
12.3 Other distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
12.4 Setting the random seed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4
1 INTRODUCTION

1 Introduction

This document has mainly two purposes: first, it should help new users to get started with Stata and
second, it should serve more experienced users as a look-up document. In order to comply with the
first goal of the document, I start with a general introduction to the software package and introduce
then chapter by chapter more complicated notions in order to familiarize the reader with the software

first and then introduce him to the possibilities Stata offers. The second goal should be achieved by
the use of a clear structure and an extensively detailed index in the end of the document. Like soft-
ware development, this document will never achieve a final version and comments and suggestions are
always welcome. Even though I refer to some econometric models, this document is NOT a reference
for econometric analysis. The reader is supposed to understand the models I present here and know
how and when to use them.
The document is under constant review and subject to changes and extensions. Please check for
updated versions frequently. Please report all errors to [email protected].

Check your version online

Click here or go to http://www.econ.chavezjuarez.com/vcheck.php?i=stata&v=1.7 to check if
you have the newest version.
This document is freeware. You can find the newest version online at http://econ.chavezjuarez.
com/stata.php for free.

1.1 Why to use Stata?

Stata is one of the most used software packages in applied econometrics, since it is fast, flexible and
relatively easy to understand. In addition to the very long list of included packages to run econometric

analysis, Stata benefits from a large user base and high quality user written commands that can
be installed and used very easily. For almost any econometric application that can be found in the
literature, there is a package. See section 10 for an explanation on how to add user written packages
to Stata .

Besides the wide-ranging possibilities in terms of econometric analysis, Stata has, in my opinion,
two very important advantages. First, it is not only a statistical package, but also a data-handling

package. Merging, manipulating, extracting etc. of databases is very easy and very fast in Stata . A
second plus of the software is the possibility to interact with other software packages, especially with
LATEX , allowing to produce directly result displays without a need to copy them to the LATEX .

1.2 Alternative tutorials

There are many tutorials about Stata available in the internet. Definitely the most complete way to

learn about Stata is the Stata -help, integrated in your program or online at http://www.stata.
com/help.cgi?contents. Other good addresses are:
http://www.econ.uiuc.edu/~econ472/tutorial1.html

http://www.cpc.unc.edu/research/tools/data_analysis/statatutorial/index.html

http://data.princeton.edu/stata/ (Very nice tutorial)

http://www.ats.ucla.edu/stat/stata/

5
 2 INTERFACE, COMPONENTS AND DATA STRUCTURE

2 Interface, components and data structure

2.1 Interface

The Stata interface might be somewhat scary at first glance, but it becomes very useful once the
user understands it well. Figure 1 represents a typical setting of the interface, where the big black

part is the output panel of Stata , all results will be displayed there1 . On the left side you have the
list of variables of the currently loaded database (see 3.3). Below the output panel, you will find the
command line, which should be used only under some circumstances I will present afterwards. Below

do-File
editor

Variables in
active data

Output panel

Command line
Current working
directory

Figure 1: Stata Interface (Version 11.2)

the command line you see the current working directory, which is by default always the directory

where Stata is installed. Among the button in the menu, you find the do-file editor, which will
probably be the most used button. The do-file editor is a simple text-editor with syntax highlighting
1
In the new versions of Stata, the black screen is actually white and the results in black. Personally I prefer the black
version (you can activate it under Edit Preference General preferences Color scheme: Classic.) It looks a little
bit more old-fashioned but highlights very well the different results.

6
2 INTERFACE, COMPONENTS AND DATA STRUCTURE 2.2 Command based

(since Version 11). The remaining parts of the interface will be introduced later on, but they will not
play a crucial role in the way I suggest to use Stata .
2.2 Command based

Unlike other statistical packages, Stata is mostly command based and the use of user interfaces
is relatively limited. Generally, all commands could be initiated by the use of the mouse and user
interfaces, however, it is not an efficient way to do things. My suggestion is to take the time to

learn correctly the syntax-based use of Stata and to work exclusively with the do-files. Do-files
are like m-files in Matlab or Syntax-files is SPSS and allow to write down a series of commands and
programming code. Using do-files to do all - starting from loading the database, carrying out the
analysis and storing the results - permits to save time and avoid errors. Moreover, the results are
easily reproducible by you and other researchers.
Normally, every line in the do-file corresponds to one command, which is generally what you want and
avoids the need of finishing every line with a special character like in many programming languages.

However, Stata offers also the possibility to change this temporarily or permanently if you like it.
This feature is useful when you have an extremely long command which does not fit the screen at all
(see 3.1.1 for more details).

2.3 Help files

All official Stata and most of the user written commands are accompanied by a so-called help file.
These files generally include a description of what the command is performing, an explanation of the
syntax including mandatory and optional arguments and some examples.
To display the help file just type into the command line
help commandname
where commandname is simply the name of the command you would like to see the corresponding help
file.
At the beginning of a help file, the syntax is explained. The bold elements refer to the name of the
command and in some cases to the options. The underlined part of the word is the part you have to
write in the command, the remaining characters need not be written. For instance, the command for
a simple OLS regression is regress, however it is sufficient to write reg. All elements in brackets [
] refer to optional elements. The syntax is generally divided into two parts by a comma (,) where all
behind the comma are options.
Some elements are present in almost all commands:
depvar The dependent variable
indepvar The independent variables
varlist A list of variables, just write one after the other
[if] If condition you might include to limit the command to a sample
[in] In condition you might include to limit the command to a sample
[weight] Possibility to include a variable with sample weights

2.4 File types

A Stata -project is generally composed of at least two files. The database is where your data is
stored and has the file type extension .dta. The do-files containing the commands to be performed

7
2.5 Data types: database and variables
2 INTERFACE, COMPONENTS AND DATA STRUCTURE

in the analysis end with the extension .do and is a simple text file you can also manipulate in other
text editors like Notepad or Notepad++.
Besides these two main file types there exists the help files (ending with .hlp) and the ado-files ending
with .ado. Both are used especially when you write your own commands, a topic I will discuss in
section 11.

Finally Stata is not limited to these file types, you can theoretically read and write any kind of files.
I will discuss a couple of examples like the output of graphics, the automatic generation of LATEX files
or the import of SPSS and SAS data (section 9)
The following table displays the most common file extensions used in Stata :
Extension Description
.dta database file in Stata format
.do Executable Stata syntax file
.ado Stata command file (each command is written in an ado-file)
.hlp Stata help file (has the same name as command)

2.5 Data types: database and variables

When thinking about statistical analysis the main data type is obviously the database containing all
the information you want to analyze. This kind of data is stored in the .dta files and can be loaded

(see section 3.3) with the command use. However, there are other data types used in Stata as well.
When you write a do-file, you probably need from time to time saving some information in a variable
to use it again later on. It would be very inefficient to save such information to the database, thus
you can store them in other places.

type supported format Way to define it Way to invoke value

local numeric and string local name=55 `name'
global numeric and string global name=55 $name
scalar numeric scalar name=55 name
matrix numeric matrix see matrix

The first three types are generally interchangeable, however I suggest to use the local whenever possible.
The local is preferable to the global since it remains set only until the end of the do-file, while the
globals are not cleared automatically in the end. This feature of the globals can be useful in some
cases, however, there is always a risk of having a global defined in a previous do-file and affecting later
on when it should not. Moreover, it could be argued that the locals have a clearer invoking command
than the scalars, since they need to be in between two very specific apostrophes always. On the other
hand, scalars have the advantage that we can display easily all stored scalar values at once using the
command scalar list.

Globals
To define a global and giving it the value of 888 you can just write
global myname=888
. If you would like to define the global based on the results of another command which stores the
value for instance as r-class-scalars, you might use

8
2 INTERFACE, COMPONENTS AND DATA STRUCTURE 2.6 Data type and format

global N=r(N)
for instance. To use them later on you have to use the dollar-symbol before the name:
display "The total number of observations is: $N"

Locals
The use of locals is essentially identical to the use of globals with a small difference in the way you
invoke the variable. Before the name, you have to write a single opening apostrophe (ASCII symbol
96) and right after the name a simple apostrophe (ASCII symbol 39). The following example defines
first a numeric local containing the age of a person and then a string local with the name. In the third
line, a small text with the information will be displayed:

local age=39
local name="Peter"
di "The age of name is age years"

This will produce the text The age of Peter is 39 years.

Scalars
The scalars can only take numerical values and the way to invoke them is at a first glance easier,
since you just write their name. However, this might lead to confusions with the variable names, a
potentially problematic issue.

2.6 Data type and format

The variable format is an important issue in Stata like in all application dealing with data of this
type. The goal is to use the least possible amount of memory for the variables. A particular issue

in Stata is that we have to distinguish between the actual format of the variable when it is saved
and the output format, thus the way values are displayed. Hence, the data type refers to the way the
information is stored in the database and data format is the way the information is displayed.

2.6.1 Data type

Data can be either numerical or string, which is probably the most basic distinction.

Numerical types
The following table gives a short overview of the different numerical data types in Stata :
Storage type Minimum Maximum Closest to 0 bytes
(but 6= 0)
byte -127 100 1 1
int -32767 32740 1 2
long -2147483647 2147483620 1 4
float 1.701 1038 1.701 1038 1038 4
double 8.988 10307 8.988 10307 10323 8

9
2.6 Data type and format 2 INTERFACE, COMPONENTS AND DATA STRUCTURE

String types
Regarding string variables it becomes very easy, since they are simply going from str1 being one
letter up to str244 containing a maximum of 244 letters. The format is always at least as high as the
largest entry in the database. Assume that in all but one observation, we do not have a string and
in the one observation, we have 66 symbols. In this case, 66 bytes of memory are needed for every
observation, hence it is worthy to avoid unnecessary string variables in a database.

How to choose the best type? This relative technical information on the storage type and used
bytes might frighten a bit, but in practice this is hardly a topic, since there is a wonderful command
called compress which analyzes every variable and converts it to the best storage format.

2.6.2 compress - Optimize the memory use of your database

The command compress is a very powerful tool to reduce the size of your database to the minimum
needed. The command simply analyzes each variable and sets its format to the best fitting.

Hint 1. When you work with large datasets, combine several of them and create new variables, it
is very recommended to include the command compress just before saving your database in order to
avoid wasting memory for nothing.

2.6.3 Further compressing by using a ZIP-file

Especially when working with micro data the databases can become very large. If you have to send
them to other people, you might want to consider putting them into a zip-file, which you can do under
Windows in the Explorer for instance. The size reduction you can achieve depends a lot on your
data, but it is not uncommon to reduce the size by up to 90%!! Such high compressions are especially
possible when you have a lot of missing values or even more when many long string variables are
empty.

2.6.4 Display format

Besides the format in which a value is stored on your disk, Stata has a display format, which can
differ from the first. The percentage sign is used to declare formats. For numerical values, there are
mainly two types: the general (g) and the fixed (f) format. The general format depends on the number,
while the fixed format has for instance a fixed number of decimals, no matter what the number is.
For instance:

Complete number %6.2g %6.2f

99.1234 99 99.12
.0159 .016 0.02
500 500 500.00
.001 .001 0.00

Stata supports also date and time formats. See the help file for details.
Knowing these format types can be very useful in commands like est table, estout or tabout

10
2 INTERFACE, COMPONENTS AND DATA STRUCTURE 2.7 Missing values

2.6.5 format - Change the display format of a variable

You can also assign to a variable a certain display format. For instance, if you wish to see the income
always with two decimals, you can use the command
format income %6.2f

2.7 Missing values

Like in any other statistical package, missing values have a special code in Stata . The normal
coding of a missing variable is a dot (.), for instance you might want to set a value to missing
replace age=. if age<0
if the indicated value is not plausible. Alternatively, you can also condition your commands on missing
or non-missing values2 :
tab education if age==.
This example would give you the frequency tables of the variable education for all people whose age
we do not know.
In some cases it can be interesting to distinguish different cases of missing values (e.g. refusal, not
applicable, etc), where you can extend the dot by a letter. Imagine that in the original data refusals
are coded as -999 and not applicable as -998. We can then use for instance the recode command to
set different types of missing values:
recode myvariable (-999=.a)(-998=.b)
You can even label these types of missing values like any other value label. When running econo-
metric analyses, these missing values are considered like the standard missing values, hence they are
not taken into account.

2
Normally non-missing values are excluded by default for obvious reasons

11
 3 BASIC COMMANDS

3 Basic commands
Before starting with commands allowing to perform econometric analysis, it is important to un-

derstand the generic Stata -command and to know several relevant commands to customize the

Stata environment and to load and save the data.

3.1 do-files
Commands should be written in a do-file, even though you could also write them directly in the com-
mand line and then pressing ENTER. The problem by doing so would be that you could hardly repro-

duce what you did before. Using a do-file is like writing a programming script, Stata goes through

it and performs each command you write in the do-file. An important point is that Stata stops the
execution on error, meaning that you can be sure that everything went well if the do file is executed
until the end. To start a new do-file simply click on the button for the do-file editor (see figure 1) and
start writing.

To execute the file you can click on the corresponding symbol ( ) you find in the editor, go through
the menu or click [CTRL]+[D] on your keyboard, which is obviously the most convenient way to do it.

Hint 2. The combination of pressing first [CTRL]+[S] and then [CTRL]+[D] might help you to save
a lot of time and nerves. The first simply saves your do-file and the second executes. By doing it

always like that, you can be sure your do-file is saved on the hard drive and even if Stata crashes
.

once, you will have your file

3.1.1 Line breaks

If you have very long commands that do not fit on the screen and you do not want them to be in one
line, you can generate a line break with three slashes ///:

mycommand depvar indepvar1 indepvar2 indepvar3 indepvar4 indepvar5 indepvar6 ///

indepvar7 indepvar8 indepvar9 indepvar10

alternatively, you can temporarily activate the active line break, meaning that you have to break the
lines manually by the semicolon symbol (;) like in many other programming languages. To activate
the manual line break type
# delimit ;
and to come back to the normal line break
# delimit cr;

3.1.2 Comments

You can and should comment your code which can be done with a double slash // or a star *.
The double slash works at the beginning of the line or in the middle, while the star only works to
declare a whole line as comment. For longer comments you can use the combination /* my comment */

12
3 BASIC COMMANDS 3.2 Prepare Stata for your analysis

* This is a comment on the whole line

reg y x //here I can comment the command

/* Here you can write a

comment on several
lines */
est store myreg //and yet another in-line comment

Hint 3. This is probably the most common hint: do comment your code as much and as clear as
possible! This is not only useful when working with colleagues, but also when coming back to your
do-file after a while. It might be difficult to understand your own code when it is not commented!!

3.2 Prepare Stata for your analysis

In a do-file the first things to do, even before loading the data, is to prepare the memory and the

Stata environment for it. Typically, we want to clean the memory before starting the analysis in
order to have a common starting point every time the do-file is used. Using clear all erases all

Stata -relevant information from your memory. Once you did that you can change the memory

size at the disposal of Stata by using the command set mem 100m to give Stata 100 megabytes of

memory. Generally, it is advised to provide Stata with sufficient memory, at least 100 megabytes
more than the size of the database.
A next command you might want to use is the cd to change the working directory. The working

directory is the place Stata looks for databases if you do not indicate the whole path. Hence, it
might be useful to change the working directory to the place where your data is stored (or where you
want to store the output). This helps you to avoid writing every time the whole path. For instance,
if you want to change the working directory to the folder data in your C drive you write:
cd C:/data
An alternative to change the directory is to define variables containing the path of your data and to
use then these variables. Both methods have the enormous advantage over writing the whole path
every time you load or save something, that when running the file on a different computer, you simply
adapt one line of your code and you do not have to change it all over in your file.
Another command I suggest to use in the beginning of the do-file is
set more off
which avoids breaks in the execution of the do-file when the output of one single command is longer

than the screen size. Normally Stata stops allowing you to see all the results, but this is not very
convenient when working with do-files. Rather than indicating this command in all do-files, you can
execute once the command
set more off, permanently
in order to change it permanently (this changes the default value from on to off).
Hence, the head of a do-file could look like:

1: clear all

13
3.3 Loading and saving data 3 BASIC COMMANDS

2: set more off

3: set mem 250m
4: cd C://data
5: global source="C:/data/analysis1"
6: global output="C:/data/analysis1/results"

where line one deletes all from your memory, line 2 disables the break in the output, line 3 increases
or decreases the memory to 250 megabytes, line 4 changes the working directory to C://data and the
lines 5 and 6 define two variables (source and output) containing the information of the source and
output folder you will use.

3.3 Loading and saving data

Loading data in Stata means to load the database into the RAM-memory of your computer. Only
one database can be active at a time. The command is extremely simple
use C:/data/analysis1/mydata
to load for instance the database mydata.dta located in folder C://data/analysis1. Note that with
the global variable defined in the short example before, we could simply write
use $source/mydata
since in the variable source we stored the whole path. Yet another possibility is to use only the rela-
tive path from the current working directory. As we are in the working directory C://data (indicated
in line 4 of the example before) we can simply write
use analysis1/mydata

Hint 4. Always load and save databases with a do-file to avoid overwriting a database or to work on
the wrong one.

Hint 5. Always save the database in a do-file with another name than the database you open at the
beginning. Otherwise, you could not run the do-file twice, since the changes would now be in the
initial database.

If needed, you can only load a part of a database, say variables x1 x2 and x3. The command becomes
simply
use x1 x2 x3 using $source/mydata
Moreover, you can use conditions on the data, for instance an if (see section 3.6)

3.4 The generic Stata command

All Stata commands have a common structure:
name varlist conditions, options

The command starts with the name of the command in order to tell Stata what to do. The name is
followed by the list of variables you would like to use in the command. Depending on the command, this
list might include dependent and independent variables individually or together. The list of variables
is followed by the sample conditions allowing you to perform the command only on a subsample of your

14
3 BASIC COMMANDS 3.5 Types of commands

data. All you find after the comma are options, some of them might be mandatory. Understanding

the logic of a Stata command is crucial and it allows to fully understand the Stata -help I present
in section 2.3

3.5 Types of commands

Stata commands are classified into different classes, depending on their content. Knowing about the
exact definition of the classes is not absolutely needed, but it might be important to know the two
most common for posterior use of stored results. The two main classes are:

r-class This is the most general class, including most of the commands for descriptive
statistics such as summarize or tabstat. Results of this type of commands are
stored in r(). To display the whole set of results, type return list after the
command. Note that these results remain active until the next r-class command.
e-class The e-class commands are normally econometric estimations such as regress. The
results are stored in e() and can be displayed typing ereturn list. As for the
r-class, the values in e() are stored until the next e-class command is executed.

Example
For instance, if you would like to use the mean of a variable as a local you could use the following
code:

summarize income
local meanincome=r(mean)
di The mean income is: income $

where the first line is the r-class command summarize displays a series of summary statistics such
as the mean, the standard deviation and others. The second line recovers the stored data3 in r()
and saves it in the local meanincome. Finally, in the third line of this not very practical example, a
short text indicating the mean income is displayed. The following example recovers the adjusted R2
statistic from a simple OLS regression and stores it in a scalar:

regress income education experience

scalar rsquared=e(r2_a)

Besides the e-class and r-class there exist also n-class and s-class commands, however, they are used
very rarely. An easy way to find out if a command is r-class or e-class is to look in the help file how
the results are stored. Normally towards the end of the help file a list of stored results are indicated.

3.6 Conditions

Generally, Stata commands can be conditioned on a subsample of the dataset. The condition might
take different forms and is normally introduced after the varlist and before the comma separating
the command from its options. The most important condition is the if-condition starting simply with
the word if followed by a logical condition. For instance, if you want to run a regression only for
women and you have a dummy variable female taking the value of 1 if the person is a woman, the
simple regression command becomes
3
To see all the stored values, type return list just after the summarize command

15
3.7 Relational and logical operators 3 BASIC COMMANDS

regress y x1 x2 x3 if female==1
A second way to limit the sample is the in-condition, where you can run the command for instance
for the first 100 observations (regress y x1 x2 x3 in 1/100).

Hint 6. The in-condition might be very helpful when you write a do-file containing a very computation
intensive command and you would like to run the do-file in order to check if it works. Limiting the
command to a small number of observations avoids loosing time when checking the do-file.

Hint 7. A nice and sometimes more elegant alternative to the if condition is to multiply your
expression with a logical statement. Imagine you want to compute the value of a variable only for
|Z| < 1, you can use the following command
gen kernel=(1-abs(Z)) * (abs(Z)<1)
where (abs(Z)<1) is a logical element returning 1 if the condition is satisfied and 0 otherwise, hence
the variable kernel will take the value 0 (not missing!) whenever the condition is not satisfied.

3.7 Relational and logical operators

Like in every programming language, logic operators are crucial for a correct programming.

+ Addition > greater than

- Subtraction >= greater or equal than
* Multiplication < smaller than
/ Division <= smaller or equal than
^ Power == equal than
& and != not equal
| or

Stata is very liberal in the use of logical operators and allows much more than other languages4 .
For instance = and == are not supposed to mean the same thing: gen v1=1 generates a variable
containing the value 1 and ttest v1==1 performs a t-Test if the variable v1 has mean 1. However, you

could also write ttest v1=1 and Stata automatically considers = as ==. This is sometimes
useful, however I strongly suggest the correct distinction between the two equal signs, since some
commands do not support both notations!

3.8 Placeholders and selection of multiple variables

Placeholders can become very important when you work a lot with Stata . Imagine the quite
academic case in which you have a database with one dependent variable y and 25 independent
variables named x1 to x25. Without placeholder you would have to write each x-variable individually
in a command, however, using the * placeholder you can shortcut your command by a lot:
reg y x*
would be a simple OLS regression (see regress) of the variable y on all x-variables. Now assume that
for some reason you only want to use x12 to x18. In this case the * placeholder is not very useful, but
4
One could also say that Stata is somewhat sloppy with the logical operators

16
3 BASIC COMMANDS 3.9 Matrices

you can use the - sign:

reg y x12-x18
will include all the variables starting from x12 up to x18. However, this gives you only the correct
result if the variable order is alphabetical or if at least only the desired variables are in between x12 up

to x18. Hence, the - sign simply tells Stata to take all variables between the two indicated variables.

Hint 8. The order of variables in Stata has generally no specific logic and new variables are simply
added in the end of the table. Using the command order you can order the variables according to
your needs and aorder can be used to order all variables alphabetically.

3.9 Matrices

Stata has two matrix systems, one that is implemented directly in the Stata environment and
Mata, which is to some extend a language apart. In this section, I deal only with the standard matrix

package of Stata . It has to be noticed that for matrix algebra Stata is probably not the best
software, so do not expect the quality of MATLAB or R when dealing with matrices.

3.9.1 matrix define - Generic command for matrices

Even though it is rather unusual to input a matrix by hand, in some cases it might be necessary (e.g.
for very simple user written command).

matrix input A=(1,2,3\4,5,6)

defines
" #
1 2 3
A=
4 5 6

Combining matrices
You can also create a matrix as a function of other matrices (as long as the dimensions match):
matrix define C = A + B
where A and B are two matrices of the same dimensions. The following table provides an overview of
matrix operations available:

Operator Symbol
transpose
negation -
Kronecker product #
division by scalar /
multiplication *
subtraction -
addition +
column join ,
row join \

17
3.9 Matrices 3 BASIC COMMANDS

3.9.2 Some basic matrix functions

Some basic matrix functions include the following:

Function Description
colsof(M) Returns the number of columns in matrix M
rowsof(M) Returns the number of rows in matrix M
issymmetric(M) Returns 1 if M is symmetric, otherwise 0
det(M) Returns the determinant of matrix M
trace(M) Returns the trace of matrix M
I(n) Returns the entity matrix of dimension n
inv(M) Returns the inverse matrix of M

3.9.3 matrix list - Displaying matrices

To display matrix mymat you simply use
matrix list mymat

3.9.4 matrix rownames/colnames - Renaming matrix columns and rows

To rename the column and row names you can use the following commands:
matrix colnames A="Column 1" "Column 2" "Column3"
for the columns, and
matrix rownames A=A B C
for rows. As you can see, whenever the name consists only of one word, you do not need the quotes.
3.9.5 matrix get - Getting system matrices
Many estimation commands store important information in matrices, for instance the estimated coeffi-
cients of a regression are generally stored in the matrix e(b) and the corresponding variance-covariance
matrix in e(V). For further use of this information it is needed or at least recommended to extract
these matrices to a user-defined matrix with the command
matrix b = get( b)
for the betas for instance. Learn more about extracting results from estimations in section 7.6
3.9.6 mkmat - Data to Matrix conversion
It might be necessary in some cases to convert your data into a matrix or vice versa. This can be done
very easily with
mkmat x1 x2 x3 x4
to convert the variables x1,x2,x3 and x4 into four column vectors with the same names or with
mkmat x1 x2 x3 x4, matrix(mymat)
to convert the four variables into a x 4 matrix called mymat.
3.9.7 svmat - Matrix to Data conversion
To perform the reverse operation, you use
svmat mymat

18
3 BASIC COMMANDS 3.10 Factor variables and time series operators

which will create four new variables named mymat1, mymat2, mymat3 and mymat4 respectively.
3.10 Factor variables and time series operators
Many times, you do not want to enter variables in regressions just like they are, but in some specific
form. This is when the so-called factor variables enter the game. Basically, we are talking about

prefixes for the variables to tell Stata what to do with it.

3.10.1 i. - Categorical variables in the regression

Imagine you have a categorical variable and would like to include each category as a dummy variables
(for category fixed effects). Defining for each outcome a dummy variable (using tabulate might be
one solution, but probably not the most efficient one. You can simply write i. before the variable

and Stata will understand that all but one of these values should be used as dummy variables. This
is not simply saving time in coding dummy variables but also reducing memory needs by a lot.

3.10.2 l. - Lagged variables

In time series or panel analysis, you might want to use lagged variables. This is easily done with the
prefix l., for instance
reg y l.y x1 x2
will perform an OLS estimation (regress) of y on the lagged value of y and two control variables x1
and x2 . If you need more than one lag you can different choices:

Syntax Variables Description

ll.x xt2 Double lagged variable used
l(1/2).x xt1 ,xt2 Lagged and double lagged variables
l(0/5).x xt ,xt1 , ... xt5 From non lagged to 5 periods lagged variables

3.10.3 f. - Forwarded variables

Just like lagged variables, it might be of need to use forwarded variables using the f. prefix. It works
exactly the same way as l..
3.10.4 d. - Difference
Just like lagged variables, you can create difference variables, for instance yt = yt yt1 is created
by
gen dy=d.y
. Higher interval differences can be created with the same logic seen for l..
3.10.5 c. - Continuous variables

Sometimes you have to specifically tell Stata that a variable should be considered to be continuous,
e.g. when you use interaction terms. However, normally this does not have any impact:
reg y c.x1
is exactly the same as
reg y x1

19
3.11 Loops and programming conditions 3 BASIC COMMANDS

3.10.6 # - Interaction terms

Instead of defining interaction terms in a new variable using the product of the two in case of having
continuous variables or combinations for categorical data, you can use the symbol #. However simply
writing
reg y x z x#z
works only if x and z are categorical variables. In this case, all possible combinations are included as

a dummy variable. To be clear about what Stata does, I would however always write
reg y x z i.x#i.z
which is not absolutely needed but recommendable. This is especially the case because the #-symbol
can also be used to create squared values and continues interaction terms. Assume now that both, x
and z are continues and you would like to estimate the model

y = + 1 x + 2 z + 3 xz + 

This can be done by writing

reg y x z c.x#c.z

where the c. tells Stata that you would like to create an interaction by multiplication of two
variables and not with all possible combination dummies.

3.11 Loops and programming conditions

A very important notion in any programming language or syntax are the loops and conditions. It is
essential to understand them and to use them wherever it is possible. To understand this section on
conditions and loops, it is necessary to be familiar with the data type local explained in section 2.5

Hint 9. Using loops instead of writing several times the same code with only slight differences helps
reducing errors. Generally the shorter your code is, the better you did your job!

3.11.1 if/else - if/else/elseif clause

The most basic clause is the if clause. In Stata it is implemented in a rather standard way:
local random=uniform()
if(random>0.99){
display "This is rather an unlikely event"
}
else if(random<0.01){
display "This is also an unlikely event"
}
else{
display "This is pretty normal"
}

20
3 BASIC COMMANDS 3.11 Loops and programming conditions

The example first defines a local variable drawn form a uniform distribution. Then it displays a text
in function of the value. The first if condition is TRUE when the value is higher than 0.99, the second
element is a else if (with a space) condition being TRUE when the value is below 0.01 and the last
element is the else-statement if the two conditions before returned FALSE.

3.11.2 while - while loop

The while condition is implemented in a very similar way to the if statement:
local i=1
while(i<10){
di "i"
local i=i+1
}

This example simply displays all numbers from 1 to 10 on the output screen.

3.11.3 foreach - Looping through elements

Things are getting more interesting with the foreach statement, which loops through a predefined set
of elements. For instance you can loop through a local variable:
local text="Hello Stata users"
foreach word of local text{
di "word"
}

or

local myarray="1 2 5 9 7 15"

foreach i of local myarray{
di "i"
}

The example takes every element of the local, stores it temporarily in the new local called word and
uses it in the commands. You can also loop through variables:

foreach var of varlist x*{

sum var
}

or

foreach var of varlist x1 x2 x3{

sum var
}

where the loop performs the summarize command for each variable starting with the letter x. The
general syntax of foreach is

foreach runner of arraytype array

21
3.11 Loops and programming conditions 3 BASIC COMMANDS

where the italic elements can be changed. The runner refers to the local variable that changes in
every loop, the arraytype indicates what kind of array the loop should go through (for instance local,
global, varlist ) and finally the array is the array containing the elements to loop through.

3.11.4 forvalues - Looping through numerical values

The last loop type is the forvalues which is contrary to the foreach limited to numerical values.
Here are two examples:
forvalues i=1/4{
display "i"
}
forvalues i=4(0.5)5{
display "i"
}

where the first example displays the series 1,2,3,4 and the second 4,4.5,5

22
4 SYSTEM COMMANDS

4 System commands
4.1 System commands

There are some Stata commands, which you will not use very frequently, however, they might be of
interest in some situations. The system commands are related neither directly to the data nor to the

econometric analysis, but they allow you to adapt Stata to your needs or update it.

4.1.1 exit - Leaving the do-file or Stata

The command exit has mainly two purposes. When running a do-file you might want to stop it at a
given point (breaking point). By including the command exit the do-file will stop executing without

an error message. If you use exit in the command line, Stata will be closed if you do not have
unsaved data in your memory. If you do have unsaved data, then you can add the option clear in

order to force Stata to shut down without saving the data.

4.1.2 query - Displaying Stata options

Stata has many default settings you can change temporarily or permanently. To display all the
settings the user can change just type query or query topic where topic is the topic for which you
would like to the see the settings. For instance, query memory displays all the settings related to the
memory. To change the settings, use the command set.

4.1.3 set - Change settings

Use set to change the default settings of Stata , with or without the option permanently to keep

the new settings for all future uses of Stata . For instance, to change the memory allocated to

Stata to 300mb you can simply write
set mem 300m

Hint 10. The settings are saved locally on the machine, thus changing them permanently on a server

version of Stata might be useless.

4.1.4 about - Getting information of the version and license

By typing about Stata displays information on your license and the version of your Stata . Addi-
tionally information on the available memory of the computer is given.

4.1.5 update - Update your Stata

Like every software package Stata needs to be updated from time to time. By typing update all

Stata will update all possible files where an update is available. Updates are made on executables,
utilities and especially ado-files. If you simply wish to see if there are updates available type update

query. Normally Stata performs checks for updates every week, you can get the information by
typing query update.

23
4.1 System commands 4 SYSTEM COMMANDS

Hint 11. Updating a server based Stata must be done on the server and not on the local machine

4.1.6 findit - Find ado-files online

findit is a very powerful command permitting you to search for ado-files (commands) on the internet.
If you know the name of an ado-file then type it, otherwise use keywords to describe the command

you are looking for. The search is performed in well-known Stata -repositories and by clicking on

install Stata will automatically download and install the package. For instance, knowing that
the package nldecompose performs Oaxaca decomposition for non-linear models, you can find and
download using
findit nldecompose
Otherwise, if you want to find this package but you do not know its name, just write
findit oaxaca
and a list of potentially interesting packages will be displayed.

4.1.7 search - Search ado-files online

see findit

4.1.8 help - Display help

The Stata -help is very useful for new and experienced users. I guess nobody knows all the commands

and especially the options of Stata -commands by heart, therefore the help is your best friend. Just
type
help commandname
to display the help file of the command commandname. See section 2.3 for more information on the
Stata-Help.

24
5 DATA HANDLING

5 Data handling
5.1 Variable manipulation

In Stata it is very easy to create, label and rename variables allowing you to understand your data
afterwards much better. In this section, I present some useful commands.

5.1.1 generate - Generate a new variable

To create a new variable, you use generally the command generate (gen). For instance you can
generate a variable by summing up two other variables
gen v3 = v1 + v2
To generate variables that contain summary statistics of the sample, for instance the sample mean,
you have to use oftentimes egenerate (egen)

Generating dummy variables

To generate dummy variables you can use the generate command together with some conditions,
for instance:

gen teenager=0
replace teenager=1 if age>10 & age <=20

generates a dummy variables taking the value of 1 if a persons age is bigger than 10 and smaller or
equal 20. A simpler way to define this variable is just by including the condition in the first statement

gen teenager=(age>10 & age<=20)

5.1.2 replace - Replacing a variable

When a variable already exists you cannot use the command generate, but you might use replace,
which works essentially the same way. For instance, if you want to replace negative income of people
by a zero (or missing value) you could do
replace income=0 if income<0
or
replace income=. if income<0

5.1.3 egenerate - Generate a new variable with summary statistics

The command egenerate or simply egen is similar to generate but is used normally to generate a
new variable out of statistical operations. For instance, if you want to compute the average of several
variables, lets say x1 x2 x3, you could use
gen xmean=(x1+x2+x3)/3
however this gives you a missing value if in any of the three variables you have a missing value, which
might be exactly what you are looking for. If this is not the case and you would like to get a value
even if one variable has missing values, then you should use

25
5.1 Variable manipulation 5 DATA HANDLING

egen xmean=rowmean(x1 x2 x3)

Experience working with Stata will help you to understand in which situation to use egen rather
than gen.
5.1.4 recode - Recoding a variable
Recode is a powerful command to change the values of a categorical variable in particular. Imagine
we get a database where the variable indicating the gender is coded as follows: 1=man and 2=woman
and 3=unknown. We would like to have a variable where man=0 and woman=1 and .=unknown (see
2.7 for a discussion of missing values). We could use three times the replace command:
replace gender=0 if gender==1
replace gender=1 if gender==2
replace gender=. if gender==3

This is probably even a good solution in this case, but imagine you want to change the values of a
categorical variable with 10 values. This is when recode becomes more efficient. To make the before
mentioned change, the command would be
recode gender (3=.)(2=1)(1=0)

Actually, the command recode can much more than that:

recode x1 x2 x3 (-999 -998 -997 = .)(1 2 3 = 1)(5 6 7 = 2)
recodes the three variable x1 , x2 and x3 simultaneously and put the negative values -999, -998 and
-997 to missing, 1,2 and 3 to 1 and 4, 5 and 6 to 2.
It is also possible to generate new variables (in this example x new and y new) out of the recoding and
holding the original variables at their initial values:
recode x y (1=0)(2=1), gen(x new y new)
Note that the order of the varlist in the option gen must correspond to the order in the main
varlist of the command.

5.1.5 label - Label your variables

Stata know two types of labels: value labels and variable labels.

Variable labels The variable label is simply a text describing a variable in general. It appears in
the variable overview and helps you to understand the content of a variable. To define or redefine
(overwrite the old) a variable label simply write
label variable income "Income of the person in US$"
to label the variable income with the text between the two quotes.

Hint 12. Try to keep variable names short and describe the content in the variable label. For
instance if you have a variable with the log annual income per capita avoid variable names like
log annual income per capita and use rather lincpc with a corresponding variable label Log annual
income per capita

26
5 DATA HANDLING 5.1 Variable manipulation

Value labels

Besides the variable label Stata is also capable to assign value label, meaning that each value of
a variable is labeled and in the database the label is shown instead of the actual value. This is useful
in the case of categorical data. Let us illustrate the value labels with a small example. Imagine a
variable lstatus taking three values: 1 for people still in education, 2 for people in the active labor
force and 3 for retired people. Using for instance the command tab we could display the frequencies,
but the table would not be very self-explaining:

lstatus | Freq. Percent Cum.

------------+-----------------------------------
1 | 357 35.70 35.70
2 | 324 32.40 68.10
3 | 319 31.90 100.00
------------+-----------------------------------

Therefore it is useful to label the values. The first step is to define the label. This would be
label define lstatlabel 1 "In education" 2 "Labor force" 3 "Retired"
where we name the label lstatlabel and then indicate the value and in quotes the corresponding label.
This command only saves the label, we still have to assign it to the variable.
To assign the variable we use again the command label:
label value lstatus lstatlabel
Now the table with the frequencies becomes:

lstatus | Freq. Percent Cum.

-------------+-----------------------------------
In education | 357 35.70 35.70
Labor force | 324 32.40 68.10
Retired | 319 31.90 100.00
-------------+-----------------------------------

which looks already much nicer.

Hint 13. It is possible to assign the same label to different variables, for instance you can define a
label called dummy and assign it to all the dichotomous variables. Let us say we want to label the
variables d1, d2 and d3 with the label dummy we would use label value d1 d2 d3 dummy

Once the variables are labeled, it is also possible to automatically extract these labels. For more
details see section 11.2.6.

5.1.6 rename - Rename a variable

Variables in micro data have oftentimes quite reasonable but not very intuitive names, thus it is

important to rename variables in order to have intuitive names. Renaming in Stata is extremely
easy, let us say we want to rename the variable st98q01 in indigenous. The command is simply given
by:
rename st98q01 indigenous

27
5.2 Describing and sorting the data 5 DATA HANDLING

Note that this does not alter the content of the variable nor any kind of labels.
5.1.7 drop - Deleting variables
To delete one or various variables use
drop id gender age
The example above simply drops the three variables id, gender and age.
5.2 Describing and sorting the data

5.2.1 describe - Describe the dataset

describe is used to display a description of the dataset, not descriptive statistics of the data! By
typing desc without a variable all the variables will be described, while the inclusion of a variable list
will limit the description to those variables. Here is the example for only one variable:
Contains data
obs: 1,000
vars: 1
size: 8,000 (99.9% of memory free)
-----------------------------------------------------------------
storage display value
variable name type format label variable label
-----------------------------------------------------------------
lstatus float %12.0g lstatlabel Labor market status
-----------------------------------------------------------------

where Stata indicates you first how many observations are available in the dataset (independent
of missing values!) and how many variables you have. The value size refers to the memory use of
the database (in this case almost nothing is used). After this information on the whole dataset, the
information of each variable is displayed. First the name, then the format and the display format,
followed by the assigned value label and the variable label.

5.2.2 codebook - Display the codebook

codebook is similar to the variable related part of describe, however, with much more details. Besides
the format and some information on the labels, the units, the range, frequencies etc. are displayed.
5.2.3 sort - Sorting the data
You can easily sort your data according to one or more variables with the command sort. For instance

sort IDhousehold IDindividual

would make the following change in the database:
IDhousehold IDindividual Age IDhousehold IDindividual Age
2 1 56 1 1 33
1 2 34 becomes 1 2 34
1 1 33 2 1 56
2 2 61 2 2 61

5.2.4 gsort - Sorting the data

gsort works basically like sort, but you can sort the data in the opposite direction by including a
minus sign (-):
gsort -year

28
5 DATA HANDLING 5.3 Joining several databases

will sort the data according the variable year starting with the highest value.
5.2.5 order - Sorting the variables
While sort sorts your data vertically, the command order allows you to sort the data horizontally,
meaning you can change the order of the variables. Normally this is not a very important feature, but
there are situations when it might be necessary (e.g. to have the identifier of the observations at the
beginning for easier use or to use the placeholder symbol - in estimation commands). By writing
order id name country
the variables id, name and country will be put in the beginning of the dataset, while all the remaining
variables remain unchanged!
5.2.6 aorder - Sorting the variables alphabetically
Like the command order, aorder allows you to change the order of the variables in your dataset,
however, in the alphabetical order. By simply writing
aorder
without any varlist , all the variables will be ordered alphabetically, where special symbols like
underlines ( ) come first, followed by capital letters and lower case letters.
If you indicate a varlist like
aorder ID name country
then the variables id, name and country will come first, followed by all the remaining variables in
alphabetical order.
5.3 Joining several databases
Especially when working with micro data, it is oftentimes needed to merge several databases into one.

Stata is very efficient in this kind of data handling. Two main ways of merging/appending two or
more datasets into one are to be considered. The first situation is when we have two datasets with
different variables for the same individual, firm or household. In this case we use the command merge.
The second situation is when we have two databases with the same variables but not for the same
people, thus we would like to add one to the other and we use append

5.3.1 merge - Merging/appending two databases (horizontally)

Before starting to describe the command merge there must be made a difference between the version
11 and higher and the previous versions, since the command changed slightly. Both versions perform
the same action in the end, but the syntax became more explicit in the newer version. In what follows
I refer to both versions.

Prepare the data for merge

Before being able to merge we have to prepare the data. A necessary condition is that in both
databases we have an identifier of the observation, which might consist of a single variable (e.g.
individual ID number), or a series of variables (e.g. family ID number + intra family rank). In both
databases the variables must be coded the same way and should have the same format. We have to
sort the databases by these identifiers first. Since only one database can be active at a time, there
is always the master -database (the active one) and the using-database, the one we integrate in the
master.

29
5.3 Joining several databases 5 DATA HANDLING

Merging under version 10 or lower

In version 10 or lower of Stata the command to use when ready is:
merge id1 id2 using myusingdata
where id1 and id2 are two variables that identify the observation and that are present in both databases.

The word using is used to indicate Stata that the following part will be the database(s) to be joined.
The last part refers to our using-data called myusingdata in this example.

Merging under version 11 and higher

Basically the syntax in the newer version is the same, with the small exception that we have to
indicate now where we can find multiple entries. Imagine that you have individuals in the master
data and household information in the using data, thus the same household data must be added to

all members of the same household. In the earlier versions Stata just tried to understand it at the
risk to misunderstand it. Now, you must indicate it here. The example would then be for instance:
merge [n:1] householdID using householdinfo
In this example the master data contains the individuals and the using-data the household informa-
tion. The term n:1 means that multiple observations in the master -data are to be merged with one
observation in the using-data. If you have individual data on both sides, you would use 1:1 and in
case of merging multiple to multiple the term becomes n:n.

Merging result

Once the merge of the databases performed, Stata displays the results and stores information in
the new variable merge by default5 . The generated variable is coded in the following way:

1 the observation was only found in the master data (no merge)
2 the observation was only found in the using data (no merge)
3 merge successful, observation found in both
4 observation found in both, missing values updates
5 observation found in both, conflict in some variables

5.3.2 append - Appending/merging two databases (vertically)

Appending data is generally easier than merging, since you dont need identifiers. The command
append simply adds the rows using-database (see merge) to the current database. In case of having
the same variable in both databases (the normal case), the data is just added. If in one of the two
databases the variable was not present, the column is added and in the database where it was not
present, missing values are added. The command is simply
append using myusingdata
where myusingdata is the database to add to the active database.
In case of not having the same format in the variables (same variable in the two databases), you should

use the option force in order to enable Stata to convert the formats where necessary. This might
be especially needed for string variables, since they are generally defined to be as long as the longest
5
with the option generate(myvarname) you can change merge to myvarname

30
5 DATA HANDLING 5.4 Changing structure of a database (panel data)

entry (to save memory) and these values might differ from one the other database.
5.4 Changing structure of a database (panel data)

5.4.1 reshape - Reshape your panel data

There are two ways to save panel data in a database, the long and the wide form. Depending on
the amount of variables that are constant over time (e.g. gender) the one or the other form are more
suitable.

long:
id year income wide:
id income2000 income2005
1 2000 5.6
1 2005 7.2 1 5.6 7.2
2 2000 8.3 2 8.3 9.1
2 2005 9.1

The command reshape can be used to change your data easily from one to the other format. If you
want to reshape from wide to long, then use
reshape long income, i(ID)
where a new variable j will be created with the years. If you want to call it years right away, you
can include additionally the option j(year).
To get back to the wide form, write
reshape wide income,i(id) j(year)

31
 6 SUMMARY STATISTICS AND GRAPHICS

6 Summary statistics and graphics

6.1 Descriptive statistics

6.1.1 summarize - Summary statistics

The command summarize or just sum provides you with the essential summary statistics of the variable.
Without options or varlist the command provides the number of non-missing observations, the
average, the standard deviation and the minimum and maximum value. By using the option detail
additional summary statistics like the percentiles, skewness and kurtosis are displayed. summarize is
a r-class function and stores all the displayed valued in r(name) which you can consult typing return
list

6.1.2 tabstat - More flexible summary statistics

An alternative and oftentimes more flexible way to get summary statistics is the command tabstat.
Its basic syntax is
tabstat varlist [if] [in],[ stats(statisitics) by(byvar)]
where the varlist contains all variables you would like to analyze. The if and in conditions are
straightforward. The true utility of the tabstat command comes with the two non-mandatory options
stats(statisitics) and by(byvar). The first allows you to specify the list of statistics you would like
to compute. The most common options include the mean (mean), median (median or p50), standard
deviation (sd), maximum (max), minimum (min) or the number of observations (n or count). The
whole list of available statistics can be found in the helpfile. The by(byvar) allows you to compute
these statistics by subpopulations defined by the variable byvar.
Let us consider two examples. First,
tabstat income, stats(mean median min max) by(region)
will give you the mean, median, minimum and maximum of the variable income for each region as

defined in the variable region. For each region, Stata will display a row with the four specified
statistics. Adding the option column(variable) would change the direction of the table, putting the
variable in the columns and the statistics in the rows.
tabstat income wage consumption, stats(mean median min max) column(stats)
will do the same but instead of displaying a row for each region, you will get a row for each variable

specified. You can also combine the two! Without the option column(stats) Stata will display the
variables in the columns and the statistics in the row.

6.1.3 tabulate - One- and twoway tables of frequencies

While summarize is very useful for continuous variables, tabulate is the most useful command for
categorical data. It provides you with a table of frequencies of each possible outcome. You can also
create a 2-way table where every possible combination of two variables are presented. For instance
tab gender indigenous, mi
gives something like
| indigenous
gender | 0 1 . | Total
-----------+---------------------------------+----------
0 | 2,608 669 54 | 3,331
1 | 2,627 635 74 | 3,336
. | 48 12 2 | 62

32
6 SUMMARY STATISTICS AND GRAPHICS 6.2 Graphs and plots

-----------+---------------------------------+----------
Total | 5,283 1,316 130 | 6,729

where the option mi (missing) indicates that you also want the missing values to be considered as
a category. The total number of observations is 6729, 2608 non-indigenous men and for instance 12
indigenous people whos gender we dont know. About 2 people we do not have any information.

Generate dummies of each category

The tabulate command has an interesting side-feature that allows you to create easily a dummy
variable for each possible outcome, by simply adding the option gen(name). For example
tab gender, gen(d)
creates a variable d1 for men and d2 for women. If we would have specified the option missing
an additional variable d3 for the missing values would have been created. The naming of the new
variables is simply the name indicated and an increasing integer at the end.

6.1.4 correlate - Correlation and Covariance

The simple correlation of two (continuous) variables is probably one of the most useful descriptive
statistics. To display the correlation matrix of a set of variables simply use

corr x1 x2 x3
and Stata will provide you with a result like
| x1 x2 x3
-------------+---------------------------
x1 | 1.0000
x2 | -0.0185 1.0000
x3 | 0.7128 0.6881 1.0000

In case of preferring the covariance matrix instead of the correlation matrix, add the option covariance
to the command:
corr x1 x2 x3, cov
which will display you the following matrix:
| x1 x2 x3
-------------+---------------------------
x1 | 1.05657
x2 | -.018846 .987245
x3 | 1.03772 .9684 2.00612

6.2 Graphs and plots

6.2.1 plot - Easy scatter plot

A very easy, even though somewhat outdated, way to visualize your data is the plot command, which
provides you in the output window an approximate graph of your data. For instance,
plot x1 x3
would display something like:
5.10264 +
| *
| *
| * *
| *
| ** * *
| * ** *
| * ** *
| * ** * *
x | * ****** *

33
6.2 Graphs and plots 6 SUMMARY STATISTICS AND GRAPHICS

1 | * * *
| ******* *
| * * ***
| ** * *
| * * **
| * ** ***
| **
| * *
| ** *
|
.057795 + *
+----------------------------------------------------------------+
4.16933 x3 31.902

I agree that this way of visualizing data is probably not the state-of-the art in the century of vectorial
graphics, however, it is a fast way to get a first impression. A newer - and much more advanced - way
to visualize data is the graphics command, which I will explain in the next section.

6.2.2 graphics - General graphics

Stata has an enormous amount of possible graphics you can create, however, the precise syntax is

not always easy to obtain. The Stata graph website provides a very useful visual overview of the
different graphs that can be produced with corresponding commands. The following table presents
some example, hoping that the reader can extract from the proposed syntax some useful information
and tricks.

Output Syntax
kdensity x1, normal
Kernel density estimate
.15
.1
Density
.05
0

0 5 10 15 20
x1

Kernel density estimate

Normal density
kernel = epanechnikov, bandwidth = 0.9835

34
6 SUMMARY STATISTICS AND GRAPHICS 6.2 Graphs and plots

Output Syntax

histogram x1, normal width(1)

frequency title(Histogram of variable
Histogram of variable x1
x1) scheme(economist)
20

15

Frequency
10

0
0 5 10 15 20
x1

graph pie , over(female)

Gender distribution of natives and foreigners by(foreigner,title(Gender
Natives Foreigner
distribution of natives and
foreigners, span) note(Note:
Artificial data, span)) plabel( all
42.31%
45.45% percent) scheme(sj)
54.55%
57.69%

Men Women
Note: Artificial data

twoway (lfitci x1 x2)(scatter x1 x2),

Relationship between x1 and x2 title(Relationship between x1 and x2)
20

ytitle(x1) xtitle(x2) legend(order(1

"95% confidence interval" 2 "Linear
15

fit" 3 "Obseravtions"))
10
x1
5
0

5 10 15 20 25
x2

95% confidence interval Linear fit

Obseravtions

35
6.2 Graphs and plots 6 SUMMARY STATISTICS AND GRAPHICS

Output Syntax

kdensity alpha, title(Kernel density

Kernel density of of {&alpha}) xtitle({&alpha})
.1

ytitle(Density) bw(1.25) note(Notes:

.08

Epanechnikov kernel using bandwidth

h=1.25)
.06
Density
.04 .02
0

0 5 10 15 20 25

Notes: Epanechnikov kernel using bandwidth h=1.25

Hint 14. Using the combination of curly brackets {} and the &-symbol, you can use greek letters in
the text you add to graphics. In the last example, is written in greek letters. Here are some examples:

Symbol Stata-Code
{&gamma}
{&phi}
{&Phi}

6.2.3 spmap - Build vectorial maps with shapefiles (shp)

spmap is a user written package to generate very nice looking vectorial maps based on the data in

your Stata database. You need a shapefile6 (shp) containing the GIS data of the country of interest
with the same identifier for entities (e.g. states) as you have in your data7 . First at all, you have to
convert the shp-file to a Stata database using the command shp2dta:
shp2dta using CHE adm, data(CHE-d) coor(CHE-c) genid(id) replace
This is an example to convert the file CHE adm.shp into two databases with names you can freely
choose. Here I chose CHE-d for the database containing the data (all type of information from the
shp file) and CHE-c for the database with the coordinates. Moreover, the option genid(id) generates
a new variable id with the identifier of the unit.
In a second step we have to merge the CHE-d (data database) with our database of the analysis, e.g.
your working database on any economic phenomenon. It is important to check if the identifier used
in the shp-file corresponds to the one you use in your file, otherwise you have to change the it in your
file, as it would become very hard to change in the shp-file. Once you have merged your data, it is
very simple to create a map, using for instance the following code
1: spmap myvalue using "CHE-c.dta", id(id) clmethod(custom) ///
2: clbreaks(0 0.2 0.4 0.6 0.8 1) fcolor(Blues) ///
3: ocolor(black ..) plotregion(icolor(none)) legenda(on) legstyle(0) ///
4: legend(order(1 "0 - 0.2 " 2 "0.2 - 0.4" 3 "0.4 - 0.6" 4 "0.6 - 0.8" 5 "0.8 - 1.0") position(11) )
6
Free shapefiles can be found under http://www.mapcruzin.com and on the websites of some government agencies
depending on the country
7
If the numbers are different, then change the identifier in your data, this is easier than in the shp-files

36
6 SUMMARY STATISTICS AND GRAPHICS 6.2 Graphs and plots

produces the graphic hereafter. The first line indicates on which variable we want to perform the
graphic (myvalue) and which database contains the coordinates. The id(id) indicates the identifier
of the unit (here Swiss cantons) and clmethod(custom) is used to customize the thresholds between
categories. These are indicated in clbreaks(....) and the option fcolor selects the color set to be
used. An overview of the color sets can be found in the help file. ocolor is used to define the color
of the border and plotregion(icolor(none)) defines the background (here empty). The following
commands are used to customize the legend: first its set to be displayed, then the style is selected and
finally the values are changed to whatever text you want.

Random values in Switzerland

0 0.2
0.2 0.4
0.4 0.6
0.6 0.8
0.8 1.0

This fully vectorial map was exported from Stata with graph export as explained in sections 6.2.4
and 8.1.1.

6.2.4 Export graphs

When you produce a graphic in Stata it is generally displayed in a new window. You can easily save
graphics in various formats using the command graph export, followed by the name of the file (with
extension!). Use the option as(format ) to indicate the format to export and replace to overwrite
an old graphic if you wish to do it. If you dont use the as() option, the file extension will be used to
determine the format. The supported graphic formats under windows are:

.ps PostScript
.eps EPS (Encapsulated PostScript)
.wmf Windows Metafile
.emf Windows Enhanced Metafile
.png PNG (Portable Network Graphics)
.tif TIFF

I suggest the use of PNG for the standard use, since the graphics are relatively small and all standard
programs can read them.

37
6.2 Graphs and plots 6 SUMMARY STATISTICS AND GRAPHICS

Hint 15. If you work with LATEX and would like to use Stata -graphics, you should export the
graphics as vector graphics in order to get the best possible quality. See section 8 on how to export
graphs to LATEX and how to use them in LATEX without loss of quality.

38
7 ECONOMETRIC ANALYSIS

7 Econometric analysis
The goal of this section is not to describe all possible estimation commands in detail, but rather to
give a short overview of commands, helping to find the needed routine. If you wish to learn more
about the command and its options, you should refer to the help file, which includes in many cases
examples. Type
help mycommand
to display the help file of the command mycommand.

7.1 Continuous outcome

7.1.1 regress - OLS estimation

The most basic regression is econometrics is the OLS estimation and the regress command might be
the most used. To perform an OLS estimation of y1 on x1 , x2 and x3 you simply write
reg y x1 x2 x3
or using placeholders you can even reduce it to
reg y x*
if there are no other variables starting with the letter x.
All regression commands include a variety of possible options, starting from more technical settings

for Stata and going to relatively complicated versions of the estimated model. A commonly used
option is used to change the way standard errors are computed. In a general way, the option is called
vce(vcetype ), where vcetype can take many different values8 . Common options include:

option Description
robust Computes heteroskedasticity-consistent standard errors according to
White (1980). This option is available in many estimation commands
and can be invoked directly by typing for instance reg y x1 x2,
robust instead of reg y x1 x2,vce(robust).
noconstant Performs the OLS estimation without constant term.
beta Provides standardized coefficients defined as
x
=
y

where is the estimated coefficient,x and y the standard deviation of

the independent and dependent variable respectively.

A typical output of a simple regression analysis looks like:

Source | SS df MS Number of obs = 1000
-------------+------------------------------ F( 2, 997) = 996.87
Model | 2034.95712 2 1017.47856 Prob > F = 0.0000
Residual | 1017.61138 997 1.0206734 R-squared = 0.6666
-------------+------------------------------ Adj R-squared = 0.6660
Total | 3052.56851 999 3.05562413 Root MSE = 1.0103
8
see help vce option for more details

39
7.1 Continuous outcome 7 ECONOMETRIC ANALYSIS

------------------------------------------------------------------------------
y | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
x1 | 1.058603 .0324676 32.60 0.000 .9948909 1.122316
x2 | .9816983 .0326528 30.06 0.000 .9176222 1.045774
_cons | .9089836 .031962 28.44 0.000 .8462631 .971704
------------------------------------------------------------------------------

where the upper left panel provides ANOVA-like information of the sum of squares, degrees of freedom
and mean sum of squares. The upper right panel provides some statistics of the model fit and the
main panel thereafter is the actual estimation. Each row refers to one regressor, starting with the
coefficient (coef.), followed by the standard errors (Std. Err.), the t-statistic (t), the p-value (P > |t|)
and the 95% confidence interval.

7.1.2 Other estimators of continuous outcome

The following table provides a short overview of other estimators used for continuous variables.

Command Description
cnsreg Constrained linear regression model. Define first a constraint using the command
constraint:
constraint 1 x1=x2
and then use the command for the constrained linear regression model:
cnsreg y x1 x2 x3,constraints(1)
gmm Generalized method of moments estimation. See the help file for details.
heckman This command allows you to perform the OLS estimation with selection following
Heckman (1976). The command is used similar to regress, but you include the
option select(model ) where model refers to the selection model, having a dummy
variable as dependent variable, taking the value of 1 when the observation is selected
in the main equation and 0 otherwise.
ivreg Instrumental variable (IV) estimator. The syntax is relatively easy, assume that you
regress y on x1 , x2 and x3 , where x1 and x2 are assumed to be endogenous and
therefore will be instrumented by z1 ,z2 and z3 . Then the command is simply
ivreg y (x1 x2=z1 z2 z3) x3
Using the options 2sls, liml and gmm you can choose the estimator to be the two-
stage least squares, the limited-information maximum likelihood or the generalized
method of moments estimator respectively.
reg3 Three-stage least squares (3SLS) estimator for simultaneous equations. See the help
file for details.
sureg Zellners seemingly unrelated regression. See the help file for details.
tobit Tobit estimation for censored dependent variable. The syntax is similar to the
regress where you add two options ul(#) and/or ll(#) to indicate the upper and
the lower limit respectively. For example if you have data on income that is censored
due to data collection at 999999 then the command would be
tobit income educ exper, ul(999999)
treatreg Treatment-effects model
truncreg Truncated regression model. Similar to the tobit model but for truncated variables.
For instance the income is truncated at zero, since we cannot have negative incomes:
truncreg income educ exper, ll(0)
- Continued on next page -
40
7 ECONOMETRIC ANALYSIS 7.2 Categorical outcome

selmlog This is a user written command to perform OLS estimation with Selection bias correc-
tion based on the multinomial logit model. The command features correction terms
according to Lee (1983), Dubin and McFadden (1984) and Dahl (2003).

7.2 Categorical outcome

7.2.1 probit - Probit estimation

probit estimates the standard probit model for dichotomous dependent variables. The syntax is
exactly the same as for regress. If you want to obtain rather the marginal effects than the raw
coefficients, the command dprobit is more suitable. See also section 7.9 for more details on marginal
effects using the newly introduced command margins.

7.2.2 dprobit - Probit estimation with marginal effects

dprobit is based on probit but reports the marginal effects. If you want to publish marginal effects
in your work and use estout to produce your estimation table, then you have to use dprobit instead
of probit, since otherwise estout cannot report the marginal effects.

7.2.3 logit - Logit regression

The logit command is used exactly the same way as the probit command. See the help file for
details.

7.2.4 mlogit - Multinominal logit

To perform a multinominal logit model, you can use the command mlogit where the syntax is very
easy as well:
mlogit y x1 x2 x3, base(2)
where y is a categorical variable with integer values. The option base(2) in the example is not
necessary but permits you to define the base outcome, here the value 2. See the help file for details.

7.2.5 oprobit - Ordered probit model

oprobit performs an ordered probit model (Greene, 2008, Ch. 23.10.1, page 831). The dependent
variable must be an ordered categorical variable, where higher values refer to better outcomes.

7.2.6 ologit - Ordered logit model

ologit performs an ordered logit model, see oprobit for more details.

7.3 Count data

7.3.1 poisson - Poisson regression

poisson performs a Poisson regression on count data. The dependent variable must contain positive
integer values only. See the help file for details.

41
7.4 Panel data 7 ECONOMETRIC ANALYSIS

7.3.2 nbreg - Negative binomial regression

nbreg performs a negative binomial regression on count data. The dependent variable must contain
positive integer values only. See the help file for details.

7.4 Panel data

7.4.1 xtset - Set-up the panel data

Before being able to perform panel data analysis, you have to tell Stata about the panel structure
of the data. Normally the data should have the long form (see reshape for details on the form and
how to change it). Typically you have a variable identifying the individual (the unit more generally)
and a variable describing the time dimensions (e.g. the year). Assume now that these two variables
are id and year, then the command to define the panel data is given by
xtset id year
which provides you directly with some information on the data enabling you to verify if everything
went well:

panel variable: id (unbalanced)

time variable: year, 2000 to 2004, but with gaps
delta: 1 unit

To get a more detailed description of the datas pattern, consider the command xtdescribe

7.4.2 xtdescribe - Describe the pattern of the panel data

xtdescribe allows you to get some more information on the structure of your panel data. You must
have used xtset before. The just type
xtdescribe
and you will get a detailed description like the following:
id: 0, 1, ..., 99 n = 100
year: 2000, 2001, ..., 2004 T = 5
Delta(year) = 1 unit
Span(year) = 5 periods
(id*year uniquely identifies each observation)

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

2 3 4 4 5 5 5

Freq. Percent Cum. | Pattern

---------------------------+---------
49 49.00 49.00 | 11111
8 8.00 57.00 | 11.11
7 7.00 64.00 | 111.1
6 6.00 70.00 | .1111
6 6.00 76.00 | 1.111
6 6.00 82.00 | 1111.
4 4.00 86.00 | 1.11.
3 3.00 89.00 | ..111
2 2.00 91.00 | .1.1.
9 9.00 100.00 | (other patterns)
---------------------------+---------
100 100.00 | XXXXX

42
7 ECONOMETRIC ANALYSIS 7.5 Time series

First general information on the panel structure is indicated, for instance the variables identifying
both dimensions with their respective value pattern. You will also be informed if the combination of
dimension identifiers identifies each observation uniquely, a very important issue for further analysis.
Finally the most common patterns in the data are displayed, where the 1 indicates that the value
is present and the dot (.) refers to missing data. In our example we have 49 observations where
information is available for all 5 years in the sample, followed by the second most common pattern
where in all but the third year information is available.

7.4.3 xtreg - Panel regression: fixed and random effects

Once your data is set up, you can start running regression models. In this section I only present the
command xtreg which is able to perform fixed-, between- and random-effects and population-average
linear models.
The syntax is similar to the syntax used in regress by simply putting
xtreg depvar indepvars, model
where depvar and indepvar are the dependent and independent variables respectively and model refers
to the type of model that you want to estimate:
re Random effects model (default), GLS estimation
fe Fixed effects model
be Between effects model
mle Maximum likelihood random effects model
pa Population average model

See the help file for details.

Hint 16. See the section 3.10 on factor variables and timeseries operators to learn more about the
use of lagged, forwarded and differenciated variables in panel models

7.5 Time series

In general you can refer to what was written about panel data in section 7.4 and apply it directly to
time series data. Instead of the prefix xt you should now use ts. For instance to set up your data,
you use
tsset year

For the moment, no more time series specific information is available in this document. For further
details refer to section 7.4 or check out http://dss.princeton.edu/online_help/stats_packages/
stata/time_series_data.htm

7.6 Extracting estimation results

Almost every Stata routine saves results in macros names r(name), e(name), s(name) or c(name).
You can easily access this data either directly of saving them to your own macros. Look at the following
example:

sum x1

43
7.7 Post estimation 7 ECONOMETRIC ANALYSIS

local mean=r(mean)
local sd=r(sd)
local t=mean/sd
display "The t-statistic is: t"

First, the summary statistics of the variable x1 are computed with the command summarize. This
command is a so-called r-class command, meaning that the data is stored in r(name). To see all
saved values in r() type return list9 . The second and third line simply copies the values into used
defined local variables. The fourth line then computes a new value out of the stored data and the last
line displays the result.
Some results are stored in matrices (e.g. estimation coefficients and variance-covariance matrices. See
section 3.9.5).

7.7 Post estimation

Once a regression is performed, it is oftentimes needed to perform some post-estimation tests. In this
section I present a few of them, however, in general you find in the help files of the command a link
to the post-estimation commands of this specific command. This is useful, since not all estimations
allow for the same post-estimation commands.

7.7.1 hettest - Breusch-Pagan / Cook-Weisberg test for heteroskedasticity

After regress you can simply type hettest to perform a Breusch-Pagan test on heteroskedastic error
terms. In the output the H0 is explained, being a constant variance. Hence, finding a p-value lower
than the threshold would indicate heteroskedastic error terms and you might want to consider the
White (1980) correction using the option robust (see section 7.1 on page 39)

7.7.2 test - Test linear hypotheses after estimation

test allows you to make tests on the coefficients of a regression. For instance, to test whether the
coefficient of the variable x1 is equal to 2, type
test x1==2
or to test whether the coefficients of x1 and x2 are equal put
test x1==x2

This command will be performed on the active estimation, which is in general the last estimation
performed.

7.8 Saving and reusing estimations

7.8.1 est store - Save an estimation

Like for the database, you can always have at most one estimation active, but estimations can be
stored and used later on as well. To store an estimation simply write
est store myestimation1
9
Type ereturn list for e-class commands, sreturn list for s-class commands and creturn list for c-class com-
mands

44
7 ECONOMETRIC ANALYSIS 7.8 Saving and reusing estimations

to save the current estimation under the name myestimation1. This is very useful to produce estimation
tables afterwards (see est table (section 7.8.4) and estout (section 8.2.1)).

7.8.2 est restore - Restore an estimation

You can restore (re-activate) an estimation previously saved at any moment by simply writing
est restore myestimation1
This will not display the regression again, but the data off the regression will be available again in the
e-class-scalars (type ereturn list to see them).s

7.8.3 est replay - Replay an estimation

If you want to display an earlier regression again, then you can use
est replay myestimation
Important: this command does not reactivate the estimation, it simply displays it. To reactive an
earlier regression, use the command est restore.

7.8.4 est table - Display an estimation table of several regressions

A very nice feature of the general command estimates is the sub-command estimates table, which
allows you to display several regressions at a time in order to compare them. I illustrate the use of
this command with a short example using the system-database auto.dta. First, we perform and save
to regressions:
sysuse auto, clear
reg price weight length
est store reg1
reg price weight length trunk
est store reg2

This example simply performs two OLS regressions with the command regress on two and three
independent variables respectively and saves the results in reg1 and reg2. Now we would like to
display the two regressions with
est table reg1 reg2
which yields to the following display:
----------------------------------------
Variable | reg1 reg2
-------------+--------------------------
weight | 4.6990649 4.7215989
length | -97.960312 -102.66518
trunk | 28.376437
_cons | 10386.541 10812.329
----------------------------------------

This is not yet the nicest estimation table ever seen! With a few additional options, you can make it
much better:
est table reg1 reg2, stats(N r2 a) b(%6.3f) star(0.1 0.05 0.01)
you already get a useful display like:
----------------------------------------
Variable | reg1 reg2
-------------+--------------------------
weight | 4.699*** 4.722***
length | -97.960** -102.665**
trunk | 28.376
_cons | 1.0e+04** 1.1e+04**

45
7.9 Marginal effects 7 ECONOMETRIC ANALYSIS

-------------+--------------------------
N | 74 74
r2_a | 0.329 0.320
----------------------------------------
legend: * p<.1; ** p<.05; *** p<.01

where you immediately see the significance levels of the coefficients and some statistics of the estima-
tion. Let us have a look at the different options:

star(n1 n2 n3) Allows you to define significance levels of the stars. In our example the
standard 10%, 5% and 1% were chosen.
b(format) Allows you to format the output of the coefficients and to limit the
amount of digits after the coma. See 2.6.4
stats(args) Enables you to display some estimation relevant statistics:
N Number of observations
aic Akaikes information criterion
bic Schwarzs Bayesian info. criterion
chi2 Value of 2
ll log likelihood
se, t, p You can display in addition to the coefficients the standard errors (se),
the t-statistics (t) and the p-value (p). Simply indicate the desired
statistic together with its format, e.g. se(%6.3g).
keep(coeflist) Only the coefficients indicated in coeflist will be shown
drop(coeflist) All coefficients in coeflist will be dropped from the table

7.9 Marginal effects

In many model the estimated coefficients of a model cannot be interpreted directly and one needs to
compute the marginal effects first. In the case of a probit, the usual way to do this was to use dprobit

instead of probit. However, Stata introduced recently a completely new command margins that is
intended to compute marginal effects for a very wide range of models. Since I believe that this will be
the command to be used in the future, I will discuss it with some more details hereafter.

7.9.1 margins - Compute marginal effects (post estimation)

margins is a relatively new command to compute marginal effects for a large range of models. It is
a somewhat complicated command with a rather difficult syntax at first sight. On the other hand, it
offers many very nice options to estimate not only marginal effects at the mean, but basically wherever
you want. In this explanation I will focus on the very basic capabilities of margins and I encourage the
reader to have a look at the official help file of the command for more advanced uses of the command.
I will explain the command using a simple probit model, since most of the reader are familiar with it
and we will be able to compare it easily to the older dprobit command. For the following discussion
I use the freely available dataset margex which you can load by typing
webuse margex
Let us consider the simple probit model where outcome is explained by two continuous variables age
and distance and the gender dummy called sex. The simple estimation of

46
7 ECONOMETRIC ANALYSIS 7.9 Marginal effects

probit y sex age distance

will provide us with the estimated coefficients that are not directly interpretable. An easy way to get
the marginal effects instead would be to use
dprobit y sex age distance
Unfortunately, this solution does not store the marginal effects in the system matrices. Moreover, the
newly introduced command margins is more powerful, so let us discuss how it works.
Since it is introduced as a post-estimation command, it has to be always in combination with a prior
estimation command. To get the same marginal effects as in the example with dprobit one would
have to type:

probit outcome i.sex age distance

margins , dydx(*) atmeans

where it is important to use the factor variable (i.sex instead of sex) in the probit. Otherwise, the
command margins does not recognize sex as being a dichotomous variable and uses the procedure

for continuous variables. The option dydx(*) is used to indicate Stata that we want to estimate
the marginal effect for all regressors. By replacing the * by some regressors, margins computes the
marginal effects only for the indicated variables. The option atmeans permits to compute the marginal
effect at the mean instead of the average marginal effect (see next paragraph).

ME at the mean vs. mean ME

Marginal effects are not uniquely defined in non-linear models. One option is to compute the
marginal effect for an average person (this is what the example above does) or the marginal effect for
each observation at their observed values and then average the different effects. Formally and with
some slight abuse of notation, the average marginal effect (AME) and the marginal effect at the mean
(MEAM) can be defined as follows for the probit model:
" #
(X )
AM E = E (1)
xj
(E[X])
M EAM = (2)
xj

where (.) is the cumulative normal distribution, X is the whole matrix of covariates and xj indicates
the variable for which we are computing the marginal effect.
It is also possible to compute the marginal effect at different points of the distribution. Let us consider
some examples, always assuming that the model we estimated before the margins command was
probit outcome i.sex age
In this case
margins, dydx(*)
computes the marginal effects for each observation at its observed levels of the regressors. In a second
step the average of all these marginal effects is computed giving you the average marginal effect. In
contrast,
margins, dydx(*) atmeans
defines a person that has the average characteristics and computes for this person the marginal effect
giving you the marginal effect at the mean.

47
7.9 Marginal effects 7 ECONOMETRIC ANALYSIS

Marginal effects at different points of the distribution

Sometimes it is interesting to look at the marginal effects at different points of the distribution. For
instance, if we would like to compute the marginal effect for people at the age of 50, we would use
margins , dydx(*) at(age=50)
To limit it further to men at the age of 50, we would use
margins , dydx(*) at(age=50 sex=1)
It is also possible to compute the marginal effects for several values. For instance, to compute them
simultaneously for the age of 20, 40 and 60, we would use
margins , dydx(*) at(age=(20 40 60))
Finally, we can automate these intervals using the following notation
margins, dydx(*) at(age=(20(5)60)
where the marginal effects are computed for all multiple of 5 starting from 20 and going up to 60.

Marginal effects for different groups/populations

It could also be interesting to compute the marginal effects for different populations. These groups
can be defined by the covariates of the model or variables that were not used in the model. To activate
this computation by groups you can use the over(varlist ), where the whole computation is done for
each value of the variables in varlist 10 . It is important not to confound this option with the previously
explained at() option. While the at() option refers to the position in the distribution of a covariate
for which the marginal effect should be computed, the option over(varlist ) computes the marginal
effect at a given position (can be specified as well) in the distribution for different groups. For instance,

margins , dxdy(*) over(sex) atmeans

computes the marginal effect at the mean for men and women separately.

Computational issues
Sometimes the computation of marginal effects can be very time consuming. If the computation
takes too much time, one might consider not to compute the standard errors (if they are not absolutely
needed), which should make the computation much faster. This can be done by adding nose as an
option.

Export marginal effects

Most of the time, we do not want to limit the computation to the display on the screen but would
like to export them. margins might surprise you negatively when you first try to do this. For instance,
if you use the following code

probit outcome i.sex age distance

margins , dydx(*) atmeans
est store reg1
estout reg1, margin

you will get the estimated coefficients of the probit and not the marginal effects. This is because
margins does not overwrite the e-values of the probit estimation. If you want margins to do so, you
10
Alternatively, you can also use the option by(varlist ), giving you exactly the result

48
7 ECONOMETRIC ANALYSIS 7.9 Marginal effects

just have to add the option post. Hence, the correct syntax for the above example would be

probit outcome i.sex age distance

margins , dydx(*) atmeans post
est store reg1
estout reg1, margin

Actually, the option margin is no longer needed in the command estout, however, I leave it here,
because you might have other regressions in the estout that still need the option.

Predicted values at different levels of covariates

An additional nice feature of margins is the ability to compute predicted values (or expected
values) conditioned on some specific values of covariates. Imagine for instance that we would like to
have the marginal effect of an interaction of two dummy variables. For non-linear models the standard
procedure cannot be used. However, we can simply compute the predicted outcome conditional on
all possible combinations of the two dummy variables. Then we can compute the marginal effect and
using the test command we can test its significance. Here a very simple example:

sysuse nlsw88, clear

probit union i.south##i.married age
margins, predict() at(south=(0 1) married=(0 1)) atmeans post
test 4._at-3._at-2._at-1._at == 0

The first line loads the database, on the second line I estimate a probit model with an interaction
term between two dummy variables (see 3.10 for details on factor variables). The third line computes
the four possible conditional probabilities, where the option atmeans set all other control variables to
their respective mean and the option post saves the coefficient for posterior use. Finally, the fourth
line performs a test whether the marginal effect of the interaction term is significant or not11 .

Final remarks on margins

I presented here only the most basic features of margins, however, there are many more. The
command is very complete and usable for a wide range of estimations. Even though the syntax is
somewhat heavy at first sight, I encourage the reader to invest some time in this command, because
you can really get very nice results out of it.

11
See Ai and Norton (2003) to see why the test on the coefficient is not useful in non-linear models

49
 8 STATA MEETS LATEX

8 Stata meets LaTeX

Besides the high quality features in econometric analysis and data handling, Stata is also a very
powerful tool thanks to its ability to communicate with other software package. More precisely, it
is relatively easy to export econometric results and graphics to formats that can be read by other
packages, especially LATEX . In this section I present some of the possible ways to export your results
to LATEX and also how to import them into your LATEX project. The ultimate goal of this section
is to enable you to automate the process completely. Changing your model slightly in Stata will
automatically change the tables and graphs in your LATEX document, thus running the Stata code
and compiling the LATEX document will suffice!12

8.1 Exporting high quality graphics to LaTeX

8.1.1 graph export - Graph export to eps

Once you have your graph in Stata you can export it into the .eps-format. This format is generally
not very easily readable under Windows, but the import into LATEX is easy. The normal command
includegraphics can be used in combination with the package epstopdf. The package converts the
eps graphic into a pdf graphic which LATEX can use. Hence the complete LATEX code to use is:
In the preamble: \usepackage{epstopdf}
and in the body: \includegraphics{mygraphic.eps}
The following two graphs were built with the same code, once exported as PNG (left) and once as
EPS (right). I suggest you to zoom in the two graphs to a maximum and then I let you choose your
preferred graph:

Gender distribution in sample data

10.7%

44.4%

44.9%

male female
unknown

8.2 Exporting estimation results to LaTeX

8.2.1 estout - Creating estimation tables

Different genuine Stata - and user written-commands are available to export your estimation results
to LATEX . Here I focus on the command estout written by Ben Jann. In my opinion it is the most
flexible command and offers great possibilities. The basic syntax is similar to est table, but it offers
much more features. I illustrate its use with a small example based on the two regressions used in the
demonstration of est table in section 7.8.4:
12
The part of interpreting the results and writing the paper cannot be automated yet ;-)

50
8 STATA MEETS LATEX 8.2 Exporting estimation results to LaTeX

1: estout reg1 reg2 using regtable.tex, replace style(tex) margin cells(b(fmt(%9.3f) star) se(par) ) ///
2: starlevels(* .1 ** 0.05 *** .01) ///
3: stats(N r2_a , labels(N "Adj. $ R^2$ ") fmt(%9.0f %6.3g)) mlabels("Model 1" "Model 2") ///
4: collabels(none) varlabels(_cons "Constant"weight "Weight" length "Length" trunk "Trunk space") ///
5: order(length weight trunk _cons) ///
6: prehead(\begin{center}\begin{tabular}{l*{@M}{l}} "\hline") ///
7: posthead("\hline") ///
8: prefoot("\hline") ///
9: postfoot("\hline\end{tabular}\end{center}")

produces the following table:

Model 1 Model 2
Length -97.960** -102.665**
(39.175) (42.587)
Weight 4.699*** 4.722***
(1.122) (1.132)
Trunk space 28.376
(97.058)
Constant 10386.541** 10812.329**
(4308.159) (4574.211)
N 74 74
Adj. R2 .329 .32
The command might look a bit scary at first, but you will get used to it very quickly. Moreover,
once you defined your style correctly, you can copy-paste large parts of the code from one table to
the next. Lets go through the code with detail. The first line corresponds mainly to what we saw
for the command est table. First we include with reg1 reg2 the names of the regressions we want
to display, followed by the file to create. After the comma we use the option replace in order to be
able to re-run the script (otherwise there is an error) and the style(tex) to declare that we want
a LATEX output. The option margin does not make much sense in this example, since in the OLS
estimation all estimated effects are marginal effects. However, if you estimate for instance a dprobit
model, this option is needed to display marginal effects rather than the coefficient estimates. The
option cells is used to choose all the coefficient-related information to be displayed. In this example
the betas (b) is formatted with the 9.3f format with stars included and for the standard errors (se)
I use the option par to get the standard errors in parentheses.
The second line simply defines the significance levels of the stars and in contrast to the option saw in
est table, you can also choose other symbols than the stars.
The third line first declares the statistics of the estimation to be displayed, followed by the label of
the statistic and the format. The option mlabels allows you to give each regression a name; if you do
not specify this option, reg1 and reg2 will be displayed.
Line 4 formats the first column where the variables are indicated. Without any of these option the

variable name, as it appears in Stata , will be displayed. The first option, which is not active here,
would allow you to use the labels of the variable rather than the names and the option varlabels
allows you to change each variable name separately for the table output without changing anything
in your data. In this example I just changed all the names to capital letters and the constant term to
a nicer looking word.
The option order on line 5 allows you to change the order of the variables in the table and the remaining
4 lines are all related to the LATEX code. They allow you to put some free text (or LATEX commands)

51
8.2 Exporting estimation results to LaTeX 8 STATA MEETS LATEX

at very specific places in the table. The prehead is before the name of the regressions. I use it here
to include directly the tabular environment of LATEX using the variable @M, which is the number of
models, directly computed by the command. Hence, you do not have to adapt the number of columns
if you add a regression, this will be done automatically. If you want to include your table in a table-
environment of LATEX , you could start it here as well. All the information will simply be transferred
to the generated LATEXcode. The posthead is between the regression names and the first coefficients
and the prefoot will be placed between the coefficients and the statistics, while the postfoot is after
the statistics.
This is only one of many possible examples of the command estout and I encourage you to consult
the manual of the command.
To include the command to your latex file, simply use

\input{regtable}

The big advantage is that you can make small changes directly in your do-file and the table will be
adapted in your LATEX -paper. This command should definitely help you to avoid copying estimation
results by hand to you LATEX file, a process with some risk of making errors.

8.2.2 tabout - Creating descriptive statistics tables

Similar to the command estout you can export frequency tables with the command tabout, written
by Ian Watson. This command has extremely many options, which makes it a little bit difficult at the
beginning. A very simple example of the code is:

tabout female indigenous using crosstable.tex , replace ///

style(tex) cl1(2-4) c(cell) layout(rb) f(1p) h3(nil) font(bold)

which makes a cross frequency table of the variables gender and indigenousas follows:

indigenous
female No Yes Total
No 44.4% 5.8% 50.2%
Yes 45.7% 4.1% 49.8%
Total 90.1% 9.9% 100.0%

The using crosstable.tex defines the file to be written, the replace options allows you to overwrite
an earlier version of the file. style(tex) indicates that you want to have a LATEX file, the cl1(2-4)
add a horizontal line between column 2 and 4 right after the variable indigenous. c(cell) indicates
that you want to have the percentage by cell, alternatives are freq for frequencies and col and row
for percentages according to the column and row respectively. The f(1p) option indicates the format
of the table, here the p stands for percentage. Simply f(1) would create a number with one decimal.
The h3(nil) option avoids that at the top of each column a %-sign is shown (N in case of frequencies)
and finally font(bold) makes some parts of the table bold. You could also include some pretable and
posttable LATEX -code, but this must be done with external files. Therefore I prefer not to do it in my
latex-code, thus the import becomes then:

52
8 STATA MEETS LATEX 8.2 Exporting estimation results to LaTeX

\begin{tabular}{llll}
\input{crosstable}\endrule
\end{tabular}

53
 9 IMPORTING DATA IN OTHER FORMATS

9 Importing data in other formats

9.1 usespss - Read SPSS data
Oftentimes micro data is made public in the SPSS (or PASW) format (file extension .sav), however, it

is very easy to import such data into Stata . Using the command usespss you can open SPSS-data

just like the normal Stata format. Labeled data will be imported with labels. The command works
like use:
usespss $source/myspssdata.sav
will import the data and then you can use it directly or save (convert) it into the .dta-format by using
the command save.

9.2 fdause - Read SAS XPORT data

Exactly like the usespss-command, you can import data from SAS using the command fdause.

54
10 STATA IS NOT ENOUGH? ADD USER-WRITTEN MODULES TO ENHANCE STATA

10 Stata is not enough? Add user-written modules to enhance Stata

Stata comes with a quite complete set of commands and modules to perform most of modern econo-
metrics analyses. However, some very specific analyses might still be missing. If you are not eager
to program the modules yourself, you can search for a user-written command in the internet. For
econometric analyses the best platform is SSC (Statistical Software Components) on IDEAS13 . To
find the correct module, use the findit command. For instance, if you would like to run a treatment
effect analysis, type
findit treatment effect
Several propositions are made. An alternative way to find user written modules is to search on the
internet and most of the time you will end up at the SSC on IDEAS.
If you already know the name of the command and it is available on SSC, you can install it very easily

ssc install NAMEOFMODULE

Other SSC related commands that might be useful are:

Command Description
ssc new Displays the newest user-written command
available on SSC.
ssc hot Displays the most downloaded modules.
ssc uninstall mypackage Uninstalls the package named mypackage
ssc install mypackage, replace Updates the package mypackage
ssc describe mypackage Displays the description of the package

Hint 17. To check which version of a command you have installed on your machine, you can simply
type
which packagename
and Stata will show you the version of the module packagename

Very useful user-written commands include: ivreg2, estout, tabout. Many of the user-written com-
mands do not necessarily help you estimating very complicated model, but they might be very useful
for some basic tasks like converting data to other formats, exporting nice-looking tables, performing
some basic statistical tests, etc...

13
http://ideas.repec.org/s/boc/bocode.html

55
 11 PROGRAMMING YOUR OWN COMMAND

11 Programming your own command

When you start using Stata a lot and perform many times the same type on analysis it might be
interesting to program your own routine. This is fairly easy and not much more knowledge is needed
than for a standard do-file. In this section I will just write how to start programming, but I will not
enter a long discussion about specific commands for programmers.

1: capture program drop iop

2: program define iop , rclass
3: version 8.2
4: syntax varlist [if] [in] [, BOOTstrap(integer 0) DECOMPosition
GRoups(varlist max=1) PRopt(str) BOOTOPT(str) PROBIT*]
5: marksample touse
6: quietly{

7: tokenize varlist
8: local depvar 1
9: macro shift
10: local xvars *

11: // BOOTSTRAP
12: if(bootstrap & bootstrap>0){

13: forvalues i=1/bootstrap{

[SOME LINES OF THE CODE ARE NOT REPORTED HERE]

14: end

The first line drops the program (if already existent) from the memory. This is needed to enable
us to define it again. The word capture is included to avoid an execution stop in case of an error,
typically when the program does not exist yet. The second line then defines the program with the
name iop, which must be equal to the name of the ado-file (see later on). The option after the coma

refers to the class of the routine, r-class is generally a good option. Line 3 declares Stata that the
ado-file should work from version 8.2 on, hence trying to run it on an earlier version causes a problem.
The fourth line is the most important for the moment, since we declare the syntax - this is basically
the same as in the help files. After the word syntax you write varlist if you will enable the user
to provide a varlist. You can include the possibility to offer the user the if and in statements. If

you do so, you have to consider this after in the code, since Stata does not limit the routine to the
limitations given by the user automatically! After the coma you might include all the options. In this
case all options are non-mandatory, since the opening bracket is before the comma. The capital letters
indicate the minimum amount of letters the user has to write, for instance the option bootstrap will

be understood by Stata whether the user writes boot or bootst or bootstrap, but not if he writes
boo. Two types of options are available: with and without arguments. Those without arguments
generate simply a local containing the complete name of the option when the option is chosen. The
option with arguments save the arguments in a local variable with the name of the option. You have to

56
11 PROGRAMMING YOUR OWN COMMAND 11.1 Where and how to save your routine?

indicate always the nature of a argument, being for instance str for a string or varlist for a varlist .
Line 5 then converts the if and in conditions into a temporary variable I call here touse. In any
routine you use afterwards you will have to indicate
commandname varlist if touse
in order to limit your routine to the sample. An alternative is to keep only the sample you need:

marksample touse
preserve //saves the current state of the DB for later
keep if touse
[ALL YOUR PROGRAM]
restore //restores the database as it was before the preserve command

The remaining lines of the code are then more or less standard. You can use all the routines available

in Stata . The do-file end with the command end (line 14).

11.1 Where and how to save your routine?

The routine can be written like a normal do-file but must be saved under the extension .ado under the
form myroutinename.ado. It is important that the file has the same name as your command, otherwise

Stata cannot execute it. Under Windows it is most recommended to have a folder C://ADO where

Stata saves all updates anyway and there to save your ado-files in C://ADO/personal in order to
have them at a save place. Once you saved and executed your ado-file once, it is available for the use

in Stata as long as you dont delete the file!

11.2 Useful commands for programming

Some commands are especially useful when programming ado-files, allowing you to achieve the highest
possible flexibility. Here are some examples:

11.2.1 tokenize - Tokenize a string

Many commands include a varlist where the first variable is the dependent variable and the remain-

ing independent. Technically the varlist is just a string stored in a local variable, thus Stata does
not know a priori what to do with it. An easy way is to tokenize the string. Assume to have a
varlist in you syntax and that the string will be available for you in the form of a local variable
called varlist. By using
tokenize varlist
you can split up your string into its elements. To illustrate this point consider the following example:

local varlist="income educ exper gender"

tokenize varlist

which splits up the text income educ exper gender into 4 elements stored in local variables called 1,
2, 3 and 4. For instance, you can now change the order:

display="3 2 4"

57
11.2 Useful commands for programming 11 PROGRAMMING YOUR OWN COMMAND

would produce exper educ gender. Additionally the local named simply * contains the whole initial
string:
di "*"
produces income educ exper gender. The interesting feature of * is the use with macro shift.

11.2.2 macro shift - Shifting the content of a local

The command macro shift makes a very simple but effective operation on the local * (see tokenize)
by eliminating the first element. Assume that the local * contains the string from section 11.2.1:
income educ exper gender, then
macro shift
convert the string into educ exper gender.
Consider now the combination of macro shift with tokenize. Assume that the varlist contains as
fist element the dependent variable and the remaining elements represent the explanatory variables,
then the following code help you to divide them:
tokenize `varlist' // Splits up the local varlist
local depvar=`1' // Stores the first element in the local depvar
macro shift // Drop the first element of the local *
local indep=`*' // Stores the remaining elements as local indep
11.2.3 marksample - Selecting the sample considering if and in
As already mentioned in the beginning of this section, the command marksample allows you to retake
the information of the if and in option of your syntax.
marksample touse
simply generates a temporary variable touse taking the value of 1 if the observation should be included
and zero otherwise.
You can then use in every command of your ado-file the condition
my command if `touse'
or simply limit the sample from the beginning. If choosing the second option, do not forget to preserve
the data at the beginning and restore it in the end, otherwise the user will have bad surprises after
using your command!
11.2.4 preserve - Making an image of your current data
The command preserve is not particularly a command for ado-files, but I present it here, since in the
programming of ado-files its most often used. preserve simply makes an image of the current state
of your data, allowing you to manipulate it thereafter and being able to recover it again, as it was,
with the command restore
11.2.5 restore - Restoring an image of your data
restore can only be used when preserve was used before and allows you to recover the image of the
data created earlier.
11.2.6 local:extended - Extract labels from variables and value labels
When programming a module it is sometimes useful to get the main labels and the value labels of
the variables. For instance, instead of displaying the variable names in the output, it might be nicer
to display the variable label. Using the extended functions for local variables (it works also with

58
11 PROGRAMMING YOUR OWN COMMAND 11.2 Useful commands for programming

global variables), you can easily extract this information from the data and use it afterwards. For a
discussion on the labels and how to create them, see section 5.1.5.
Extracting the variable label
Let us start by extracting the variable label from a variable female. Simply type
local mylocal:variable label female
to save the variable label of the variable female as local called mylocal14 . To use the extracted text,
you can simply include the local, for instance by typing:
display "The variable label is:`mylocal'"

Extracting the value label

Besides the label for the variable, the value labels might be of interest for the programmer. To
extract the value labels we have to proceed in two steps. First we have to extract the name of the
value label for a given variable and we can then extract the label for every possible value. Let us
consider again the variable female and assume that it has a value label called gender where 0 is
labeled Male and 1 Female. To get this information we simply extract the name of the value label by
writing
local foo:value label female
We have now stored the name of the value label in the local foo and can proceed to the extraction of
the individual labels, for instance for the value of 1:
local label2:label `foo' 1
We have now stored the value label for 1 in the local label2 and could use if for instance by typing
display "The value label for female=1 is:`label2'"
Of course, if we already know the name of the value label, we could directly request the value by
typing:
local label2: label gender 1
However, this is rarely the case when we program a routine which should be applicable for different
datasets.

14
Of course, you can use whatever name for your local.

59
 12 SIMULATION

12 Simulation

The discussion of simulation in this document is limited to the discussion of generating random
variables.

12.1 Uniform distribution

The easiest way to create a random variable is definitely the uniform variable, which can be combined
with the generate command:
gen x=uniform()
which gives a variable x U[0,1] To generate rather U[10,20] you only need some basic algebra:
gen x=10 + uniform()*(20-10)

12.2 Normal distribution

To generate normal variable a special command is available besides the one related to a normal
generate. I suggest the use of drawnorm:
drawnorm x1 x2
will create two i.i.d variables with standard normal distribution.
drawnorm x1 x2, means(1 2) sds(4 3)
will generate two unrelated random variables:

x1 N (1, 4)
x2 N (2, 3)

One can go further by simulating random variables form a higher dimension normal distribution.
To do this, you will need to indicate either the covariance or the correlation matrix. The following
example uses the correlation matrix:

matrix define R=(1,0.9\0.9,1)

drawnorm x1 x2, means(0 0) sds(2 3) corr(R)

The first line defines the correlation matrix (see section 3.9.1) and the second line perform the random
draw. Note that by default as many observations as the dataset has will be drawn. This can be
changed with the option n(#)

12.3 Other distributions

Like the uniform distribution in section 12.1, random variables can be drawn from many different
distributions. Here is a short overview:

60
12 SIMULATION 12.4 Setting the random seed

runiform() = uniform()
rbeta(a,b) Beta distribution
rbinomial(n,p) Binomial distribution
rchi2(df) 2 distribution
rgamma(a,b) Gamma distribution ()
rhypergeometric(N,K,n) Hypergeometric distribution
rnbinomial(n, p) Negative binomial distribution
rnormal() Standard normal distribution (drawnorm preferable)
rnormal(m,s) Normal distribution (drawnorm preferable)
rpoisson(m) Poisson distribution
rt(df) Student distribution

To generate such a variable, you can simply use the generate command:
gen myvar=rpoisson(4)
generates a variable called myvar drawn from a Poisson distribution with parameter = 4.

12.4 Setting the random seed

By default no fixed random seed is set, thus every time you run you simulation another result will
appear. To control the random seed, you can set it with the following statement:
set seed n
where n can take any integer value, which allows you also to repeat the same simulation with different
seeds, using for example loops (section 3.11).

61
REFERENCES REFERENCES

References
Ai, Chunrong and Edward C. Norton, Interaction terms in logit and probit models, Economics
Letters, 2003, 80 (1), pp. 123129.

Dahl, G. B., Mobility and the Returns to Education: Testing a Roy Model with Multiple Markets,
Econometrica, 2003, 70, 23672420.

Dubin, J.A. and D.L. McFadden, An Econometric Analysis of Residential Electric Appliance
Holdings and Consumption, Econometrica, 1984, 52, 345362.

Greene, William H., Econometric Analysis, 6 ed., Pertinence Hall, Upper Sadle River, New Jersey,
2008.

Heckman, J., The common structure of statistical models of truncation, sample selection, and
limited dependent variables and a simple estimator for such models., Annals of Economic and
Social Measurement, 1976, 5, 475492.

Lee, L.F., Generalized Econometric Models with Selectivity, Econometrica, 1983, 51, 507512.

White, Halbert, A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test

for Heteroskedasticity, Econometrica, 1980, 48 (4), 817838.

62
Index
.do, 7 describe, 28
.dta, 7 descriptive statistics, 32
#, 20 dichotomous, 19
display
about, 23 matrix, 18
aorder, 29 display format, 10
append, 30 distribution
betas, 43 normal, 60
Breusch-Pagan test, 44 uniform, 60
do-file, 7
c., 19 comments, 12
change working directory, 13 execute, 12
classes of commands, 15 linebreak, 12
clear all, 13 Download user written, 55
clear memory, 13 dprobit, 41
cnsreg, 40 drawnorm, 60
codebook, 28 drop, 28
coefficients dummy, 19
extract, 18 define, 25
test, 44
command e-class, 15
classes, 15 egenerate, 25
types, 15 est replay, 45
comments, 12 est restore, 45
compress, 10 est store, 44
compressing database, 10 est table, 45
condition, 15 estimation
programming, 20 extract results, 43
conversion estout, 50
data to matrix, 18 exit, 23
matrix to data, 18 extract
Cook-Weisberg test, 44 betas, 43
correlate, 33 estimation results, 43

d., 19 f., 19
data factor variables, 19
description, 28 fdause, 54
format, 9 File
load, 14 File types, 7
save, 14 findit, 24
data to matrix, 18 float, 9
database foreach, 21
compressing, 10 format, 9

63
INDEX INDEX

format, 11 System matrix, 18

forvalues, 22 matrix define, 17
matrix get, 18
generate, 25 matrix list, 18
global, 8 matrix rownames/colnames, 18
gmm, 40 matrix to data, 18
graph export, 50 memory
graphics, 34 change, 13
greek letters, 36 clear, 13
gsort, 28 merge
heckman, 40 horizontal, 29
help, 24 prepare the data, 29
heteroskedacity, 39 vertical, 30
hettest, 44 merge, 29
Missing values, 11
i., 19 mkmat, 18
if-condition, 15 mlogit, 41
if/else, 20
import nbreg, 42
SAS, 54 normal distribution, 60
SPSS, 54 numerical variables, 9
in-condition, 15 ologit, 41
ivreg, 40 Operators
l., 19 logical, 16
label relational, 16
value, 26 oprobit, 41
variable, 26 order, 29
label, 26 plot, 33
lagged value, 19 poisson, 41
linebreak, 12 Postestimation, 44
local, 8, 9 Marginal effects, 46
local:extended, 58 Predicted values, 49
logical, 16 preserve, 58
logit, 41 probit, 41
loops, 20
query, 23
macro shift, 58
Marginal effects, 46 r-class, 15
non-linear models, 49 random
margins, 46 seed, 61
marksample, 58 variables, 60
matrix, 8, 17 recode, 26
display, 18 reduce size, 10
functions, 18 reg3, 40
rename columns&rows, 18 regress, 39

64
INDEX INDEX

relational, 16 usespss, 54
rename, 27
replace, 25 value
reshape, 31 label, 26
restore, 58 variable
robust, 39 label, 26
robust standard errors, 39 squared, 19
variables
SAS XPORT, 54 factor, 19
save, 14 random, 60
scalar, 8 VCE options, 39
scaler, 9 VCE types, 39
search, 24 Version of module, 55
seed, 61
while, 21
selmlog, 41
working directory
set, 23
change, 13
set memory, 13
simulation, 60 xtdescribe, 42
sort, 28 xtreg, 43
spmap, 36 xtset, 42
SPSS, 54
squared variable, 19 zip file, 10
SSC, 55
string, 9
string types, 10
summarize, 32
summary statistics, 32
sureg, 40
svmat, 18
symbols
greek, 36

tabout, 52
tabstat, 32
tabulate, 32
test, 44
tobit, 40
tokenize, 57
treatreg, 40
truncreg, 40
tsset, 43

uniform distribution, 60
update, 23
use, 14
User written command, 55

65