Applied Econometrics Using Stata

Applied Econometrics using Stata
Index
Chapter 1. First steps in Stata ..................................... ??
Chapter 2. Least Squares ........................................... ??
Chapter 3. Instrumental Variables................................. ??
Chapter 4. Panel Data ............................................... ??
4.1 Static ................................................... ??
4.2 Dynamic ................................................ ??
4.3 Nonlinear Models ...................................... ??
Chapter 5. Binary Logit/Probit..................................... ??
Chapter 6. Hazard Models .......................................... ??
Chapter 7. Count-Data Models ..................................... ??
Chapter 8. Selection Models........................................ ??
Chapter 9. Partially Continuous variables ........................ ??
Chapter 10. Logit/Probit Models................................... ??
10.1 Multinomial ........................................... ??
10.2 Ordered ............................................... ??
Chapter 11: Quantile Regression................................... ??
Chapter 12. Robust Inference ...................................... ??
12.1 Clustered Standard Errors .......................... ??
12.2 Bootstrap ............................................. ??
12.3 Two-Stage Models ................................... ??
Chapter 13: Matching................................................ ??
Chapter 1
First steps in Stata
Introduction
The Stata screen is divided in 4 parts. In "review" you can see the last commands that
have been executed. In "variables" you can see all the variables in the current
database. In "results" you can see the commands' output. Finally, in the window
"command" you can enter the commands.
Data input and output
Stata has its own data format with default extension ".dta". Reading and saving a
Stata file are straightforward. If the filename is "sales.dta" (and if it is located in the
Stata's directory), the commands are:
. use sales
. save sales
Notice that if you don't specify an extension, Stata will assume that it is ".dta". If the
file is not stored in the current directory (e.g. in the folder "c:\user\data"), then the
complete path must be specified:
. use c:\user\data\sales
Nevertheless, the easiest way to work is keeping all the files for a particular project
in one directory, and then changing the "default" directory of Stata to that folder.
For instance, if you are working in the folder "c:\user\data":
. cd c:\user\data
. use sales
Insheet (importing from MS Excel)
There are two simple ways to transform an Excel database (or similar formats) into a
Stata database. For instance, please create the following table in MS Excel:
name
account
money
John Doe
1001
55
Tom Twain
1002
182
Tim Besley
1003
-10
Louis Lane
1004
23
Save as a text file (tab delimited or comma delimited) by selecting "File" and
choosing "Save As", under the name "bank.txt" in the Stata folder. Notice that saving
as "txt" (only text) you will loose all the information on formats, formulaes, etc. Then
import the data to Stata using the command "insheet":
. insheet using bank.txt
Para ver si las variables se cargaron exitosamente, abra el browser:
. browse
Alternatively, in order to import the data you can highlight the cells under
consideration in MS Excel, and then select "Edit" and choose "Copy". With an empty
dataset (i.e. first use the command "clear"), enter in Stata:
. edit
And a spreadsheet will appear (similar to what appeared with the command
"browse", now you can modify the database). Then click the right button on the
editor and press "Paste".
The command "save" let us save the data in Stata format (".dta"). The option
"replace" replaces the old version if the file already exists:
. save prueba.dta, replace
The command "clear" clears the current database:
. clear
The option "clear" can be used simultaneously with the command "use" to open a new
database:
. use sales, clear
By default Stata separates 1mb of memory for loading the database. However, some
databases demands more than 1mb, and then (before opening the file) we need to
indicate how much space is needed (using the command "set mem"). For instance, if
we want to allocate 5 megabytes of memmory:
. set mem 5m
Preserve and Restore
As you may have noticed, there is no "Undo and Redo" in Stata. But you can use
"preserve" to "save provisionally" a database. Then, if you want to "undo", you can
execute "restore" and go back to the previous state.
. preserve
. drop
Ups! You dropped all the observations. However, you can go back to "preserve":
. restore
Log-File
The log-files are useful to keep record of everything that appears in the "results"
window. A log-file records both the history of commands and the history of outputs.
. log using test, replace
The option "replace" replace the existing file. When the session is finished, you must
close the log-file:
. log close
You can open a log file using the Notepad or using the option "File Log View".
Do-File
The do-files are extremely useful in Stata. A do-file is an unformatted text file
(ASCII) that contains a sequence of Stata commands. Stata interprets them exactly as
if they were entered in the command window. Then you can save code lines and time
when you want to repeat an entire piece of code with a minor modification.
Let's begin with a classis of programming: making Stata say Hello. The
corresponding command is:
. display "Hello"
Then, if you want to create a do-file with the above code, simple open the do-file
editor (you can even use the Notepad) and enter:
. display "Hello"
Remember always to add an "Enter" at the end! Then save the file as "hello.do" in the
Stata flder. If you want to execute the do-file, you must use the commmand "do":
. do hello
You will understand the advantages of using do-files as soon as you begin working
with them for the Problem Sets. Sometimes you want other people to understand or
use your piece of code, or maybe you know that you may access to it in the future.
For that sake it is vary useful to include comments in the do-file for describing what
you are doing in each group of lines. For inserting an entire line of comments you
must use an asterisk at the beginning of the command line:
. * This is a comment, write whatever you want
. * And if you need a further line, just begin it with an asterisk
The "/* text */" comment delimiter has the advantage that it may be used in the
middle of a line. What appears inside /* */ is ignored (it cannot be used in the
command window, as it will only work in a do-file). The "//" comment indicator may
be used at the beginning or at the end of a line:
. describe */ text /* var1-var10
. describe // test
Exercise 1.1: After finishing this week's notes open the do-file editor and create a
do-file called "week1" to reproduce all the commands.
Help and search
As in every programming environment, the command "help" has a special importance,
as it gives detailed information about the commands under consideration. For
instance, write the following and press Enter:
. help summarize
A window with information on the command "summarize" will appear. If you want to
do something, but you do not know which the right command is, you should use
"search". For instance, if you want to find the mean of a variable, you may enter the
following:
. search mean
Then just choose the appropriate option.
Commands
As indicated by the help command, there is a generic command structure for the
mayority of the Stata commands.
[by varlist:] command [varlist] [=exp] [if exp] [in range]
[weight] [using filename] [, options]
For any given command, some of these components may not be available. In the
help-file you may find links to information on each of the components:
[by varlist:] instructs Stata to repeat the command for each combination of values in
the list of variables varlist. For instance, "by location" would repeat the command for
the set of observations with each value for the variable "location".
[command] is the name of the command and can be abbreviated. For instance, the
command "summarize" can be abbreviated as "sum" and the command "regress" can
be abbreviated "reg".
[varlist] is the list of variables to which the command applies. There are some
shortcuts. For example, instead of writing "var1, var2,, var9" you can write "var*",
or "var1-var9". Alternatively, if you are interested in listing the variables from
"var1john" through "var9john" you can use simply "var?john" (as in the old DOS).
[=exp] is an expression.
[if exp] restricts the command to that subset of the observations that satisfies the
logical expression "exp". For instance, "height>170" restricts the command to those
observation with "height" greater than 170.
[in range] restricts the command to those observations whose indices lie in a
particular range. The range of indices is specified using the syntax f/l (for first to
last) where f and/or l may be replaced by numerical values if required, so that 5/12
means fifth to twelfth and f/10 means first to tenth. Negative numbers are used
to count from the end, for example: list var in -10/l lists the last 10 observations.
[weight] allows the weighting of observations.
[using filename] specifies the filename to be used.
[options] are specific to the commands.
Version
Some commands change a lot from version to version. If you want to execute a code
from a previous version (e.g. Versin 7.0), you can do so by using the command
"version" at the beginning of the code:
. version 7.0
Data Management
Variables
In Stata there are two types of variables: string and numeric. Subsequently, each
variable can be stored in a number of storage types (byte, int, long, float, and
double for numeric variables and str1 to str80 for string variables).
If you have a string variable and you need to generate numeric codes, you may find
useful the command "encode". For instance, consider a variable "name" that takes
the value "Nick" for the observations belonging to Nick, "John" for the observations
belonging to John, and so forth. Then you may find useful the following:
. encode name, gen(code)
A numeric code-variable will be generated (e.g. it takes the value 1 for the
observations belonging to Nick, the value 2 for the observations belonging to John,
and so forth).
Missing values in numeric variables are represented by dots. Some databases have
other special characters for missing values, or maybe particular numbers (9, 66, 99,
etc.). Missing value codes may be converted to missing values using the command
"mvdecode". For instance, if the variable "gender" takes the value 9 if missing value,
then enter:
. mvdecode gender, mv(9)
It will replace by dots all values of variable gender equal to 9.
Let's work
In the website of the course (??) you will find some databases (grouped by weeks)
that will be used throrought this notes. I will not mention any more that you first
have to download the database (and save it in the Stata folder or whatever folder
you are working in).
Now we will use the database "russia.dta". It is a compilation of health, economic
and welfare variables from the Russian Longitudinal Monitoring Survey (RLMS;
information at the official website: www.epc.unc.edu/projects/rlms).
. use russia.dta, clear
Describe, list, browse and edit
In order to know basic information on the variables allocated in memmory, use:
. describe
You can see the names, storage types, display formats, labels (e.g. "For how many
years" for the variable yr) and value labels (e.g. "Male" if gender==1 and "Female" if
gender==0). If you want to see information for only some variables, simply list them:
. describe totexpr gender
You can list some observations. For instance, the following code list the gender and
total expenditure for the first 10 observations:
. list gender totexpr in 1/10

The sequence "Ctrl+Alt+Break" interrupts the execution of a command (you can also
use the red button with a white cross in the top toolbox).
If you want to see the database in a spreadsheet-fashion, you may try:
. browse
As you might notice, you cannot edit the information. But if you use the command
"edit" you will be able to do so:
. edit
In a while we will study some commands meant to describe the data. But in the
meantime you can try the command "summarize"(or "sum"):
. summarize
It gives means, standard deviations, maximum and minimum values for each variable.
As always, if you want to see the information for a particular group of variables, then
you simply have to list them:
. summarize gender belief
Sort
The command "sort" sorts the observations according to one or more variables
(numerically and/or alphabetically). Some commands require the data to be sorted
first (such as "by" and "merge"). For instance, you can sort the individuals by their
real expenditures and real income, in that order:
. sort totexpr totincm_r
Drop and keep
The command "drop" eliminates either variables or observations. For instance:
. drop belief
That drops the variable "belief". If you want to delete observations, you can enter:
. drop if gender==1
That will delete observations for males. Alternatively:
. drop in 1/10
That will delete the first ten observations. The command "keep" is just the opposite:
it deletes everything but what you list. For instance, "keep if gender==1" would
delete every observation for females.
Rename and label

The command "rename" change the names of variables:
. rename gender sex
The command "label" allows us to label variables, where we can indicate more
precisely their descriptions:
. label variable ortho "=1 if individual professes Orthodoxy Religin, =0 otherwise"
Gen and replace
If you want to generate a dummy variable, you may find the following piece of code
very useful:
. gen midget=(height<150)
If you put a condition in parenthesis, then the parenthesis takes the value 1 if the
condition is true, and 0 otherwise. Then the above code line assign a 1 to the
variable "midget" if "height"<150, and zero otherwise (if "height" is missing, then
"midget" will be missing as well).
The logical operators are "&" (and) and "|" (or), and the conditions may involve "=="
(equal), "~=" or "!=" (different), ">" (greater), and ">=" (equal or greater). You can
create variables as complex as you want:
. gen midget1 =((height<150 & gender==1) | (height<140 & gender==0))
The command "replace" replaces values of existing variables. For instance, the
following line of code put a 0 in "midget" if the individual is obese:
. replace midget=0 if obese==1
Simple transformations
You can generate variables using mathematical and statistical functions: cos, ln,
log10, sqrt, max, min, floor, round, sum, etc. For instance, generate a variable
containing the natural logarithm of the household total expenditure:
. gen lntotexpr=ln(totexpr)
Extended Generate
The command "egen" (extended generate) is useful when you need to create a
variable that is the mean, meidan, standard deviations, etc. of an existing variable.
For instance, I can create a variable that takes the mean life satisfaction over the
entire sample:
. egen mean_satlif = sum(satlif)
10
Or I can create variable that takes the mean life satisfaction over their geographical
sites:
. egen site_mean_satlif =mean(satlif), by(site)
The command "egen" have other useful option. We can use the option "group" to
group the population according to different combinations of some variables. For
instance, I would like to identify the individuals according to if they "smokes" and if
they are "obese". As both categories are binary, the command "group" will generate
four possible categories (people that smoke and are obese, people that don't smoke
and are obese, people that smoke and are not obese, and people that don't smoke
and are not obese):
. egen so=group(smokes obese)
Let's see the results:
. browse smokes obese so
Label values
We can put labels to the different values of "so":
. label values so solabel
. label define solabel 1 "Smokes:NO Obese:NO" 2 "Smokes:NO Obese:YES" 3
"Smokes:YES Obese:NO" 4 "Smokes:YES Obese:YES"
You may see the results using the command "tabulate":
. tab so
If you want to modify the value labels, you need to drop the old label and create a
new one:
. label drop solabel
. label values so solabel
. label define solabel 1 "Skinny and clean" 2 "Clean but fatty" 3 "Skinny smoker " 4
"Fatty smoker"
Let's see the results again:
. tab so
By and bysort
The option "by" indicates that the command must be run for many groups of
variables. Some commands use by as "by xx: command", and other command use it as
11
in "command, by xx". Before running "by" the data must be sorted by the variable
after the "by":
. sort geo
. by geo: count if gender==1
The command "count if gender==1" counts the number of observations for men. Then,
adding "by geo" make Stata count the number of men inside each geographical area
(for geo==1, geo==2 and geo==3). If you find uncomfortable to sort the data first, you
can use "bysort" directly:
. bysort geo: count if gender==1
Append
Using "append" you can add observations to database using another database (i.e. you
can "append" one to another). In the database "week1_2" there are 100 observations
on individuals from the Russian Household Survey, and in the database "week1_3"
there are 100 additional observations. You can browse both databases:
. use week1_2, clear
. browse
. browse
We want to "paste" one database below the other. You must open the first database:
And then using "append" you add the observations of the second database:
. append using week1_3
Finally, you can see the results:
. browse
Collapse
The command "collapse" generates a "smaller" database contaning the means, sums,
standard deviations, etc. of the original dataset.
. use russia.dta, clear
We can take the means of life satisfaction ("satlif") and economic satisfaction
("satecc") within geographical sites ("site"):
. collapse (mean) satlif satecc, by(site round)
12
You can now see the final product:

. describe
. browse
Instead of having thousands of observations on several variables now we have one
only observation per geographical site (the mean) for only two variables (life
satisfaction and economic satisfaction).
If one variable has more missing values than the other, then the means, standard
deviations, etc. will be based in different sample sizes. If you want to avoid that, you
can use the option "cw" (casewise deletion), which forces Stata to eliminate any
observations without data for every variable involved in "collapse".
Reshape
Suppose we have j=1J variables on each individual i=1I. This information may be
viewed in two ways. Firstly, each variable j may be represented by a variable xj and
the individual identifier may be a variable "individual_id". However, we may need
one single response vector containing the responses for all variables for all subjects.
These two "data shapes" are called wide and long, respectively.
For instance, the database "week1_3" has a wide shape:
name
money1990
money1991
money1992
John Doe
10
12
15
Tom Twain
Tim Besley
25
20
18
Louis Lane
14
14
11
Use the following code to transform the database into a long shape:
. reshape long money, i(name) j(month)
13
. browse
name
year
money
John Doe
1990
10
John Doe
1991
12
John Doe
1992
15
Louis Lane
1990
14
Louis Lane
1991
14
Louis Lane
1992
11
Tim Besley
1990
25
Tim Besley
1991
20
Tim Besley
1992
18
Tom Twain
1990
Tom Twain
1991
Tom Twain
1992
You can also reshape it again back to wide:

. reshape wide money, i(name) j(year)
Merge
If you want to join two files for the same individuals but with different sets of
variables, then you must use the command "merge":
. merge using filename
Instead of adding new observations (as in "append"), you add new variables joining
corresponding observations from the dataset currently in memory (called the master
dataset) with those from Stata-format datasets stored as "filename" (called the using
dataset). As always, if filename is specified without an extension, then ".dta" is
assumed.
A "one-to-one merge" simply "pastes" both datasets side by side. If you want Stata to
match each observation in the master dataset with the corresponding observation in
the using dataset (e.g. observations for the same individual), you must perform a
"match merge". You should have a key variable (e.g. the individuals' names). After
merging the datasets, Stata will create a variable called "_merge" to indicate the
result of the merging for each observation: 1 for the observations from the master
dataset that did not match with the using dataset; 2 for the observations from the
using dataset that did not match with the master dataset; 3 for the successful
matches. Enter "help merge" for further details.
14
Exercise 1.2: Generate a do-file to carry out the following: using the dataset
"russia.dta", generate an id-variable for the observations (individuals); divide it in
two separate databases, each one with different sets of variables; merge those
datasets back together.
Descriptive Analysis
We will see some commands to describe the database, which will involve creating
complex tables and figures. There may many objectives in doing so. For instance, you
can show the motivations of a study, a difficult and important task. In the Problem
Sets we will ask you to reproduce a lot of tables and figures from published papers.
It is also important to show descriptive statistics for letting the reader have a
quantitative idea of the parameters estimated by the econometric model. Using
descriptive statistics you can also check the internal consistency of the data (e.g. can
an individual be 453 years old? Can an individual consume a negative number of
cars?). Additionally, you can provide evidence in favor of the external validity of the
model (e.g. are the individuals surveyed representative of the entire population?).
Summarize and tabstat
The command "summarize" shows basic descriptive statistics:
. summarize totexpr
The option "detail" adds different percentiles to the table:
. summarize totexpr, detail
The option "pweight" weights the observations by using the inverse of the probability
of entering the sample:
. summarize totexpr [pweight=inwgt], detail
The command "tabstat" shows specific statistics:
. tabstat hswrk, stats(mean range)
Some arguments for "stats()" are: mean, count, sum, max, min, range (max-min), sd,
var, etc. The option "by(varname)" makes "tabstat" build a table with descriptive
statistics for each value of "varname":
. tabstat totexpr, stats(mean range) by(gender)
The command "ci" makes an confidence interval for the mean of a variable at a given
statistical level of confidence:
. ci totexpr, level(95)
15
Tabulate, tab1 and table

The command "tabulate" produces frequency tables for numeric variables:
. tabulate econrk
Using "tab2" you can produce two-way tables of frequency counts, along with various
measures of association:
. tabulate econrk powrnk
Some option for "tabulate" are cell (shows percentages for each cell), column (show
percentages by columns), missing (includes missing values as a separate value).
You can create a set of dummies starting from a discrete variable. For example, the
variable "belief" takes the values from 1 to 5. Enter:
. tabulate belief, gen(belief)
That generates 5 dummies: belief1 (that takes value 1 if belief=1 and 0 otherwise) to
belief5 (that takes value 1 if belief=5 and 0 otherwise).
You can also produce three-way table of frequency, using the option "by":
. bysort gender: tabulate econrk powrnk
The command "table" is to "tabulate" what "tabstats" is to "summarize". You can ask
for specific information for the tables of frequency:
. table econrk, contents(freq)
. table econrk powrnk, contents(freq)
. table econrk powrnk gender, contents(freq)
And you can also create "super-tables":
. table econrk powrnk gender, contents(freq) by(smokes)
Some option for "contents()" are: "freq" (frequency); "mean varname" (mean of
varname), "sd varname" (standard deviation of varname), and so on.
Assert
We can detect errors using the assert command. For instace, we know that gender
can only take values 0 or 1 (besided missing values):
. assert gender==0| gender==1| gender==.
Scatterplot
You graph a "scatterplot" entering "graph twoway scatter" followed by the y-variable
and the x-variable, respectively:
16
. graph twoway scatter htself height, by(gender)

If you enter more than one x-variable, then Stata will generate in the same graph
two scaterplots (in different colors) with the two combinations between the yvariable and both x-variables:
. graph twoway scatter waistc htself height, by(gender)
There are many options common to all graphs (i.e. options for "graph twoway"). For
instance, the option "title()" creates a title for the graph, and there are many
additional labels to be defined:
. graph twoway scatter htself height, title(This is the title) b2(This is bottom 2)
l1(This is left 1) l2(This is left 2) t1(This is top 1) t2(This is top 2) r1(This is right 1)
r2(This is right 2)
Enter "search axis" or "search title" for further details. You will discover much of
those options when starting to work with graphs seriously. Enter "help graph twoway"
to see other two-way graphs available.
Histogram
Use the command "hist":
. hist height, title(Heights) by(gender)
Nonetheless, the histograms are very sensitive to the parameters used for their
construction. Thus, you are strongly encouraged to (carefully) provide them by
yourselves: "bin(#)" (number of bins), "width(#)" (width of bins) and "start(#)" (lower
limit of first bin). The option "discrete" indicates to Stata that the variable under
consideration in categorical (it is very important to do so):
. hist belief, title(Trust in God) discrete
If you didn't, your graph would look like this:
. hist belief, title(Trust in God)
The option "normal" adds a normal density to the graph, and the option "freq" shows
the frequencies instead of the percentages:
. hist height, width(2) start(140) norm freq
Boxplot
. graph box height, by(site)
Pie chart
17
. graph pie totexpr, over(site)
Symmetry Plots
. symplot height
QQ Plots
. qnorm height
Some Programming in Stata

Foreach and forvalues
Stata has the typical commands from standard programming. For instance,
sometimes you have to repeat the same piece of code several times, which usually
would imply a lot of typing. The command "for" is very useful in those ocations.
Suppose you want to run several mean-tests for smokes. You may write the test
several times:
. ttest smokes==0.1
. ttest smokes==0.15
. ttest smokes==0.2
. ttest smokes==0.3
. ttest smokes==0.4
. ttest smokes==0.5
Or you can use the "for" syntax:
. forvalues k = 0.1(0.05)0.5 {
. ttest smokes==`k'
.}
The `k' in the "for" syntax represents the values of the parameter within the given
range (from 0.1 to 0.5, taking 0.05-steps). You can also use the "for" syntax for a list
of names (using "foreach"):
. foreach file in file01.dta file02.dta file03.dta {
18
use `file', clear
sum smokes
.}
Ado-files
We can make Stata execute a list of command using do-files. For instance, we made
Stata say hello running the do-file "hello.do". Alternatively, you can create a program
to run a code. For instance, let's create a program called "hello.ado" to make Stata
say hello:
. program define hello
1. display Hello
2. end
Nothing happened, because we have to run the new program first:
. hello
Some Macros
The "macros" stores many useful values. For instance, the macro "_n" stores the
number of the current observation. For instance, the following line of code lists the
first nine observations:
. list totexpr if _n<10
On the other hand, the macro "_N" stores the total number of observations. For
instance, the following line of code shows the last observation within each
geographical site:
. bysort site: list totexpr if _n==_N
You can create a variable based on the macros. For instance:
. clear
. set obs 100
. generate index = _n
. browse
Generate a variable "x" with random numbers distributed uniformly between zero and
one:
. gen x=uniform()
Then, you can generate lagged values:
19
. generate xlag = x[_n-1]

. browse
It may be interesting to obtain lagged values in a panel of individuals. For instance,
see the following database:
. browse
There are 5 (already sorted) observations for 5 different individuals. You must
generate the lagged values for "x" entering:
. bysort N: generate xlag = x[_n-1]
. browse
Please convince yourself that you cannot use "generate xlag = x[_n-1]".
Return and ereturn
After "summarize" and many other commands, the results are stored in "macros". You
can retrieve them entering "return":
. sum smokes
. return list
For example, the mean for "smokes" has been stored in "r(mean)". We can use that
information in the future:
. display r(mean)
. gen mean_smokes = r(mean)
For some commands you must enter "eretunr list" (estimation commands) or "sreturn
list" (commands that assist in parsing).
Quietly, capture and more off
The capture prefix makes Stata continue running the do-file even if the command
throws an error. On the other hand, the prefix quietly eliminates all output but error
messages. For instance, if you want to retrieve the mean for smokes, you may enter:
. quietly: sum smokes
. quietly: return list
. display display "The mean for smokes is "r(mean)
Exercise 1.3: Try each one of the following inside a do-file, and comment the
results:
20
. capture: blablabla // This is wrong

. capture: sum smokes // This is right
. quietly: blablabla
. quietly: sum smokes
Set more
The command "set more off" causes all the output to scroll past automatically instead
of waiting for the user to scroll through it manually. The command "set more on"
reverses it.
21
Chapter 2
Least Squares
Creating Random Data and Random Samples

Before starting the exercises on least squares, you must learn how to create random
samples. To "generate" observations from nothing, you must use "set obs":
. set obs 100
If you want to generate a variable "number" containing random numbers distributed
uniformly between 0 and 1, enter:
. gen number=uniform()
The function "uniform()" returns uniformly distributed pseudorandom numbers on the
interval [0,1). Entering either "browse" or "list" you will be able to see the generated
observations. If you want to generate uniformly distributed pseudorandom numbers
on the interval [a,b), you must enter "gen var=uniform()*(b-a)+a". For instance, to
generate numbers on the interval [3,7):
. gen number_1=uniform()*4+3
Check the desired properties (max, min and mean) entering "sum number_1". If you
want to generate integers, then you must multiply the numbers by a multiple of ten
and then truncate them towards zero (using "int"). For instance, to create integers
between 0 and 9:
. gen number_2=int(uniform()*10)
Use "list" to see the results. Alternatively, if you want to generate random numbers
with different distributions, you have to evaluate the inverse density function at
"uniform()". For instance, using "invnormal(uniform())" returns normally distributed
random numbers with mean 0 and standard deviation 1. Sum and multiply
accordingly in order to obtain different means and deviations.
Furthermore, if you want to repeat the same "succession" of pseudo-random numbers
each time that you execute a do-file, then you must initialize the "seed":
. set seed 339487731
For instance, generate ten random numbers twice:
. clear
. set obs 10
22
. gen n=uniform()
. gen n1= uniform()
. list
And then repeat the process using the same seed:
. clear
. set obs 10
. set seed 11223344
. gen n=uniform()
. set seed 11223344
. gen n1= uniform()
. list
A "fictional" example
We will create values for some variables, using the "actual" values of the linear
parameters involved. Then we will try to retrieve those parameters using OLS, what
will let us experiment with some basic properties.
Let's generate i.i.d. data on wages, education, intelligence, two explanatory
variables uncorrelated with education and intelligence but correlated with wages (a
and b), and finally a variable (c) totally uncorrelated with all the former variables.
. clear
. set obs 100
The variable intelligence will be the IQ of the individuals. IQs have approximately a
normal distribution centered in 100 with a standard deviation of 20:
. gen intelligence=int(invnormal(uniform())*20+100)
Notice that we have truncated the decimal part of the numbers. Since more
intelligent people is expected to study more (see the original model of Spence on the
signaling purpose of education), the years of education will be equal to the
intelligence (over 10) plus a normally distributed noise with mean 0 and deviation 2.
Finally, we will keep only the integer part of the numbers:
. gen education=int(intelligence/10+invnormal(uniform())*2)
I will stop repeating "enter browse to see the results". Then, feel free to do so
whenever you want. Variable a (b) will be normally distributed with mean 10 (5) and
23
standard deviation 2 (1). Variable "c" will be normally distributed with mean 15 and
standard deviation 3.
. gen a=int(invnormal(uniform())*2+10)
. gen b=int(invnormal(uniform())*1+5)
. gen c=int(invnormal(uniform())*3+15)
Finally, the unobserved error term "u" will be normally distributed with mean 7 and
standard deviation 1:
. gen u=int(invnormal(uniform())*1+7)
Wages will be the result of "intelligence" multiplied by 3, plus variables "a" and "b"
multiplied by 1 and 2 respectively, plus the unobserved error term "u":
wagei = 3intelligencei + 1ai + 2bi + ui

. gen wage=3*intelligence+a+2*b+u
We estimate the "right" equation. The command for OLS is "reg" followed by the
dependent variable and then the list of explanatory variables. We will include the
option "robust", which indicates the use of robust variance estimates:
. reg wage intelligence a b, robust
The estimated coefficients are accurately near the true values. Notice that
"education" does not "affect" wages. Then, if we included "education" and
"intelligence" in the regression, then the former should not appear with a
significative coefficient:
. reg wage education intelligence, robust
Notwithstanding, education is correlated with intelligence. Thus, if we forgot to
include "intelligence" then the coefficient on "education" would be different from
zero at reasonable confidence levels:
. reg wage education, robust
The reason is that in the last equation "intelligence" is in the error term (because it
"causes" wages but it is not included in the regression), and "intelligence" is
correlated with "education". Thus, the orthogonality condition is not satisfied.
= ( X ' X )1 X ' Y
= ( X ' X )1 X '( X + u )
= + ( X ' X ) 1 X ' u
E = + E ( X ' X )1 X ' u
24
= + E ( X ' X )1 X ' E (u | X )
Which equals because E (u | X ) = 0 . A similar result holds with consistency (i.e.
for large sample).
Let's see that the exclusion of "a" and "b" does not violate the exogeneity condition.
Since "intelligence" is not correlated with "a" and "b", its coefficient should remain
consistent and unbiased:
. reg wage intelligence, robust
Nonetheless, including "a" and "b" should decrease the standard deviation of the
coefficient on "intelligence":
Conversely, due to their independence, including "a" and "b" but excluding
"intelligence" should not affect the consistence of the coefficients on the former:
. reg wage a b, robust
Finally, let's see the effect of including an "irrelevant" variable ("c") in the "right"
equation:
. reg wage intelligence a b c, robust
Compared to the "right" equation", the loss of one degree-of-freedom is irrelevant in
this case:
Taking advantages of do-files and macros
We can create a do-file including all the previous exercise:
. clear
. set obs 100
. gen intelligence=int(invnormal(uniform())*20+100)
. gen education=int(intelligence/10+invnormal(uniform())*2)
. gen a=int(invnormal(uniform())*2+10)
. gen b=int(invnormal(uniform())*1+5)
. gen c=int(invnormal(uniform())*3+15)
. gen u=int(invnormal(uniform())*1+7)
. gen wage=3*intelligence+a+2*b+u
. ()
25
You can also set a seed in order to keep your results "tractable".
Exercise 2.1: Repeat the do-file including minor modifications to see the following:
a. That an increase in the sample size implies a decrease in the standard errors.
b. What happens if you increase the variance of "u".
c. In the real world "u" is by definition something we cannot measure nor observe.
We estimate the coefficients using that "u" is orthogonal to the included regressors. If
we estimated "u" (as the residual of the regression), we would find that it is exactly
orthogonal to the included regressors. But in this fictional world you know "u", and
then you can calculate the error term in each equation and then you can test the
orthogonality condition.
d. Include a measurement error in the "observed" intelligence (e.g. a normally
distributed noise). Then observe that the estimated coefficient is downward biased.
A "real" example
It is time to use the command "regress" with real data. We will use the database
"russia.dta" we used last week. It is a compilation of health, economic and welfare
variables from the Russian Longitudinal Monitoring Survey for 2600+ individuals in
2000. At the end of the first Problem Set you may find data definitions.
. use russia, clear
Suppose that you want to explain the Health Self-Evaluation indexes ("evalhl"; the
larger the healthier) using the following variables:
. reg evalhl monage obese smokes
Notice that Stata automatically puts the constant. If you wanted to exclude it, you
would have to enter "nocons" as option:
. reg evalhl monage obese smokes, nocons
As any other command in Stata, "regress" can be applied to a subset of the
observations. Suppose you want to run two separate regressions, one for males and
the other for females, respectively:
. reg evalhl monage obese smokes if gender==1
. reg evalhl monage obese smokes if gender==0
There are many options for "regress" (enter "help regress"). Possibly the most
important is the option "robust", which uses a robust estimate for the variances-andcovariances matrix:
. reg evalhl monage obese smokes if gender==0 & round==9, robust
26
Exercise 2.2: Why "robust" is not the default option? Since homocedasticity is a
particular case: Does it seem logic to you? Experiment with some regressions: using
the "homocedastic coefficients", are you over-rejecting or under-rejecting the null
hypotheses?
Besides "robust", there are other options regarding the estimation of standard errors:
"cluster()" (adjusts standard errors for intra-group correlation) and "bootstrap"
(Bootstrap estimation of standard errors).
Tests
After each estimation Stata automatically provides t-tests (for linear regressions) or
z-tests (for nonlinear models) of the null hypothesis whether the coefficients are
zero. Notwithstanding, other hypotheses on the coefficients can be tested.
For instance, using the command "test" (after the regression was run) you can test
whether the effect of being obese equals -0.05:
. reg evalhl monage obese smokes gender, robust
. test obese=-0.05
We can test hypotheses that involve more than one variable. For instance, we can
test if the coefficients on "smokes" and "obese" are both null, or if the sum of the
coefficients on "obese" and "gender" equals 1:
. test smokes obese
. test obese + gender == 1
Predictions and residuals

After every estimation command (e.g. reg, logit, probit) some "predicted" values can
be calculated: fitted values, residuals, influence statistics, etc. Those values will be
stored in a new variable using the command "predict". For instance, in the last model
the predictions are:
. reg evalhl monage obese smokes gender, robust
. predict yhat
Enter "browse evalhl yhat" to appreciate the fit of the model. Alternately, the
residual can be obtained:
. predict res, residual
The relationship between the actual values, the predicted values and the residuals
can be shown in a scatter plot:
27
. graph twoway scatter yhat res evalhl
Partial regression plot

The command "avplot" graphs the relationship between the dependent variable and
one of the explanatory variables conditional on the rest of the regressors. It is very
useful to identify outliers. For instance, we will collapse the database to obtain the
means for some variables within each geographical site (enter "tab site" to see the
geographical distribution of people). We will use "preserve and restore" to avoid
loading the database again:
. preserve
. collapse (mean) satlif totexpr satecc powrnk, by(site)
We run a regression of life satisfaction on three variables:
. reg satlif totexpr satecc powrnk, robust
Finally, using the command "avplot", we show the partial relationship between
"satecc" and "satlif", identifying each geographical site:
. avplot satecc, mlabel(site)
. restore
Normality of the residuals

The multiple regression model does not require normality of the residuals to make
inference with large samples (depending on the version of the limit central theorem
used, it only needs error terms i.i.d.). Nonetheless, normality is needed to make
inference in small samples.
After running a regression, we can predict the residuals:
. reg waistc monage height hipsiz gender
As a first step, we can graph the nonparametric estimate of the density function of
the residuals using Kernels:
. kdensity res, norm
The option "norm" graphs a Gaussian bell with the sample mean and variance. We
could also use the commands "qnorm" and "pnorm" to evaluate graphically the
normality of the residuals. Some tests for normality, such as the Shapiro-Wilk and
Shapiro-Francia ("swilk"), can also be useful.
Tests for heteroskedasticity

28
The command "imtest" performs the White's test, and the command "hettest"
performs the Breusch-Pagan's test. In both tests the null hypothesis is whether the
variance of residuals is homogeneous.
Nevertheless, you should be very careful. These tests are pretty sensitive to the
assumptions of the model (for instance, they suppose normality for the error term).
"By hand" regression

Since we will use mathematic elements, we need to maximize the room reserved for
those objects:
. set matsize 800
Recall the OLS estimator. Write the exogeneity condition:
1
N
u x
i i
=0
Then use the definition:
yi = xi + ui ui = yi xi
1
( yi xi ) xi = 0
N
( yi xi xi xi ) = 0
And finally you get:
= ( xixi )1 xi yi
= ( X ' X ) 1 X ' Y
As Stata do not allow for more than 800 rows or columns for matrixes, it would be
impossible to work directly with X or Y (as they have +2600 rows). But there is a
rather simple trick: the command "mat accum". It executes an intermediate step,
X'X, which creates a "small" matrix. First we have to eliminate the observations with
missing values in the variables that will be included in the model (why?):
. drop if evalhl==. | monage==. | obese==. | smokes==.
Then we run the regression in the "traditional" way:
. reg evalhl monage obese smokes
Let's begin with calculating X'X and storing it in the matrix "XpX":
. mat accum XpX = monage obese smokes
You can see the result entering:
. mat list XpX
Calculate X'Y and store it in the matrix "XpY":
29
. mat vecaccum YpX = evalhl monage obese smokes

Then transpose it (using an aphostrophe) to obtain X'Y:
. mat XpY = YpX'
. mat list XpY
Finally, we can get using the above formula:
. matrix beta = invsym(XpX)*XpY
. mat list beta
Exercise 2.3: Calculate "by hand" the R-squared, the standard deviations and the
corresponding p-values.
Regressions' output
You need to show the regressions in tables similar to those utilized in most
economics papers (e.g. one column per specification, and standard errors in
parentheses). Use then the command "outreg". As it is not a "default" command in
Stata, you need to install it first:
. search outreg
Select "sg97.3" and then press "click here to install".
We run a regression:
. reg evalhl monage obese smokes satlif totexpr
And then we save the output in the file "regresion.out":
. outreg using regresion, replace
The option "replace" indicates to Stata that if the file "regresion.out" exists, then it
must be replaced. Go to the Stata folder and open "regresion.out" (using MS Excel or
MS Word).
By default "outreg" shown in parentheses the t-values and puts an asterisk if the
coefficient is significative at the 5%, and two asterisks if it is significative at the 1%.
We can ask for the standard errors in parenthesis, and one asterisk if the coefficient
is significative at the 10%, two asterisks if it is significative at the 5%, and three
asterisks if it is significative at the 1% (options "se" and "3aster" respectively):
. outreg using regresion, se 3aster replace
The command can also show various regressions in the same table (as columns). We
must add regression outputs using the option "append" instead of "replace". Let's run
30
three regressions: one with only "monage" and "obese" as explanatory variables;
another with only "smokes", "satlif" y "totexpr"; and finally a regression with all them.
. reg evalhl monage obese
. outreg using regresion, se 3aster replace
. reg evalhl smokes satlif totexpr
. outreg using regresion, se 3aster append
. reg evalhl monage obese smokes satlif totexpr
. outreg using regresion, se 3aster append
Open using MS Excel the file "regresion.out" and see the results. If you do not want to
show the coefficient for a set of dummy variables, you will find the command "areg"
very useful.
31
Chapter 3
Instrumental Variables
Open the database:

. use russia, clear
Instrumental Variables
Consider a least squares regression on the determinants of self-reported health
evaluation:
. reg evalhl alclmo cmedin belief operat obese smokes totexpr monage if gender==0,
robust
Many variables can be endogenous. Focus on the exogeneity of a particular variable:
"obese" (a dummy indicating whether the individual is obese or not). For instance,
maybe there is some simultaneous causality: having an "upward biased" health self
evaluation (e.g. I feel healthier than I am) may justify being a little more obese.
Then I propose the individual's height ("height"), waist circumference ("waistc") and
hip circumference ("hipsiz") as instruments for "obese". Take as an example the waist
circumference. A person (conditionally on obesity) should not feel less healthy if he
has a larger waist circumference. If he has a larger waist circumference simply
because he has a tougher body, then it should not affect his health. However, if the
larger waist circumference responds to an obesity problem, then it will (only
indirectly) impact health evaluation. I try to exploit (supposedly) exogenous
variations in body characteristics. The relevance condition obviously holds: obesity is
significantly correlated with the three variables proposed as instruments.
For instrumental-variables (two-stage least-squares) regression the command is
"ivreg", followed by the dependent variable and the list of regressors. However, the
endogenous regressors must be put in parentheses along with the instruments
(separated by an "="). For instance:
. ivreg evalhl alclmo cmedin belief operat (obese =height hipsiz waistc) smokes
totexpr monage if gender==0, robust
For showing the first stage, add "first" as an option:
totexpr monage if gender==0, robust first
32
There we can see that taller people and people with larger waists and hips are
probably more obese.
The two-stages "by hand"

Let's write the GIVE (Generalized Instrumental Variables Estimator) making the usual
assumption that the disturbances have a scalar covariance matrix:
1
b2 SLS = X ' Z ( Z ' Z ) 1 Z ' X X ' Z ( Z ' Z ) 1 Z ' Y

Call X * = Z ( Z ' Z ) 1 Z ' X to the fitted values from an OLS regression of X on Z. Notice
that X ' Z ( Z ' Z ) 1 Z ' X = X ' X * = X *' X * and X ' Z ( Z ' Z ) 1 Z ' y = X *'Y . Thus, the IV
estimator can also be written:
b2 SLS = ( X *' X ) 1 X *'Y = ( X *' X * ) 1 X *'Y

The last expression is an OLS regression of Y on X * (the latter being the residual
from a previous regression). This explains why this estimator is also called two-stage
estimator (2SLS). We will perform the two stages step by step. Firstly, we must drop
every observation with missing values in at least one of the variables utilized. This
avoids using information from an individual only in the first stage:
. drop if evalhl==. | alclmo==. | cmedin==. | belief==. | operat==. | obese==. |
height==. | hipsiz==. | waistc==. | smokes==. | totexpr==. | monage==.
Run the first stage:
. reg obese alclmo cmedin belief operat smokes totexpr monage height hipsiz waistc
if gender==0, robust
Then save the fitted values as "obese_hat":
. predict obese_hat
And then run the second stage:
. reg evalhl alclmo cmedin belief operat smokes totexpr monage obese_hat if
gender==0, robust
Comparing with the table obtained with the command "ivreg", you will see that (as
expected) the values for the slopes are the same. However, the estimates for the
standard deviations of those slopes are inconsistent (you cannot use directly the
information on "obese_hat").
Hausman Test
33
You must notice that you cannot test the instruments' exogeneity (because they are
assumptions!). But given that the instruments are valid, you can test the
instrumented variables' exogeneity.
You must save the IV estimates:
. est store iv
You must also save the OLS estimates:
. reg evalhl alclmo cmedin belief operat obese smokes totexpr monage if gender==0,
robust
. est store ols
Then use the command "hausman" indicating first the consistent estimates and then
the efficient estimates:
. hausman iv ols
The null hypothesis is that there is not systematic difference between the estimates.
Exercise 3.1: If you reject the null, which regression should you run? Why? What if
you do not reject the null hypothesis? If I already knew for sure that the instruments
are valid, why should I care for this?
There is an alternative way to run the Hausman test. You must run the fist stage, and
then you must save the residuals:
Then run the original least squares regression, but including the residuals:
. reg evalhl alclmo cmedin belief operat smokes totexpr monage obese res if
gender==0, robust
First notice that the coefficients (though not the standard errors) are the same than
those obtained using "ivreg":
34
Finally, the Hausman test consists in testing if the coefficient on the residuals ("res")
is null: if rejected, then (given that the instruments are valid) "obese" are
endogenous.
Test of Overidentifying Restrictions

If the model is overidentifyed (the number el nmero de condiciones de momentos es
mayor al nmero de coeficientes a ser estimados), then we can test whether some of
the moment conditions are invalid (although the test will not indicate which
conditions are invalid). There are two well-known tests: Sargan test and J-test.
The Sargan test
Estimate the equation by IV and obtain the residuals. Then, regress those residuals
on all the exogenous variables (instruments and controls):
. predict resid, residual
. reg resid alclmo cmedin belief operat height hipsiz waistc smokes totexpr monage,
robust
Finally, obtain the R2 of the last regression and use it to compute the following
statistic: S=nR2, where n is the number of observations. Under the null hypothesis
that all instruments are exogenous, S is distributed m2 k , where (m-r) is the number
of instruments minus the number of endogenous variables.
. quietly: ereturn list
. display chi2(2,e(N)*e(r2))
The null hypothesis is that all moment conditions are valid. If the test is rejected,
you cannot determine which the invalid moment conditions are.
The J-test
Estimate the equation by IV and obtain the residuals. Then, regress those residuals
on all the exogenous variables (instruments and controls). Lastly, compute the Fstatistic from testing that all instruments are jointly zero:
. predict resid1, residual
35
. reg resid1 alclmo cmedin belief operat height hipsiz waistc smokes totexpr monage,
robust
. test height==hipsiz==waistc==0
The overidentifying restriction test statistic is J=nF. Under the null hypothesis that
all instruments are exogenous, J is distributed m2 k , where (m-r) is the number of
instruments minus the number of endogenous variables:
. quietly: return list
. display chi2(2,e(N)*r(F))
Weak Instruments
If the correlation between the instruments and the endogenous variables is poor,
then the asymptotic distribution of the IV standard errors is anything but normal. A
widely spread "rule of thumb" to test whether there might be a weak instruments
problem was proposed by Stock and Watson ("Introduction to Econometrics", Chapter
10). In the first stage regression, if the F-statistic corresponding to testing whether
all the instruments are conjunctly null is greater than 10, then you should care about
weak instruments.
Let's estimate the first stage and perform the test:
. test height==hipsiz==waistc==0
Ivreg2
You can download the command "ivreg2" (with the commands "overid", "ivendog",
"ivhettest" associated), which provides extensions to Stata's official "ivreg". It
supports the same command syntax as official "ivreg" and supports (almost) all of its
options. Among the improvements you may find some very useful as the enhanced
Kleibergen-Paap and Cragg-Donald tests for weak instruments.
The post-estimation command "overid" computes versions of Sargan's (1958) and
Basmann's (1960) tests of overidentifying restrictions for an overidentified equation.
You may find details in the help file. The Durbin-Wu-Hausman test for endogeneity
("ivendog") is numerically equivalent to the standard Hausman test obtained using
"hausman" with the "sigmamore" option.
36
Chapter 4
Panel Data
4.1. Static Panels

We will begin using the database ("crime.dta") from Rafael Di Tella and Ernesto
Schargrodsky "Do Police Reduce Crime? Estimates Using the Allocation of Police
Forces after a Terrorist Attack" (American Economic Review, 2004). Please read the
paper carefully first. In the Problem Set you will be asked to explain the estimation
design, present some descriptive statistics and replicate much of the regressions.
. set mem 5m
. use crime, clear
The syntax for panel data regressions is "xtreg depvar indepvars, i(unit) model",
where "unit" must be replaced by the variable that identifies the different units in
the panel, and "model" must be replaced by "fe" (fixed effects model), "re" (random
effects model), or "be" (between effects model), among others. The default option
is "re".
In Di Tella et al. (2004) the block constitutes the unit of observation for their study.
They obtained information about each auto theft in 876 blocks for the nine-month
period starting April 1, 1994 and ending December 31, 1994. The difference-indifference model is:
CTit = 0 SBit + 1OBit + 2TBit + M t + Fi + it

Where CTit is the number of car thefts in block i for month t (variable ""); SBit is a
dummy variable that equals 1 for the months after the terrorist attack (August,
September, October, November, and December) if there is a protected institution in
the block, and 0 otherwise (variable "instp"); OBit is a dummy variable that equals 1
after the terrorist attack if the block is one block away from the nearest protected
institution, and 0 otherwise (variable "inst1p"); TBit is a dummy variable that equals
1 after the terrorist attack if the block is two blocks away from the nearest
protected institution, and 0 otherwise; M t is a month fixed effect (dummy variables
"month5- month12"); Fi is a block fixed effect (variable "blockid"); and it is the
error term.
Let's run different estimates (you should compare the fixed effects estimate to the
second column from Table 3). First we have to eliminate the observations after the
terrorist attack and before the police intervention (months 72 and 73):
37
. drop if month==72 | month==73

. xtreg cartheft instp inst1p month5-month12, fe i(blockid) robust
. xtreg cartheft instp inst1p month5-month12, re i(blockid) robust
. xtreg cartheft instp inst1p month5-month12, be i(blockid)
Exercise 4.1: Why did the between estimator dropped the monthly effects? In the
fixed effects model, can you include dummy variables for neighborhood or streets?
Why didn't we include a dummy "month4"?
The prefix "xt" is common to every panel data model. For example, the IV estimator
with panel data is "xtivreg", and the logit model with panel data is "xtlogit".
Hausman test
There is a test useful to choose between the fixed effects estimates, which can be
consistent if the individual effects are correlated with the included variables, and
the random effects estimates, which are consistent and efficient if the individual
effects are correlated with the included variables, but inconsistent otherwise.
First estimate your model using fixed effects and save the estimates:
. est store fixed
Then run the random effects estimation:
. xtreg cartheft instp inst1p month5-month12, re i(blockid) robust
Finally, enter:
. hausman fixed
The null hypothesis is that the differences in the coefficients are not systematic. As
we cannot reject such hypothesis, we can use the random effects estimates to
improve the efficiency.
Fixed effects estimates "by hand"
There are many ways to obtain the fixed effects estimates. For instance, we could
have included a set of dummy variable identifying each unit, what is known as the
LSDV (Least Squares Dummy Variables) estimator. You can use the command "areg":
. areg cartheft instp inst1p month5-month12, absorb(blockid) robust
The option "absorb" generates a dummy variable for each value of "blockid" (i.e. a
dummy for each block), but it does not show their estimates (as that would involve
38
estimating and showing 876 coefficients that are not interesting by themselves).
Notice that the estimates are the same than those obtained using "xtreg".
Exercise 4.2: A less-known way to obtain fixed effects estimates is running a
regression for each unit, and then averaging all the estimates obtained. Carry out
such estimation.
Within Transformation
Consider the following model:
y it = X it + i + it
It cannot be consistently estimated by OLS when i is correlated with it. The within
transformation proposes to take out individual means in order to obtain the
following:
( yit yi ) = (xit xi ) + ( it i )
The over-bar denotes deviations from the mean for a given i:
xi
t =1
xit
Such transformation eliminated i and then made OLS consistent. Let's replicate such
transformation. First generate the deviations from the mean using the command
"egen". For instance, for the dependent variable you should enter:
. bysort blockid: egen mean_cartheft= mean(cartheft)
. gen dm_cartheft= cartheft - mean_cartheft
And for each independent variable X you should enter:
. bysort blockid: egen mean_X= mean(X)
. gen dm_X= X - mean_X
First we should eliminate observations with missing values in any of the variables
included in the model. However, there are no missing values in this database. In
order to save code lines and time, we will use the "foreach" syntax:
. foreach var of varlist cartheft instp inst1p month5-month12 {
. bysort blockid: egen mean_`var'=mean(`var')
. gen dm_`var'= `var' - mean_`var'
.}
Then we run the regression:
39
. reg dm_cartheft dm_instp dm_inst1p dm_month*, nocons robust

Compare the results with the original estimates using "xtreg":
Exercise 4.3: Show that the within transformation is more efficient than first
differences, because the former uses more information.
Furthermore, show that
they are numerically identical when T=2.

Between estimator
By definition, the between estimator is obtained regressing by OLS the means of
every variable at the unit-level. In "mean_var" we have already stored such means.
Thus, we are ready to obtain the between estimates by hand:
. reg mean_cartheft mean_instp mean_inst1p mean_month*, robust
Compare it with the "xtreg" command using the "be" option:
Exercise 4.4: In both "by hand" versions of the within and between estimators the
standard errors were not the same than those showed by the "xtreg" command. Why?
4.2. Dynamic Panels

Consider a dynamic model with fixed effects where the lagged values of the
dependent variable appear as a regressor (for the sake of simplicity, we omit further
exogenous regressors):
yi ,t = yi ,t 1 + i + i ,t
Consider < 1 . We have observations on individuals i=1,,N for periods t=1,,T.
The within estimator for is:
FE
( y y )( y y )
=
(y y )
with yi =
i =1
t =1
N
i =1
t =1
i ,t
i ,t 1
i , 1
i ,t 1
i , 1
1 T
1 T
y and yi ,1 = t =1 yi ,t 1
t =1 i ,t
T
T
Substituting by the definition of yi ,t , we can obtain:
(1/ NT ) i =1 t =1 ( i ,t i )( yi ,t 1 yi ,t 1 )
N
FE = +
(1/ NT ) i =1 t =1 ( yi ,t 1 yi ,t 1 )
N
This estimator is biased and inconsistent for N and fixed T. It can be shown
that (Hsiao, 2003, Section 4.2):
40
plim
N
1
NT
(
N
i =1
t =1
i ,t
i )( yi ,t 1 yi ,t 1 ) =
2 (T 1) T + T
0
T2
(1 ) 2
Thus, for fixed T we have an inconsistent estimator. One way to solve this
inconsistency problem was proposed by Anderson and Hsiao (1981). Take first
differences:
yi ,t yi ,t 1 = ( yi ,t 1 yi ,t 2 ) + ( i ,t i ,t 1 )
If we estimated the above model by OLS we would yield inconsistent estimates, since
yi ,t 1 and i ,t 1 are correlated by definition. However, notice that yi ,t 2 is correlated

with ( yi ,t 1 yi ,t 2 ) but not with ( i ,t i ,t 1 ) (unless i ,t exhibits autocorrelation).
Then, we can use yi ,t 2 as an instrument.
Anderson and Hsiao also proposed
(y
i ,t 2
yi ,t 3 ) as an instrument. Furthermore,
Arellano and Bond (1991) suggested that the list of instruments can be extended by
exploiting additional moment conditions, and then using a GMM framework to obtain
estimates.
Exercise 4.5: What do you think about the orthogonality condition between yi ,t 2
and i ,t i ,t 1 ? (Hint: using the same argument, one could argue to use yi ,t 1 as an
instrument for yi ,t in non-dynamic models).
Xtabond
The Arellano-Bond estimator is obtained with the command "xtabond".1 Nevertheless,
we need first to declare the dataset to be panel data using "tsset":
. tsset panelvar timevar
This declaration is also needed for other commands, such "stcox" and "streg" for
duration models. Once you have done such declaration, you can refer to lagged
values using "L.variable" and first difference using "D.variable". You can also utilize
the commands "xtsum" and "xttab", similar to "summarize" and "tabulate" but
designed specifically for panel data.
Then you can implement the Arellano and Bond estimator ("xtabond"), which uses
moment conditions in which lags of the dependent variable and first differences of
the exogenous variables are instruments for the first-differenced equation. The
syntax is:
. xtabond depvar indepvar, lags(#)
This command has been updated on May 2004. Then, you should enter "update all" to install
official updates to Stata.
41
The option lags(#) indicates the number of lags of the dependent variable to be
included in the model (the default being 1). The option "robust" specifies that the
Huber/White/sandwich estimator of variance must be used in place of the traditional
calculation. The option "twostep" specifies that the two-step estimator has to be
calculated.
For instance, consider a dynamic version of Di Tella et al. (2004):
. tsset blockid month
. xtabond cartheft instp month5-month12, lags(1)
The moment conditions of these GMM estimators are valid only if there is no serial
correlation in the idiosyncratic errors. Because the first difference of white noise is
necessarily autocorrelated, we need to focus on second and higher autocorrelation.
You can see the error-autocorrelation tests at the end of the table generated by
"xtabond". In this case we reject the possibility of second degree autocorrelation.
Notice that Stata also calculates the Sargan test of over-identifying restrictions.
We can see that the coefficient on "instp" is not statistically different from zero any
more. There are two probable reasons. The first has to do with the exogeneity
assumption. In the difference-in-difference estimator we were pretty convinced
about the internal validity of the model. However, the Arellano-Bond estimator
involves a great set of moment conditions, and we do not have serious evidence to
think that there are valid. Then, the Arellano-Bond estimate for "instp" may be
inconsistent.
On the other hand, there is a simple loss of power related to performing first
differences instead of using the within transformation (because we are employing
less information). Indeed, if we had used first differences in the original model we
would have found a coefficient on "instp" statistically not different from zero:
. reg D.cartheft D.instp D.month5 D.month6 D.month7 D.month8 D.month9
D.month10 D.month11 D.month12, robust
Monte Carlo Experiment
We will implement a Monte-Carlo experiment to evaluate the seriousness of the bias
in "short" dynamic panels. Consider the Data Generating Process of the following
growth model:
GDPi ,t = GDPi ,t 1 + 1insti ,t + 2entri ,t + i + i ,t

Where GDPi ,t is the real gross domestic product per capita, insti ,t is an index of the
quality of institutions, entri ,t is an index of the entrepreneurship spirit in the country,
42
i 's are the country fixed effects, and i ,t is the error term. The information is
collected for i=1,,N individuals, during t=1,,T periods.
Suppose N=50 and T=5. Generate variables "year" and "country":
. clear
. set obs 250
. gen year=mod(_n-1,5)+1
. gen country=ceil(_n/5)
See the results using "browse". Let's generate data on "inst" and "entr":
. gen inst = 0.2 + year*0.1 + uniform()/2
. gen entr = 0.4 + year*0.08 + uniform()/3
Generate the fixed effects as correlated with both "inst" and "entr":
. gen fe_aux=(inst + entr + uniform())/3 if year==1
. bysort country: egen fe = mean(fe_aux)
. drop fe_aux
Finally, generate data on "GDP":
. gen GDP = 0.2*inst+0.3*entr+fe+uniform()
. bysort country: replace GDP = GDP+ 0.3*GDP[_n-1] if _n>1
. drop fe
See the results for the first 6 countries:
. graph twoway line GDP inst entr year if country<7, by(country)
Antes de continuar, declare "tsset":
. tsset country year
We know that the "true" values of the parameters are: = 0.3 , 1 = 0.2 and
2 = 0.3 . Firstly, run a pooled OLS regression:

. reg GDP inst entr, robust
As you can see, the estimates are seriously biased. Now add the lagged dependent
variable as a regressor:
. reg GDP L.GDP inst entr, robust
The estimates are still very biased. Run a regression including only fixed effects:
. xtreg GDP inst entr, fe i(country) robust
43
The estimates are also wrong. Run a regression including both fixed effects and
lagged dependent variable, but inconsistent:
. xtreg GDP L.GDP inst entr, fe i(country) robust
Although considerably less biased than the previous estimates, they are still
inconsistent. Finally, run the Arellano-Bond estimator:
. xtabond GDP inst entr, robust
The bias is even smaller. We can repeat the whole experiment using different values
for N and T. In particular, keep N fixed in 100 and vary T=5, 10, 15, , 35. We will
show the estimates of the coefficients and their standard values. Run the following
do-file:
. * Generate data for T=25 and N=50
. clear
. set mat 800
. set obs 1250 // 25*50
. gen year=mod(_n-1,25)+1
. gen country=ceil(_n/25)
. gen inst = 0.2 + year*0.05 + uniform()/2
. gen entr = 0.35 + year*0.08 + uniform()/3
. gen fe_aux=(inst + entr + uniform())/3 if year==1
. bysort country: egen fe = mean(fe_aux)
. drop fe_aux
. gen GDP = 0.2*inst+0.3*entr+fe+uniform()
. bysort country: replace GDP = GDP+ 0.5*GDP[_n-1] if _n>1
. drop fe
. tsset country year
. * Repeat the experiment for different T's
. forvalues T=3(2)25 {
. preserve
. quietly: keep if year<=`T'
. display "T="`T'
. quietly: xtreg GDP L.GDP inst entr, fe i(country) robust
. matrix b = e(b)
44
. matrix se = e(V)
. display "XTREG: gamma=" b[1,1] " (" se[1,1] ")"
. quietly: xtabond GDP inst entr, robust
. matrix b = e(b)
. matrix se = e(V)
. display "XTABOND: gamma=" b[1,1] " (" se[1,1] ")"
. restore
.}
Run the do-file a couple of times. You will notice that for T=3 the "xtabond" may
yield more inconsistent coefficients than the original "xtreg". For T>5 the coefficient
yielded by "xtabond" is always more close to the real value (0.5) than that yielded by
"xtreg". In these simulations the bias becomes relatively insignificant for T>21.
Exercise 4.6: Repeat the experiment changing the GDP error term (e.g.
"uniform()*1.5" instead of "uniform()"). Recalling the formula for the inconsistency
(Hsiao, 2003), comment the results.
Exercise 4.7: Repeat the last experiment including second order autocorrelation in
the error term. Comment the results.
Xtabond2
If you are using Stata 7 to 9.1, you can install "xtabond2". You might not find the
command using "search xtabond2". Then, to download it type the following
command: ssc install xtabond2, all replace. If that does not work, then download
the ado- and help-files from "http://ideas.repec.org/c/boc/bocode/s435901.html",
and copy them in "StataFolder\ado\base\?\" (where "?" must be replaced by the first
letter of each command). This procedure is particularly useful in computers with
restrained access to the Web.
The command "xtabond2" can fit two closely related dynamic panel data models. The
first is the Arellano-Bond (1991) estimator, as in "xtabond", but using a two-step
finite-sample correction. The second is an augmented version outlined in Arellano
and Bover (1995) and fully developed in Blundell and Bond (1998).
A problem with the original Arellano-Bond estimator is that lagged levels are often
poor instruments for first differences, especially for variables that are close to a
random walk. Arellano and Bover (1995) described how, if the original equations in
levels were added to the system, additional moment conditions could be brought to
45
bear to increase efficiency. In these equations, predetermined and endogenous

variables in levels are instrumented with suitable lags of their own first differences.
However, if you are using Stata 10, you may use the command "xtdpdsys".
4.3. Nonlinear Models

Logit with fixed effects
We can see a logit model with fixed effects for T=2. First remember that in a logit
estimation we cannot have perfect predictors. Then, if an individual repeated its
outcome every time, then its fixed effect would be a perfect predictor. As a
consequence, we must eliminate those observations with constant outcomes.
In the T=2 case, the possible values for ( y1 , y 2 ) will be (0,1) or (1,0) . The
conditional probability for the first is:
P{ y i = (0,1) | y i = 1 2 , i , } =
P{ y i = (0,1) | i , }
P{ y i = (0,1) | i , } + P{ y i = (1,0) | i , }
Use that:
P{ y i = (0,1) | i , } = P{ y i1 = 0 | i , }P{ y i 2 = 1 | i , }
And:
P{ y i 2 = 1 | i , } =
exp{ i + xi2 }
1 + exp{ i + xi2 }
P{ y i1 = 0 | i , } = 1
exp{ i + xi1 }
1 + exp{ i + xi1 }
Then, after some algebra steps detailed at the end of this document, it follows that
the conditional probability will be given by:
P{ yi = (0,1) | yi = 1 2, i , } =
exp{( xi 2 xi1 ) ' }

exp{( xi* ) ' }
=
exp{( xi 2 xi1 ) ' } + 1 exp{( xi* ) ' } + 1
Which does not depend on i . The last looks exactly as a simple logit regression,
where xi* is in fact the first difference for xi . In an analogous way you may obtain
P{ y i = (1,0) | y i = 1 2 , i , } .
In summary, the estimator consists in the following: keep only the observations with
( y1 , y 2 ) equal to (0,1) or (1,0). Then generate a dependent variable taking the value
1 for positive changes (0,1), and the value 0 for negative changes (1,0). Then, regress
(by an ordinary logit) the transformed dependent variable on the first differences for
the regressors ( xi* = xi 2 xi1 ). You can obtain a similar result for T>2.
46
An example
For this exercise we will use a small panel (T=3) extracted from a Russian database.
. use russia1, clear
The
command
xtlogit
fits
random-effects,
conditional
fixed-effects,
and
population-averaged logit models:

. xtlogit smokes monage highsc belief obese alclmo hattac cmedin totexpr tincm_r
work0 marsta1 marsta2 round2 round3, fe i(eid)
You can use the post-estimation command mfx (presented in the notes on the
binary logit/probit models):
. mfx, predict(pu0)
It calculates the marginal effects where the explanatory variables are set at their
mean values and the fixed effect is set at zero.
We can compare the results to the pooled logit:
. xtlogit smokes monage highsc belief obese alclmo hattac cmedin totexpr tincm_r
work0 marsta1 marsta2 round2 round3, fe i(eid)
. logit smokes gender monage highsc belief obese alclmo hattac cmedin totexpr
tincm_r work0 marsta1 marsta2 round2 round3
You will notice that several variables that in the simple model appeared as
significant, now including fixed effects are not statistically different from zero any
more. Before we studied the association between some variables and the fact that
some individuals smoked and others did not. However, in the fixed effects model we
are studying the association between the inter-temporal change in some variables
and the fact that some individuals have decided to either stop or start smoking. Now
you can see that this fixed effects estimator is equivalently to the first-differences in
the OLS model.
However, some variables stopped being significant just because they have almost no
intertemporal variation (such as "monage" and "highsc").
Probit with random effects

The latent variable specification is:
y it = xit + u it
Where uit have mean zero and unit variance, and it can be further decomposed as
it + i . Additionally:
47
y it = 1
if y it > 0
y it = 0
if y it 0
The likelihood contribution of individual i will be the joint probability of observing

outcomes ( y i1 ,..., y iT ) . Such probability is derived from the joint distribution of
( y i1 ,..., y iT ) . Recall that in the one-period case:

P( y i = 1) = P(u i > xi ) =
xi
f (u )du
In the T-periods case, integrating over the appropriate intervals would involve T
integrals (which in practice is done numerically). When T 4 , the maximum
likelihood estimation is made infeasible. This curse of dimensionality may be
avoided by using simulation-based estimators (Verbeek, 2004, Chapter 10.7.3).
If u it could be assumed as independent:
f ( y i1 ,..., y iT | xi1 ,..., xiT , ) = t f ( y it | xit , )

The estimation procedure would involve only one-dimensional integrals (as in the
cross-section case). Thus, if it is assumed to be independent over time:
f ( y i1 ,..., y iT | xi1 ,..., xiT , ) =
f ( y i1 ,..., y iT | xi1 ,..., xiT , i , ) f ( i )d i
f ( y it | xit , i , ) f ( i )d i
We can use arbitrary assumptions about the distributions of it and i . If we

assumed that u it is distributed as a multivariate normal, then we would obtain the
random effects probit model.
An example
The command "xtprobit" does all the work. Lets continue with the previous example:
. xtprobit smokes gender monage highsc belief obese alclmo hattac cmedin totexpr
tincm_r work0 marsta1 marsta2 round2 round3, re i(eid)
We can compare the results to the pooled logit:
. probit smokes gender monage highsc belief obese alclmo hattac cmedin totexpr
tincm_r work0 marsta1 marsta2 round2 round3
Notice that if the explanatory variables were exogenous, both estimators would be
consistent, but the former would be more efficient.
You can compute the marginal effects where the explanatory variables are set at
their mean values and the fixed effect is set at zero:
. mfx, predict(pu0)
48
Tobit with random effects

This model is very similar to the probit with random effects. Rewrite the model:
y it = xit + i + it
Where i and it are i.i.d. normally distributed, independent of ( xi1 ,..., xiT ) with
zero means and variances 2 and 2 . As in the original model:
y it = y it
if y it > 0
if y it 0
y it = 0
As in the random effects probit model, the likelihood contribution of individual i will
be the following:
f ( y i1 ,..., y iT | xi1 ,..., xiT , ) =
[ f ( y
t
it
| xit , i , ) f ( i )d i
We can obtain the desired model simply by replacing f () by the normal density
function, and integrating over i numerically.
An example
Use the dataset "hinc.dta", which is a quasi-generated sample of individuals based on
a British household panel:
. clear
. set mem 50m
. use hinc
See the data definitions:
. describe
Estimate the model using a pooled Tobit:
. tobit hincome educ exper expersq married widowed divorcied, ll(0)
And now estimate the model using a Tobit with random effects:
. xttobit hincome educ exper expersq married widowed divorcied, i(pid) ll(0)
You can use almost all the features available for the pooled Tobit model. You will be
required to do so in the Problem Set.
49
exp{ i + xi2 } exp{ i + xi1 } exp{ i + xi2 }
1 + exp{ i + xi2 } 1 + exp{ i + xi1 } 1 + exp{ i + xi2 }
P{ yi = (0,1) | yi = 1 2 , i , } =
exp{ i + xi2 } exp{ i + xi1 } exp{ i + xi1 } exp{ i + xi2 }
+
2
1
+
exp{
+
x
}
1
+
exp{
+
x
}
1
+
exp{
+
x
}
1
+
exp{
+
x
}
i
i2
i
i1
i
i1
i
i2
exp{ i + xi2 }(1 + exp{ i + xi1 }) exp{ i + xi1 } exp{ i + xi2 }
(1 + exp{ i + xi1 })(1 + exp{ i + xi2 })
P{ yi = (0,1) | yi = 1 2 , i , } =
exp{ i + xi2 }(1 + exp{ i + xi1 }) + exp{ i + xi1 }(1 + exp{ i + xi2 }) 2 exp{ i + xi1 } exp{ i + xi2 }
(1 + exp{ i + xi1 })(1 + exp{ i + xi2 })
P{ yi = (0,1) | yi = 1 2 , i , } =
exp{ i + xi2 }(1 + exp{ i + xi1 }) + exp{ i + xi1 }(1 + exp{ i + xi2 }) 2 exp{ i + xi1 } exp{ i + xi2 }
P{ yi = (0,1) | yi = 1 2 , i , } =
exp{
}(
1
exp{
})
exp{
}(
1
exp{
})
2
exp{
}
exp{
}
+
+
+
+
+
+
+
+
+
i
i2
i
i1
i
i1
i
i2
i
i1
i
i2
exp{ i + xi2 }
P{ yi = (0,1) | yi = 1 2 , i , } =
exp{
+
x
}
+
exp{
+
x
}
i
i
2
i
i
1
P{ yi = (0,1) | y i = 1 2 , i , } =
exp{ i
exp{ i + xi2 }
exp{ i + xi1 }
+ xi2 } + exp{ i + xi1 }
exp{ i + xi1 }
exp{xi2 xi1 }
exp{( xi 2 xi1 )' }
=
P{ yi = (0,1) | yi = 1 2 , i , } =
exp{xi2 xi1 } + 1 exp{( xi 2 xi1 )' } + 1
50
Chapter 5
Binary Logit/Probit
Binary regression
When you try to estimate by OLS a model with binary dependent variable, you may
encounter some of the following problems: i. The marginal effects are linear in
parameters; ii. There is heterocedasticity (though it would be enough to use robust
standard errors); iii. The model might predict values below zero and above one
(though you may use set of dummies as regressors, and then that problem would be
solved).
In summary, the main concern is that in some economic models (i.e. from the
aprioristic analysis) you may expect to find nonlinearities. For instance, consider a
model for the probability of approving an exam. For the most prepared students
(those with high probabilities of passing), an additional hour of study will have
virtually no effect on their probabilities of approving. The same is valid for the
students that have not paid any attention since the first day of class: the impact of
an additional hour of study on their probabilities of approving is almost zero.
However, consider a student right in the middle. As she is in the borderline, she
might pass the exam just because what she studied in the last hour: the marginal
effect of an additional hour of study on her probability of approving will be
considerable. Then, we will introduce a model that captures exactly that kind of
nonlinearities. Consider a vector x of explanatory variables, and a vector of
parameters. The probability (p) that an event y happens is:
p = F ( x )
Where F () has the following properties:
F () = 0, F () = 1, f ( x) = dF ( x) dx > 0
For instance, in the probit model F () is the accumulated distribution of a standard
normal:
2
F ( x ) =
1 s2
e ds
2
On the other hand, in the logit model F () is the accumulated distribution of a

logistic random variable:
F ( x ) =
51
e x
1 + e x
Notice that the marginal effects of the explanatory variables are not linear:
p
= k f ( xi )
xk
The sign of the derivative is the same than the sign of the coefficient k .
sgn(p xk ) = sgn( k )
However, the value of the coefficient lacks direct quantitative interpretation. The
parameter k is multiplied by f ( xi ) , which is maximum when xi = 0 and
decreases (in absolute value) when xi goes towards (indeed, the difference
between the models logit and probit is the weight of their tails). We will discuss
thoroughly how to present marginal effects.
The Maximum-Likelihood estimator
We have an i.i.d. sample of binary events and explanatory variables ( yi , xi ) , for i
=1,,n,. The random variable yi follows a Bernoulli distribution with pi = P( yi = 1) .
The likelihood function is then:
n
1 yi
L( ) = pi (1 pi ) = piyi (1 pi )
yi =1
yi = 0
i =1
And the log-likelihood function:

n
l ( ) = [ yi ln( pi ) + (1 yi ) ln(1 pi ) ]
i =1
n
= [ yi ln( F ( xi )) + (1 yi ) ln(1 F ( xi ))]

i =1
The first order conditions are:

n
i =1
( yi F ( xi )) f ( xi ) xki
= 0 k = 1,..., K
F ( xi )(1 F ( xi ))
It is a system of K non-linear equations and K unknowns. There is a solution for if

the regressors are linearly independent and if there is no a perfect classifier.
However, it is impossible to find it explicitly.
You can find it solving numerically the maximization problem. Although Stata has a
simple command to run both models, we will utilize a Stata routine to solve the
maximum likelihood problem "by hand".
Maximum likelihood estimation in Stata
Recall the Russian database from the first two weeks. We have a variable called
"smokes" that indicates whether the individual smokes or not. We have plenty of
52
interesting dependent variables (income-variables, health-variables, etc.), so we will

try to estimate a model of the decision of smoking:
. use russia.dta
The command "ml" indicates to Stata each element of the maximum-likelihood
estimation. Its syntax is:
. ml model lf name equations
Where "lf" is the method (we will discuss this soon), "name" is the name of the
program where we will save the log-likelihood function, and "equations" is/are the
specification/s (i.e. the list of the independent variable plus the dependent
variables).
Then, the first task is to generate a program with the log-likelihood function:
. program define probit_ml
.
version 1.0
args lnf XB
quietly replace `lnf' = $ML_y1*ln(norm(`XB'))+(1-$ML_y1)* ln(1-norm(`XB'))
. end
Where "args lnf XB" indicates the value of the likelihood function ("lnf") and xi
("XB"). The expression "$ML_y1" is the convention for yi . The program you write is
written in the style required by the method you choose. The methods are "lf", "d0",
"d1", and "d2". See "help mlmethod" for further information. However, some global
macros are used by all evaluators. For instance, "$ML_y1" (for the first dependent
variable), and "$ML_samp" contains 1 if observation is to be used, and 0 otherwise.
Then, the expression "`lnf' = $ML_y1*ln(norm(`XB'))+(1-$ML_y1)* ln(1-norm(`XB'))" is
exactly the formulae of the log-likelihood function:
l ( ) = yi ln( F ( xi )) + (1 yi ) ln(1 F ( xi ))
Notice that we used the normal accumulated distribution, and then we are
estimating a probit model. Now we can use the command "ml":
. ml model lf probit_ml (smokes = gender monage highsc belief obese alclmo hattac
cmedin totexpr tincm_r work0 marsta1 marsta2)
Once you have defined the maximum likelihood problem, you can verify that the loglikelihood evaluator you have written seems to work (strongly recommended if you
are facing a problem not covered by any standard command):
. ml check
53
And you can also display a description of the current problem:

. ml query
There are a bunch of additional commands to use after "ml model" (see "help ml").
For instance, "ml init" provides a way to specify initial values and "ml search"
searches for (better) initial values.
As you can see, we have included demographic, economic and health variables as
regressors. But "ml" does not perform the maximization nor obtain the estimates. To
do so, we need to invoke the command "ml maximize":
. ml max
Finally, we got the estimates. We can obtain a robust variance-covariance matrix or
even clustered standard errors adding "robust" and "cluster()" as options when
invoking "ml model". After invoking "ml max" you can use "ml graph" to graph the loglikelihood values against the iteration number, and "ml display" to redisplay the final
results.
You can choose a particular maximization algorithm. If you are estimating a (not
necessarily binary) logit model you know that the maximization problem is a concave
one, and then you do not have to worry about this. Nonetheless, in some problems
there might be a lot of local maximums (and sensitiveness to the choice of initial
values), and you have to be careful.
The option "technique(nr)" specifies Stata's modified Newton-Raphson (NR) algorithm,
"technique(bhhh)"
specifies
the
Berndt-Hall-Hall-Hausman
"technique(dfp)"
specifies
Davidon-Fletcher-Powell
(DFP)
(BHHH)
algorithm,
algorithm,
and
"technique(bfgs)" specifies the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm.

The default is "technique(nr)". For a detailed explanation of the NR and BHHH
algorithms we suggest Kenneth Train's "Discrete Choice Methods with Simulation"
(2003), Chapter 8 ("Numerical Maximization").
Probit and logit
As we previously introduced, there is a simple command to run a logit/probit model.
We can check that if the above estimates are exactly the same:
tincm_r work0 marsta1 marsta2
As always, we have the options "robust", "cluster()", and so on. Notice that we can
only infer qualitative information from the raw coefficients (e.g. there is a positive
54
association between smoking and being male, and it is statistically different from
zero at the 1% level of confidence).
The binary logit and probit models are particular cases of more general models. For
instance, we have the multinomial logit ("mlogit"), nested logit ("nlogit"), mixed logit
("mlogit"), ordered logit ("ologit"), and so forth. In the last week we will explain
those models. The literature is known as "discrete choice", and it is particularly
useful to model the consumer behavior. At the end of this section we will develop a
logit model with fixed effects ("xtlogit").
Marginal Effects
As stated previously, the marginal effects depend on the value of x. You may choose
the mean values for every element of x, and evaluate the marginal effects there:
p
xk
= k f ( x )
x= x
This is the "marginal effects at the mean". You can compute them using the command
"dprobit" for the probit model:
. dprobit smokes gender monage highsc belief obese alclmo hattac cmedin totexpr
tincm_r work0 marsta1 marsta2, robust
At the mean values, the marginal effect of changing the value of obese from 0 to 1
on the probability of smoking is -14 percentage points. You can evaluate the marginal
effects at a point different than the mean values:
. matrix input x_values = (1,240,1,3,0,1,0,1,10000,10000,1,1,0)
. dprobit smokes gender monage highsc belief obese alclmo hattac cmedin totexpr
tincm_r work0 marsta1 marsta2, robust at(x_values)
You can also use the command "mfx":
. mfx
For the logit model you may only use the command "mfx":
. mfx
55
However, analyzing the marginal effects at the mean values of x could be

troublesome. Maybe there is no such thing as the "mean" individual (e.g. those values
are not representative of any single individual).
In those cases you might be interested in evaluating the "mean marginal effects",
that is, the mean of the marginal effects evaluated at every single observation. As
you started interested in nonlinearities, this may sound contradictory. However, the
"mean marginal effects" give an accurate idea of the actual impact of marginal
changes in a variable over the entire sample:
1 n p
n i =1 xk
=
x = xi
1 n
k f ( xi )
n i =1
You can use the command "margeff", which you have to install in advance ("search
margeff"):
. margeff
It works both with logit and probit models. Moreover, you can calculate these
marginal effects "by hand". Estimate the model, predict each xi and store them in
the variable "xb":
. predict xb, xb
We need to retrieve the coefficients on the variable under consideration (for
instance, on "monage"):
. ereturn list
. matrix coef=e(b)
And we can calculate the marginal effect at every observation (to be stored in
variable "me"):
. gen me=coef[1,2]*normden(xb)
Now we can calculate the mean of the marginal effects, or even their median:
. tabstat me, stats(mean median)
Alternatively, we can estimate nonparametrically the density distribution of the
marginal effects over the sample (using the command "kdensity", for obtaining a
Kernel estimator):
56
. kdensity me
There is no "right" or "wrong" way to evaluate the marginal effects. The important
thing is to give the right econometric interpretation for the strategy chosen.
Goodness of Fit
There is a simple way to imitate the R2 from OLS, denominated pseudo-R2:
LR = 1
ln L
ln L0
Where L is the maximum value of the log-likelihood function under the full
specification, and L0 is the maximum value of the log-likelihood function in a model
with only a constant. Then LR measures the increase in the explanatory power from
considering a model beyond a simple constant.
As you can see in Menard (2000), there are some "desirable" properties for a
goodness-of-fit index that the R2 complies. There is no such a good index for the
logit/probit model, but you have a relatively wide set of opportunities. Once again,
there is no "right" or "wrong" way to measure the goodness of fit. However, you must
read carefully the econometric meaning of your choice.
Another strategy is to generate a "percentage of right predictions". As you probably
noticed, the predictions are in the (0,1) open interval, and then you cannot compare
them with the actual outcomes (either zero or one). But you can use cutoff points (c)
to transform the (0,1) predictions in either 0 or 1:
yi 1[ p i > c ]
yic 1[ yi = yi ]
H=
n
i =1
yic
Where H is the "percentage of right predictions". The index is pretty sensitive to the
choice of c. Furthermore, a trivial model has H 1 2 . For instance, consider a
model explaining the decision to commit suicide. Since less than 0.1% of the people
commit suicide, a trivial model predicting "no person commit suicide" would obtain a
H = 0.999 .
We will create a table of "Type I Type II" errors. Suppose c=0.5. Let's generate y i
and then yic :
. predict smokes_p
57
. gen smokes_hat = (smokes_p>0.3)

And then just employ "tabulate":
. tabulate smokes smokes_hat
The right predictions are on one of the diagonals, and the Type I and Type II errors
are on the other diagonal. The H can be obtained dividing the observations on the
"right" diagonal over the sample size.
Exercise 5.1: Repeat the exercise varying the cutoff point (c). Suppose that you are
hired by a company that is deeply worried about committing a Type-I-error: would
you increase or decrease c? Explain carefully.
58
Chapter 6
Hazard Models
Introduction to Hazard functions

Sometimos you are interested in explaining the duration of a certain event. For
instance, you may be interested in explaining the time it takes for a bank to be
either dissolved or acquired by another bank, the duration of unemployment, and so
on.
Let T denote the time spent in the initial state. In the above example, T would be
the number of months until a bank is dissolved or acquired. Since T is continuous,
the distribution of T is expressed by the following cumulative density function:
F(t) = P(T t)
The survivor function is the probability surviving past t and is defined as:
S(t) = 1 - F(t) = P(T > t)

Having survived up to time t , the likelihood of leaving the initial state within the
time interval t until t + h may be written as follows:
P(t T < t + h | T t)
Dividing by h we can obtain the average likelihood of leaving the initial state per
period of time over the interval [t, t + h) . Making h tend to zero we get the hazard
function (defined as the instantaneous rate of leaving the initial state):
(t) = lim
h 0
The
hazard
and
survival
P(t T < t + h | T t)
h
functions
provide
alternative
but
equivalent
characterizations of the distributions of T. To see this, rewrite the conditional

probability as follows:
P(t T < t + h | T t) =
P(t T < t + h) F (t + h) F (t )
=
P(T t)
1 F (t )
And then notice that:
lim
h 0 +
F (t + h) F (t )
= F (t ) = f (t )
h
59
Finally:
(t) =
f (t )
f (t )
=
1 F (t ) S (t )
Additionally, it is easy to show that (Verbeek, 2004):

s
F(s) = 1 - exp(- (t)dt )

0
For instance, consider the simplest case: the hazard rate is constant, (t) = . It
implies that T follows the exponential distribution: F(t) = 1 - exp(-t) .
Let x i be a vector of explanatory variables. A widespread group of models are the
so-called proportional hazard models. The idea behind those models is that the
hazard function can be written as the product of a baseline hazard function that does
not depend on x i , and a individual-specific non-negative function that describes the
effect of x i :
(t, x i ) = 0 (t)exp(x i )
Where 0 is the baseline hazard function (i.e. that of an hypothetical individual with
x i = 0 ), and exp(x i ) is the proportional term that stretches and shrinks the
baseline function along the y axis (because the adjustment is the same for every t ).
Take logarithm and then take the derivative with respect to x ik :
ln (t, x i )
= k
x ik
The coefficient k is the (proportional) effect of a change in x k on the hazard
function. When x k is increased, the hazard function is stretched if k > 0 (i.e. the
instantaneous likelihood of leaving the initial state increases for every t), and if
k < 0 it is shrunk.
An example
Open the database lung.dta and see its content:
. use lung, clear
. describe
. browse
First of all, use stset to declare the data to be survival-time data. You have to
indicate the permanence variable ("time") y and the failure variable ("dead"):
60
. stset time, failure(dead)

We can already draw a Kaplan-Meier non-parametric survival estimator (see for
example Lancaster, 1990):
. sts graph, by(sex)
We are going to estimate non-proportional models (notice that x i cannot be timevarying). The models with a proportional hazard ratio parameterization included in
the command streg are those with the following baseline hazard functions:
exponential, Weibull, and Gompertz. For instance, if 0 (t ) follows a Weibull
distribution:
. streg age sex wt_loss, nohr distribution(weibull)
The nohr option means that coefficient estimates must be shown. This option
affects only how results are displayed, not how they are estimated. As we already
explained, when x k is increased the hazard function is stretched if k > 0 and
shrunk if k < 0 . If we had not included the nohr option, the hazard ratios
( = exp( k ) ) would have been shown instead of the coefficients:
. streg, hr
The estimates imply that, at each survival time, the hazard rate for females is only
60% of the hazard rate for males.
We can also consider a non-parametric baseline hazard, which would suggest the
non-parametric Cox model:
. stcox age sex wt_loss, robust basesurv(base_survival)
The command stcurve plots the survival, hazard, or cumulative hazard function
after stcox or streg:
. stcox age sex wt_loss, basesurv(hazard_aux1)
. stcurve, survival
. stcox age sex wt_loss, basehc(hazard_aux2)
. stcurve, hazard
. streg age sex wt_loss, distribution(weibull)
. stcurve, survival
. stcurve, hazard
Additionally, we can graph the survival function evaluated at a particular point of the
explanatory variables:
61
. streg age sex wt_loss, distribution(weibull)

. stcurve, survival at(age=62.44, sex=1, wt_loss=9.83)
. stcurve, survival at1(sex=0) at2(sex=1)
Estimating hazard functions using the logit model
Jenkins (1995) points out that the likelihood functions of a logit model and a duration
model (under some sampling schemes) are pretty similar. As a consequence, the
discrete-time duration model based on data derived from some sampling schemes
can be estimated as a regression model for a binary dependent variable.
We only need to rearrange the database. Each individual in the database must have
as many observations as time intervals that it spent in the sample. We can use the
command "expand":
. clear
. set mem 10m
. use lung
. expand time
We need a time variable (i.e. the first observation for individual i will correspond to
its first day in the simple, the second observations will correspond to its second day
in the sample, and so on).
. sort id
. quietly by id: gen day = _n
Now we must simple generate the new dependent variable ("death"). For every
observation different than the last one (for each individual), it must take the value
0. For the last observation of each individual, death must take the value 0 if the
individual did not die (i.e. it was censored) or 1 otherwise (if died).
. gen death=0
. quietly by id: replace death = dead if _n==_N
We can now see the results:
. browse id day death
We need to create a time-control. If we wanted to create the correlate for the nonparametric hazard model, we would need to create a set of dummies for time. Or we
can simply use a log specification:
. gen ln_t=log(time)
62
Finally, we can estimate the model:

. logit death age sex wt_loss ln_t
63
Chapter 7
Count-Data Models
Count-data Models
The central characteristics of a count-dependent-variable are: i. It must be a positive
integer, including zero; ii. It has not an obvious maximum or upper limit; iii. Most of
the values must be low, and particularly there must be lots of zeros.
Consider the following example, illustrated by the same database that we will use
thoroughly the notes: the visits to the doctor in the last two weeks ("count.dta").
. use count, clear
. tab doctorco
. hist doctorco, discrete
Visits to the doctor in the last two weeks (doctorco) - Actual frequency distribution
Count
Total
Frequency
4,141
782
174
30
24
12
12
5,190
Percent
79.79 15.07
3.35
0.58
0.46
0.17
0.23
0.23
0.1
0.02
100
Notice that the "count nature" is clear: i. The variable only takes integer values,
including the zero; ii. There is no obvious upper limit; iii. The 95% of the
observations take values either 0 or 1, and 80% of the sample is composed by zeros.
The Poisson Regression Model
We need to estimate an appropriate model for E ( y | x) , where x is a vector of
explanatory variables. The model will comply:
E ( y | x ) = exp( x )
Which guarantees E ( y | x) > 0 and provides a rather simple interpretation for the
marginal effects:
ln E ( y | x )
= k
xk
Then, the coefficients are semi-elasticities (notice that the coefficients on binary
variables will be read as proportional changes).
Remember that a random variable has a Poisson distribution if:
f ( y ) = P (Y = y ) =
e y
y!
64
y = 0,1, 2,...
It is straightforward to show that:
E (Y ) = V (Y ) =
The property that the expectation equals the variance is called equidispersion. The
Poisson regression model corresponds to:
f ( y | x) =
e ( x ) ( x ) y
y!
y = 0,1, 2,...
Where ( x) = exp( x ) and x is a vector of K explanatory variables (including an

intercept). Consequently: E (Y | x) = exp( x ) .
Maximum likelihood estimation of the Poisson Model
Suppose y | x
Po( = exp( x 0 )) , and that we have a random sample i.i.d. ( yi , xi )
with i=1,,n. The maximum-likelihood function is:

n
L( ) =
i =1
e i i yi
yi !
With i = exp( xi ) . Take logarithm:

n
l ( ) = [ yi xi exp( xi ) ln yi !]
i =1
Then is such that:

n
y exp( x )x
i
=0
i =1
Like in the logit model, there is no explicit solution and you must find the estimates
numerically. Under the right specification the estimator is consistent, asymptotically
normal and efficient.
Furthermore, the Poisson estimator is a Quasi-MLE. If E (Y | x) = exp( x ) holds, then
is consistent and asymptotically normal for any underlying distribution. However,

we will need to use the robust estimate of the variance-covariance matrix to make
valid inference.
Exercise 5.3: Obtain the maximum-likelihood estimates "by hand" (i.e. using the
command "ml").
An example: Visits to the doctor
We will use the database on visit to the doctor in the last two weeks, along with
demographic, economic and health variables. Enter "describe" to see the variables'
label with a description of the data:
. describe
65
The "poisson" command performs the estimation:

. poisson doctorco sex age agesq income levyplus freepoor freerepa illness actdays
hscore chcond1 chcond2, robust
As always, we utilize the option "robust" for obtaining robust standard errors. As we
anticipated, the marginal effects are easily interpretable. For instance, the
coefficient on "income" indicates that an increase in ten thousands dollars in annual
income is associated with a 20% decrease in the visits to the doctor. Remember that
we are always speaking about associations. We would need an extremely exhaustive
structural model or a clever natural experiment to infer causality.
The post-estimation command "estat gof" performs a goodness-of-fit test of the
model. If the test is significant, the Poisson regression model is inappropriate. In this
case, you could try a negative binomial model.
. estat gof
Negative Binomial Regression Model
The Poisson distribution is a particular case of the Negative Binomial distribution.
Equivalently, the Poisson Regression Model is a particular case of the Negative
Binomial Regression Model. In the negative binomial (maximum-likelihood) regression
model the count variable is believed to be generated by a Poisson-like process,
except that the variation is greater than that of a true Poisson. This extra variation is
referred to as overdispersion. You can refer to Rainer Winkelrmann's "Econometric
Analysis of count data" to see some theoretical and practical details.
The command for the NB regression is "nbreg":
. nbreg doctorco sex age agesq income levyplus freepoor freerepa illness actdays
hscore chcond1 chcond2
If the parameter "alpha" is null, then the Poisson model is right. In order to test the
Poisson suitability, Stata construct an LR-test comparing the log-likelihood values
attained in each model. The result seems natural if you see the difference between
the mean and the variance of the (raw) dependent variable:
. sum doctorco
See "xtpoisson" and "xtnbreg" for closely related panel estimators.
66
Chapter 8
Selection Models
The internal validity of the econometric models is commonly threatened by selection
problems. Consider the following example:
y = x + u
Call s to the selection variable, which takes the value 1 if y is observed, and 0
otherwise. Imagine a super-sample ( yi , xi , si ) of size N, and that we only observe the
sub-sample ( yi , xi ) for those with s = 1 . Consider as an example a wage equation for
women: we only observe wages for those women who work.
If we had an i.i.d. sample ( yi , xi ) , the consistency would depend on E (u | x) = 0 . The
problem is that we now have a sample conditional on s = 1. Take expectation:
E ( y | x, s = 1) = x + E (u | x, s = 1)
Then OLS using the sub-sample will be inconsistent unless:
E (u | x, s = 1) = 0
Notice that not any selection mechanism makes OLS inconsistent. If u is independent
from s, then OLS is consistent. Additionally, if the selection depends only on x then
OLS is also consistent.
A selectivity model
Consider the following system of equations:
y1i = x1i 1 + u1i

y2i = x2 i 2 + u2i
Called the regression equation and the selection equation, respectively. Define the
binary variable y2i 1 y2i > 0 . The pair ( y2i , x2i ) is observed for the entire sample.
However, the pair ( y1i , x1i ) is observed iff y2i = 1 (called the selected sample).
For instance, in our previous example y1i would be the wages for women, x1i would
be the determinants of wages, y2i would be the net utility from working, and x2i
would be the determinants of the latter. The y2i = 1 would indicate whether the
woman decided to work (if she had positive net utility from working), in which case
we would observe her wage (which would be unobservable otherwise).
Let's assume that (u1i , u2i ) are independent from x2i and have mean zero. In
addition, suppose that u2i
N (0, 22 ) . Finally, assume:
E [u1i | u2i ] = u2i
67
This allows the non-observable from both equations to be related. Take expectation:
E ( y1i | x1i , y2i = 1) = x1i 1 + E [u1i | x1i , y2i = 1]

= x1i 1 + E [ E (u1i | u2i ) | x1i , y2i = 1]
= x1i 1 + E [ u2i | x1i , y2i = 1]
= x1i 1 + E u2i | x1i , y2i > 0
= x1i 1 + E [u2i | x1i , u2i < x2 i 2 ]

= x1i 1 + ( x2 i 2 / 2 )
= x1i 1 + zi
x1i 1
Where zi = ( x2 i 2 / 2 ) . If we regressed by OLS using the selected sample we would
be including zi in the error term. If zi and x1i were correlated, and if in addition
u1i and u2i were correlated ( 0 ), then OLS would be inconsistent.

Two-stage consistent estimator
Define:
u1i y1i x1i 1 zi

And write:
y1i x1i 1 + zi + u1i

Where, by definition:
E u1i | x1i , y2i = 1 = 0

If x1i and zi were observable when y2i = 1 , then regressing by OLS y1i on x1i and zi
(using the selected sample) would yield consistent estimates for 1 and . The
problem is that zi is not observable, though it could be calculated from 2 and 2 .
Given that u2i
N (0, 22 ) :
P ( y2i = 1) = P ( y2i = 1) = P (u2i 2 < x2 i 2 2 ) = ( x2 i )
Then P( y2i = 1) corresponds to a probit model with coefficient . If x2i and y2i are
observed for the complete sample, then can be estimated consistently using a
probit model (notice that despite we can identify , we cannot identify 2 and 2
separately).
Finally, 1 and
can be consistently estimated using the following two-step
procedure:
68
First stage: obtain an estimate for using the probit model P( y2i = 1) = ( x2 i ) for
the complete sample. Then estimate zi using zi = ( x2 i) .
Second stage: Regress y2i on x1i and zi utilizing the censored sample, which should
yield consistent estimates for 1 and .
It can be shown that the second stage is heterocedastic by construction. You can
derive robust standard errors, though the adjustment is not as simple as in the
sandwich estimator (it would be if zi were observable). Nevertheless, Stata
computes it automatically.
An example: wage equation for women
The objective of this example is estimating a simple wage equation for women. The
database is the same presented for the Tobit example. Enter "describe" to see the
data definitions:
. use women, clear
. describe
The variable "hwage" represents the hourly salary income (in US dollars), which must
be set to missing if the person does not work (since the "heckman" command
identifies with missing values those observations censored):
. replace hwage=. if hwage==0
Then we can use the command "heckman". Inmmediately after it you must specify
the regression equation, and in the option "select()" you must specify the
explanatory variables for the selection equation:
. heckman hwage age agesq exp expsq married head spouse headw spousew pric seci
secc supi supc school, select(age agesq exp expsq married head spouse headw
spousew pric seci secc supi supc school) twostep
With the option "twostep" it fits the regression model using the Heckman's two-step
consistent estimator. Additionally, you can use different specifications for the
regression and selection equations:
. heckman hwage age agesq exp expsq pric seci secc supi supc, select(age agesq exp
expsq married head spouse headw spousew pric seci secc supi supc school) twostep
Selection Test
You can compare the estimate with the (seemingly inconsistent) OLS regression:
. reg hwage age agesq exp expsq pric seci secc supi supc, robust
69
The differences are indeed minuscule. This is not surprising, since the inverse Mills
ratio term is statistically not different from zero. Testing H 0 : = 0 provides a
simple test for selection bias. Under H 0 the regresin model with the selectioned
samples is homocedastic, and then you can perform it without correcting for
heteroskedasticity.
. test lambda
Two stages "by hand"
. gen s=(hwage!=.)
. probit s age agesq exp expsq married head spouse headw spousew pric seci secc
supi supc school
. predict xb, xb
. gen lambda = normden(xb)/normal(xb)
. regress hwage age agesq exp expsq pric seci secc supi supc lambda if hwage>0,
robust
And compare the estimates to those obtained using "heckman":
Notice that obtaining the two-stage estimates "by hand" you cannot use the standard
errors directly.
Maximum-likelihood estimation
Under the (more restrictive) assumption that (u1i , u2i ) follows a bivariate normal
distribution it is possible to construct a maximum-likelihood estimator. If you do not
specify the "twostep" option, then Stata fits such an estimator:
. heckman hwage age agesq exp expsq married head spouse headw spousew pric seci
secc supi supc school, select(age agesq exp expsq married head spouse headw
spousew pric seci secc supi supc school)
This strategy is usually discarded. See for example Nawata y Nagase (????). In the
Problem Set you will be asked to compare both estimations.
Drawbacks
The classical problem with the Heckman model is the high correlation between x1i
and zi . Since the function () is monotonously increasing, if its argument has little
70
variation then it may resemble a linear function. If x1i is similar to x2i , then the
correlation between x1i an zi may be high. Indeed, the identification of needs
strictly () not to be linear. If x1i and x2i have a lot of variables in common, then
the second stage will be subject to a problem of high multicollineality. The final
result are insignificant coefficients on and 1 .
We can obtain a measure of how bad this problem could be. First estimate the
Heckman model using the option "mills(newvar)", which stores () in "newvar":
mills(mills)
And then regress () on x1i :
. reg mills age agesq exp expsq pric seci secc supi supc
The higher the resulting R2, the higher the multicollineality problem. In this
particular example high multicollineality does not seem to be an issue.
Marginal effects
The marginal effects for the expected value of the dependent variable conditional on
being observed, E(y | y observed), are:
. mfx compute, predict(ycond)
The marginal effects for the probability of the dependent variable being observed,
Pr(y observed), are:
. mfx compute, predict(psel)
71
Chapter 9
Partially continuous variables
Some random variables are discrete: each value in the support can happen with
positive probability (e.g. bernoulli, poisson). Some random variables are continuos:
each point in the support happens with probability zero, but intervals happens with
positive probability (e.g. normal, chi-squared). But some variables are partially
continuous: some points in the support have positive probability.
Consider expenditure in tobacco as an example: an individual may not smoke with
positive probability; but given that he/she smokes, then the tobacco expenditure is a
continuous variable (i.e. the probability that he/she will expend exactly $121,02 is
null).
This discontinuity usually responds to either censorship or corner solutions. Consider
for example the following simple model:
y = x + u
y = max(0, y )
With E (u | x) = 0 . The variable y may be expenditure in food, and y (the latent

variable) food consumption. If we observed an i.i.d. sample of ( yi , xi ) ; i = 1, , n,
then we would be able to estimate consistently regressing yi on xi by OLS.
However, we must find out what happens when we have data only on ( yi , xi ) .
Following our previous example, the surveys measure expenditures in different
goods, but not consumption. We may regress yi on xi for the entire sample, or we
may regress yi en xi only for those observations with yi > 0 . We will show that both
strategies lead to inconsistent estimators.
Consider first the estimation of by OLS regressing yi on xi only for the
observations with yi > 0 (the truncated sample). We need to check out what happens
with E ( y | x, y > 0) . If ui
N (0, 2 ) :
y | x, y > 0 = x + ( u | x, y > 0 )
E ( y | x, y > 0) = x + E (u | x, y > 0)
= x + E (u | x, u > x )
= x +
72
( x )
1 ( x )
= x +
( x / )
( x / )
= x + ( x / )
Where ( z ) = ( z ) ( z ) is known as the inverse Mills ratio. If we regressed yi on xi
then ( x / ) would appear in the error term. Since the latter is correlated with
xi , the OLS estimate on would be inconsistent.

We must prove that OLS also yields inconsistent estimates if using the censored
sample:
E ( y | x) = E ( y | x, y > 0) P ( y > 0 | x) + E ( y | x, y 0) P ( y 0 | x)
= E ( y | x, y > 0) P ( y > 0 | x)
= [ x + ( x / )] P (u > x )
= [ x + ( x / )] ( x )
Maximum-likelihood estimation
The model with censored data with normal errors is known as Tobit Model. Under the
normality assumption it is relatively easy to obtain a consistent maximum-likelihood
estimator. Let's begin by dividing the sample ( yi , xi ) in pairs with yi = 0 and pairs
with yi > 0 . The first happen with probability P( yi = 0 | xi ) , while the second follow
the density distribution f ( yi | xi , yi > 0) . Define wi 1[ yi > 0] . Then:
L( ) =
P( y
i| yi = 0
= 0) f ( yi | xi , yi > 0)
i| yi > 0
= P ( yi = 0)(1 wi ) f ( yi | xi , yi > 0) wi
i =1
l ( ) = i =1 (1 wi ) ln (1 ( xi ) ) + wi ln ( (1 ) (( yi xi ) )
n
Under standard conditions the maximization problem has a unique solution, which
can be easily retrieved using maximization algorithms.
Interpreting the results
The maximum-likelihood coefficients have the following interpretation:
E ( y | x )
=
x
They are the partial derivatives on the latent variable (e.g. consumption). But
sometimes the interest must be focused on different marginal effects. Recall that:
E ( y | x) = E ( y | x, y > 0) P( y > 0 | x) = [ x + ( x )] ( x )
Using the chain-rule:
73
E ( y | x) E ( y | x, y > 0)
P ( y > 0 | x)
=
P( y > 0 | x) + E ( y | x, y > 0)
xk
xk
xk
The above is the Donald-Moffit decomposition. It is straightforward to show that:
E ( y | x)
= k ( x )
xk
This is a rather simple re-scaling of the original estimate. However, as in the logit
model, the marginal effects depend on x. In our example, this would be the marginal
derivative on expenditures: part of the effect comes from the families that start
expending because of the change in x, and the other part comes from the increase in
expenditures for those families that were already expending.
We can also obtain the following:
E ( y | x, y > 0)
= k 1 ( x ) [ ( x ) + ( x )]
xk
It is also a re-scaling. This would be the effect on expenditures only for those
families who were already expending something.
An example: wage determination for women
We want to estimate a rather simple model on the determination of income for
women. We have data ("women.dta") on 2556 women between 14 and 64 years old
living in Buenos Aires in 1998.2 Enter "describe" to see the data definitions:
. use women, clear
. describe
Then, regressing wages on some individual characteristics we can estimate the effect
of those variables on the wages' determination. The variable "hwage" stores the
hourly wage in US dollars, and it takes the value zero if the woman does not work. As
potential explanatory variables and controls we have variables on education, age,
number of kids, marital status, etc.
Regarding our theoretical model, the y would be the "potential" wage (e.g. if
compelled to work, a woman may be paid less than her reservation wage), and y*
would be the "actual" wage (e.g. if a woman is offered less than her reservation
wage, then she will not work and she will earn nothing).
Let's estimate the following wage determination equation (inconsistently) by OLS
using the censored sample and the truncated sample:
2
Extracted from an Argentinean household panel called "Encuesta Permanente de Hogares",

and provided by Walter Sosa Escudero.
74
. reg hwage age agesq exp expsq married head spouse children hhmembers headw
spousew pric seci secc supi supc school, robust
. reg hwage age agesq exp expsq married head spouse children hhmembers headw
spousew pric seci secc supi supc school if hwage>0, robust
Now we will retrieve the Tobit coefficients:
. tobit hwage age agesq exp expsq married head spouse children hhmembers headw
spousew pric seci secc supi supc school, ll(0)
With the option "ll(0)" we are indicated that the sample is left-censored at zero. We
know that = E ( y | x) x is the marginal effect of x on the (conditional expected
value of) "potential" wages. For instance, the impact of being married on the "latent"
wage is -8.6. Following the previously introduced formulas, we can also estimate the
marginal effect of x on the expected "actual" wage (at the mean x ), and the
marginal effect of x on the expected "actual" wage conditional on being uncensored
(at the mean x ).
We will take advantage of the command "dtobit", which you must install first. For
instance:
. version 6.0
. dtobit
. version 9.1
Notice that we need to specify "version 6.0" before running the Tobit model. The
reason is that in version 6.0 Stata stores the estimate for sigma, which is necessary
to carry on "dtobit". If you do not want to calculate the marginal effects at the mean
x, you may specify the values for x using the option "at()".
For instance, the effect of being married on the "actual" hourly salary income is -3,
while the effect on the "actual" hourly salary income for working women is only 2.49. The differences with respect to the latent variable are considerable.
Notice that the latter two estimates are respectively the "consistent" versions of the
OLS regressions performed at the beginning of the example. Comparing them you can
make up an idea of how much biased were those early estimates (in this example,
considerably).
Exercise 6.1: Reproduce the latter two marginal effects "by hand" (i.e. as if the
command "dtobit" did not exist).
75
Exercise 6.2: Obtain the maximum-likelihood estimates "by hand" (i.e. using the
command "ml").
Exercise 6.3: Choose a continuous explanatory variable. Then graph a scatterplot of
that variable and hourly wages, and add the regression lines for each one of the
estimated coefficients showed above (including the inconsistent OLS estimates).
For the marginal effects, you can use the command mfx as well. The marginal
effects for the probability of being uncensored are obtained in the following way:
mfx compute, predict(p(a,b)), where a is the lower limit for left censoring and
b is the upper limit for right censoring. In our example:
. mfx compute, predict(p(0,.))
The marginal effects for the expected value of the dependent variable conditional on
being uncensored and the marginal effects for the unconditional expected value of
the dependent variable are obtained respectively:
. mfx compute, predict(e(0,.))
. mfx compute, predict(ys(0,.))
LR Test
Let L ( ) be the log-likelihood function, let be the unrestricted estimator, and let
be the estimator with the Q nonredundant constraints imposed (e.g. with less
explanatory variables). Then, under the regularity conditions, the likelihood-ratio
(LR) statistic LR = 2[ L() L ( )] is distributed asymptotically as Q2 under H 0 .
For instance, the following model reaches a log-likelihood of about -4513.77:
. tobit hwage age agesq exp expsq married head spouse headw spousew pric seci secc
supi supc school, ll(0)
If the variables "hhmembers" and "children" are added, then the log-likelihood
becomes about -4496.83:
The likelihood ratio statistic is about 2(-4513.77 (-4496.83)) = -33.88. As it follows a
22 , the implied p-value is practically zero (enter "display 1-chi2(2,33.89)").

Therefore, these two variables are jointly significant.
As in every maximum-likelihood model, "lrtest", which does all the calculation above,
is offered as a postestimation command:
76
. est store A
. tobit hwage age agesq exp expsq married head spouse headw spousew pric seci secc
supi supc school, ll(0)
. lrtest A
77
Chapter 10
Logit/Probit Models
10.1. Multinomial Logit (Probit)

The logit model for binary outcomes can be extended to the case where the response
has more than two outcomes. For instance: political views (Republican, Independent
or Democrat), occupational choice (to work, study or stay at home), the commuting
to work (by bus, car, train or bicycle), etc. The individual has to choose only one
alternative from the group of choices.
Let y denote a random variable taking on the values {0; 1; . . . ; J} for J a positive
integer, and let x denote a set of conditioning explanatory variables (a 1xK vector
including a constant). For example, in the commuting to work example y would be
{j=0: bus; j=1: car; j=2: train; j=3: bicycle}, and x would probably contain things like
income, age, gender, race, marital status and number of children. As usual, the
( xi , yi ) are i.i.d.
We are interested in how ceteris paribus changes in the elements of x affect the
response probabilities, P ( y = j | x) , j = 0, 1,, J. Since the probabilities must sum to
unity, P ( y = 0 | x) will be determined once we know the probabilities for j = 1,, J.
We have to think on the multinomial model as a series of binary models. That is,
evaluate the probability of the alternative j against alternative i for every i j . For
instance, consider the binary model P ( y = j | y {i,j}, x) :
Pj
P ( y = j | x)
=
= F(X j )
P( y = i | x) + P( y = j | x) Pi + Pj
Obtenemos:
Pj = F ( X j )( Pi + Pj )
Pj
Pi
= F(X j )
Pi + Pj
Pi
F(X j ) F(X j )
F(X j )
=
=
= G( X j )
Pi
Pj
1 F ( X j )
1
Pi + Pj
Pi + Pj
Notice that:
Pj
P
j i
j i
Pi
1 Pi 1
= 1
Pi
Pi
P
1
= 1+ j
Pi
j i Pi
78
Using the expression for Pj Pi obtained above:
1
= 1 + G( X j )
Pi
j i
Pi =
1
1 + G( X j )
j i
Finally, solve for Pj :
Pj =
G( X j )
1 + G( X i )
i j
To find an explicit form for Pj we only have to substitute the G () by exp() , and
then we obtain the multinomial logit model:
P( y = j | x) =
exp( X j )
1 + exp( X i )
j = 1,..., J
i j
As the response probabilities must sum to 1, we must set the probability of the
reference response (j=0) to:
P( y = 0 | x) =
1
1 + exp( X i )
i 1
The ML are obtained through maximum-likelihood estimation:

n
L( ) = P ( yi = j | x)1[ yi = j ]
i =1 j = 0
McFadden (1974) has shown that the log-likelihood function is globally concave, what
makes the maximization problem straightforward.
The partial effects for this model are complicated. For continuous xk, we can
express:
ik exp( X i )
P( y = j | x)
i 1
= P ( y = j | x) jk
xk
1 + exp( X i )
i 1
Where ik is the k-th element of i . Notice that even the direction of the effect is
not entirely revealed by ik . You may find other ways to interpret the coefficients.
Conditional logit model
79
McFadden (1974) showed that a model closely related to the multinomial logit model
can be obtained from an underlying utility comparison. It is called the Conditional
logit model. Those models have similar response probabilities, but they differ in
some key regards. In the MNL model, the conditioning variables do not change across
alternative: for each i, xi contains variables specific to the individuals but not to the
alternatives. On the other hand, the conditional logit model cannot have variables
varying over i but not over j. However, using an appropriate transformation you can
obtain the MNL model using the conditional technique, as it turns out to actually
contain the MNL model as a special case.
An example: a simple model of occupational choice
Utilizaremos la base de datos "status.dta". Enter "describe" to see the data
definitions. It has been extracted from Wooldridge "Econometric Analysis of Cross
Section and Panel Data", example 15.4 (page 498). It is a subset from Keane and
Wolpin (1997) that contains employment and schooling history for a sample of men
for 1987. The three possible outcomes are enrolled in school ("status=0"), not in
school and not working ("status=1"), and working ("status=2"). As explanatory
variables we have education, past work experience, and a black binary indicator.
Open the database:
. use status, clear
. describe
Now we can enter the command "mlogit":
. mlogit status educ black exper expersq, b(0)
With "b(0)" we indicate that the base category is "enrolled in school" ("status=0").
Marginal Effects
We can calculate the marginal effects "by hand" using the formula derived above, or
we can simply take advantage of the command "mfx":
. mfx compute, predict(outcome(0))
Where "outcome()" denotes the response under consideration. For example, an
addition year of education (at the mean x) changes the probability of going to school
by +0.01, the probability of staying home by -0.05, and the probability of working by
+0.03. This is completely logical: in general people invest in education in order to get
further education (e.g. going to college in order to get a Ph.D. in the future) or they
80
invest in education in order to enter the job market. Thus, investing in education
reduces the probability of staying home.
We can also obtain predicted probabilities to provide some useful comparisons. For
instance, consider two non-black men, each with 3 years of experience (and then
"expersq=9"). Calculate the three predicted probabilities for a man with 12 years of
education:
. mfx, predict(p outcome(0)) at(12 0 3 9)
And calculate the predicted probabilities for a man with 16 years of education:
You can see that the 4 years of additional schooling changed the probability of going
to school by +0.06, the probability of staying home by -0.1, and the probability of
working by +0.04.
Tests for the multinomial logit model
You may install the command "mlogtest", which computes a variety of tests for
multinomial logit models.
You may select the test you want by specifying the
appropriate option. For each independent variable, "mlogtest" can perform a

likelihood-ratio or Wald test of the null hypothesis that the coefficients of the
variable equal zero across all equations. It can also perform Wald or LR tests of
whether any pair of outcome categories can be combined. In addition, "mlogtest"
computes both Hausman and Small-Hsiao tests of independence of irrelevance
alternatives (IIA) assumption.
For instance, let's perform the Wald tests for combining outcome categories. The null
hypothesis is that the difference between all coefficients except intercepts
associated with given pair of outcomes are 0 (i.e. the categories can be collapsed).
. mlogtest, combine
Exercise 7.1: Predict the probabilities of each response and observation: p ij . Then
generate an "educated guess" based on those predicted probabilities:
p% ij = max { p ij }
j
81
Then enter "tab2 p ij
p% ij " and calculate the proportion of "right" and "wrong"
predictions from the model.

Nested logit
The conditional logit model has as assumption the independence from irrelevant
alternatives (IIA). This means that relative probabilities for any two alternatives
depend only on the attributes of those two alternatives. Thus, it implies that adding
another alternative or changing the characteristics of a third alternative does not
affect the relative odds between two alternatives. This implication is implausible for
applications with similar alternatives. For more details, see Wooldridge "Econometric
Analysis of Cross Section and Panel Data", page 501.
A different approach to relaxing the IIA assumption is the specification of a
"hierarchical model". The most popular of these is the nested logit model. Suppose
that the total number of alternatives can be put into S groups of similar alternatives
Gs . Thus the first hierarchy corresponds to which of the S groups y falls into, and the
second corresponds to the actual alternative within each group. We can propose
separate models for those probabilities:
P( y Gs | x) and P( y = j | y Gs , x)
The response probability P ( y = j | x) , which is ultimately of interest, is obtained by
multiplying both equations. The problem can be solved either in a two-step fashion
or using the full maximum likelihood.
For instance, consider the model of commuting to work. First you might want to
decide whether to travel by car, to travel "naturally", or by public transportation.
Once you decided to travel by car, you have to decide whether to travel alone or
carpooling. Once you decided to travel "naturally", you have to decide whether to
travel by foot or in bicycle. Once you decided to travel by public transportation, you
have to decide whether to travel by bus or by train.
10.1. Multinomial Logit (probit)

Another kind of multinomial response is an ordered response. As the name suggests,
if y is an ordered response, then the values we assign to each outcome are no longer
arbitrary. For example, y might be a credit rating on a scale from zero to six, with y
= 6 representing the highest rating and y = 0 the lowest rating.
The fact that 6 is a better rating than 5 conveys useful information, even though the
credit rating itself only has ordinal meaning. For example, we cannot say that the
difference between 4 and 2 is twice as important as the difference between 1 and 0.
82
Consider for instance the ordered probit model. Let y be an ordered response taking
on the values {0, 1,, J} for some known integer J. And define the latent variable as:
y = x + e, e | x
N (0,1)
Where x does not contain an intercept. Let 1 < 2 < ... < J be unknown cut points
(a.k.a. threshold parameters), and define:
0 if y 1
1 if 1 < y 2
y=
M
J if y >
J
Finally, compute each response probability:
P( y = 0 | x) = P( y 1 | x) = P ( x + e 1 | x ) = (1 x )
P( y = 1| x) = P(1 < y 2 | x) = ( 2 x ) (1 x )
M
P( y = J 1| x) = P( J 1 < y J | x) = ( J x ) ( J 1 x )
P( y = J | x) = P( y > J | x) = 1 ( J x )
The parameters and can be estimated by maximum likelihood. The magnitude
of the ordered probit coefficient does not have a simple interpretation, but you can
retrieve qualitative information directly from its sign and statistical significance. We
can also compute marginal effects with respect to xk.
As always, replacing the normal cumulative distribution by the logistic yields the
ordered logit model.
Example: life satisfaction in Russia
Recall the Russian database used in the first two weeks ("russia.dta"). We will
estimate a model explaining life satisfaction using health and economic variables. It
will also include dummies for geographical areas (area*). The variable "satlif"
measures life satisfaction, and takes values from 1 ("not at all satisfied") to 5 ("fully
satisfied").
First open the database, generate the geographical dummies, and transform the
household expenditures to thousand rubles:
. use russia, clear
. gen area1=(geo==1)
. gen area2=(geo==2)
83
. replace totexpr= totexpr/1000

Then estimate the ordered probit model:
. oprobit satlif monage obese smokes operat hattac totexpr econrk resprk powrnk
work0 area1-area2, robust
We can interpret directly the sign and statistical significance of the coefficients.
Only the slopes on expenditures, power ranking, respect ranking and economic
ranking are statistically different from zero, and they are all positive. As expected,
more power, respect, status and consumption raises life satisfaction.
We can use "mfx" to retrieve marginal effects at the mean x:
. mfx, predict(outcome(1))
An increase in household expenditures of ten thousand rubles would change the
probability of being not at all satisfied by -0.04, the probability of being less than
satisfied by -0.02, the probability of being both yes and no satisfied by 0.02, the
probability of being rather satisfied by 0.03, and the probability of being fully
satisfied by 0.01. Once again: more consumption implies a lower probability of being
relatively dissatisfied and than a higher probability of being satisfied.
Exercise 7.2: Predict the probabilities of each response and observation: p ij . Then
generate an "educated guess" based on those predicted probabilities:
p% ij = max { p ij }
j
Then enter "tab2 p ij
p% ij " and calculate the proportion of "right" and "wrong"
predictions from the model.
84
Chapter 11
Quantile Regression
An Introduction
Notice that we can define the sample mean as the solution to the problem of
minimizing a sum of squared residuals, we can define the median as the solution to
the problem of minimizing a sum of absolute residuals (Koenker et al., 2001). Since
the symmetry of the absolute value yields the median, minimizing a sum of
asymmetrically weighted absolute residuals yield other quantiles. Then solving:
min ( y i )
Yields the -th sample quantile as its solution, where the function
is:
( z ) = 1[z < 0] z (1 ) + 1[z 0] z . Con esa optimizacin podemos estimar

quantiles de la unconditional mean, pero it is easy to define conditional quantiles in
an analogous fashion. Given a random sample {y1 , y 2 ,..., y n }, we solve:
min ( y i ( xi , ) )
If ( xi , ) is formulated as a linear function of , then the problem can be

efficiently solved by linear programming methods.
An example
Open the database:
. use russia, clear
Command qreg fits quantile regression models. Option quantile(x) estimate
quantile x. The default is quantile(.5). Por ejemplo, vamos a correr una median
regression of real household expenditures contra real household income y una
constante:
. qreg totexpr tincm_r, quantile(.5)
Using other quantiles:
Command bsqreg estimates a quantile regression with bootstrap standard errors.
Option reps() specifies the number of bootstrap replications to be used. The
default is 20, which is arguably too small.
85
. bsqreg totexpr tincm_r, quantile(.5) reps(100)

Graphical comparisons
Solo para mejorar el grfico (aunque genere un bias) vamos a dropear los incomes y
expenditure mayores a 50,000 rubles:
. drop if totexpr>50000 | tincm_r>50000
Vamos a hacer el grfico de regresin parcial para los quantiles .25, .5 y .75, y para
la regresin OLS:
. predict fitted25
. predict fitted50
. predict fitted75
. reg totexpr tincm_r
. predict fittedols
. graph twoway scatter totexpr tincm_r || line fitted25 fitted50 fitted75 fittedols
tincm_r, legend(label(1 "Expenditures") label(2 "Quantile .25") label(3 "Quantile .5")
label(4 "Quantile .75") label(5 "OLS"))
86
Chapter 12
Robust inference
12.1. Clustered Standard Errors

Consider the following linear regression model:
y ij = xij + i + eij
Where i denotes units and j denotes groups (e.g. individuals within households,
students within schools). Notice that panel data is a particular case. In OLS the usual
assumption is that the error term ( eij ) is independently and identically distributed.
However, this is noticeably violated in many applications. Since the consistency of
depends on E(x ij | e ij ) , the bias arises in the estimates of its standard errors.
One of the common violations to the i.i.d. assumption is that errors within groups are
correlated in some unknown way, while errors between groups are not. This is known
as clustered errors, and OLS estimates of the variance can be corrected following a
Eicker-Huber-White-robust treatment of errors (i.e. making as few assumptions as
possible). We will keep the assumption of no correlation across groups, but we will
allow the within-group correlation to be anything at all.
First recall the OLS variance estimator:
VOLS = s 2 ( X X ) 1 , where s 2 =
1
N
e2
i =1 i
N k
And the heterocedasticity-robust (unclustered) variance estimator is:
V R = ( X X ) 1
i =1
(ei xi )(ei xi ) ( X X ) 1
Finally, the robust cluster variance estimator is:
VC = ( X X ) 1
NC
j =1
u j u j ( X X ) 1 , where u j = j =C1 ei xi
N
Being N C the total number of clusters. For simplicity, I omitted the multipliers
(which are close to 1) from the formulas for robust and clustered standard errors.
Over-rejection
When you compare the first and the second estimates, generally the second yield
greater standard errors. The, you have heterocedasticity and you do not use the
robust se, then you are over-rejecting the null hypothesis that the coefficients are
zero.
87
But when you compare the robust (unclustered) y and the clustered variance
estimators, there is not a general result. If the within-cluster correlation is negative,
then within the same cluster there will be big negative eij along with big positive eij ,
and small negative eij with small positive eij . This would imply that the cluster sums
of eij xij have less variability than the individual eij xij , since within each cluster the
eijs will be cancelling each other. Then the variance of the clustered estimator
will be less than the robust (unclustered) estimator. In this case, using cluster-robust
standard errors will not only make your inference robust to within-cluster
correlation, but it will improve your statistical significance.
You repeat the reasoning for the opposite case (i.e. with positive within-cluster
correlation), and where there is no clustered errors. The following is an applied
example.
An example
Open the database:
. use education, clear
This is a database on test scores that was generated including within-school
correlation in the error term. The estimate of the coefficient on treatment is
consistent. Now we can estimate the unclustered and clustered standard errors:
. xtreg test treatment female educ income, fe i(school) robust
. xtreg test treatment female educ income, fe i(school) cluster(school)
In this example not accounting for within-cluster correlation would have led to wrong
inference at the 10% for the treatment coefficient.
Should I always use clustered standard errors?
If the assumptions are satisfied, and the error term is clustered, you will get
consistent standard error estimates if you use the cluster-robust estimator. On the
other hand, if the assumptions are satisfied and errors are not clustered, you will get
roughly the same estimates as if you had not specified cluster.
Why not always specify cluster? Well, the cluster-robust standard error estimator
converges to the true standard error as the number of clusters M approaches infinity,
not the number of observations N. Kezdi (2004) shows that 50 clusters (with roughly
equal cluster sizes) is often close enough to infinity for accurate inference.
Moreover, as long as the number of clusters is large, even in the absence of
clustering there is little cost of using clustered-robust estimates.
88
However, with a small number of clusters or very unbalanced cluster sizes, inference
using the cluster-robust estimator may be very incorrect. With finite M, the clusterrobust estimator produces estimates of standard errors that may be substantially
biased downward (i.e. leading to over-rejection). See Wooldridge (2003) and
Cameron et al. (2006) for further discussions and suggestions on the matter.
Right specification
The within-cluster correlations may disappear with a correctly specified model, and
so one should always be alert to that possibility. Consider as an example a model
where the dependent variable is cell phone expenditure. If you only included
explanatory variables at the individual level (pooled OLS), then you would find
serious within-cluster correlation: since many of the calls are between household
members, the errors of the individuals within the household would certainly be
correlated to each other. However, a great deal of such within-cluster correlation
would disappear if you used a fixed-effects model. Furthermore, by adding the right
predictors the correlation of residuals could almost disappear, and certainly this
would be a better model.
For instance, you can see whether in the above model the difference between the
unclustered and clustered standard errors is magnified or not after controlling for
fixed effects:
. reg test treatment female educ income, robust
. reg test treatment female educ income, cluster(school)
. xtreg test treatment female educ income, fe i(school) robust
Nested multilevel clustering
Beyond the basic one-dimensional case, one may consider a multiple-level clustering.
For instance, the error term may be clustered by city and by household, or it may be
clustered by household and by year. In the first example the levels of clustering are
called nested. To estimate cluster-robust standard errors in the presence of nested
multi-level clustering, one can use the svy suite of commands. However, specifying
clustering solely at the higher level and clustering at the higher and lower level is
unlikely to yield significantly different results.
Nonlinear models
In Wooldridge (2006) there is an entire section of nonlinear clustered Standard
errors, with examples and Stata commands in the Appendix.
89
12.2. Bootstrapped standard errors

While a consistent estimator may be easy to obtain, the formula for the variancecovariance matrix is sometimes more difficult, or even mathematically intractable.
As a consequence, some applications need a lot of assumptions to construct an
estimate for that matrix.
Bootstrapping is a nonparametric approach for evaluating the distribution of a
statistic based on random resampling. It is very useful because it only assumes that
the sample is representative of the population. To retrieve such estimates, you have
to follow three steps: 1. Draw random samples with replacement repeatedly from the
sample dataset; 2. Estimate the desired statistic corresponding to these bootstrap
samples; 3. Calculate the sample standard deviation of the sampling distribution.
Stata regression commands usually have bootstrapped standard errors as an option.
For instance, in the above model:
. xtreg test treatment female educ income, fe i(school) vce(boot)
Please notice that the bootstrapped estimates vary among every estimations. For
instance, run the above code again:
. xtreg test treatment female educ income, fe i(school) vce(boot)
The bootstrapped option has some useful options. The "reps()" option specifies the
number of bootstrap replications to be performed. The default is 50, and 50-200
replications are generally adequate for estimates of standard errors. The "size()"
option specifies the size of the samples to be drawn. The default is the same size as
the data. For further information see for instance Andrews et al. (2000), who
proposed a three step method of choosing the number of bootstrap repetitions for
bootstrap standard errors, confidence intervals, confidence regions, hypothesis tests
and p-values.
If the option "strata()" is specified, bootstrap samples are taken independently within
each stratum. Finally, the "cluster()" option specifies the variables that identify
resampling clusters. This is the cluster-robust version for bootstrap. If this option is
specified, the sample drawn during each replication is a bootstrap sample of clusters.
This resampling method is called a pairs cluster bootstrap. In this case the default
size is the number of clusters in the original dataset. For instance:
. xtreg test treatment female educ income, fe i(school) vce(boot) cluster(school)
In Guan (2003) you can find a Montecarlo simulation that compares the power of
conventional and bootstrapped versions of the clustered estimates. Guan (2003) also
90
compares standard ML estimates versus bootstrapped in a Nonlinear least squares

regression, and it gives an example on 2SLS.
T-test bootstrapping
Following Cameron et al. (2006), the bootstrap-t procedure, proposed by Efron
(1981) for confidence intervals, computes the following Wald statistic for each
bootstrap replication:
wb = 1,b 1 s , b = 1,..., B
1, b
Where b denotes each resampling, s is a cluster-robust standard error for 1,b ,

1 ,b
and 1 is the original OLS coefficient. As we are interested in cluster-robust

estimates, the resampling is done at the cluster-level. Note that wb is centered on
1 . Then obtain the original Wald estimate:
w = 1 0 s
Obtaining s 1
from the CRVE standard errors, which controls for both error
heteroskedasticity
across
clusters
and
quite
general
correlation
and
heteroskedasticity within cluster (obtained using jointly the robust and cluster()
options).
Finally, we test HO: 1 = 0 against Ha: 1 0 . We reject H0 at level if w < w
1 2
or w > w
th
B.
, where w denotes the q quantile of w ,..., w
Obtain w and 1 :
. use education, clear
. matrix V=e(V)
. matrix b=e(b)
. scalar se=sqrt(V[1,1])
. scalar b1=b[1,1]
. scalar wfirst=b1/se
Then obtain wb (say) 100 times:
. forvalues j=1(1)100 {
. quietly: preserve
. quietly: bsample , cluster(school)
91
. quietly: xtreg test treatment female educ income, fe i(school) cluster(school)

. quietly: matrix V=e(V)
. quietly: matrix b=e(b)
. quietly: scalar se=sqrt(V[1,1])
. quietly: scalar beta=b[1,1]
. quietly: scalar w`j'=(beta-b1)/se
. quietly: restore
.}
Mostremos el w:
. display wfirst
Save the 100 variables wb in 100 observations of a new variable, wstar:
. gen wstar=.
. forvalues t=1(1)100 {
. quietly: replace wstar= w`t' if _n==`t'
. }
We need to see if w < w
or w > w1 2 . Then show the quantiles /2% and
(1 ) /2%, and compare it to w (variable wfirst):

. tabstat wstar, stat(p1 p5 p95 p99)
12.3. Two-Stage Models

Thorough this notes we will follow Hardin (2002)3 and Greene (2000, Chapter 17.7).
Just start noticing that some econometric models involve a two-step estimation
procedure:
First stage:
E [ y1 | x1 ,1 ]
Second stage:
E [ y 2 | x 2 , 2 , E ( y1 | x1 , 1 )]
There are two standard approaches to such estimation. The first is a full information
maximum likelihood model (FIML), which consists in specifying the joint distribution
He also presents a sandwich estimate of variance as an alternative to the MurphyTopel
estimate.
92
f (y1 , y 2 | x 1 , x 2 , 1 , 2 ) and maximizing its corresponding log-likelihood function. The

second approach is denominated the limited information maximum likelihood (LIML)
two-step procedure. Since the first model only involves estimating the first
parameter vector ( 1 ), we can estimate it separately. Then we can estimate the
second parameter vector conditional on the results of the first step estimation. This
is performed by maximizing the following conditional log-likelihood:
n
L = ln f y 2i | x 2i , 2i , ( x 1i ,1i )
i =1
The Murphy-Topel estimator

For the second approach, Greene (2000) briefly summarizes the results from Murphy
and Topel (1985). Assume that 1 is a q1 vector of unknown parameters associated
with an nq matrix of covariates X . Additionally, 2 is a p1 vector of unknown
parameters associated with an n p matrix of covariates W . The MurphyTopel
formula for the variance estimate of 2 is then given by:
V2MT = V2 + V2 (CV1C RV1C CV1 R )V2

Where:
V1 (a q q matrix) is the Asymptotic variance matrix of 1 based on L1 (1 )
V2 (a p p matrix) is the Asymptotic variance matrix of 2 based on L2 ( 2 | 1 )

L L2
C (a p q matrix) equals E 2
2 1
L L
R (a p q matrix) equals E 2 1
2 1
Matrices V1 and V2 can be estimated in many ways. For instance, in the following
example we will use the robust versions.
An example
We will consider an example from Hardin (2002), using data from Greene (1992). The
interesting dependent variable is the number of major derogatory reports recorded in
the credit history for the sample of applicants to a credit card. There are 1319
observations in the sample for this variable (10% of the original dataset), 1060 of
which are zeros. A Poisson regression model will be then a reasonable choice for the
second stage.
In a first stage we will estimate of a model of credit card applications outcome,
using a logit model. Then we will include the predicted values of the latter model as
explanatory variable in the Poisson regression. Notice that using the full information
93
approach would involve writing a very complex joint distribution. Furthermore,

maximizing such likelihood might yield inconsistent estimates. Then, we will pursue
the two-step approach.
From the original study, the author makes only 100 observations available for our
use. The following are the description of the variables involved:
Variable Name
Description
=1 if application for credit card accepted
Number of major derogatory reports
Age
Age in years plus twelfths of a year
Income
Yearly income (divided by 10,000)
Expend
Average monthly credit card expenditure
Ownrent
=1 if own their home
Selfemp
=1 if self employed
Open the database:

. use credit, clear
Entonces podemos correr los modelos:
. logit z age income ownrent selfemp, robust
. predict zhat
. poisson y age income expend zhat, robust
. drop zhat
dep var: z
age
income
ownrent
selfemp
cons
First Stage
coef
std error
-0.073
0.219
0.189
-1.944
2.724
0.035
0.232
0.628
1.088
1.060
dep var: y
p-value
0.035
0.345
0.763
0.074
0.010
age
income
expend
zhat
cons
Second Stage
coef
std error
0.073
0.045
-0.007
4.632
-6.320
0.048
0.178
0.003
3.968
3.718
p-value
0.125
0.800
0.022
0.243
0.089
This output shows the "naive" (robust) standard errors (i.e. which assume that there
is no error in the generation of z i in the first-stage).
Within Stata it is not too difficult to obtain the MurphyTopel variance estimates for
a two-stage model. We need to gather all the information for calculating the matrix
V2MT . First we have to obtain V1 and V2 (V1r and V2r in the following code). We will
94
also obtain z i , y i and the coefficient on z i in the second stage (zhat, yhat and zz in
the code), as we will need them later:
. logit z age income ownrent selfemp, robust
. matrix V1r = .99 * e(V)
. predict zhat
. poisson y age income expend zhat, robust
. matrix V2r = .99 * e(V)
. predict yhat
. scalar zz = _b[zhat]
You may notice that we have multiplied the robust variance-covariance matrices by
0.99. This is because Stata applies a small sample adjustment n (n 1) to the robust
variance estimates, and then we can undo that adjustment by multiplying
accordingly.
Then we have to calculate R and C . First we have to do a little bit of algebra. A
logistic model where z is the outcome (whether the application is successful) and
X is the matrix of covariates has log likelihood given by:

n
L1 = ( z i xi 1 ln{1 + exp( xi1 )})

i =1
A Poisson model where y is the outcome (number of derogatory reports) and W is

the matrix of covariates has log-likelihood given by:
n
L2 = ( y i wi 2 exp( wi 2 ) ln ( y i + 1) )
i =1
Differentiating we obtain:
n
L1
= xi ( z i zi )xi = X diag ( z i zi ) X
1 i =1
n
L2
= xi ( y i y i )zi (1 zi )2 xi = X diag (( y i y i )zi (1 zi )) X
1 i =1
n
L2
= wi ( y i y i )w i = W diag ( y i y i )W
2 i =1
Where 2 is the estimate obtained from the two-step procedure. The "matrix accum"
command calculates XX. Lets obtain C and R :
. gen byte cons = 1
95
. matrix accum C = age income ownrent selfemp cons age income expend zhat cons
[iw=(y-yhat)*(y-yhat)*zhat*(1-zhat)*zz], nocons
. matrix accum R = age income ownrent selfemp cons age income expend zhat cons
[iw=(y-yhat)*(z-zhat)], nocons
And keep only the desired partition:
. matrix list C
. matrix C = C[6..10,1..5]
. matrix list R
. matrix R = R[6..10,1..5]
Finally, we can calculate the MurphyTopel estimate:
. matrix M = V2r + (V2r * (C*V1r*C - R*V1r*C - C*V1r*R) * V2r)
. matrix list M
To show the MurphyTopel estimates, Hardin (2002) suggested a rather useful way. In
case it already exists, drop the program "doit":
. capture program drop doit
Then define it:
. program define doit, eclass
. matrix b = e(b) // stores the coefficients in b
. ereturn post b M // saves the old coefficients along with the new standard errors
dep var: y
age
income
expend
zhat
cons
Second Stage
coef
std error
0.073
0.045
-0.007
4.632
-6.320
0.314
1.483
0.019
30.591
24.739
p-value
0.816
0.976
0.723
0.880
0.798
. ereturn local vcetype "Mtopel" // names the new standard errors

. ereturn display // shows the regression table
. end
Thus, after running a regression and calculating "M", just run "doit":
. doit
96
Hole (2006) shows how the calculation of the Murphy-Topel variance estimator for
two-step models can be simplified in Stata by using the scores option of predict.
97
Chapter 13
Matching
98
References
[1] Verbeek, Marno (2000), A Guide to Modern Econometrics - John Wiley & Sons
Eds.
[2] Jenkins, S. (1995), "Easy Estimation Methods for Discrete-time Duration Models"
- Oxford Bulletin of Economics and Statistics, Vol. 57 (1), pp. 129-138.
[3] Lancaster, Tony (1990), The econometric analysis of transition data
Econometric Society Monographs.
[4] Koenker, Roger and Hallock, Kevin F. (2001), Quantile Regression - Journal of
Economic Perspectives, Volume 15 (4), pp. 143156.
[5] Cameron, Colin A.; Gelbach, Jonah and Miller, Douglas L. (2006), BootstrapBased Improvements for Inference with Clustered Errors NBER Technical
Working Paper No. 344.
[6] Efron, Bradley (1981), Nonparametric Estimates of Standard Error: The
Jackknife, the Bootstrap and Other Methods - Biometrika, Vol. 68 (3), pp. 589599.
[7] Guan, Weihua (2003), From the help desk: Bootstrapped standard Errors - The
Stata Journal, Vol. 3, pp. 7180.
[8] Kezdi, Gabor (2004), Robust Standard Error Estimation in Fixed-Effects Panel
Models - Hungarian Statistical Review Special, Vol. 9, pp. 96-116.
[9] Nichols, Austin and Schaffer, Mark (2007), Clustered Errors in Stata Mimeo.
[10]Wooldridge,
Jeffrey
M.
(2006),
Cluster-Sample
Methods
in
Applied
Methods
in
Applied
Econometrics: An extended analysis - Mimeo.

[11]Wooldridge,
Jeffrey
M.
(2003),
Cluster-Sample
Econometrics - The American Economic Review, Vol. 93 (2), pp. 133-138.

[12]Greene, W. (2000), Econometric Analysis- 4th ed., Upper Saddle River, NJ:
PrenticeHall.
[13]Greene, W. (1992), A statistical Model for Credit Scoring Working Paper No.
EC-92-29, New York University, Stern School of Business.
[14]Hardin, James W. (2002), The robust variance estimator for two-stage models
- Stata Journal, Vol. 2 (3), pp. 253-266.
[15]Hole, Arne Risa (2006), Calculating Murphy-Topel variance estimates in Stata:
A simplified procedure - Stata Journal, Vol. 6 (4), pp. 521-529.
99
[16]Murphy, K. M. and Topel, R. H. (1985), Estimation and inference in two-step

econometric models - Journal of Business and Economic Statistics, Vol. 3 (4),
pp. 370379.
[17]Sophia Rabe-Hesketh and Brian Everitt, "A Handbook of Statistical Analices Using
Stata" - Chapman & Hall/CRC, 2007.
(http://www.stata.com/bookstore/statanalyses.html)
[18]Norman R. Draper and Harry Smith, "Applied Regression Analysis" Wiley, 2007.
(http://www.stata.com/bookstore/ara.html)
[19]Lawrence
C.
Hamilton,
"Statistics
with
(http://www.stata.com/bookstore/sws.html)
100
Stata"
Brooks/Cole,
2006.

Applied Econometrics Using Stata

Uploaded by

Copyright:

Available Formats

Applied Econometrics Using Stata

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Applied Econometrics Using Stata

Uploaded by

Copyright:

Available Formats

What topics are covered in the book?

What topics are covered in the book?

What are some of the econometric techniques discussed?

What are some of the econometric techniques discussed?

Applied Econometrics using Stata

. list gender totexpr in 1/10

Rename and label

You can now see the final product:

You can also reshape it again back to wide:

Tabulate, tab1 and table

. graph twoway scatter htself height, by(gender)

. graph pie totexpr, over(site)

Some Programming in Stata

use `file', clear

. generate xlag = x[_n-1]

. capture: blablabla // This is wrong

Creating Random Data and Random Samples

wagei = 3intelligencei + 1ai + 2bi + ui

Predictions and residuals

. graph twoway scatter yhat res evalhl

Partial regression plot

Normality of the residuals

Tests for heteroskedasticity

"By hand" regression

Then use the definition:

. mat vecaccum YpX = evalhl monage obese smokes

Open the database:

The two-stages "by hand"

b2 SLS = X ' Z ( Z ' Z ) 1 Z ' X X ' Z ( Z ' Z ) 1 Z ' Y

b2 SLS = ( X *' X ) 1 X *'Y = ( X *' X * ) 1 X *'Y

Test of Overidentifying Restrictions

4.1. Static Panels

CTit = 0 SBit + 1OBit + 2TBit + M t + Fi + it

. drop if month==72 | month==73

. reg dm_cartheft dm_instp dm_inst1p dm_month*, nocons robust

Furthermore, show that

they are numerically identical when T=2.

4.2. Dynamic Panels

Substituting by the definition of yi ,t , we can obtain:

yi ,t 1 and i ,t 1 are correlated by definition. However, notice that yi ,t 2 is correlated

official updates to Stata.

GDPi ,t = GDPi ,t 1 + 1insti ,t + 2entri ,t + i + i ,t

2 = 0.3 . Firstly, run a pooled OLS regression:

bear to increase efficiency. In these equations, predetermined and endogenous

4.3. Nonlinear Models

exp{( xi 2 xi1 ) ' }

population-averaged logit models:

Probit with random effects

The likelihood contribution of individual i will be the joint probability of observing

( y i1 ,..., y iT ) . Recall that in the one-period case:

f ( y i1 ,..., y iT | xi1 ,..., xiT , ) = t f ( y it | xit , )

f ( y i1 ,..., y iT | xi1 ,..., xiT , ) =

f ( y i1 ,..., y iT | xi1 ,..., xiT , i , ) f ( i )d i

We can use arbitrary assumptions about the distributions of it and i . If we

Tobit with random effects

f ( y i1 ,..., y iT | xi1 ,..., xiT , ) =

exp{ i + xi2 } exp{ i + xi1 } exp{ i + xi2 }

1 + exp{ i + xi2 } 1 + exp{ i + xi1 } 1 + exp{ i + xi2 }

exp{ i + xi2 }(1 + exp{ i + xi1 }) exp{ i + xi1 } exp{ i + xi2 }

b2 SLS = ( X ' X ) 1 X 'Y = ( X ' X ) 1 X *'Y

quietly replace `lnf' = $ML_y1ln(norm(`XB'))+(1-$ML_y1) ln(1-norm(`XB'))