Microeconometrics

M1 APE
Adam Zylbersztejn
[email protected]
 I. Introduction
Roadmap

- Discrete choice models

(binar/multiple outcomes)
- Censored/truncated models
- Selection models

2
Binary information
Multiple examples of economic variables
represented as indicator variables:
 Active (Act=1) / Inactive (Act=0)
 Higher eduction (Educ=1) / High school or less (Educ=0)
 High quality product (Qual=1) / Low quality (Qual=0)
 Married (Mar=1) / Unmarried (Mar = 0) ; etc.

Information on the realization of a continuous

variable may be reduced to an indicatior variable:
 Full time job (FT=1) / Part time (TP=0) → working time?
 Adult (Adult=1) / Minor (Adult=0) → age?
3
Encoding binary information
Various ways to encode the very same information:
 Active (Act=1) / Inactive (Act=0)
 Inactive (Inact=1) / Active (Inact=0)
 Active (Act=3) / Inactive (Act=999)
 Active (Act= « active») / Inactif (Act= « inactive »)

→ They are all equivalent!

→ Chosen values have no particular meaning per se.

4
Encoding binary information
Binary information → Indicator variable (0/1)
 Active (Act=1)
 Inactive (Act=0)

Advantage: a direct link between sample mean and

empirical probability of the outcome of interest:
mean(Act) = Empirical probability of being active

5
Multinomial variable
Another common form of discrete variables are the
multinomial ones:
 Inactive (Emp=0) / Unemployed (Emp=1) / Student
(Emp=3) / Employed (Emp=4) / Retired (Emp=5)
 Degree: MA/MSc or more (Ed=1) / BA/BSc (Ed=2) / High
school (Ed=3) / None (Ed=4)

Sometimes, continous information is only available

as a multinomial variable:
 Full time job (FT=1) / Part time job, at half time or more
(FT=2) / Less than half time job (FT=3)

6
Encoding multinomial
information
Different encoding yet same information:
 Inactive (Emp=0) / Unemployed (Emp=1) / Student
(Emp=3) / Employed (Emp=4) / Retired (Emp=5)
 Inactive (Emp=66) / Unemployed (Emp=23) / Student
(Emp=54) / Employed (Emp=12) / Retired (Emp=47)
 Inactive (Emp= « inact ») / Unemployed
(Emp= « unem ») / Student (Emp= « stud ») / Emplyed
(Emp= « empl ») / Retired (Emp= « reti »)

→ Strictly equivelent ways of encoding information

→ Once again, values have no instrinsic meaning
7
Encoding multinomial
information
Same information with n binary variables?
 Inactive = 0 / 1
 Unemployed = 0 / 1
 Student = 0 / 1
 Employed = 0 / 1
 Retired = 0 / 1

→ All categories must be mutually exclusive

→ Then, mean value of each variable gives the
relative probability of an underlying categories

8
Ordinal variable
There are cases in which absolute values have no
interpretation, but their relative values do:
 Full time job (FT=1) / Part time job, at least half time
(FT=2) / Less than half time job (FT=3)
→ working time: « 1 » > « 2 » > « 3 »

Although we have no precise information on the

realization of the underlying continuous variable
(working time), different categories (FT) can be
ordered in a meaningful way
9
Censoring/truncation/sample
selection
Somtimes, the qualitative dimension may not be
directly of interest, but rather define the scope of
available information
 Model of working hours (h)

 Censoring: h known excactly if high enough

(h>c), otherwise h=c (left-censoring)
 Truncation: only observation with h>c (left-
truncation)
 Wage equation

 Restricted to the observed wages: only

emplyed individuals (sample selection) 10
II. Preliminaries:
Qualitative explonatory variables

11
II.1. Indicator explonatory
variable
a.k.a. « dummy » variable

12
Dummy variables
An indicator variable (dummy) can only
take either of the two values: 0 or 1

Examples : sex (1 for males, 0 for

females), geographical location (1 for Paris,
0 otherwise), ethnicity (1 for Caucasian, 0
otherwise), etc.

13
Dummy explanatory variable
Consider a simple model with one continuous
variable (x) and one dummy variable (d)
y = α0 + α1 x + d0d + u

Geometrically, the OLS dummy coefficient is

interpreted as a translation of the model (i.e., a
change in the intercept, but not the slope)
 If d = 0, then y = α + α x + u
0 1

 If d = 1, then y = (α0 + d0) + α1x + u

d = 0 is the reference category

14
 Example: wage equation
Modèle 18: MCO, utilisant les observations 1-526
Variable dépendante: wage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const −1.65055 0.652317 −2.530 0.0117 **
educ 0.506452 0.0503906 10.05 7.56e-22 ***
male 2.27336 0.279044 8.147 2.76e-15 ***

Moy. var. dép. 5.896103 Éc. type var. dép. 3.693086

Somme carrés résidus 5307.161 Éc. type de régression 3.185520
R2 0.258819 R2 ajusté 0.255985
F(2, 523) 91.31542 p. critique (F) 9.66e-35

15
Example (assuming d0 > 0)
y y = (α0 + d0) + α1x In both cases,
slope = α1
d=1

d0 { d=0
y = α0 + α1x
}α 0

x
16
 Example
wage= 01 male 2 educu

This model measures mean wage gap between

males and females while controlling for years of
eduction

→ For a given education level (E), we have:

̄ Male =β0+ β1 + β2 E
wage
wage =β + β E  wage
 Male − wage
 Female = 1
̄ Female 0 2
17
Example
wage= 01 male 2 educu
Note that this model may also be estimated with a
dummy variabe « female » (i.e. taking males as
reference category)

wage=01 female 2 educu

This yields:

α0=β0−α1  1=−1  2 = 2
18
Example
wage=β0 +β1 +β 2 educ
wage=β0 + 2.27+β 2 educ Males (male = 1)

Females (male = 0)

wage= 0 2 educ

0 1= 02,27
0

19
 Example

wage=0 2 educ Males

Females

wage=01 2 educ
NB : ici α1 < 0
 0=0 1
 0 1=0

20
Example
Attention: one cannot jointly include an intercept
and both dummies female et male in a regression model

Reason: perfect multincolinearity between explanatory

variables (« dummy variable trap » ), since:

female + male = 1

Implications: matrix X does not have full rank, X'X is

not invertible, the OLS estimator cannot be computed

21
Modèle 18: MCO, utilisant les observations 1-526
Variable dépendante: wage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const −1.65055 0.652317 −2.530 0.0117 **
educ 0.506452 0.0503906 10.05 7.56e-22 ***
male 2.27336 0.279044 8.147 2.76e-15 ***

Moy. var. dép. 5.896103 Éc. type var. dép. 3.693086

Somme carrés résidus 5307.161 Éc. type de régression 3.185520
R2 0.258819 R2 ajusté 0.255985
F(2, 523) 91.31542 p. critique (F) 9.66e-35

Modèle 19: MCO, utilisant les observations 1-526

Variable dépendante: wage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const 0.622817 0.672533 0.9261 0.3548
educ 0.506452 0.0503906 10.05 7.56e-22 ***
female −2.27336 0.279044 −8.147 2.76e-15 ***

Moy. var. dép. 5.896103 Éc. type var. dép. 3.693086

Somme carrés résidus 5307.161 Éc. type de régression 3.185520
22
R2 0.258819 R2 ajusté 0.255985
Interpretation
wage= 01 male 2 educu
The dummy variable coefficient is interpreted
as an absolute difference between the two groups
(i.e., in montaryvalue)

β1 = 2.27
→ Mean difference in houly wage between
males and females equals $2.27

23
Alternative specification
Wage and income equations are generally
estimated in a semilogarithmic form :

ln wage= 01 male 2 educu

24
Alternative specification
ln(w)=β0+ β1 male+ β2 educ+ u
If the dependent variable is in logarithm, the
dummy variable coefficient is interpreted as the
relative mean difference between the two group (i.e.,
in percentage)
Exemple : β1 = 0,2 → mean difference of 20%
w Male
β1=ln(w Male )−ln(w Female )=ln( )
w Female
w Male −w Female w Male
= −1=exp(β1)−1≈β1 ( for β1≈0)
w Female w Female
25
 Semi-logarithmic specification
Modèle 15: MCO, utilisant les observations 1-526
Variable dépendante: lwage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const 0.465404 0.0912268 5.102 4.72e-07
***
educ 0.0772033 0.00704715 10.96 2.76e-25
***
male 0.360865 0.0390245 9.247 5.83e-19
***

Moy. var. dép. 1.623268 Éc. type var. dép. 0.531538

Somme carrés résidus 103.7982 Éc. type de régression 0.445496
R2 0.300220 R2 ajusté 0.297544
F(2, 523) 112.1887 p. critique (F) 2.87e-41

26
II.2. Indicator variables and
complex information

27
Dummy variables and multiple
categories
One can use dummy variables to model discrete
information with multiple categories
Example : geographical variable may take 4 values
(North, South, East et West)
 Create three dummies to capture this

information:
North = 1 if the region is North and 0 otherwise
South = 1 if the region is South 0 otherwise
West = 1 if the region is West and 0 otherwise
 Reference category: East
28
Dummy variables and multiple
categories
Every multinomial variables can be transformed
into a set of dummies
The reference category is captured by the
intercept For n categories, we only need n – 1
dummy variables
Risk of model saturation: the number of variables
may quickly become substantial (e.g., socio-
economic information) which requires redefining
the categories of interest
29
Dummy variables and multiple
categories
With multiple sets of dummies in the model (for
instance, one set for sex and one set for
geography), the reference category are
observations belonging to all the omitted
categories (e.g., females living in the East)

30
Example
Geographical information with 4 categories: East,
North (+ Centre), South et West:
→ have 3 dummies: east, northcen et south

ln  Incearn= 01 male 2 educ

3 east 4 northcen5 southu

 Reference: Female + West

31
 Logarithmic specification
Modèle 24: MCO, utilisant les observations 1-526
Variable dépendante: lwage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const 0.575731 0.101308 5.683 2.21e-08 ***
male 0.366500 0.0389607 9.407 1.63e-19 ***
educ 0.0761099 0.00707167 10.76 1.59e-24 ***
east −0.0713549 0.0623620 −1.144 0.2531
northcen −0.131511 0.0608907 −2.160 0.0312 **
south −0.142092 0.0573876 −2.476 0.0136 **

Moy. var. dép. 1.623268 Éc. type var. dép. 0.531538

Somme carrés résidus 102.3574 Éc. type de régression 0.443668
R2 0.309934 R2 ajusté 0.303298
F(5, 520) 46.71017 p. critique (F) 7.18e-40

32
Dummy variables and ordinal
information
If qualitative information is ordinal, a one-unit increase
in X may not have a constant (linear) effect on y
→ use a set of dummy variables
→ no ordinal structure per se, butordinal
interpretation
Example: classify insurance buyers according to risk
 1: negligible risk
 2: low risk
 3: medium risk
 4: high risk
 5: substantial risk
33
Dummy variables and ordinal
information
Sometimes, one can recover categorical information
from quantitative data in order to pin down non-
linearity
Examples:
 Age category

 Income category
 Education level as the number of years of

schooling
Trade-off: flexibility vs. greater number of
coefficients to estimate
34
II.3. Dummy variables and
interactions

35
Interactions between dummy
variables
Allows to measure the impact of one qualitative
variable on the effect that another qualitative variable
has on the dependent variable
Example : The effect of marital status on wage can
vary between males and females
 If we are only interested in the difference between
males and females, or between married and unmarried,
we only need to include two variables in the model:

ln wage= 0 1 male 2 married ...

36
 Interactions between dummy
variables
If we also want to account for the possibility that a
person's sex affects ways in which marital status
changes the expected wage, we need to include
interaction terms:

ln wage= 0  1 male 2 married.male 3 married.female...

λ 0 :Unmarried female(reference)
λ 0 +λ 1 :Unmarried male
λ 0 +λ 3 : Married female
λ 0 +λ 1 +λ 2 : Married male
37
 Marital status, sex and wage
Modèle 25: MCO, utilisant les observations 1-526
Variable dépendante: lwage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const 0.586773 0.0993957 5.903 6.44e-09 ***
educ 0.0733479 0.00675332 10.86 6.75e-25 ***
east −0.0527639 0.0594836 −0.8870 0.3755
northcen −0.122215 0.0580648 −2.105 0.0358 **
south −0.152881 0.0547493 −2.792 0.0054 ***
male 0.109183 0.0598959 1.823 0.0689 *
male_married 0.406780 0.0552726 7.360 7.29e-13 ***
fem_married 0.0384117 0.0534270 0.7190 0.4725

Moy. var. dép. 1.623268 Éc. type var. dép. 0.531538

Somme carrés résidus 92.59097 Éc. type de régression 0.422785
R2 0.375776 R2 ajusté 0.367341
F(7, 518) 44.54722 p. critique (F) 2.86e-49
38
Testing for difference
We can directly test for significance of the male-female
difference in the effect of marital status on wage:
ln wage= 0  1 male 2 married  3 married.male...
λ 0 :Unmarried female(reference)
λ 0 +λ 1 :Unmarried male
λ 0 +λ 2 : Married female
λ 0 +λ 1 +λ 2 +λ 3 : Married male
→ t-test : significativity of λ3: significativity of the
difference of the effet of being married on wage between
males and females
39
 Marital status, sex and wage
Modèle 26: MCO, utilisant les observations 1-526
Variable dépendante: lwage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const 0.586773 0.0993957 5.903 6.44e-09 **
educ 0.0733479 0.00675332 10.86 6.75e-25 **
east −0.0527639 0.0594836 −0.8870 0.3755
northcen −0.122215 0.0580648 −2.105 0.0358 **
south −0.152881 0.0547493 −2.792 0.0054 **
male 0.109183 0.0598959 1.823 0.0689 *
married 0.0384117 0.0534270 0.7190 0.4725
male_married 0.368368 0.0767202 4.801 2.06e-06 **

Moy. var. dép. 1.623268 Éc. type var. dép. 0.531538

Somme carrés résidus 92.59097 Éc. type de régression 0.422785
R2 0.375776 R2 ajusté 0.367341
F(7, 518) 44.54722 p. critique (F) 2.86e-49
40
Interactions between dummy and
continuous variables
Allows to grasp the difference in slopes across groups:

ln(wage)=β0+ β1 male+ β2 educ+ β3 exp+ β4 male.exp+ ...

β3: return on professional experience for females
β3+β4: return on professional experience for males

t-test on β4 : test of equality of returns on professional

experience for males and females

41
Modèle 28: MCO, utilisant les observations 1-526
Variable dépendante: lwage

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const 0.372205 0.107983 3.447 0.0006 ***
educ 0.0871848 0.00705587 12.36 6.62e-31 ***
east −0.0552700 0.0579905 −0.9531 0.3410
northcen −0.142256 0.0568711 −2.501 0.0127 **
south −0.155444 0.0533661 −2.913 0.0037 ***
male 0.0423943 0.0669275 0.6334 0.5267
married 0.00152592 0.0544990 0.02800 0.9777
male_married 0.292925 0.0791116 3.703 0.0002 ***
exper 0.00426238 0.00201803 2.112 0.0352 **
exp_male 0.00644014 0.00283404 2.272 0.0235 **

Moy. var. dép. 1.623268 Éc. type var. dép. 0.531538

Somme carrés résidus 87.56200 Éc. type de régression 0.411939
R2 0.409680 R2 ajusté 0.399384
F(9, 516) 39.78916 p. critique (F) 9.57e-54

42
 β4 > 0
Males
y y = β0 + β1 + (β3 + β4).exp

Females
y = β0 + β3.exp

exp
43
II.4. Dummy variables and
program evaluation

44
Program evaluation
Dummy variables seem particularly hand when it
comes to program evaluation
 Example: Income of people participating in a
training program vs. income of non-participants

Unfortunately, the participation status dummy

variable typically stems from a non-random
selection process (people self-select to the training
program)
45
Program evaluation
ln( Incearn)=β0 +β1 training +β 2 educ+...

Important notice:
Some individual determinants of selection to the
program (e.g., motivation) may be unobserved and
correlated with the unobserved determinants of
wage
 Dummy variable is not exogeneous
 The OLS estimator is biased
46
III. Qualitative information as
dependent variable

Binary choice model

47
III.1. Introduction

48
A little bit of history
Relatively recent models
 First contributions date back to the 1940's
 Berkson (Logit et Probit)
 Economic application even more recent (starting
in the 1970's)
 MacFadden et Heckman (Prix Nobel 2000)

49
Areas of application
In various economic settings (and not only –
psychology, sociology, etc.) the variable of interest
in binary:
 Choosing to pursue higher education (e.g., works

by Radner et Miller in the 1970's)

 What determines this choice?
 Activity or not on the labor market

 Why do people decide to enter the market?

 Credit default

 Determinants? Risk estimation?

50
III.2. Why not basic OLS?

- Binary dependent variable (LPM)

- Multinomial dependent variable

51
Linear Probability Model (LPM)
Formalization :

N observations of y i such that : y i =0 ∨ y i =1

K explanatory variables : vector X i of dimension K

→ y i = X i β+εi

{β : column vector of dimension K

εi i.i.d.

52
Interpretation of coefficients
y i = X i β+ εi
The model contains:

y i : qualitative information

X i β+εi : quantitative information
Encoding of the qualitative information is arbitrary
Different ways of encoding yield different values of β
→ β is generally not interpretable
53
Interpretation of coefficients
Binary encoding 1/0 allows for modeling the
probability that dummy yi equals 1 since:

P(y= 1|X) = E(y|X)

→ The OLS estimator of β k measures the

variation in the probability of success (y=1) as xk
changes

54
Linear approximation not
suitable
Simple case: univariate model
y i =β0+ β1 x 1+ εi

With qualitative dependent variable

→ Very peculiar scatterplot
→ Makes it impossible to fit a regression line in a
reasonable manner
→ LPM may still be useful for data
exploration/descriptive analysis

55
Non-normality of residuals
y i = X i β+ εi → εi = y i− X i β
With binary yi:

{
εi =1− X i β
ou
εi =0−X i βi=−X i β

→ εi may only take 2 possible values, conditional on

xi
→ εi follows a discrete distribution and thus cannot
be normally distributed 56
Violation of model's assumptions
We know that:
Prob( y i =1)= p i = X i β
→ 0 ≤ X iβ ≤ 1
however, ex ante we cannot be sure if the OLS
estimator satisfies this constraint

If ex post this constraint is not safisfied for some

values of xi then the estimation procedure is not
compatible with the underlying model assumptions
57
Heteroscedasticity
y i = X i β+ εi → εi = y i− X i β

Variance of the residuals:

2
V (εi )=E ((εi−E (εi )) )
2
=E (εi ) NB : E (εi )=0
2 2
= pi (1−X i β) + (1− pi )(− X i β)
2 2
= X i β(1− X i β) + (1−X i β)(− X i β)

→ V (εi )= X i β(1−X i β) ∀i∈[1... N ]

58
 Illustration: credit card and
income

. reg card income

Source | SS df MS Number of obs = 100

-------------+------------------------------ F( 1, 98) = 34.38
Model | 5.55556122 1 5.55556122 Prob > F = 0.0000
Residual | 15.8344388 98 .161575906 R-squared = 0.2597
-------------+------------------------------ Adj R-squared = 0.2522
Total | 21.39 99 .216060606 Root MSE = .40197

------------------------------------------------------------------------------
card | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
income | .0188458 .003214 5.86 0.000 .0124678 .0252238
_cons | -.1594495 .089584 -1.78 0.078 -.3372261 .018327
------------------------------------------------------------------------------

59
Illustration: credit card and
income

60
Illustration: credit card and
income

61
Linear model for multinomial
dependent variable
Employed / Self-employed / Retired / Inactive
1/2/3/4?
2/4/1/2?
4/1/3/2?
NB: For OLS, 3 = 3x1 = 1.5x2 , 4 = 4x1=2x2 !

→ Linear model completely useless

62
III.3. Solutions

63
More suitable estimation
methods
Binary variables
 Logit / Probit

Multinomial variables
 Multinomial Logit / Probit

Ordinal qualitative variables

 Ordered Logit / Probit

Censored/truncated/self-selected dependent variables

 Tobit / Heckit

→ Maximum likelihood estimator (MLE)

64
Example
Question: evaluation of a program to improve
learning economics
 Based on the article by Spector et Mazzeo (Journal
of Economic Education, 1980) as cited by par W.H.
Greene, Econometric Analysis, 5th ed.
Variables:
 grade: 1 if the grade in econ improved, 0 otherwise
 gpa: average grade in other course
 psi: 1 if a student took part in the program, 0 otherwise
 tuce: econ grade from the entry exam
65
 LPM (OLS - GRETL)
Modèle 1: MCO, utilisant les observations 1-32
Variable dépendante: GRADE

coefficient erreur std. t de Student p. critique

-----------------------------------------------------------------
const −1.49802 0.523889 −2.859 0.0079 ***
GPA 0.463852 0.161956 2.864 0.0078 ***
TUCE 0.0104951 0.0194829 0.5387 0.5944
PSI 0.378555 0.139173 2.720 0.0111 **

Moy. var. dép. 0.343750 Éc. type var. dép. 0.482559

Somme carrés résidus 4.216474 Éc. type de régression 0.388057
R2 0.415900 R2 ajusté 0.353318
F(3, 28) 6.645658 p. critique (F) 0.001571

66
LPM (OLS - Stata)

67
Logit (GRETL)
Modèle 2: Logit, utilisant les observations 1-32
Variable dépendante: GRADE
Écarts type basés sur la matrice hessienne

coefficient erreur std. z p. critique

------------------------------------------------------------
const −13.0213 4.93132 −2.641 0.0083 ***
GPA 2.82611 1.26294 2.238 0.0252 **
TUCE 0.0951577 0.141554 0.6722 0.5014
PSI 2.37869 1.06456 2.234 0.0255 **

Moy. var. dép. 0.343750 Éc. type var. dép. 0.482559

R2 de McFadden 0.374038 R2 ajusté 0.179786
Log de vraisemblance −12.88963 Critère d'Akaike 33.77927
Critère de Schwarz 39.64221 Hannan-Quinn 35.72267

Nombre de cas 'correctement prédis' = 26 (81.2%)

f(beta'x) à la moyenne des variables indépendantes = 0.189
Test du ratio de vraisemblance: Chi-deux(3) = 15.4042 [0.0015]

Prédit
0 1
Actuel 0 18 3
1 3 8 68
Logit (Stata)

69
Probit (GRETL)
Modèle 3: Probit, utilisant les observations 1-32
Variable dépendante: GRADE
Écarts type basés sur la matrice hessienne

coefficient erreur std. z p. critique

------------------------------------------------------------
const −7.45232 2.54247 −2.931 0.0034 ***
GPA 1.62581 0.693882 2.343 0.0191 **
TUCE 0.0517289 0.0838903 0.6166 0.5375
PSI 1.42633 0.595038 2.397 0.0165 **

Moy. var. dép. 0.343750 Éc. type var. dép. 0.482559

R2 de McFadden 0.377478 R2 ajusté 0.183225
Log de vraisemblance −12.81880 Critère d'Akaike 33.63761

Nombre de cas 'correctement prédis' = 26 (81.2%)

f(beta'x) à la moyenne des variables indépendantes = 0.328
Test du ratio de vraisemblance: Chi-deux(3) = 15.5459 [0.0014]

Prédit
0 1
Actuel 0 18 3
1 3 8 70
Probit (Stata)

71