Econometrics 2

4. Binary data methods

Laurent Davezies & Elia Lapenta

ENSAE, 2024/2025

1 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

2 / 47
Motivation

▶ The objective is to explain a binary variable Y by

X = (1, X1 , ..., XK −1 )′ ∈ RK .
▶ The two possible values that Y can take are arbitrary, we assume
without loss of generalization that Y ∈ {0, 1}.
▶ There are many contexts where the dependent variable Y is binary:
▶ In micro-economics: participant or non-participant in the labor
market, employment vs unemployment, purchase or non-purchase of
a durable consumption good.
▶ In credit risk analysis: borrower defaults or not.
▶ In assurance: insurance claim or not.
▶ In biostatistics: individual recovers from a disease or remains ill,
medical treatment is effective or not.
▶ In social sciences: graduation or not, living in couple or single, vote
or abstention, etc.

3 / 47
General ideas

▶ Linear models are not well adapted to analyze binary dependent

variables ...
▶ ... But we will assume that a latent (unobserved) variable Y ∗ is
linear in the parameters and determine how Y is defined through
this latent variable.
▶ Since the resulting models for Y are non-linear, it is crucial to think
carefully about what are the parameters of interest.
▶ The fact that Y ∗ is not observed implies that certain parameters
need to be normalized.
▶ Parameters are estimated using the maximum likelihood method.
▶ Some of these ideas reappear in chapter 5.

4 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

5 / 47
Modelization
▶ Linear models are not well suited to analyze binary dependent
variables. Indeed, if Y ∈ {0, 1}, then
E (Y |X ) = P(Y = 1|X ) ∈ [0, 1]. (1)
In a linear model (under the exogeneity assumption E (ε|X ) = 0), it
follows that E (Y |X ) = X ′ β0 . But nothing guarantees that
X ′ β0 ∈ [0, 1].
▶ In order to satisfy the restriction (1), the following assumption is
made:
E (Y |X ) = F (X ′ β0 ), (2)
where F (.) is a strictly increasing bijective (and known) function
from R to ]0, 1[, that is to say F satisfies the properties of a
distribution function.
▶ N.B.: equation (2) corresponds to the equation of a so-called
Generalized Linear Model (GLM), i.e., a model of the form:
h(E (Y |X )) = X ′ β0 ,
where h is a known function called the link function.
6 / 47
Non-linear models and latent variables

▶ Model (2) can be interpreted in terms of latent variables.

▶ Suppose there exists a continuous variable Y ∗ ∈ R such that

Y = 1{Y ∗ ≥ 0}.

▶ Suppose in addition that Y ∗ is generated by a linear model:

Y ∗ = X ′ β0 + ε, (3)

where −ε is independent from X and has the distribution function

F . Then

P(Y = 1|X ) = P(X ′ β0 + ε ≥ 0|X ) = P(−ε ≤ X ′ β0 |X ) = F (X ′ β0 ).

▶ So we obtain again equation (2).

▶ The latent variable interpretation is, often, quite natural.

7 / 47
Latent variable interpretation

Example 1: microeconomics.
▶ Y = agent’s choice between two options. Let U1 be the (expected)
utility of this agent when choosing option Y = 1, and U0 the utility
when choosing Y = 0.
▶ Then define the difference between the utilities of the two choices as
Y ∗ = U1 − U0 .
▶ If the agent is rational the chosen option is the one that generates
the highest utility:

Y = 1{U1 ≥ U0 } = 1{Y ∗ ≥ 0}.

Example 2: corporate finance.

▶ The default (Y = 1) of a company occurs if its debt, denoted D, is
larger then some threshold value S (possibly random).
▶ The latent variable is then Y ∗ = D − S.

8 / 47
Latent variable interpretation

Example 3: biostatistics.
▶ Y =1 if an individual is ill, 0 otherwise.
▶ Un individual has recovered if the number of bacteries N (for
example) has descended below a certain threshold S, which may
possibly be individal-specific.
▶ On a alors Y ∗ = N − S.
Example 4: education.
▶ Y = 1 if a student graduates and obtain a diploma, 0 otherwise.
▶ The diploma is obtained if the student’s average grade M exceeds a
fixed threshold s.
▶ Then Y ∗ = M − s.

9 / 47
Two important examples: the probit and logit models

▶ A priori any choice for F is possible.

▶ The two most common choices are
▶ F = Φ, distribution function of a N (0, 1) random variable: probit
model;
▶ F (x ) = Λ(x ) = 1/(1 + exp(−x )), the logistic distribution function:
logit model.
▶ The difference between the two distribution functions is quite small.
2
When |x | → +∞, φ(x ) = Φ′ (x ) ∝ e −x /2 and
Λ′ (x ) = Λ(x )(1 − Λ(x )) = O(e −|x | ).
⇒ the logistic density function has fatter tails.

10 / 47
Parameters and marginal effects

▶ Qualitatively, the j−th component of X , Xj , has a positive effect

on P(Y = 1|X ) iff β0j > 0.
▶ Quantitatively, the interpretation of β0j is more subtle.
▶ In a standard linear model E (Y |X ) = X ′ β0 , the parameter β0j
corresponding to Xj can be interpreted as the marginal effect of Xj :

∂E (Y |X1 = x1 , ..., XK −1 = xK −1 )
= β0j .
∂xj

▶ This effect is independent of x = (x1 , ..., xK −1 ).

11 / 47
Parameters and marginal effects

▶ But in binary models (as generally in non-linear models), the

marginal effect of Xj is no longer β0j , and it depends on x :

∂E (Y |X = x ) ∂F (u) ∂x ′ β0
= = f (x ′ β0 )β0j with f = F ′ .
∂xj ∂u ′
u=x β0 ∂x j

▶ Remark 1 : if f is symmetric and unimodal, the effect of a variable

on P(Y = 1|X ) is larger the closer x ′ β0 is to 0, that is
P(Y = 1|X ) ≃ 0.5.
▶ Remark 2 : we always have

β0l ∂E (Y |X = x )/∂xl
=
β0j ∂E (Y |X = x )/∂xj

The ratio of parameters can thus be interpreted as in a standard

linear model.

12 / 47
Parameters and marginal effects
▶ Besides estimating β0j, it is of interest to estimate the average
marginal effect:
∆j = E [f (X ′ β0 )] β0j .
This is the expected marginal effect of Xj across the whole
population.
▶ It is also possible to focus on marginal effects calculated across
sub-populations, by calculating the expected marginal effect for units
verifying X ∈ A (for example): E [f (X ′ β0 )|X ∈ A] β0j , or the
marginal effect for the average (representative) unit, f (E (X )′ β0 ) β0j .
▶ When an explanatory variable Xj is discrete (dichotomous), it is
more appropriate to replace marginal effects by
′ ′
F (x−j β0−j + β0j ) − F (x−j β0−j ),
where x−j = (1, x1 , ..., xj−1 , xj+1 , ..., xK −1 )′ . The average effect of
globally switching Xj from 0 to 1 is then:
′ ′
 
E F (X−j β0−j + β0j ) − F (X−j β0−j ) .
13 / 47
A specificity of the logit model: the odds-ratios.

▶ Define the risk (or odd) as the ratio of probabilities

P(Y = 1|X )/P(Y = 0|X ).
▶ In the case of a logit model:
′
P(Y = 1|X = x ) 1/(1 + e −x β0 ) ′
= −x ′ β0 = e x β0
P(Y = 0|X = x ) e /(1 + e −x ′ β0 )
▶ Consider now a binary explanatory variable Xj ∈ {0, 1}. We then get:

P(Y = 1|X−j = x−j , Xj = 1)/P(Y = 0|X−j = x−j , Xj = 1)

e β0j = .
P(Y = 1|X−j = x−j , Xj = 0)/P(Y = 0|X−j = x−j , Xj = 0)

▶ Therefore e β0j equals the ratio of risks (odds-ratio) corresponding to

Xj = 1 and Xj = 0. This odds-ratio is independent of the value of
X−j .

14 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

15 / 47
Identification
▶ Let us return to the equation:
Y = 1{X ′ β0 + ε ≥ 0}.
▶ Two questions: (i) why is the threshold fixed at 0 ? (ii) why is the
variance of ε fixed (at 1 for the probit model, at π 2 /3 for the logit
model) ?
▶ Reason: the model parameters are not identified otherwise. Indeed,
we have:
′ ′
Y = 1{β01 +X−1 β0−1 +ε ≥ s} ⇐⇒ Y = 1{β01 −s+X−1 β0−1 +ε ≥ 0}.
▶ Put differently, it is not possible to identify separately the constant
β01 and the treshold s. The threshold s is therefore (arbitrarily)
fixed at 0.
▶ Similarly, it is not possible to identify separately β0 and the variance
σ02 of the error term. ε. Indeed,
Y = 1{X ′ β0 + ε ≥ 0} ⇐⇒ Y = 1{X ′ (β0 /σ0 ) + ε/σ0 ≥ 0}.
▶ The variance σ02 is therefore arbitrarily fixed.
16 / 47
Identification
When s and σ0 are fixed and if E (XX ′ ) is an invertible matrix, the model
is identified.
Proof: letting Pβ be the distribution of the observations when the true
parameter is β, it is necessary to show that the function β 7→ Pβ is
injective. In our conditional binary model, the identifiability of the model
is equivalent to

Pβ (Y = 1|X ) = Pβ ′ (Y = 1|X ) ⇒ β = β ′ ∀(β, β ′ ).

But
(E (XX ′ ) is invertible ) ⇐⇒ (X ′ λ = 0 =⇒ λ = 0)
Consequently,

Pβ (Y = 1|X ) = Pβ ′ (Y = 1|X ) ⇐⇒ F (X ′ β) = F (X ′ β ′ )
⇐⇒ X ′β = X ′β′
⇐⇒ β = β′ 2

17 / 47
Estimation of the model: the maximum likelihood method

▶ Let us now consider the estimation of β0 using a sample of i.i.d.

observations ((Y1 , X1 ), ..., (Yn , Xn )).
▶ Since the model is fully parametric, the unknown parameters can be
estimated by the method of maximum likelihood.
▶ The density of Y conditional on X is
y 1−y
P(Y = y |X = x ) = [P(Y = 1|X = x )] [P(Y = 0|X = x )]
= F (x ′ β)y (1 − F (x ′ β))1−y .

▶ The maximum likelihood function of the i.i.d. observations

(Y, X) = ((Y1 , X1 ), ..., (Yn , Xn )) conditional on X can then be
written as
n
Y
Ln (Y|X; β) = F (Xi′ β)Yi (1 − F (Xi′ β))1−Yi .
i=1

18 / 47
Estimation of the model: the maximum likelihood method

▶ The maximum likelihood estimator is then defined as:

βb ∈ arg max Ln (Y|X; β).

β∈RK

▶ Note that this estimator is not necessarily unique, and it might not
even exist.
▶ Also note that Ln (Y|X; β) is the likelihood conditional on X.
Denoting g(Xi ) the density of Xi , the unconditional likelihood can
be written
n
Y
Ln (Y, X; β) = F (Xi′ β)Yi (1 − F (Xi′ β))1−Yi g(Xi ).
i=1

Since in practice the distribution of X is typically of no particular

interest, the focus is on the conditional likelihood.

19 / 47
First order conditions
▶ It is easier to maximize the log-likelihood function:
n
X
ℓn (Y|X; β) = Yi ln (F (Xi′ β)) + (1 − Yi ) ln (1 − F (Xi′ β))
i=1

▶ We have ∂F (Xi′ β)/∂β= f (Xi′ β)Xi . Therefore:

n 
f (Xi′ β) −f (Xi′ β)

∂ℓn X
(Y|X; β) = Yi + (1 − Yi ) Xi .
∂β i=1
F (Xi′ β) 1 − F (Xi′ β)

Reorganizing terms gives

n
∂ℓn X f (Xi′ β) ′
∂β
(Y|X; β) =
F (X ′ β)(1 − F (X ′ β)) [Yi − F (Xi β)] Xi . (4)
i=1 i i

▶ The first order conditions can hence be written as:

n
X f (Xi′ β)
b h i
Yi − F (Xi′ β)
b Xi = 0 (5)
i=1 i
b − F (X ′ β))
F (X ′ β)(1 b
i

which do not imply, in general, a simple analytical solution.

20 / 47
Existence and unicity of the solution

▶ If a dichotomous variable Xj is such that: if xij = 1 ⇒ yi = 1 for all

i (or xij = 1 ⇒ yi = 0 for all i), the the estimator does not exist.
▶ Indeed, ∂ℓn /∂βj can then be written as (considering the case xij = 1
⇒ yi = 1 for all i)
n
X f (xi′ β) X f (x ′ β)
[yi − F (xi′ β)] xij = i
>0 ∀β
i=1
F (xi′ β)(1 ′
− F (xi β)) i:x =1
F (x ′
i β)
ij

▶ This example shows that, among observations such that xij = 1, the
variable yi should vary across i in order to estimate β0j . In the
absence of variation, economtrics/statistics software packages like
Stata indicate that β0j is not identified and will automatically
“expel” xj from the list of explanatory variables.

21 / 47
Existence and unicity of the solution

▶ In the case of a logit model, we have Λ′ = Λ(1 − Λ), so

n
∂ 2 ℓn X
(Y|X; β) = − Λ′ (Xi′ β)Xi Xi′ << 0.
∂β∂β ′ i=1

The matrix of second derivatives is a negative definite matrix. The

log-likelihood is then strictly concave ⇒ the first order conditions
contain at most one solution, and this solution correspond to a
global maximum.
▶ In the case of a probit model, it can also be shown that the
log-likelihood is strictly concave.
▶ In general, the maximisation program is not necessarily concave and
multiple solutions may exist. Ideally it must then be verified that the
solution corresponds to a global maximum.

22 / 47
Remarks on the optimisation procedure

▶ Unlike the OLS estimator, the maximum likelihood estimator

cannot, in general, be expressed explicitly.
▶ The estimator can be obtained numerically using optimisation
algorithms such as Newton-Raphson algorithm (there are other
algorithms as well).

Starting with an initial value β (0) , define the
(m)
sequence β m∈N
:
−1
∂ 2 ℓn

∂ℓn
β (m+1) = β (m) − (Y|X; β (m) ) (Y|X; β (m) )
∂β∂β ′ ∂β

▶ Under strict concavity of ℓn (Y|X; β), the sequence β (m) , necessarily

converges to the the maximum likelihood estimator.
▶ In the cases of the logit and probit models the iterations typically
converge very quickly to the optimum.

23 / 47
Asymptotic properties

Under various technical conditions (see the Statistics 1 course), it can be

P √ d
shown that βb −→ β0 and n(βb − β0 ) −→ N 0, I1−1 (β0 ) , where I1 (β0 )

is the Fisher information associated with one observation. Moreover,

f 2 (X ′ β0 )
 
′
I1 (β0 ) = E XX .
F (X ′ β0 )(1 − F (X ′ β0 ))

This Fisher information can be consistently estimated by:

n
1X f 2 (Xi′ β)
b
I\(β
1 0 ) = Xi Xi′ .
n i=1 F (Xi β)(1
′ b − F (X ′ β))
i
b

Recall that the maximum likelihood estimator is asymptotically the best

“regular” estimator: if another estimator βe verifies
√ e 
d
n β − β0 −→ N (0, V ), then V >> I1−1 (β0 ).

24 / 47
Asymptotic properties
Proof of the formula for I1 (β0 ): we have
 
∂ℓ1
I1 (β0 ) = V (Y |X ; β0 )
∂β

where ℓ1 (Y |X ; β0 ) is the log-likelihood (evaluated at β0 ) of one

observation:

ℓ1 (Y |X ; β0 ) = Y ln (F (X ′ β0 )) + (1 − Y ) ln (1 − F (X ′ β0 ))

Using the law of total variance gives:

     
∂ℓ1 ∂ℓ1
I1 (β0 ) = E V (Y |X ; β0 ) X +V E (Y |X ; β0 ) X .
∂β ∂β

Using equation (4)) gives

∂ℓ1 f (X ′ β0 )
(Y |X ; β0 ) = [Y − F (X ′ β0 )] X .
∂β F (X β0 )(1 − F (X ′ β0 ))
′

25 / 47
Asymptotic properties

Therefore  
∂ℓ1
E (Y |X ; β0 ) X =0
∂β
because E (Y − F (X ′ β0 )|X ) = 0, and
    ′ 
∂ℓ1 ∂ℓ1 ∂ℓ1
V (Y |X ; β0 ) X = E (Y |X ; β0 ) (Y |X ; β0 ) X
∂β ∂β ∂β
 
f (X ′ β0 )2 2
Y − F (X ′ β0 ) XX ′ X

= E ′ ′
F (X β0 ) (1 − F (X β0 ))
2 2

f 2 (X ′ β0 )XX ′
=
F (X ′ β ′
0 )(1 − F (X β0 ))
 
since E (Y − F (X ′ β0 ))2 X = F (X ′ β0 )(1 − F (X ′ β0 )). The result is
then obtained.

26 / 47
Hypothesis testing

▶ We wish to test a hypothesis of the form

H0 : Rβ0 = 0 against H1 : Rβ0 ̸= 0 (R matrice p × K , p ≤ K ).

▶ For example, β0j = 0 or β02 = ... = β0K −1 = 0 (that is β0−1 = 0).

▶ We use one of the following three maximum-likelihood based tests:
the Wald test, the score test, and the likelihood ratio test. The test
statistics associated with 3 tests are:
h −1
i−1
ξnW = nβb′ R ′ R I\ 1 (β0 ) R′ R βb
1 ∂ℓn −1 ∂ℓn
ξnS = (Y|X; βbC ) I\
1 (β0 ) (Y|X; βbC )
n ∂β ′ ∂β
h i
ξnR = 2 ℓn (Y|X; βb) − ℓn (Y|X; βbC )

where βbC is the constraint maximum likelihood estimator, i.e., the

one obtained under H0 .

27 / 47
Hypothesis testing
▶ Concerning the statistic ξnW , I\
1 (β0 ) corresponds to the formula on
page 24.
▶ Concerning the statistic ξnS , it is the same formula except that βb is
replaced by βbC .
▶ Note that the three statistics tend to be “small” under the
hypothesis H0 .
d
▶ Under H0 , ξnT −→ χ2 (p) (T = W , S or R).
▶ The critical region of a test of asymptotic level α therefore takes the
2 2
form {ξnT > q χ (p) (1 − α)} where q χ (p) (y ) is the quantile of order y
of a χ2 (p).
▶ To test H0 : β0j = 0 against H1 : β0j ̸= 0, the usual t-test is
mostly used. This test produces the same result as the Wald test,
 2
because (you should verify this) ξ W = βbj /se(βbj ) ≡ t 2 and
n j
|tj | > q N(0,1) (1 − α/2) ⇔ ξnW > q χ2 (1) (1 − α) where
q N(0,1) (1 − α/2) is the quantile of order 1 − α/2 of a 0, 1 .
28 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

29 / 47
Explanatory power of the model

▶ Let us define, similarly to the R 2 in linear model, the pseudo-R 2 as:

ℓn (Y|X; β)
b
pseudo-R 2 = 1 −
ℓn (Y|X; βbC )

where βbC is the parameter estimate obtained under the null

hypothesis β0−1 = 0.
▶ Since 0 > ℓn (Y|X; β) b ≥ ℓn (Y|X; βbC ), the pseudo-R 2 belongs to
]0, 1]. It is close to 1 when

Yi = 1 and F (Xi′ β)
b ≃ 1 or Yi = 0 and F (X ′ β)
i
b ≃ 0.

▶ Like the R 2 , the pseudo-R 2 augments mechanically with the number

of variables in the model.
▶ Other indicators: concordance table, score, percentage of concordant
pairs...

30 / 47
Choice of variables
▶ Trade off between:
▶ increase of explanatory power of model;
▶ loss of precision due to estimating a large number of parameters.
▶ One can test the null hypothesis that variables have no effect,
possibly through sequential procedures (forward, backward, ...).
▶ Drawback: when n tends to infinity, such procedures entail that the
null is rejected for most of the explanatory variables.
▶ one can also use the information criteria AIC (Akaike Information
Criterion, Akaike, 1973) or BIC (Bayesian Information Criterion,
Schwarz, 1978).
▶ Such criteria are used to determine which model to select. Suppose
there are J possible parametric models:

(Pβ (1) )β (1) ∈B (1) , ..., (Pβ (J) )β (J) ∈B (J) .

The objective is to select the “true” data-generating model.

31 / 47
Choice of variables

▶ Akaike’s criterion for model j having pj parameters:

AIC(j) = ℓn (Y|X; βb(j) ) − pj

The chosen model is then model j0 = arg maxj AIC(j).

▶ This criterion does not necessarily lead to a correct selection when n
tends to infinity. Indeed, it does not sufficiently the number of
parameters of each model.
▶ To account for that drawback, Schwarz (1978) proposed the
following criterion:
pj
BIC(j) = ℓn (Y|X; βb(j) ) − ln(n)
2

32 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

33 / 47
Advantage of a linear model
▶ Sometimes, for simplicity reasons, a linear probability model is
estimated instead of a logit or probit model:
E (Y |X ) = X ′ β0 .
▶ Example: panel data. Suppose that
E (Yit |Xit , αi ) = Xit′ β0 + αi ,
where αi is the unit-specific fixed which is possibly correlated with
Xit .
▶ This unobserved fixed effect can be eliminated by applying the first
difference or within operator:
E (Yit − Yit−1 |Xit , Xit−1 ) = (Xit − Xit−1 )′ β0 .
▶ In non-linear models, this trick does not work since
E (Yit − Yit−1 |Xit , Xit−1 , αi ) = F (Xit′ β + αi ) − F (Xit−1
′
β + αi ).
▶ Furthermore, maximum likelihood estimation of (β, α1 , ..., αn ) does
not produce consistent estimators of the parameters because of the
so-called incidental parameter problem: the number of parameters
tend to infinity with n. 34 / 47
Modelization and estimation
▶ The linear probability model can be rewritten as Y = X ′ β0 + ε, with

1 − X ′ β0 with (conditional) probability X ′ β0

ε=
−X ′ β0 with probability 1 − X ′ β0
▶ Therefore:

V (ε|X ) = E (ε2 |X ) = X ′ β0 (1 − X ′ β0 )2 + (1 − X ′ β0 )(X ′ β0 )2

= X ′ β0 (1 − X ′ β0 ).
▶ The model is heteroscedastic. It can be estimated by OLS but also
by Generalized Least Squares (GLS):
▶ First estimate β0 by OLS: βbOLS .
▶ Then reestimate β0 by
n
X 1 2
Yi − Xi′ β

βbGLS = arg min
β
i=1
Xi′ βbOLS (1 − Xi′ βbOLS )

▶ The GLS estimator more precise in theory, but not necessarily in

practice if Xi′ βbOLS ≃ 0 or ≃ 1 for certain observations i.
35 / 47
Comparaison logit/probit/linear model

▶ The difference between logit, probit and linear models is that not the
same distribution function F (.) is chosen in E (Y |X ) = F (X ′ β0 ):
F = Λ, Φ, or the identity function, depending on the type of model.
▶ There are semi-parametric models wherein it is assumed that
P(Y = 1|X ) = F (X ′ β0 ) with F and β0 unknown. This is equivalent
to considering a latent model Y ∗ = X ′ β0 + ε where ε ⊥⊥ X and with
an unknown distribution function.
▶ Such model are less rectrictive but are harder to estimate.
▶ The results obtained using a logit model, a probit model, or a linear
model, are often quite similar in terms of their marginal effects.
▶ In terms of the coefficients, we have in general

βblogit ≃ 1.6βbprobit ≃ 4βblinéaire .

36 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

37 / 47
Potential Outcomes

Consider a bianry outcome Y and an experimental framework where

Y (0), Y (1), X ⊥
⊥ D and a Logit regression of Y on (1, D, X ):
▶ Experimental framework ensures E (Y (d)|X ) = E (Y |D = d, X ).
▶ First order conditions of the MLE are
n
X
Yi − Λ(βb0 + βbD Di + Xi′ βbX ) = 0
i=1
n
X  
Di Yi − Λ(βb0 + βbD Di + Xi′ βbX ) = 0
i=1
n
X  
Xi Yi − Λ(βb0 + βbD Di + Xi′ βbX ) = 0
i=1

38 / 47
Logit and Potential Outcomes

The two first equations are equivalent to:

n
1 X   
Pn (1 − Di ) Yi (0) − Λ βb0 + Xi′ βbX =0
i=1 1 − Di i=1

n
1 X   
Pn Di Yi (1) − Λ βb0 + βbD + Xi′ βbX =0
i=1 Di i=1

Let Zi = (1, Di , Xi′ )′ and assume E (Zi Zi′ ) non singular. Then
(b0 , bD , bX ) 7→ E (ℓ1 (Y|X; (b0 , bD , bX′ )′ )) admits a unique maximum
denoted β0 , βD , βX .
We can show that (βb0 , βbD , βbX ) converges to (β0 , βD , βX ) and the first
order conditions ensures the non random equations:

E (Yi (0) − Λ(β0 + Xi′ βX )|Di = 0) = 0, (6)

E (Yi (1) − Λ(β0 + βD + Xi′ βX )|Di = 1) = 0. (7)

39 / 47
Robust estimation of the Average Treatment Effect
Even if E (Y (d)|X ) ̸= Λ(β0 + βD d + x ′ βX ) (misspecification):
Independence of D with Y (0), Y (1), X ensures that the average
treatment effect is (!):

δ := E (Y (1) − Y (0))
(1)
= E (Y (1)|D = 1) − E (Y (0)|D = 0)
(2)
= E (Λ(β0 + βD + Xi′ βX )|D = 1) − E (Λ(β0 + Xi′ βX )|D = 0)
(3)
= E (Λ(β0 + βD + Xi′ βX ) − Λ(β0 + Xi′ βX )) ,

(1) and (3) are consequences of D ⊥ ⊥ (Y (0), Y (1), X ), (2) is a

consequence of (6) and (7).
δ can be estimated consistently by:
n
1X b
δb = Λ(β0 + βbD + Xi′ βbX ) − Λ(βb0 + Xi′ βbX )
n i=1

40 / 47
Robust estimation of the Average Treatment Effect
In an experimental design, with a binary outcome:
Pn Pn
Yi Di (1−Di )Yi
1. Naive estimator δ =
b Pi=1
n
D
− P
i=1
n
1−D
of the ATE is obtained
i i
i=1 i=1
by (check this !):
▶ Linear regression of Y on (1, D): b
δ = βbD
▶ Logit regression of Y on (1, D): δ = Λ(βb0 + βbD ) − Λ(βb0 )
b
▶ δ = Φ(βb0 + βbD ) − Φ(βb0 )
Probit regression of Y on (1, D): b
▶ Considering any MLE based on P(Y = 1|D) = F (β0 + βD D) for F
continuous with continuous inverse: bδ = F (βb0 + βbD ) − F (βb0 )
2. If some X also affect potential outcomes Y (0), Y (1), more efficient
estimators to consider:
▶ Linear regression on (1, D, X ): e
δ = βbD
▶ Logit regression on (1, D, X ) (robust to misspecification):
Pn
δ̆ = n1 i=1 Λ(βb0 + βbD + Xi βbX ) − Λ(βb0 + Xi βbX )
▶ but estimators derived from Probit of other model for binary
outcome are not robust to misspecification!
▶ Choice between e δ or δ̆ depends on their respective asymptotic
variance (use robust variance estimator based on sandwich formula!).

41 / 47
Conditional Average Treatment Effect in Expe.

Moreover if E (Y (d)|X ) = Λ(β0 + βD d + x ′ βX ) (no misspecification):

n
1X b
δ(x ) =
b Λ(β0 + βbD + x ′ βbX ) − Λ(βb0 + x ′ βbX )
n i=1
converges to the conditional average treatment effect:

δ(x ) := E (Y (1) − Y (0)|X = x )

This estimator is not robust to misspecification.

Competition with linear regression on (1, D, X ): E (Y (d)|X = x ) can not
be equal to β0 + x ′ βX if x ′ βX has a support larger than [−β0 ; 1 − β0 ] for
any β0 ∈ R.
Similar consistency for the Probit (or model based on other link function)
but again these estimators of the CATE are not robust to
misspecification.

42 / 47
Outline

Introduction

Modelization and parameters of interest

Identification and estimation

Model quality, selection of variables

Linear probability models

Potential Outcomes

Application

43 / 47
Example: labor market participation of women

▶ The objective is to explain labor market participation (Y = 1, Y = 0

otherwise) of women according to their age, their diploma and
family situation (living in couple or not, number of children below
the age of 3 years).
▶ The 2023 French labor force survey (Enquête Emploi) is used,
focusing on women aged less than 65, and who have terminated
their studies.
▶ Modalities of the variable measuring education level (DIP7):
1 Advanced diploma (Master level university or Grande Ecole)
2 Baccalauréat + 3/4 years
3 Baccalauréat + 2 years
4 Baccalauréat or brevet professionnel or equivalent diploma
5 CAP, BEP or equivalent diploma
6 Brevet des collèges
7 No diploma or CEP

44 / 47
Code R

45 / 47
Results: logit model coefficients

46 / 47
Results: probit model coefficients

47 / 47