University of Bonn SoSe 2023

Department of Economics Christoph Breunig

Theoretical Problems 2 – Solution

Problem 1. An advantage of logit models is that the logistic pdf. can be written as a
combination of cdfs. Please show this claim!

Solution: Recall that Λ(z) = (1 + e−z )−1 . Now we evaluate

λ(z) = Λ0 (z) = (−1)(1 + e−z )−2 (−1)e−z = e−z (1 + e−z )−2

= (1 + e−z )−1 [e−z (1 + e−z )−1 ]
= (1 + e−z )−1 [(1 + e−z )(1 + e−z )−1 − (1 + e−z )−1 ]
= (1 + e−z )−1 [1 − (1 + e−z )−1 ]
= Λ(z)[1 − Λ(z)]

Problem 2. Show that for a standard normally distributed random variable zi we have

φ(a)
E[z | z > a] = ,
1 − Φ(a)

Solution: Note that R∞

a
zf (z)dz
E[z | z > a] =
P (z > a)
Consider the numerator, we calculate:
Z ∞
1 − z2 1 h − z2 i∞ 1  a2
 e− a22
z · √ e 2 dz = − √ e 2 = −√ 0 − e− 2 = √ = φ(a)
a 2π 2π a 2π 2π
Also, for the denominator we evaluate

P (z > a) = 1 − P (z ≤ a) = 1 − Φ(a)

Problem 3. Consider the binary response logit model. Assume that xi contains an inter-
cept. Show that y then coincides with the average estimated success probabilities based on
maximum likelihood estimator β.b

1
University of Bonn SoSe 2023
Department of Economics Christoph Breunig

Solution: From lecture notes we know that the log likelihood function is given by
X X
log L(β) = log F (x0i β) + log(1 − F (x0i β)).
yi =1 yi =0

Now the function F is given by the logistic distribution function Λ, i.e.,

X X
log L(β) = log Λ(x0i β) + log(1 − Λ(x0i β)).
yi =1 yi =0

Taking derivatives w.r.t. β:

∂ log L(β) X λ(x0 β)xi X λ(x0 β)xi

i i
= 0
− 0
∂β yi =1
Λ(x i β) yi =0
1 − Λ(x i β)
n 
λ(x0i β)xi λ(x0i β)xi
X 
= yi 0
− (1 − yi )
i=1
Λ(x i β) 1 − Λ(x0i β)
n
X
= (yi (1 − Λ(x0i β))xi − (1 − yi )Λ(x0i β)xi )
i=1
n
X
= (yi xi − Λ(x0i β)xi ) .
i=1

Consequently, the first order condition is given by

n
X
(yi − Λ(x0i β)) xi = 0
i=1

Now MLE βb satisfies

n 
X 
yi − Λ(x0i β)
b xi = 0
i=1

Since xi contains an intercept we obtain

n 
X 
yi − Λ(x0i β)
b =0
i=1

and thus n n
X X
−1 −1
y=n yi = n Λ(x0i β)
b
i=1 i=1

where Λ(x0i β)
b = P (yi\
= 1 | xi ).

2
University of Bonn SoSe 2023
Department of Economics Christoph Breunig

Problem 4. For sample selection problems we derived the equation

E[yi | xi , z i , Ti = 1] = x0i β − σuw λ(z 0i γ).

where λ(a) = φ(a)/Φ(a). Why do we need additional information on z i for small values of
z i ? Provide a theoretical argument and plot the function in R.

Solution: Note that by L’Hôpital’s rule

φ(a) −aφ(a)
lim = lima→−∞ = lima→−∞ (−a)
a→−∞ Φ(a) φ(a)

In R:
a <− seq ( −10 , 1 0 , length . out = 1 0 0 )
>
> # C a l c u l a t e t h e v a l u e s o f p h i ( a ) and Phi ( a )
> p h i a <− dnorm( a )
> Phi a <− pnorm( a )
>
> # C a l c u l a t e t h e r a t i o p h i ( a ) / Phi ( a )
> r a t i o <− p h i a / Phi a
>
> # P l o t t h e graph
> plot ( a , r a t i o , x l a b = ” a ” , y l a b = expression ( p h i ( a ) / Phi ( a ) ) ,
> main = ” P l o t o f p h i ( a ) /Phi ( a ) ” )
> grid ( )

Problem 5. Suppose that

yi∗ = x0i β + ui ,

where ui | xi ∼ N (0, σ 2 ). Define the observed outcome


 1

 if yi∗ ≤ c1
yi = 2 if yi∗ ∈ (c1 , c2 ]

if yi∗ > c2 .

 3

Suppose c1 and c2 are known constants. Write down the conditional log-likelihood function
of this model for a random sample {yi , xi }ni=1 .

Solution: Since yi is discretely distributed, the general form of the conditional log-
likelihood function is
n Y
3
! n X
3
Y X
1(yi =j)
log P (yi = j | xi ) = 1(yi = j) log(P (yi = j | xi )).
i=1 j=1 i=1 j=1

3
University of Bonn SoSe 2023
Department of Economics Christoph Breunig

Here

P (yi = 1 | xi ) = P (yi∗ ≤ c1 | xi )
= P (ui ≤ c1 − x0i β | xi )
c1 − x0i β
 
= Φ .
σ

Similarly

P (yi = 2 | xi ) = P (c1 − x0i β < ui ≤ c2 − x0i β | xi )

c2 − x0i β c1 − x0i β
   
= Φ −Φ
σ σ

and
c2 − x0i β
 
P (yi = 3 | xi ) = 1 − Φ .
σ