You are on page 1of 4

Question 1:

i)

To optimize the ordinary least squares (OLS) objective function


2
Objective function : J ( β)=(1/n)∗Σ ( yi−xiβ )
where yi is the population value, xi is the vector of predictors, β is the vector of coefficients to be
optimized, n is the number of observations, and Σ represents the sum over all observations.

To find the optimal values for β, we need to minimize the objective function J(β). taking the derivative of
J(β) with respect to β and setting it equal to zero:

∂ J (β )/∂ β =(1/n)∗Σ(−2 xi)( yi−xiβ)=0


Simplifying the equation, we get:

Σxi( yi−xiβ)=0
Expanding the sum, we have:

Σxiyi−Σxi(xiβ )=0
Rearranging the terms,

Σxiyi=Σxi(xiβ)
we can rewrite Σxi(xiβ) as β * Σxi(xi).

The equation now becomes:

Σxiyi=β∗Σxi (xi)
To solve for β:

β=Σxiyi/ Σxi(xi)
Here, β represents the vector of coefficients that minimizes the OLS objective function.

The the equation can be rewritten as:


−1
β=( X X ) X ' y
'

where X is the matrix of predictors, X' is the transpose of X, y is the vector of population values, and '^(-
1)' denotes the inverse of a matrix.

ii)

We have the population equation: yi = xi'β + ui

The objective function of GLS aims to minimize the weighted sum of squared residuals, taking into
account the covariance structure of the error terms. The objective function can be written as follows:
'
J (β )=( yi− x i β ) Σ ( yi−xi' β )
' −1

we need to take the derivative of J(β) with respect to β and set it equal to zero:
' −1
∂ J ( β )/∂ β =−2 X Σ ( y −Xβ)=0
Simplifying the equation, we have:
' −1
X Σ ( y− Xβ)=0
Expanding the equation, we get:
' −1 ' −1
X Σ y −X Σ Xβ=0
Rearranging the terms, we obtain:
' −1 ' −1
X Σ Xβ= X Σ y
To solve for β, we isolate it on one side of the equation:
( −1 ) X ( −1)
β=[ X Σ ]
' ' −1
XΣ y
Here, β represents the vector of coefficients that minimizes the GLS objective function.
( −1)
β=( X ' Σ (−1) X ) X ' Σ −1 y
In GLS, the optimal values for the coefficients β take into account the covariance structure of the error
terms and provide a more efficient estimation compared to OLS when the error terms are
heteroscedastic or correlated.

iii)

The Generalized Method of Moments (GMM) objective function as follows:

Moment Condition:
The moment condition in GMM reflects the orthogonality between the instruments zi and the residuals
from the population equation. It is given by:

E [zi( yi−xi' β )]=0


where E[.] denotes the expectation operator.

GMM Objective Function:


The GMM objective function aims to minimize the weighted sum of squared moment conditions. It is
defined as:

J (β )=[1/n]∗Σ[ zi( yi−xi ' β)]' W [zi( yi−xi' β )]


where n represents the number of observations, Σ denotes the summation over all observations, W is a
weighting matrix, and ' represents the transpose operation.
Deriving the Closed-Form Expression for the GMM Estimator:

the derivative of J(β) with respect to β and setting it equal to zero:

∂ J (β )/∂ β =−[1/n]∗Σ[ zi(xi( yi−xi ' β))]W [zi ( yi−xi ' β)]=0
Simplifying the equation, we get:

Σ[ zi(xi( yi−xi ' β))]W [ zi( yi−xi ' β)]=0


Expanding the equation, we have:

Σ[ zi( yi−xi ' β) xi' ]W [ zi( yi−xi ' β)]=0


Rearranging the terms, we obtain:

Σ[ zi(xi' β )xi ']W [zi( yi−xi' β )]=Σ[ ziyi]W [zi( yi−xi' β )]


To solve for β, we isolate it on one side of the equation:

Σ[ zi(xi' β )xi ']W [zi( yi−xi' β )]=Σ[ ziyi]W [zi( yi−xi' β )]


Multiplying both sides of the equation by the inverse of Σ[ zi(xi' β )xi ']W [zi( yi−xi' β )], we get:

β=[Σ[ ziyi ]W [zi ( yi−xi ' β)]]/[Σ [zi(xi ' β) xi' ]W [ zi( yi−xi ' β)]]
Here, β represents the vector of coefficients that minimizes the GMM objective function.

Question 2:

(i) Potential Outcomes and Observed Outcome:

We consider an evaluation with treatment T={0,1}. For each unit i:

 Y_i(1) is the potential outcome if treated (T=1)

 Y_i(0) is the potential outcome if not treated (T=0)

The observed outcome is:

Y i=T ¿ (1)+(1−T i)Y i (0)


This is the outcome we actually see for unit i, depending on its treatment status T_i.

(ii) Strong Ignorability Conditions:

For causal inferences under the potential outcome framework, we require the strong ignorability
assumption:

1. Unconfoundedness: Potential outcomes are independent of treatment assignment conditional


on covariates:
(Y i (1),Y i (0)) ⫫ T i ∨X i
2. Common support: 0 < P(T_i = 1|X_i) < 1 almost surely. Every type has a positive probability of
receiving either treatment.

These conditions ensure that, conditional on covariates X, the treatment assignment is as if randomly
assigned. Then we can use methods like matching or regression to adjust for differences in X and
identify the average treatment effect.

Strong ignorability allows the use of observational data and eliminates hidden bias, enabling consistent
causal effect estimation from samples rather than experiments. It is a key assumption in justifying most
causal inference techniques.

You might also like