PhD/MA Econometrics Examination

Part A
August, 2023
Answer any TWO of the following three questions

Q1. Statistics
2
a. Let the pdf of X be 𝑓𝑓(𝑥𝑥) = 2𝑥𝑥𝑥𝑥 −𝑥𝑥 , 0 < 𝑥𝑥 < ∞, and 0 otherwise. Find the pdf of 𝑌𝑌 = 𝑋𝑋 2 .

b. Suppose a sample of Y1, Y2, …, Y100 is drawn from a gamma distribution with E(Y)=1.5 and
Var(Y)=0.75. What’s the 95% confidence interval of 𝑌𝑌�?

c. Explain the pairwise relationship of the following distributions: normal, t, chi-square and F. Be
specific on their degrees of freedom.

d. For a random variable 𝑥𝑥 following log normal distribution with

1  (ln x − µ) 2 
= f ( x) exp  −  , x > 0.
xσ 2π  2σ 2 

Let 𝑦𝑦 = ln(𝑥𝑥), find the pdf of 𝑦𝑦.

e. Suppose a random sample of 100 observations is drawn from a normal distribution with mean
µ and variance σ2. Find the 90% confidence interval for σ2 as a function of variance 𝑠𝑠 2 . Critical
values can be denoted algebraically.
Q2. OLS estimation
For a dependent variable vector f with n observations, its corresponding independent variable
matrix is S with 𝜆𝜆 variables, parameter vector is 𝜃𝜃 and residual vector is µ.

a. Derive the OLS solution of parameter vector estimate 𝜃𝜃.

b. Explain what is the heteroskedasticity issue and its problems on estimation results.

c. Using the dataset to run a restricted regression, which is a regression without a constant. The
new residual vector is e and the goodness of fit is denoted as 𝑅𝑅12 . The original full regression
goodness of fit is denoted as 𝑅𝑅02 . Write down the formula for 𝑅𝑅12 and 𝑅𝑅02 , and compare which is
bigger.

d. Derive the distribution of estimated parameter vector, assuming the normality of residual
distribution as 𝑁𝑁(0, 𝜎𝜎 2 𝐼𝐼𝑛𝑛 ).

�3 − 𝜃𝜃3 is asymptotically N(0, 𝜎𝜎 2 /𝑛𝑛), use the delta

e. For parameter 𝜃𝜃3 , suppose we know 𝜃𝜃
method to derive a formula to test the significance level of its transformation g�θ �3 � =
�3 �−1 /2 .
�𝜃𝜃
Q3. Regression application
A multiple regression of Y on a constant, X1 and X2 produces the following results:

Regression Statistics
R Square 0.934
Adjusted R Square 0.929
Standard Error …
Observations …

ANOVA
Sum of Significance
Squared F-stat F
Regression SSE 157 191.5 0.00
Residual SSR 11
Total SST 168

Standard
Coefficients Error t Stat Lower 95% Upper 95%
Intercept 1.70 … … … …
X1 2.47 … … … …
X2 -1.32 … … … …

30 14 22
X’X = 14 50 51
22 51 107

0.0398 -0.0051 -0.0057

-0.0051 0.0397 -0.0180
(X’X)-1 =
-0.0057 -0.0180 0.0191

83
X’Y = 135
87
a. What’s the number of observations and the sample mean of y, x1 and x2?

b. What’s the sample variance of y, x1 and x2?

c. What’s the correlation between x1 and x2? What’s the unbiased sample variance estimate of
σ2 ? Fill the missing Standard Error estimate in the table.

d. Calculate the covariance matrix var(b) of the estimated parameter vector.

e. Based on d, fill in the standard errors and t-stat values (use an approximate critical value).
Evaluate their significance.
 PhD/MA Econometrics Examination

PART B

August, 2023

Answer any TWO of the following three questions

4. Consider the following model 𝑌𝑌 = 𝑋𝑋1 𝛽𝛽1 + 𝑋𝑋2 𝛽𝛽2 + 𝜀𝜀, where 𝑋𝑋1 is a matrix of 𝑘𝑘1 variables
and 𝑋𝑋2 is a matrix of 𝑘𝑘2 variables such that

 x11
1
x112  x11k1   x121 2
x21 k1
 x21 
   
X1 =       , X2 =      .
 x11n
 x12n  x1kn1   x12 n
 x22n  x2k1n 

Denote 𝑏𝑏1 and 𝑏𝑏2 as the Ordinary Least Squares (OLS) estimates for 𝛽𝛽1 and 𝛽𝛽2 , respectively.

a. Derive the expression for the OLS estimator 𝑏𝑏1 as a function of 𝑌𝑌, 𝑋𝑋1 , 𝑋𝑋2 , and 𝑏𝑏2 using
the partitioned regression model.

b. Suppose you only observe 𝑋𝑋1 but not 𝑋𝑋2 . Thus, you regress the OLS model, 𝑌𝑌 = 𝑋𝑋1 𝛽𝛽1 +
𝜀𝜀.
i. Derive the expression for the OLS estimate 𝑏𝑏1 that you would estimate under
these conditions (i.e., what is the usual OLS estimator for 𝑏𝑏1 when you regress 𝑌𝑌
on 𝑋𝑋1 ).

ii. If 𝑌𝑌 = 𝑋𝑋1 𝛽𝛽1 + 𝑋𝑋2 𝛽𝛽2 + 𝜀𝜀 is the true model, give an expression for the bias of 𝑏𝑏1
in this circumstance (that you estimated above) as a function of 𝑋𝑋1 , 𝑋𝑋2 , and 𝑏𝑏2 .

c. Now suppose you observe both 𝑋𝑋1 and 𝑋𝑋2 . Derive the OLS estimator for 𝑏𝑏2 as a function
of 𝑌𝑌, 𝑋𝑋1 , 𝑋𝑋2 , and 𝑀𝑀1 using the partitioned regression model, where 𝑀𝑀1 is the residual
maker and 𝑀𝑀1 = 𝐼𝐼 − 𝑋𝑋1 (𝑋𝑋1′ 𝑋𝑋1 )−1 𝑋𝑋1 ′.

d. Define the Frisch-Waugh-Lovell Theorem and describe its intuition.

e. Under what conditions is the bias you solved for in b.ii. equal to zero. What does this
mean in the context of the Frisch-Waugh-Lovell Theorem (i.e., what happens when you
regress 𝑋𝑋2 on 𝑋𝑋1 ).
5. You are interested in understanding the impact of early motherhood on school completion
in Madagascar. Using data on a sample of 466 Malagasy mothers in their early twenties, all
of whom gave birth between the ages of 13 and 23 years of age, you run the following OLS
regression:

𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌ℎ𝑜𝑜𝑜𝑜𝑙𝑙𝑖𝑖 = 𝛽𝛽0 + 𝛽𝛽1 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴ℎ𝑖𝑖 + 𝜀𝜀𝑖𝑖

where 𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌ℎ𝑜𝑜𝑜𝑜𝑙𝑙𝑖𝑖 is equivalent to the completed years of schooling of mother 𝑖𝑖 and
𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴ℎ𝑖𝑖 is equal to her age in years when she gave birth to her first child.

You find obtain the following results:

Age at First
Birth 0.349618
(0.067482)
Constant -0.20554
(1.245402)

a. What is the interpretation of the coefficient on 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴ℎ?

b. Perform a test of the null hypothesis 𝐻𝐻0 : 𝛽𝛽1 = 0 against the alternative hypothesis,
𝐻𝐻𝐴𝐴 : 𝛽𝛽1 ≠ 0 at the 1% significance level (i.e., for significance level 𝛼𝛼 = 0.01). Note,
the critical value for the t-distribution at the one percent level is 2.58. Show how
you calculated the test statistic. State the decision rule you use, and the inference
you would draw from the test. What would you conclude from the test?

c. Construct a 95% confidence interval for 𝛽𝛽1 .

d. Which OLS assumption likely fails in the above regression? Explain why? In other
words, give some reasons specific to this scenario as to why you believe this
assumption may fail.

e. What is the implication of this failure for your interpretation of your estimate of 𝛽𝛽1 ?

f. What is the “Fundamental Problem of Causal Inference”. Please define it and

explain what it means for empirical analysis.
6. Suppose Y1 ,..., YN has a binomial distribution such that
 1 with probability p
Y =
0 with probability 1 - p
This gives Yi the pdf

f (Yi ) = p Yi (1 − p )1−Yi

Note: the number of individuals for which y i = 0 is N − ∑yi i .

a. What is the likelihood of observing your data (i.e. what is the likelihood function for
your sample)?

b. Derive the log likelihood and score functions for estimating the parameter p .

c. Derive the Maximum Likelihood Estimate for p .

d. Derive the asymptotic variance for p̂ MLE using the information matrix method.
 PhD/MA Econometrics Examination
PART C
August, 2023
Answer any TWO of the following three questions

Q7. Consider a study where a sample of 400 households was selected to find out their
preference to buy a micro health insurance package. They were shown a randomly drawn price
from a set (Rs. 10, 40, 150, 300, 700, 1500) and were asked to record a response of Yes or No to
the offered price. The package they had been offered was to include doctor’s visit and lab tests.
In addition, their socio-economic demographic information was collected: age, income,
comorbidity, and education. The purpose of this study was to calculate their willingness to pay
value.
a. In this study, a condition called monotonicity is expected to hold. What is it and how
would you demonstrate it graphically?
b. A linear WTP function is assumed: WTP* = x’b + u

Derive the log-likelihood function so you can estimate the parameter b. After the
derivation of the log-likelihood (either for logit or probit), explain as to how you can
extract and or estimate the parameters for the WTP equation using 1) canned STATA
command (e.g., kappa etc.) b) direct estimation of the parameter b.
c. What is the mean WTP expression? What is the median formula for the WTP? Explain
how you would go about getting the CI for the Mean-WTP (delta or simulation methods).
d. Now, consider the log-linear form for the WTP function: WTP = exp(x’b + u). Derive
the log-likelihood function. (No need to discuss canned versus direct stuff.). What are
the formulas for the mean and the median WTP? Under what scenario will the median
WTP be preferred over the mean value?

Q8. Consider the following linear model

A. 𝑦𝑦(𝑡𝑡) = 𝑎𝑎0 + 𝑎𝑎1 ∗ 𝑥𝑥(𝑡𝑡) + 𝑢𝑢(𝑡𝑡)

Where u(t) is distributed as normal with mean 0 and variance Sig^2 (s2): u ~ N (0,s2)
Derive the log-likelihood function for this case. [Don’t forget to apply the change of variable rule
to go from u to y (e.g., use of Jacobian) in such likelihood derivations; show the work.]
Q: Derive the log-likelihood function for this linear case. This is a warm-up exercise!

B Now, assume the following log-linear model

 𝑙𝑙𝑙𝑙(𝑡𝑡) = 𝑎𝑎0 + 𝑎𝑎1 ∗ 𝑥𝑥(𝑡𝑡) + 𝑢𝑢(𝑡𝑡) where 𝑢𝑢(𝑡𝑡) ~̴ 𝑁𝑁(0, 𝜎𝜎 2 )
Q: Derive the log likelihood function for this log-normal case. ~̴

C. The part C has two options:

Option 1: Box-cox model (non-linear)

In the above two cases (a and b), we went from a linear form to a log-linear transformation. We
can formulate a more flexible model in the following way to incorporate both types of
transformations:

(𝑦𝑦(𝑡𝑡)𝜆𝜆 − 1) / 𝜆𝜆 = 𝑎𝑎0 + 𝑎𝑎1 ∗ 𝑥𝑥(𝑡𝑡) + 𝑢𝑢(𝑡𝑡) where 𝑢𝑢(𝑡𝑡) ~̴ 𝑁𝑁(0, 𝜎𝜎 2 )

where the value of 𝜆𝜆 determines the form of transformation. If 𝜆𝜆 = 1 => linear. If 𝜆𝜆 = 0 => log.
The question arises if there in any other possible transformation between two extreme (linear
versus log-linear). This is known as the box-cox transformation or BC model. The only way we
can find out is to estimate the lambda parameter and test if it is 0 or 1 using the t-test. This model
can be estimated in two ways – non-linear least squares or maximum likelihood. The mle version
follows.

Q: Derive the log-likelihood function. Again, pay attention to the change of variable issue while
going from u to y in deriving the likelihood function. Show the complete work. Bonus: write the
STATA script.
𝜕𝜕𝜕𝜕
Q: Write out the formula or expression for the marginal effect: 𝜕𝜕𝜕𝜕
=? Discuss how you would go
about getting the confidence interval (delta method, e.g.).

Or
Option 2: multiplicative heteroscedasticity (heteroscedastic model estimation using Mle)
Consider the following model:

𝑦𝑦(𝑡𝑡) = 𝑎𝑎0 + 𝑎𝑎1 ∗ 𝑥𝑥(𝑡𝑡) + 𝑢𝑢(𝑡𝑡) where 𝑢𝑢(𝑡𝑡) ~̴ 𝑁𝑁(0, 𝜎𝜎(𝑡𝑡)2 )

where the variance of 𝑢𝑢(𝑡𝑡) is non-constant. That is, we will make it a function of some
variables:

𝑉𝑉�𝑢𝑢(𝑡𝑡)� = 𝜎𝜎(𝑡𝑡)2 = exp(ϒ0 + ϒ1 ∗ z(t))

Q: Derive the log-likelihood function for this heteroscedastic model. Bonus: write the STATA
script.
Q9. Q7. Consider modeling an ante-natal doctor’s visit model as a function of the distance to
the clinic, annual household income, number of children at home, and education level of the
mothers.
a. Set up a poisson modelling framework, and spell out the log likelihood function.
Show all the steps.
b. Does this function look like your typical demand function? Why or why not?
c. In this case, do we need an exposure variable? Why or why not?
d. What are the expected signs on the independent variables?
e. There will be obviously many people with a 0 entry (with no visit recorded over the
last six months), leading a problem of “excess zeros”. This causes a problem known
as “over dispersion.” You have a couple of options to deal with this situation:

Zero inflated poisson framework

Negative Binomial (Type II)
Choose one of the two options and present your rationality along with the derivation of its log
likelihood function.