EC655 – Econometrics

Department of Economics
Wilfrid Laurier University
Fall 2022
Final Exam

Date and Time: Saturday, December 17, 10:00am – 12:00pm

Structure:
 Answer 4 of the following 5 questions
 If you answer all 5 questions, I will grade only the first 4.
 Each question has equal weight. The weight of subquestions are noted in the question
 It is open book and open internet
 You are prohibited from collaborating in any way with other people on the midterm. This means
no discussing the questions in person, by phone, using chat services, question boards, email, etc.
 If you draw your answer from another source (e.g. the textbook, the internet), you must reference
it. The normal rules of plagiarism apply to this exam.
 Instructor help on the test will be limited to clarification questions only

1
Questions

1) (15 points) Suppose there is a health-based intervention in Canada where people are allowed to
enroll in a program that provides free food if their income is below $20,000. You would like to use
this program as a natural experiment to measure the effect of the program on child health.

a. (5 points) Suppose that some people who are allowed to enroll end up not participating, and
some people who are not allowed to enroll do participate. Write down a regression model
you would use to estimate the effect of the program and describe any assumptions required to
estimate the causal effect

You could use a fuzzy regression discontinuity model to estimate the causal effect of the
program in this case. The bias from program non-compliance can be corrected by
using the eligibility for the food program as an instrument for actual enrolment.

In the model above, CH is a measure of child health, fp is actual participation in the

food program, cut is a dummy for having income below $20,000, and f(inc) is a function
of income (the running variable) that could be linear or non-linear depending on what
assumptions you make about that relationship.

For this to work, we would need the untreated potential outcome to be a continuous
function of income through the threshold of $20,000. Another way of saying this is that
nothing else except participation in the food program changes across the threshold.

b. (5 points) Suppose now that you learn that some people purposely misrepresent their income
so that it falls just below $20,000, which allows them to participate in the program. Would
this cause any issues in estimating the treatment effect from (a)?

This would invalidate the method, unless the people who misrepresent their income is
random. This is because it would create a discontinuity in the untreated potential
outcome, which is to say that people across the cutoff would be different from each
other at the baseline. For example, if people without any other outside food support
decided to lie about their income, then people with $19,000 in income would have
different resources than people with $21,000 in income, which could affect child health
and confound the relationship with the food program.

2
2) (15 points) In a 1990 study, economist David Card estimated the effect of an increase in low-skill
immigration on local wages in Florida. The study focuses on a natural experiment where a sudden,
large number of Cuban migrants arrived in Miami over the course of a few months in 1980 (known
as the “Mariel Boatlift,” which increased the Miami labour force by about 7%). The study uses a
difference-in-differences methodology where the change in mean wages in Miami before and after
1980 is compared to the change in wages over the same time period in a set of comparison U.S.
cities including Atlanta, Houston, Los Angeles, and Tampa.

a. (5 points) Why is it inappropriate to just compare average wages in Miami after the boatlift
versus before? Explain.

The problem is that we cannot separate the effect of the boatlift from any other trends
that were already happening. That is, we do not know what would have happened in
the absence of the boatlift, so we cannot say anything about the causal effect of the
policy without some counterfactual estimate of the pre-existing trend in wages.

b. (5 points) Imagine that you observe several years of data on wages in Miami and the
comparison cities before and after the boatlift occurred. How would you assess the validity
of the difference in differences approach? Explain,

As we learned, one of the necessary assumptions for this method to work is common
trends: the trend in the control group has to accurately measure what would have
happened in the absence of the policy. While we can never really verify this assumption
fully, we can at least see if the trends between treatment and control groups match
before the policy, since the untreated outcome for both is observed. If they don’t match
before the policy, then the common trends assumption will not be met.

So to check the validity of the method, you could examine the trends and test to see if
they are similar. You could eyeball it, or do an F-test for differences in trends prior to
the treatment, or do an “event study” by interacting the treated dummy with a full set
of time dummies (like we did in class and the assignment) and looking at the differences
pre-policy.

3
3) (15 points) Suppose you are interested in the effect of attending lectures on final exam scores. You
collect data on Laurier students in the 2021-2022 academic year in their first year of the economics
program. Assume for the purposes of this question that all year 1 economics students take the same
10 classes in year 1. This dataset is a panel, where we observe the same students over multiple
different classes (i.e. the second dimension is classes, not time). The linear unobserved effects
model in this context is

examscor e ic =β 0+ β 1 percent atten d ic + ai +uic

In this equation, examscor e ic is the final exam score (out of 100), percent atten dic is the percent of
classes attended (0-100 scale), a i is an unobserved person-specific effect, and uic is an unobserved
effect that varies across people and classes. The subscript i indexes students, and c indexes classes.

a. (5 points) Explain the strict exogeneity assumption in the context of this question.

In this context where the second dimension is classes rather than time, strict exogeneity
means that the error in class c is unrelated to “percent attended” in all classes. As we
noted in class, this is a stronger assumption than just assuming the error in class c is
unrelated to “percent attended” in class c.

You can also state this as “percent attended” in class c is unrelated to the error in all
classes.

b. (5 points) Suppose you are interested in the causal effect of attendance on final exam scores,
so you estimate the equation above using fixed effects. Can you interpret the resulting
estimate as a causal effect? Explain.

This would require the strict exogeneity assumption above, and also that all unobserved
factors that are related to attendance (and affect exam scores) only vary across students
and not across classes within students. For instance, suppose the main unobserved
factor was “natural intelligence.” This variable is likely positively correlated with both
exam scores and attendance, but does not differ from class to class. Fixed effects would
control for this. If, on the other hand, intelligence varied from class to class (depending
on the subject matter), then fixed effects would not control for that.

So at the end of the day, it really all depends on what you think are the main
unobserved factors that affect exam scores.

4
4) (15 points) Suppose that the Waterloo Region District School Board (WRDSB) is interested in
whether teacher training improves student performance. The school board offers the training to its
teachers, and participation is voluntary. You have data on teacher participation in the program, and
individual student test scores at the end of the year for each teacher. The WRDSB hires you to
estimate the causal effect of the training program on test scores.

a. (5 points) Would a regression of test scores on a dummy variable for participation in the
training program estimate the causal effect? Explain.

Probably not, in particular because participation is voluntary and teachers will likely
select themselves into the program non-randomly. If, for example, highly motivated
teachers join the training program, then treatment and control groups will be different
at the baseline, and there will be selection bias.

b. (5 points) Suppose the principal encourages participation in the program by sending a letter
to teachers known to be weak to recommend that they attend the training. Would it be useful
to use a dummy variable equal to 1 if the teacher received a principal letter as an instrument
for attending the program? Explain why or why not.

It would not. If the principal had sent the letter randomly to teachers, that would work
because the instrument would be unrelated to all baseline characteristics. But sending
the letter to weak teachers necessarily means that they are weaker than other teachers,
and so the ones who get the letter are different than the ones who do not in ways that
affect student performance. Another way of saying this is that this instrument violates
the exclusion restriction.

5
5) (15 points) For each of the following, state whether it is true, false, or uncertain, and explain your
answer.

a. (5 points) In an instrumental variables model where you are estimating the effect of a binary
treatment w on some outcome y, the monotonicity assumption guarantees that the coefficient
on w is the Average Treatment Effect (ATE)

The monotonicity assumption is that if an instrumental variable affects treatment

takeup, it does so in one way. In class, we operationalized this as a “no defiers”
assumption, meaning that assignment to treatment doesn’t force people out of
treatment, and non-assignment doesn’t force people in.

This does not guarantee an ATE, but does guarantee a Local ATE, which is the ATE
among those that comply with the policy. If the treatment effects do not vary across the
population, we would get an ATE.

b. (5 points) In the linear unobserved effects model y ¿ =β 0+ β1 x ¿ +a i+u ¿, if x ¿ and a i are

uncorrelated, then OLS and Random Effects will produce the same estimate of β 1.

They would not. Doing Random Effects would produce a Generalized Least Squares
(GLS) estimate of the slope parameter on x, which transforms the data prior to running
least squares. This will generally be a different number than the OLS estimate.