The oral exam is an approximately 10-minute, one-on-one conversation with me over Zoom. During this conversation, I will give you a brief real-world scenario and ask you to reason through the causal design and potential complications.
The exam will test the following: given a situation, can you construct a credible causal comparison, explain what makes it credible, and recognize what could go wrong? There is no coding or math.
The exam is worth 25 percent of your course grade.
Logistics
- When: July 27 or July 29. You will be assigned a 12-minute appointment. The conversation itself will take about 10 minutes.
- Where: Zoom, with me, one-on-one.
- Materials: Closed book. You may not use notes, the internet, AI, messaging, or a second device during the exam.
The setup
At the beginning of the meeting, you will be presented with a random scenario (much like the examples below).
Every scenario uses one of the three designs we discussed during the second half of the course:
- instrumental variables (IV)
- regression discontinuity (RDD)
- difference-in-differences (DiD)
I've provided three practice scenarios below. The real scenarios will use the same structure. Working through the practice scenarios should help prepare you for the exam.
How the exam will work
At the beginning of the exam, I will give you a brief opportunity to read the scenario, which will stay on screen the entire time. I will then reveal questions one at a time.
1. Identify the question
I will ask you to identify the causal question, the treatment, and the outcome.
2. Construct the comparison
I will ask:
What comparison would you make to estimate that effect?
Explain the comparison before naming the design. I am listening for the causal logic, not just whether you recognize a label.
3. Explain the logic behind the causal interpretation and identify key assumptions
I will ask why your comparison is able to identify a causal effect. How can the design help us learn what would have happened counterfactually? What assumption or assumptions are needed to ensure the comparison is causal?
State the assumption in plain language. Saying only "parallel trends" or "exclusion restriction" is not enough. Explain what the term means in the context of the scenario.
4. Identify a concrete threat
I will ask for one specific way the assumption could fail in the scenario. A strong answer names something that could actually happen and explains why it would affect the comparison.
5. Respond to new information
I will give you one additional fact about the scenario and ask whether the original design still identifies the effect we wanted. You may conclude that the design still works, that it estimates a different effect, or that it no longer supports the causal claim.
6. Interpret the effect
Finally, I will ask whose causal effect the design would estimate if it worked.
Grading
The exam has four equal components:
| Criteria | What a strong response looks like |
|---|---|
| Causal comparison and design | You construct the appropriate comparison and name the design. |
| Identifying logic | You explain each key assumption in plain language and connect it to the missing counterfactual. |
| Concrete threat | You identify a relevant threat and explain how it affects the comparison. |
| Adaptive reasoning and interpretation | You revise the claim when given new information and correctly describe whose effect the design estimates. |
You do not need to answer instantly. Think out loud and revise an answer if you notice a problem. If you get off on the wrong foot, I will help redirect you so the rest of the conversation still gives you an opportunity to demonstrate your understanding. Credit depends on the overall arc of the discussion, not on delivering a perfect or polished speech.
How to prepare
For each practice scenario below:
- Read the scenario and the questions.
- Answer each question out loud before looking at the key.
- Scroll to the answer key only after you have finished each response.
- Compare the logic of your answer with the key, not just the terminology.
- Try the scenario again if you could not explain one of the steps in your own words.
Do not memorize the answer keys. Use them to check whether your own explanation contains the same logic.
Practice scenario 1: Instrumental variables
A nonpartisan civic group wants to know whether attending a candidate town hall makes voters more knowledgeable about the candidates' positions. The group takes a list of 2,000 registered voters and randomly sends half of them text-message reminders about an upcoming town hall. After the election, everyone receives a short political-knowledge quiz. Not everyone who receives the texts attends, and a few people who do not receive a text attend anyway.
Questions
What causal effect is the civic group trying to estimate? Identify the treatment and outcome.
What comparison would you make? What design have you described, and what role do the reminder texts play?
What assumptions make that comparison credible? Explain them in plain language.
Name one concrete way an assumption could fail in this setting. Explain why it would be a problem.
Now consider this additional information:
Attendance is 31 percent among voters assigned to receive reminders and 30 percent among voters assigned not to receive them.
Does the design still provide useful evidence about the effect of attending the town hall? Why or why not?
If the design worked, whose causal effect would it estimate?
Answer each question before checking the corresponding answer below.
Answer key
1. Causal question, treatment, and outcome
- Causal question: Does attending the town hall increase political knowledge?
- Treatment: attending the town hall
- Outcome: political knowledge
2. Comparison and design
Simply comparing attendees with non-attendees would be misleading. People who are already politically engaged may be more likely both to attend and to know more about the candidates. Instead, use the difference in attendance caused by random assignment to receive the reminders. This is an instrumental-variables design, and the random reminder assignment is the instrument.
3. Assumptions
Two central requirements are:
- Relevance: receiving the reminders must meaningfully increase town-hall attendance.
- Exclusion restriction: receiving the reminders can affect political knowledge only by changing attendance, not through another channel.
4. Concrete threat
Suppose the reminder texts summarize each candidate's positions. The texts could then increase political knowledge directly, even among voters who never attend. That would violate the exclusion restriction.
5. New information
The reminders barely change attendance, so the instrument is quite weak. Although we could still estimate the effect of attending a town hall, the estimate would be imprecise given the small number of compliers.
6. Whose effect
A valid IV estimate would describe the effect for compliers: voters who attend because they receive the reminders but would not attend without them. It would not necessarily describe the effect for every voter.
Practice scenario 2: Regression discontinuity
A state gives extra election-administration funding to any city with more than 50,000 residents. Cities at or below 50,000 receive nothing. The funding pays for more polling places, shorter lines, and better-trained poll workers. A researcher wants to know whether the funding increases turnout in the next local election.
Questions
What causal effect is the researcher trying to estimate? Identify the treatment and outcome.
What comparison would you make? What design have you described, and what are the running variable and cutoff?
What assumptions make that comparison credible? Explain them in plain language.
Name one concrete way an assumption could fail in this setting. Explain why it would be a problem.
Now consider this additional information:
Cities above 50,000 residents are also required to offer weekend voting, while cities below the threshold are not.
Does the design still estimate the causal effect of the election-administration funding? What can it estimate instead?
If the design worked, whose causal effect would it estimate?
Answer each question before checking the corresponding answer below.
Answer key
1. Causal question, treatment, and outcome
- Causal question: Does the extra election funding increase turnout?
- Treatment: receiving the funding
- Outcome: turnout
2. Comparison and design
The running variable is city population, and the cutoff is 50,000 residents. The useful comparison is between cities just above and just below 50,000, not between every city above and below the threshold. Cities very close to the cutoff should be similar while funding changes sharply at the cutoff. This is a regression-discontinuity design.
3. Assumptions
Without the funding, average turnout should change smoothly with city population at 50,000. Nothing else that affects turnout should suddenly change at that exact threshold, and cities should not be able to sort precisely to one side of it.
4. Concrete threat
Suppose city officials can annex a neighborhood just before the population count because they want the funding. Cities just above the cutoff may then have more capable or motivated governments than cities just below it. The local comparison would no longer be clean.
5. New information
Two policies now change at the same cutoff. A jump in turnout could be caused by the funding, weekend voting, or both. The design could estimate the effect of crossing the threshold and receiving the combined policy package, but it could not isolate the effect of the funding alone.
6. Whose effect
A valid RDD estimate would describe the effect for cities near 50,000 residents. It would not automatically tell us the effect for very small towns or very large cities.
Practice scenario 3: Difference-in-differences
At the start of 2024, several states adopted independent redistricting commissions, taking responsibility for drawing district maps away from the state legislature. Other states kept legislative control. A researcher has data on how competitive each state's House races were in 2022, before the reforms, and in 2024, after the reforms. The researcher wants to know whether commissions make elections more competitive.
Questions
What causal effect is the researcher trying to estimate? Identify the treatment and outcome.
What comparison would you make? What design have you described, and which states are the treated and control groups?
What assumption makes that comparison credible? Explain it in plain language.
Name one concrete way the assumption could fail in this setting. Explain why it would be a problem.
Now consider this additional information:
Before commissions were adopted, House elections were consistently more competitive in the adopting states than in the nonadopting states. However, competitiveness had been changing along similar trends in both groups.
Does the difference in starting levels defeat the design? Why or why not?
If the design worked, whose causal effect would it estimate?
Answer each question before checking the corresponding answer below.
Answer key
1. Causal question, treatment, and outcome
- Causal question: Do independent redistricting commissions increase electoral competition?
- Treatment: adopting an independent commission
- Outcome: the competitiveness of House elections
2. Comparison and design
The treated group is the states that adopt commissions, and the control group is the states that retain legislative redistricting. Compare the change in competitiveness in adopting states with the change over the same period in nonadopting states. A before-and-after comparison in adopting states alone would mix the reform with national changes. A post-reform comparison alone would mix the reform with earlier differences between the groups. This is a difference-in-differences design.
3. Assumption
Without the reform, competitiveness in the adopting and nonadopting states should have followed parallel trends. The control states must provide a credible picture of how competition would have changed in the adopting states without commissions.
4. Concrete threat
Suppose competitiveness was already increasing faster in the states that later adopted commissions. A continued increase after adoption might reflect that earlier trajectory rather than the commissions. Evidence of different pre-reform trends would make the comparison less credible.
5. New information
No. Difference-in-differences does not require the treated and control groups to begin at the same level. It requires that, without the reform, the two groups would have followed parallel trends. The difference in starting levels does not defeat the design if that assumption is credible.
6. Whose effect
A valid DiD estimate would describe the average effect for the states that adopted commissions during the period studied. It would not necessarily generalize to every state or every possible reform period.
Final advice
Practice explaining each design as a comparison, not as a vocabulary term. During the meeting, say what you are thinking, connect every assumption to the scenario, and revise your conclusion when the facts require it.