Difference-in-Differences Interview Questions: Parallel Trends and Staggered Adoption

Difference-in-Differences interview questions test whether you can turn a before-and-after table into a causal claim without hiding the identifying assumption. The arithmetic is short. The hard part is choosing a credible comparison group, defining the treatment effect you want, and recognizing when staggered adoption makes a familiar two-way fixed-effects regression difficult to interpret.

This guide focuses on observational panel data: parallel trends, event time, treatment cohorts, and the comparisons that remain valid after different groups adopt at different dates.

Quick answer: In a two-group, two-period design, DiD subtracts the control group’s change from the treated group’s change. A causal interpretation requires the treated group’s untreated potential outcome to have followed the comparison group’s trend over the relevant period, plus design assumptions such as no anticipation and no spillovers. With staggered adoption and heterogeneous effects, do not assume one standard TWFE coefficient is the average effect you mean. Define cohort-by-time effects, use never-treated or not-yet-treated comparisons that match the estimand, and aggregate them deliberately.

What the interviewer is actually testing

A strong response answers six questions in order: What is the unit? When does treatment begin? Which outcome and time window define the estimand? Who supplies the untreated counterfactual? Why is parallel trends plausible for that comparison? How will you handle different adoption dates and effects that change over exposure time?

The formula is evidence only after those choices are explicit. Official method documentation from the authors of the did package describes the canonical two-period case, the staggered-adoption extension, and group-time average treatment effects. It does not certify a business analysis or provide a universal checklist for interviews. The examples below are synthetic preparation cases, not candidate or company data.

Work the two-by-two table before writing a regression

Suppose a marketplace launches a seller-training program in Region T after month 0. Region C has not launched it. The outcome is completed orders per active seller.

Group	Before	After	Change
Treated region T	40	55	+15
Comparison region C	38	46	+8

The DiD estimate is (55 − 40) − (46 − 38) = 7 completed orders per active seller. The same number can be written as (55 − 46) − (40 − 38): the post-period gap minus the pre-period gap.

Do not call 15 the treatment effect; that includes the common change visible in the comparison region. Do not call 9 the effect either; that is only the post-period difference and ignores the groups’ initial two-order gap. Seven is the sample contrast under the stated design. Calling it causal requires the counterfactual claim that Region T would have changed by about eight without training.

An interview-ready answer names the estimand: “the average effect on treated active sellers in Region T over this post period.” It then asks whether the seller population changed, whether treatment affected entry into “active” status, and whether orders moved across regions. A clean calculation cannot repair a post-treatment denominator or spillover.

Parallel trends is a counterfactual assumption

Parallel trends does not mean the groups must have equal outcome levels. It means their untreated potential outcomes would have changed similarly over the comparison period. The two groups may begin at 40 and 38; the relevant claim is about the unobserved change Region T would have experienced without treatment.

Several pre-treatment periods can make the assumption more credible or less credible. Plot group means and event-time coefficients, inspect changes in composition, and explain why the same shocks should affect both groups. A flat set of pre-treatment estimates is not proof. Tests may have low power, and a design can pass a pre-trend test while failing after a new group-specific shock.

Avoid controlling for a variable caused by treatment just to make the lines look parallel. If training changes seller activity and you condition on active days after launch, you may redefine the population through a post-treatment variable. Baseline covariates can support a conditional parallel-trends argument when chosen before treatment and tied to the assignment story; they are not a mechanical cure.

Separate identification from estimation

Identification asks what assumptions connect observed changes to an unobserved counterfactual. Estimation asks how you compute the contrast and uncertainty from the data. Start with identification.

For a canonical two-period design, a regression with group, time, and treated-by-post terms can reproduce the 2×2 DiD contrast. Cluster standard errors at the level where treatment or shocks are correlated; the right choice depends on the assignment and sampling process. Fixed effects do not create parallel trends, remove spillovers, or make standard errors independent.

Also distinguish panel data from repeated cross-sections. With a panel, you follow the same units. With repeated cross-sections, group composition can change and the assumptions needed for a population-average contrast differ. Ask which data you have before promising unit fixed effects.

Staggered adoption changes the comparison problem

Now suppose Cohort 1 adopts in period 3, Cohort 2 in period 4, and Cohort 3 in period 5. A unit stays treated after adoption. At period 3, Cohorts 2 and 3 are not yet treated and may serve as comparisons for Cohort 1. By period 5, all three are treated; comparing an early-treated cohort with a later-treated cohort can mix the early cohort’s already-realized treatment effect into the control side.

Sun and Abraham show that in event-study regressions with variation in treatment timing, a lead or lag coefficient can be contaminated by effects from other relative periods when treatment effects differ across cohorts or exposure time. Their result is about a common estimator under specified conditions; it does not mean every fixed-effects model is unusable. It means you must inspect the comparisons and the estimand instead of interpreting coefficient labels literally.

Callaway and Sant’Anna build group-time effects such as ATT(g,t): the average effect for the cohort first treated in period g, evaluated in calendar period t. These are useful building blocks because each comparison can use a valid untreated group under the design’s assumptions. You can then aggregate across cohorts or event time with weights aligned to the question.

Choose the control group to match the question

A never-treated group remains untreated throughout observation and can offer a stable reference, but it may differ structurally from adopting cohorts. Not-yet-treated groups can look more comparable because they eventually adopt, yet they stop being valid controls after their own adoption. The estimator cannot decide which identifying story is believable.

State the choice explicitly: “I will compare cohort g with units not treated by period t, then stop using a unit as control once it adopts.” Check whether anticipation begins before the recorded launch. If sellers receive training materials or incentives early, the formal start date may not mark the first behavioral response.

When every unit is eventually treated, late adopters provide only temporary controls. Effects far from adoption may be weakly supported or unidentified because no untreated comparison remains. Show the support by cohort and event time; do not draw a smooth long-run curve from cells with no credible comparison.

Define the aggregation before reporting one number

There is no automatically correct “overall DiD effect.” You might want the average effect across treated units in each calendar period, an event-study path by time since adoption, a cohort-specific effect, or a policy effect weighted by cohort size. These answer different questions.

For example, an event-time average at e = 2 includes cohorts observed two periods after adoption. If later cohorts have not accumulated two post periods, the composition differs from e = 0. A rising event-study line may reflect duration, cohort composition, or both. Report the weights and support, then run an aggregation that matches the product or policy decision.

This is where a candidate can stand out: say what would make the recommendation change. If the decision is whether to continue the program for current treated regions, an ATT weighted by current treated sellers may fit. If the decision is how a new region might respond, transportability becomes a separate assumption; the historical ATT does not automatically answer it.

Common interview traps and how to repair them

“The pre-trend test is insignificant, so parallel trends holds.” Repair it by calling the plot and test diagnostics, discussing power, and defending the comparison from the treatment assignment process.

“Unit and time fixed effects solve selection bias.” Repair it by stating the counterfactual trend assumption and identifying group-specific shocks, anticipation, compositional change, and spillovers.

“I will control for every available feature.” Repair it by separating pre-treatment confounders from variables changed by treatment. Explain why each covariate belongs in a conditional trend argument.

“The event-study coefficient two periods after launch is the effect at two periods.” Under staggered adoption, first verify how the estimator forms comparisons and whether heterogeneous effects contaminate that coefficient.

“A significant estimate means we should launch.” Statistical uncertainty is only one part of the decision. Translate the effect into the business unit, check guardrails, and state whether the magnitude clears a worthwhile threshold.

A concise answer structure

Try this sequence aloud: “I’ll define the unit, treatment cohort, outcome, and event time. For the basic table, I’ll subtract the comparison change from the treated change. The causal interpretation relies on parallel untreated trends, no anticipation, and no spillovers across the groups. I’ll inspect pre-period dynamics and composition, but I won’t treat an insignificant pre-trend test as proof. If adoption is staggered, I’ll estimate cohort-by-time effects against never-treated or not-yet-treated units and aggregate them for the decision, rather than interpreting one TWFE coefficient without checking its comparisons.”

Then calculate the 2×2 example and name one failure mode. If asked for code, sketch the data checks before the model: one row per unit-period, stable IDs, treatment date consistency, no treatment reversal under the absorbing-treatment design, outcome completeness, and counts by cohort and time.

Candidate-reported practice prompts

Question	DiD angle to practice
Design an A/B test with causal inference	Contrast randomized identification with a parallel-trends argument.
Design an experiment with marketplace network effects	Identify spillovers that would break a simple group comparison.
Recommend a Decision After a Nonsignificant Experiment	Separate uncertainty from evidence of no effect.
Defend an Experiment Decision and Its Incremental Impact	Connect an estimand and effect size to a decision.
Analyze A/B Test Results to Inform Stakeholder Decisions	Communicate assumptions, guardrails, and remaining risk.

Final checklist for a DiD interview answer

Before you finish, verify that you named the causal unit, treatment timing, comparison group, outcome window, and target effect. Show the simple contrast. State parallel trends in counterfactual terms. Discuss anticipation, spillovers, and composition. With staggered adoption, show which cohorts remain untreated in each period and which aggregation answers the question. Finally, translate the estimate and interval into a practical decision without calling association proof.

Sources and Further Reading

• Callaway and Sant’Anna: Introduction to DiD with Multiple Time Periods — canonical DiD, staggered adoption, assumptions, and group-time effects.
• Callaway and Sant’Anna: Problems with Two-Way Fixed-Effects Event-Study Regressions — simulations and limitations of common TWFE event studies.
• Callaway and Sant’Anna (2021): Difference-in-Differences with Multiple Time Periods — identification, estimation, and aggregation with multiple periods.
• Sun and Abraham: Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects — contamination in conventional event-study coefficients and an alternative estimator.
• Callaway: Difference-in-Differences chapter — unified treatment of modern DiD results and practical complications.

Research checked September 12, 2026. The numeric examples are synthetic. Method sources support the stated statistical results.