Data · Guide

Case Study Practice

The analyst interview: structuring an open question, choosing the metric, and saying what you would do about it.

— min read Data

What The Case Round Tests

"Sign-ups dropped 20% last week — what happened?" is not a SQL question. It is testing whether you can structure ambiguity, pick a defensible metric, and end with a recommendation rather than a chart.

Candidates lose these rounds in predictable ways: jumping straight to a query, computing something impressive that answers a different question, or finishing with an observation instead of a decision. The technical part is rarely what fails.

A Framework That Holds Up

StepDo
1. ClarifyWhat decision does this support, for whom, by when
2. DefineState the metric precisely — a sign-up is what, exactly
3. StructureBreak the question into parts before touching data
4. HypothesiseList candidate causes, cheapest to check first
5. AnalyseTest them, and say what would falsify each
6. RecommendWhat you would do, and what you would watch

Say the structure out loud before analysing. Interviewers are scoring the approach, and a stated plan lets them redirect you early — which is help, not a penalty.

Ask what changed on the product, marketing and tracking sides before theorising about user behaviour. A deploy, a campaign ending, or a broken analytics event explains most sudden drops, and checking them is cheap.

Choosing The Metric

Half the answer is picking the right number. A metric worth using is specific, comparable over time, and moved by things the team controls.

TrapBetter
Total usersWeekly active users — cumulative counts only go up
Average revenue per userMedian, or a distribution — one whale distorts a mean
Raw conversionConversion by cohort — mix shifts fake movement
A vanity totalA rate with a stated denominator
Simpson's paradox is the classic case-round trap: every segment improves while the total worsens, because the mix between segments changed. Always segment before concluding a trend is real.

Diagnosing A Drop

A structured decomposition beats intuition, and it is what the interviewer wants to see.

Cut byReveals
TimeA step change (a deploy) versus a slope (a trend)
Platform and deviceAn iOS release, a broken browser build
GeographyA regional outage, a campaign that ended
Acquisition channelPaid spend stopping, an SEO change
Funnel stepWhich stage lost the users
New versus returningAcquisition problem or retention problem

Always ask whether the data itself broke first: a tracking change, a failed pipeline run or a bot filter deployed last Tuesday explain a surprising share of "drops", and finding that is a good answer, not a cop-out.

Ending With A Recommendation

Lead with the answer. Stakeholders need the conclusion, its confidence, and what to do — the methodology is support, not the story.

SayNot
"Sign-ups fell 20%, all of it on Android after Tuesday's release""I joined the events table to the users table…"
"I'd roll back and confirm within a day""Here is a dashboard"
"Confident on the where, less on the why"Unqualified certainty
"If I had more time I'd check X"Silence about the gaps
Stating uncertainty honestly reads as senior, not weak. "The effect is clear; the cause is one of these two, and here is how I'd separate them" is a stronger answer than false confidence.

Interview Questions

How do you approach an open case question?

Clarify the decision it supports, define the metric precisely, structure the question into parts, list hypotheses cheapest-first, test them, and finish with a recommendation and what you would monitor.

Sign-ups dropped 20% — what do you check first?

Whether the data broke: tracking changes, a failed pipeline, a bot filter. Then a deploy, a campaign ending, or a platform-specific release — before theorising about user behaviour.

What makes a metric worth using?

It is specific, comparable over time, and moved by things the team controls. Cumulative totals and averages distorted by outliers fail all three.

What is Simpson's paradox in this context?

Every segment improves while the overall number worsens, because the mix between segments shifted. It is why you segment before concluding a trend is real.

How should a case answer end?

With a recommendation and a stated confidence level — what you would do, what you would watch to confirm it, and what you would check with more time.

Why state uncertainty explicitly?

It reads as senior. Separating what the data shows clearly from what remains ambiguous is more useful, and more credible, than unqualified confidence.

Quick Quiz

1. The first thing to check on a sudden metric drop is…
2. Total cumulative users is a poor metric because…
3. Every segment improving while the total falls is…
4. A case answer should end with…
5. Stating the structure before analysing is…