These data analyst interview questions are for analytics managers and recruiters hiring analysts, and for candidates preparing for an analyst loop. They focus on what separates good analysts from people who can only write queries: choosing the right metric, spotting a broken number, reasoning about statistics, and explaining a result to someone who will act on it. If you need to test query writing itself, pair this page with our SQL interview questions, which go deep on joins, window functions and query plans. Pick 6 to 8 questions below, and always ask "how would you check that?" after an answer.
What a strong data analyst interview checks
- Problem framing: turning a vague request into a question with a clear metric, time window and decision attached.
- Metric sense: knowing what a number measures, what its denominator is, and how it can mislead.
- Statistics in practice: averages vs medians, variance, significance and when a difference is real.
- Data quality habits: checking duplicates, missing values, time zones and definitions before sharing anything.
- Communication: leading with the answer, choosing the right chart, and stating caveats without burying the point.
Data analyst interview questions: metrics and problem framing
1. A manager asks, "How are we doing on retention?" What do you ask before you pull any data?
What good looks like: they clarify which customers count (all, paid, a cohort), what "retained" means (logged in, purchased, still subscribed), the time window, and what decision the answer will feed. Strong candidates propose a definition and confirm it rather than waiting for a perfect brief.
Red flag: jumping straight to a query and returning one number with no definition attached.
2. Daily active users jumped 20% overnight. Walk me through what you check.
What good looks like: first, is it real? They check for a tracking change, a new app release, bot or test traffic, duplicate events, or a backfill. Then they segment by platform, country and acquisition source to see where the jump lives. Only after ruling out data problems do they look for a business cause like a campaign.
3. Conversion rose from 4.0% to 5.0%. How would you describe that change?
What good looks like: it rose by 1 percentage point, which is a 25% relative increase. Strong candidates say which framing they'd use with which audience, and ask whether the denominator (visitors, sessions, sign-ups) stayed the same over both periods.
4. Average order value went up, but total revenue went down. How is that possible?
What good looks like: fewer orders. If order count falls faster than the average rises, revenue drops. A mix shift can also do it, for example losing many small buyers while a few large ones stay. They'd break revenue into orders times average value and look at each piece by segment.
5. How would you choose a north-star metric for a subscription product?
What good looks like: a metric tied to the value customers get, that the team can influence, and that predicts revenue, such as weekly active paying accounts. They pair it with guardrail metrics (churn, support tickets) so the team can't push the main number up in a harmful way.
Statistics and experiments
6. When would you report the median instead of the mean?
What good looks like: when the data is skewed or has outliers, like income, order size or session length. A few huge values pull the mean up, while the median shows the typical case. Good candidates often report both and explain the gap.
7. An A/B test shows p = 0.03. What does that mean, and what doesn't it mean?
What good looks like: if there were truly no difference, you'd see a result at least this extreme about 3% of the time. It does not mean there's a 97% chance the variant is better, and it says nothing about whether the effect is big enough to matter. They'd also look at the confidence interval and the size of the lift.
Red flag: "So we're 97% sure B wins."
8. The test was planned for two weeks. After three days the variant is "significant." Do you stop?
What good looks like: no. Checking repeatedly and stopping at the first significant result raises the false positive rate well above the stated level. Run to the planned sample size, or use a method built for sequential monitoring. Strong candidates also mention weekday effects that a three-day window misses.
9. You planned a 50/50 split, but the variant got 46% of users. Does it matter?
What good looks like: yes, this is a sample ratio mismatch, and with a large sample it usually points to a bug in assignment or logging. The results can't be trusted until the cause is found, because the two groups may no longer be comparable.
10. Treatment B wins in every region, but loses overall. How?
What good looks like: Simpson's paradox. If B was shown mostly in regions with low baseline rates, the overall comparison mixes region effects with the treatment effect. They'd report the segmented results and fix the assignment for future tests.
11. Customers who use feature X churn less. Should we push everyone to use it?
What good looks like: not on that evidence alone. Engaged customers may both use X and stay, so X could be a symptom, not the cause. They'd suggest an experiment, or at least compare similar customers before and after adoption, before claiming causation.
Data quality and tools
12. You join orders to a promotions table and revenue doubles. What happened?
What good looks like: a one-to-many join fanned out the rows, so each order was counted once per matching promotion. They'd check row counts before and after the join, aggregate the many side first, or join on a unique key. (The SQL page has deeper join questions if you want to go further.)
13. A new dataset lands on your desk. What do you check before analyzing it?
What good looks like: row counts against the source, the grain (what one row represents), duplicates on the expected key, missing values by column, date ranges and time zones, and a few hand-checked rows. They write down assumptions so others can review them.
14. How do you handle missing values in a column you need?
What good looks like: first ask why they're missing, since a field left blank on purpose differs from a logging failure. Options include excluding rows (and saying how many), filling with a sensible default, or flagging them as a separate group. The choice depends on whether missingness is related to the outcome.
15. When do you use a spreadsheet, and when do you move to SQL, Python or a BI tool?
What good looks like: spreadsheets for small, one-off work and quick sharing; SQL when data lives in a warehouse or is too large; Python or R for repeatable cleaning, statistics or modeling; a BI tool for metrics people check repeatedly. They mention version control or documented steps for anything that will be rerun.
16. A stakeholder's number doesn't match yours. How do you resolve it?
What good looks like: compare definitions, filters, date ranges and sources line by line, often by reconciling on a small sample. They agree on one definition, document it, and fix whichever report is wrong, without making it about who was right.
Communication and behavioral (STAR)
17. Explain a box plot to someone in sales.
What good looks like: the box covers the middle half of the values, the line inside is the median, and the whiskers and dots show the spread and unusual values. A strong answer uses a concrete example, like deal sizes by region, and says what decision the chart helps with.
18. How do you structure a one-page summary of an analysis?
What good looks like: the answer first, then the two or three findings that support it, a clear chart, the recommendation, and caveats at the end. They write for the reader's decision, not to show every step they took.
19. Tell me about an analysis that changed a decision.
What good looks like: a STAR story with a clear Situation, the question they were given, the Actions (including how they checked the data), and a measurable Result. Strong candidates explain why the audience trusted the finding.
20. Tell me about a time your analysis was wrong or incomplete.
What good looks like: they name the mistake plainly, say how they found it, who they told and how fast, and what check they added afterwards. Honest ownership matters more than the size of the error.
Red flag: blaming the data team or the stakeholder for the entire problem.
21. Describe a time you pushed back on a request.
What good looks like: a request that would have produced a misleading answer, such as a dashboard of vanity metrics. They explain how they offered a better question or metric, kept the relationship intact, and what happened next.
A simple scoring rubric
Running this interview with AI
Analyst pipelines are often crowded, and the first round mostly checks reasoning and communication, which suits an adaptive AI interviewer. With NTRVSTA, candidates take a live interview (video avatar or audio) whenever they're free, and you can customize the questions per role, so the metric and experiment questions above can go straight in. The interviewer adapts in real time: a candidate who handles the conversion question well gets the sample ratio mismatch follow-up, while one who struggles gets a simpler step. The hint system shows who can recover with a small nudge.
Hiring managers get a score breakdown, including communication, plus video snippets of key answers, so the second round can focus on a live data exercise. Read how AI interviews are scored before relying on any automated score, and compare vendors in our AI interview software guide. Writing the posting? Use the data analyst job description template. For neighboring roles, see business analyst and data scientist interview questions.
Questions
Should a data analyst interview include a take-home task?
A short one can help, but keep it under two hours, use a small realistic dataset, and pay attention to the write-up as much as the numbers. Many teams get the same signal from a 30-minute live exercise where the candidate explores a table and explains what they would check first.
How is this different from a SQL interview?
A SQL interview tests whether someone can write correct queries. A data analyst interview tests whether they ask the right question, pick the right metric, check their numbers, reason about statistics and explain the result so a manager can act on it. Most analyst loops need both, usually in separate rounds.
What should junior analysts be expected to answer?
Juniors should handle the metric, denominator and data-cleaning questions and explain a chart clearly. Expect lighter answers on experiment design and forecasting, and score how they reason out loud and whether they check their own work, not whether they know every term.
Curious how an adaptive first round handles analyst candidates? Book a 20-minute demo with our team, or run a sample interview yourself first.
