Data analyst interview questions: the complete guide

In this article

1. The vocabulary of data analytics
2. What data analyst rounds test
3. Start here: the learning path
4. SQL for analysts: joins, aggregation and window functions
5. Statistics and metrics
6. A/B test analysis
7. The metric-drop case
8. Dashboard design
9. Excel and Tableau
10. The fit round
11. Questions interviewers ask on data analytics
12. All 6 guides in Data analyst interviews
13. Follow-up questions
14. Sources

The vocabulary of data analytics

Data analyst

A data analyst is a person who answers business questions with data: they pull and clean tables, compute and compare numbers, build charts and dashboards, and explain what changed and what to do about it to people who do not work with data every day.

Metric

A metric is a number, defined by a precise formula over a stated population and period, that tracks one thing a business cares about, such as the share of visitors who buy. Two teams that compute "conversion" with different formulas have two different metrics under one name.

Sample ratio mismatch

A sample ratio mismatch (SRM) is the situation in which the observed split of users between variants differs from the planned split by more than chance allows, judged with a chi-squared goodness-of-fit test. Whether 52 to 48 counts depends on the sample: it is an SRM at 100,000 visitors and not at 1,000. A true SRM signals a data-quality or assignment problem that makes the result untrustworthy until it is explained.

What data analyst rounds test

Data analyst rounds test seven skills that every analysis depends on. We group them into seven topics, one section each: how to pull the right rows, how to describe and compare numbers, how to judge an experiment, how to investigate a number that moved, how to show results so people act on them, how to work in the tools the team already uses, and how to tell your own story.

Text description of the diagram

• Data analyst interviews test a small set of skills that every analysis depends on.
• The seven topics are SQL, statistics and metrics, A/B tests, metric-drop cases, dashboards, spreadsheets and Tableau, and the behavioural round.

Start here: the learning path

Take the topics in this order; each has its own section below.

1. SQL for analysts: joins, aggregation and window functions.
2. Statistics and metrics: averages, causation and what a number means.
3. A/B test analysis: sample size, randomisation and trust checks.
4. The metric-drop case: a number fell by ten percent, what do you do.
5. Dashboard design: purpose, audience and a small number of views.
6. Excel and Tableau: lookups, pivots and charts.
7. The fit round: why you, and the stories behind your work.

1. A/B test analysis questions for analysts: sample size, SRM and the ship decision
2. How do you answer a metric dropped case interview?
3. How do you answer dashboard design questions?

Where to go deeper

For more on three of these topics, read SQL, A/B testing questions and Tableau developer questions. The metric-drop case, dashboard design and Excel are covered in the sections below.

SQL for analysts: joins, aggregation and window functions

SQL rounds are often an early screen, and they test whether you return the right rows at the right grain. A window function is a function that computes a value for each row from a set of related rows without collapsing them. The PostgreSQL 18 documentation says that window functions "do not cause rows to become grouped into a single output row", and that "the rows retain their separate identities" (PostgreSQL docs, read 2026-10-24).

The difference is easy to see. Take five illustrative sales rows in two teams. A GROUP BY team average returns 2 rows, one average per team: 20 for the north and 45 for the south. The same average written as a window function, avg(amount) over (partition by team), returns all 5 rows, each carrying its team's average beside its own amount.¹ That is the tool for "show each rep next to their team's average". The usual traps are a join that multiplies rows because the key is not unique, and an aggregate computed at the wrong grain; the SQL join Venn diagram page shows where the picture misleads, and data modeling interview questions covers grain. The SQL hub goes deeper.

Statistics and metrics

Statistics questions check whether you summarise honestly and avoid claiming more than the data shows. Two questions come up often. The first is which average to report: the mean is pulled by outliers, so the median is the safer description of a typical value when a few large values exist, as explained on mean versus median. The second is whether one thing caused another. Two metrics that move together may both be driven by a third, and correlation versus causation shows how a hidden variable fakes a relationship.

The metric itself is part of the answer. Before analysing "conversion", state the formula, the population and the period, because the number depends on all three. A strong answer also names the guardrail, a second metric that must not get worse while the first improves. For the wider toolkit of distributions, sampling and p-values, read the statistics interview guide.

A/B test analysis

An A/B test is the standard way to attribute a change to a cause. The 2007 paper by Kohavi, Henne and Sommerfield says that controlled experiments "embody the best scientific design for establishing a causal relationship" and that the key is randomisation: users "cannot be distributed 'any old which way'" (Kohavi et al., 2007). Sample size is the number of users an experiment needs before a difference can be told from chance. They give a rule of thumb for it, for 95% confidence and 90% power, n = (4rσ/Δ)², and note that it "may overestimate"; here r is the number of variants, σ the standard deviation of the metric and Δ the smallest difference you want to detect.

Apply it to a 5% conversion rate with two variants, treating each user as a yes-or-no purchase, so σ² = 0.05 × 0.95. By the paper's conservative rule of thumb, detecting a 5% relative change takes a total of about 486,400 visitors, a 10% change 121,600 and a 20% change 30,400. An exact two-proportion power calculation (two-sided alpha 0.05, power 0.90) gives about 326,976, 83,624 and 21,841, so the rule of thumb runs 1.4 to 1.5 times high.¹ The square is the lesson either way: halve the change you want to see and you need about four times the sample (3.9 by the exact calculation). The questions on this are collected in A/B testing interview questions.

Text description of the diagram

• By the Kohavi rule of thumb, detecting a 5% relative change in a 5% conversion rate takes about 486,400 visitors; an exact calculation gives 326,976.
• For a 10% change the rule gives 121,600 (exact 83,624) and for 20% it gives 30,400 (exact 21,841).

The second check is trust. A sample ratio mismatch is a split that differs from the planned one by more than chance allows, and the standard check is a chi-squared goodness-of-fit test with one degree of freedom. A 52 to 48 split of 1,000 visitors (520 against 480) gives chi-squared 1.6 and p = 0.21, which is within chance. The same 52 to 48 split of 10,000 visitors gives chi-squared 16 and p of about 0.00006, and of 100,000 visitors gives chi-squared 160 and p of about 10⁻³⁶, far below 0.001; both are SRMs.¹

A 2019 paper drawing on experiments at four software companies across more than 25 products describes a mismatch as a warning sign with many possible causes, says a simple statistical check is used to detect it, and says that ignoring it can lead to shipping harmful changes (SRM taxonomy paper, 2019). Check the split before you read the result.

The metric-drop case

A metric-drop case asks you to investigate a number that fell, for example "weekly active users are down ten percent". The interviewer is testing order of thinking more than any single answer. A sound order is this method, which is our own structure and not a quoted standard. First, check the data: did logging change, did a pipeline break, is the dashboard reading the right table.

Second, locate the drop: split by time, platform, country and user type to see whether it is everywhere or in one segment. Third, list causes for that segment: a release, a seasonal pattern, an outage, a competitor, a change in who arrives. Fourth, say how you would test each cause and what you would recommend if it holds.

Always start with data quality. Some real drops are measurement faults, and an analyst who proposes a product fix before ruling that out has skipped the cheapest check. State the metric's exact formula first, which links this case to the section above.

Dashboard design

Dashboard questions test whether you build for a person, not for the data. Tableau's own best-practices page opens with "know your purpose and audience", advises placing the key view so that it "occupies or spans the upper-left corner", and suggests limiting a dashboard to "two or three" views (Tableau Help, read 2026-10-24). The page applies to Tableau Desktop, but the rules carry over to any tool.

A good interview answer walks four decisions. Who reads it and what decision do they make from it. Which three or four metrics answer that, each with its formula written next to it. Which comparison makes a number meaningful, such as against last week or against a target. And how the reader will know the data is stale or broken. A dashboard that shows twenty metrics with no comparison answers none of these.

Excel and Tableau

Some analyst roles ask about the spreadsheet and the visualisation tool the team uses. In Excel the lookup is the classic test. Microsoft describes the lookup function XLOOKUP as searching one column or row for a value and returning the matching item from another, including from a column to the left, with an exact match by default and an #N/A error if nothing is found unless you supply a fallback (Microsoft Support, read 2026-10-24). Microsoft notes that it is not available in Excel 2016 or Excel 2019, so check the version the employer uses.

Tableau questions are usually about the chart that fits the question and about how a dashboard behaves for a viewer. Prepare to explain one chart choice and one dashboard you built, including what you removed. The Tableau developer interview questions page collects the questions for that tool.

The fit round

The behavioural round tests whether you can explain your own work to someone who was not there. The opening is often an introduction, so prepare tell me about yourself and the closing question, why should we hire you, with one project told in numbers: the question, the data, your method and what the business did after. The behavioural interview hub covers the story structure, and questions to ask the interviewer covers the end of the round. If the role leans toward pipelines, data engineering interview questions cover the building side.

Questions interviewers ask on data analytics

For a worked question set on one tool see Tableau developer interview questions; for SQL see the SQL hub and for experiments see A/B testing interview questions.

Follow-up questions

Why should we hire you as a data analyst?

Name the skill the role needs most, give one result that proves it with numbers, and say what you would do in the first months. The why should we hire you page gives the structure and a weak and a strong answer side by side.

Data analyst interview questions for freshers: how do I prepare?

Learn the seven topics in this order, practise SQL against a real table until window functions feel routine, and prepare one project story with the question, the method and the outcome. Statistics questions are covered in the probability and statistics guide. Prepare your own questions for the interviewer as well.

What are the SQL interview questions for a data analyst?

Expect joins, aggregation with GROUP BY, filtering groups with HAVING, and window functions such as a running total or a rank within a group. The common failures are duplicated rows from a join on a non-unique key and counting at the wrong grain. The SQL hub has the full set. Python with pandas is also asked in some analyst rounds, mostly for the same operations: filtering, joining, grouping and reshaping a table.

What statistics interview questions are asked for a data analyst?

Expect which average to report, when a result is more than chance, sample size, and correlation versus causation. The probability and statistics guide covers them.

How do I answer a case where a metric dropped?

Check the data first, then split the drop by time and segment, then list causes for the segment where it is concentrated, and say how you would test each. Say your recommendation and what would change your mind.

Is Excel or Tableau asked in data analyst interviews?

Often, but it depends on the employer: read the job description for the tools it names. Prepare one lookup, one pivot or chart, and one dashboard you can explain choice by choice. See Tableau developer interview questions.

Sources

1. PostgreSQL 18 documentation, 3.5 Window Functions — PostgreSQL Global Development Group (date = read date), 24 Oct 2026
2. Practical Guide to Controlled Experiments on the Web: Listen to Your Customers not to the HiPPO — Kohavi, Henne and Sommerfield, ACM KDD 2007 (author's version), 12 Aug 2007
3. Diagnosing Sample Ratio Mismatch in Online Controlled Experiments: A Taxonomy and Rules of Thumb for Practitioners — ACM, DOI 10.1145/3292500.3330722 (publication page), 25 Jul 2019
4. Best Practices for Effective Dashboards — Tableau Help, applies to Tableau Desktop (date = read date), 24 Oct 2026
5. XLOOKUP function — Microsoft Support (date = read date), 24 Oct 2026