You notice our week-over-week active users dropped 15% on Tuesday. Walk me through how you investigate.

Start by ruling out data pipeline issues before assuming the product broke — check if the ingestion job ran and if historical Tuesdays show any seasonal pattern. Then segment the drop: by platform (iOS / Android / web), by geography, by user cohort (new vs returning), by acquisition channel. Next, check correlated events: was there a deploy on Monday, a pricing change, an email campaign that stopped, a competitor announcement? Close by naming the metric you would look at to confirm root cause, not just the symptom.

What is the difference between a CTE, a subquery, and a temp table — and when do you use each?

CTEs (WITH clauses) are readable, non-materialized by default in most databases, and best for multi-step logic you need to reference more than once in the same query. Subqueries are inline and fine for one-off references but become unreadable when nested more than two levels deep. Temp tables are materialized — they physically write to disk (or memory) — making them the right choice when you need to reuse a large intermediate result multiple times or when a CTE would force the engine to recompute the same expensive logic repeatedly. Mention that CTE materialization behavior differs by database engine (Postgres materializes CTEs as of version 12, SQL Server may or may not).

Our A/B test showed a 5% lift in conversion with p=0.04. The VP wants to ship immediately. What questions do you ask before agreeing?

Ask about sample size and statistical power — a 5% lift with a small sample may be noise. Ask about the test duration — did it run through at least one full weekly cycle to capture weekday/weekend behavior? Ask about secondary metrics — did the variant hurt any other metric like support tickets or refund rate? Ask about novelty effect — for some products, any change gets a temporary lift. Ask about practical significance — a 5% lift in conversion on a page that gets 100 visits a week is worth less than one that gets 100K. Finally, ask whether the p-value was the primary metric or the result of multiple comparisons.

Tell me about an analysis you did that changed a decision or strategy at your company.

Pick a project where you can name a specific decision that changed. Open with the business question (one sentence), then describe what data you used and what you found that was non-obvious. The most important part: describe the moment the stakeholder changed their position because of your finding, and what they decided to do differently. Close with the measurable outcome. Do not bury the business impact at the end — surface it early and return to it.

Design a metric framework to measure the health of a subscription product's customer base.

Organize metrics in three layers: acquisition (new subscriber count, CAC, trial-to-paid conversion rate), retention (monthly churn rate, cohort retention curves at 30/60/90/180 days, net revenue retention if expansion exists), and health indicators (engagement depth — how many features does the average active subscriber use, support ticket volume per subscriber, NPS or CSAT trend). Then call out the leading vs lagging distinction: churn rate is lagging; engagement decline is a leading indicator of future churn. Propose one alert metric that would trigger investigation if it crossed a threshold.

You have a dataset with 30% missing values in a key column. What do you do?

First, investigate the pattern of missingness: is it random (MCAR — missing completely at random), or systematic (MAR or MNAR)? If missing values cluster around a specific time range, user cohort, or event, that is a data pipeline or instrumentation problem, not a statistical problem — fix the pipeline first. If missingness is random, the choice depends on the analysis: for simple descriptive stats, exclude and note the limitation; for predictive models, impute with median for numeric or mode for categorical, or use an indicator variable for the missingness itself. For 30% missing, always report the analysis with and without imputation to check sensitivity.

Write a query to calculate 7-day rolling average of daily active users.

Use AVG(dau) OVER (ORDER BY date ROWS BETWEEN 6 PRECEDING AND CURRENT ROW) on a daily aggregated table. Walk through why ROWS BETWEEN 6 PRECEDING AND CURRENT ROW gives a 7-day window (6 previous + current = 7). Call out that this requires the table to have one row per day — you may need a calendar table join if some days have no activity. Mention that RANGE BETWEEN 6 PRECEDING AND CURRENT ROW means something different when the ORDER BY column is not unique.

Tell me about a time you had to present an analysis to a non-technical stakeholder who disagreed with your conclusion.

Set up the disagreement specifically — what did the stakeholder believe, and what did your analysis show? Describe how you communicated: did you lead with the business implication rather than the methodology? Did you offer to re-run with their assumptions to show the sensitivity? If you were overridden anyway, describe what happened next and whether the outcome validated your analysis. Be honest if you were wrong — that is also a strong answer.

What is the difference between a left join and an inner join, and when does the choice change your result materially?

Inner join: only rows with matching keys on both sides. Left join: all rows from the left table, NULLs for unmatched right-side columns. The choice matters most when you are measuring completeness — for example, joining users to their orders to calculate "users who have never ordered." An inner join silently excludes non-buyers; a left join surfaces them with NULL order values. Another common trap: if your right table has duplicates on the join key, both join types will multiply rows — always verify cardinality before joining.

How do you prioritize when a data scientist, a product manager, and a growth team all have urgent requests that exceed your capacity?

Describe a framework you actually use. A practical one: assess decision deadline (when does this analysis actually need to influence a decision — not when the requester wants it), business impact (how much revenue or risk does this decision involve), and effort (can any of these be answered with existing data or do they need new work). Surface the tradeoff explicitly to all requesters rather than silently deprioritizing someone. For chronic overload, propose a tiered SLA and get manager alignment.

Explain what a p-value actually means, without using the phrase "probability that the null hypothesis is true."

A p-value is the probability of observing a test statistic at least as extreme as the one you computed, assuming the null hypothesis were true. It measures how compatible your data is with the null — not the probability the null is correct. A p-value of 0.04 does not mean "there is a 4% chance the treatment has no effect." It means "if the treatment had no effect, there would be only a 4% chance of seeing a result this extreme or more extreme just by random chance." Distinguish it from practical significance: a tiny p-value on a large sample can represent an effect too small to care about.

Are companies in 2026 still hiring data analysts, or is AI replacing the role?

Be direct: AI tools (Claude, ChatGPT, Gemini) have automated the low-end of SQL query writing and simple dashboard creation. But the work that remains — defining the right metric, designing an experiment, interpreting a result in business context, communicating to leadership — still requires human judgment. Analysts who were doing only mechanical query work are at risk. The role is shifting toward "analytics engineer" (building pipelines, dbt models) and "business analytics" (strategy and interpretation). You should be able to say concretely which part of the work you do that AI has not yet replaced.

Do you have questions for me?

For the hiring manager: "What is the most common reason an analysis gets ignored or overridden here, and how does the team handle that?" For a data engineer peer: "How mature is the data warehouse — are analysts self-sufficient, or is there typically a backlog of data requests before you can start analysis?" For anyone: "What does a strong analyst here do that a good analyst does not?" Avoid asking things answerable from the company's website or the job description.

How important is a SQL take-home test in a data analyst interview?

Very important — most companies use it as the primary filter before the onsite loop. The take-home typically involves a messy dataset (missing values, duplicate records, ambiguous columns) and asks you to surface an insight. The best submissions combine clean SQL, a clear business interpretation, and a note on limitations. Plan for 3-5 hours of actual work.

Do data analyst interviews in 2026 test Python, or is SQL enough?

Depends on the company. At product analytics teams (Airbnb, Spotify, DoorDash type roles), Python with pandas for data cleaning and Matplotlib/Seaborn for visualization is commonly tested. At business analytics teams inside non-tech companies, SQL + Excel + Tableau is usually sufficient. Check the job description carefully — if it lists pandas or numpy, expect a Python exercise.

Should I have a portfolio for a data analyst role?

Not required, but increasingly differentiating in a crowded market. A GitHub repository with 2-3 notebooks showing end-to-end analysis (data cleaning → exploratory analysis → insight → visualization) is more useful than a data viz portfolio alone. Focus on real datasets with business-relevant questions, not just aesthetic charts. Kaggle notebooks that answer a business question work fine.

How do I answer "what BI tools do you know" if I only know one?

Be honest about your primary tool and reframe: "I have deep experience with Tableau, and I have found that BI tools share enough conceptual overlap — calculated fields, joins, filters, level-of-detail expressions — that I have gotten productive in a new tool within a few weeks in past roles." Then demonstrate you understand the core concepts, not just button locations.

Are companies still hiring data analysts given AI tools in 2026?

Yes, but the role is evolving. AI tools have automated mechanical query writing and simple dashboard creation. Demand is strong for analysts who can design metrics, run experiments, and translate data to business decisions. The new threat is to junior analysts doing only report generation — that tier of work is being absorbed by self-serve BI and AI. Build toward metric ownership and statistical fluency if you want to be durable.

Technology$60,000 - $120,000

Data Analyst Interview Questions

What 2026 data analyst interviews actually test

Data analyst interviews shifted after 2024: SQL is still the gatekeeper, but companies now weight business judgment and metric design as heavily as technical execution. These questions come from documented patterns at B2B SaaS, e-commerce, and fintech data teams — with answer frameworks that reflect how interviewers actually score candidates.

Typical rounds

End-to-end time

3-5 weeks

Questions covered

What the Data Analyst interview loop actually looks like

Recruiter Screen

• Phone call• 30 min

Tool stack familiarity, domain background, compensation alignment. They will ask about SQL and BI tools — be specific.

Hiring Manager Screen

• Video call• 45 min

Past project depth — one analysis you owned end-to-end: data sources, methodology, how it influenced a decision. They want to know if you think in business terms, not just queries.

SQL / Technical Take-Home or Live Screen

• Take-home (24-48 hrs) or live SQL in a shared editor• 90 min

Window functions, CTEs, aggregations, data cleaning logic. Many companies now send a messy dataset and ask you to surface a business insight, not just write a correct query.

Onsite / Virtual Loop

• 3 back-to-back sessions• 180 min

1 deep SQL or analytics problem, 1 metric design / case study ("design a dashboard for our ops team"), 1 behavioral with a stakeholder manager or senior analyst.

Final / Cross-functional

• Video call with data or product leader• 45 min

Business judgment, communication style, prioritization. They are evaluating whether they will trust you to present to leadership.

14 Data Analyst interview questions

Tap any question to see what the interviewer is really asking, how to structure your answer, and the red flags to avoid.

What they're really asking

Can you write a real window function under pressure, and do you know the difference between RANK(), DENSE_RANK(), and ROW_NUMBER()? Most candidates claim SQL fluency but fail on ranking with ties.

Answer framework

Start by clarifying the schema — what tables, what columns for revenue and region. Then reach for DENSE_RANK() OVER (PARTITION BY region ORDER BY SUM(revenue) DESC) in a CTE. Filter the outer query to WHERE rnk <= 3. Walk through why DENSE_RANK() is correct for ties: if two products tie at rank 1, both appear as rank 1 and the third product appears as rank 2, not rank 3. Mention you would add a date filter as a WHERE clause on the inner aggregation, not the outer.

What a strong answer signals

You distinguish DENSE_RANK from ROW_NUMBER without prompting. You write the CTE before the outer filter. You explicitly state why the date filter belongs on the inner query for performance. You clarify the schema before writing.

Red flags to avoid

•Using ROW_NUMBER() without acknowledging it breaks ties arbitrarily
•Filtering dates in the outer query — signals you do not think about query performance
•Writing the whole query without clarifying table structure first

How Data Analyst hires actually get decided

Approximate weight hiring committees place on each dimension. Use this to focus your prep on what actually moves the decision.

SQL depth and accuracy

35%

Window functions, CTEs, performance awareness, and clean query structure. This is the non-negotiable gatekeeper skill. Most companies filter here before evaluating anything else.

Business judgment and metric design

30%

Can you translate a vague business question into a measurable framework? Can you push back on a bad metric? This is what separates a senior analyst from a mid-level one and is increasingly tested explicitly.

Communication and stakeholder management

20%

How you present a finding, handle disagreement, and navigate competing priorities. Data analysts are often the most cross-functional person on a data team — poor communicators stall regardless of technical skill.

Statistics and experimental design

10%

A/B testing literacy, p-value interpretation, and knowing when not to run an experiment. Weighted lower than SQL at most companies, but a single critical error (wrong p-value definition, ignoring multiple comparisons) can fail a candidate.

Tooling and data stack fluency

Familiarity with the company's specific BI tool (Looker, Tableau, Mode), data warehouse (Snowflake, BigQuery), and workflow tools (dbt, Airflow). Rarely determinative, but a signal of whether you can contribute quickly.

How to prepare for a Data Analyst interview

Master window functions before anything else

RANK, DENSE_RANK, ROW_NUMBER, LAG, LEAD, and frame-bounded aggregations (ROWS BETWEEN) appear in 80% of data analyst SQL screens. Practice each one on a real dataset, not just in isolation. Build one end-to-end query that uses three window functions in the same CTE — that is the level of fluency interviews expect at mid-level and above.

Prepare one investigation story per domain

Almost every data analyst interview includes a "metric dropped — investigate" scenario. Before your loop, write out two or three investigation stories from your experience: what dropped, your debugging sequence, what you ruled out, and what you found. Having real examples makes case-study responses concrete and fast. If you have not seen a real drop, work through a hypothetical until the sequence is automatic.

Know your SQL engine's specific behavior

PostgreSQL, BigQuery, Snowflake, and Redshift handle CTEs, date functions, and NULL comparison differently. Know which engine the company uses before your screen (ask in the recruiter call or check their job description), then practice with that engine. Getting a Snowflake-specific syntax question wrong because you practiced on MySQL is preventable.

Practice translating analysis to business language

Take your best past analysis and explain it in two sentences to someone who does not know SQL. Then explain it in one sentence. The compression exercise forces you to identify the actual business implication. Interviewers who are PMs or VPs of Product will weight your communication clarity as much as your technical accuracy.

Learn dbt basics even if your current job does not use it

dbt (data build tool) is now the de facto standard for analytics engineering at modern data teams. In 2026, senior analyst roles almost universally list dbt in the job description. You do not need to be an expert, but being able to explain what a dbt model is, how testing works in dbt, and why it is preferable to raw SQL scripts signals that you are current with the field.

Data Analyst interview FAQs

More for Data Analysts

Data Analyst resume example Data Analyst cover letter Data Analyst salary guide Skills for Data Analysts How to become a Data Analyst

Interview prep for related roles

Product Manager interview questions

Business • $100,000 - $200,000

Software Engineer interview questions

Technology • $80,000 - $180,000