Data Scientist Interview Prep for OPT and H-1B Candidates: Product Sense, A/B Testing, SQL and Stats

SQL, stats, A/B testing, product sense and behavioral rounds, with a worked sample-size example and the exact moments the sponsorship question lands in a data science loop.

As of November 1, 2026, most data scientist interview questions in US product and analytics loops fall into five rounds: SQL, statistics and probability, A/B testing and metrics, a product sense case, and behavioral, often with a Python screen or a take-home. If you are on F-1 OPT or STEM OPT, or heading for H-1B, a second question runs under every round: will the role's degree requirement, duties and wage level hold up for STEM OPT and an H-1B petition, and how do you answer sponsorship when a data science lead, not HR, asks?

Status as of 2026-11-01. The STEM OPT 180-day automatic extension is in effect, unaffected by the 2026-10-14 EAD rule. OPT unemployment limits (90 days, 150 with STEM OPT) are in effect. Wage-weighted H-1B selection is in effect since 2026-02-27; FY2028 registration is not announced. The $100,000 H-1B payment is blocked and not collected. The OPT fee rule is proposed; comments close 2026-11-09, and the close of comments does not make it law.

What are the rounds in a data scientist interview?

A typical product or analytics data scientist loop runs a recruiter screen, a technical screen in SQL or Python, then four or five final interviews: SQL, statistics and probability, A/B testing and metrics, a product case, and behavioral. Some teams add a take-home. Visa status usually comes up at the recruiter screen and again with the hiring manager.

Data scientist interview rounds and the visa moment in each (as of 2026-11)

Machine-learning-heavy roles add modeling rounds; the visa moments stay the same.

SQL and coding: data scientist interview questions you will see

SQL rounds test whether you can turn a vague business question into a correct query against messy tables: joins, GROUP BY with HAVING, window functions such as ROW_NUMBER and LAG, CTEs, dates and NULLs. The usual failure is not syntax; it is writing before you know what one row represents.

For "compute day-7 retention by signup week," state your assumptions first: signup is a user's first event, "retained" means active on exactly day 7, timestamps are UTC. Then (Postgres syntax):

WITH firsts AS (
  SELECT user_id, MIN(event_ts::date) AS signup_date
  FROM events
  GROUP BY user_id
)
SELECT DATE_TRUNC('week', f.signup_date) AS signup_week,
       COUNT(DISTINCT f.user_id) AS signups,
       COUNT(DISTINCT CASE WHEN e.event_ts::date = f.signup_date + 7
                           THEN e.user_id END)::numeric
         / COUNT(DISTINCT f.user_id) AS day7_retention
FROM firsts f
LEFT JOIN events e ON e.user_id = f.user_id
GROUP BY 1
ORDER BY 1;

Then volunteer the flaw: the latest signup week has not reached day 7, so its retention is understated. Exclude incomplete cohorts, and say why.

Data scientist coding interview questions in Python

Product data science screens are usually pandas and plain logic: group and merge data frames, compute a rolling average, bootstrap a confidence interval, simulate a probability puzzle. State your approach's cost and test it on a three-row example by hand.

For "each user's 7-day average of minutes watched," a row-based rolling(7) averages the last seven rows, which is not seven days once a user skips a day. Say so, then use a time-based window (rolling('7D') on a datetime index) within each user. On a merge, check row counts before and after: a fan-out from a duplicated key is the pandas version of the SQL grain error above.

Statistics interview questions for data scientists

Statistics rounds check whether you can explain inference plainly: p-values, confidence intervals, Type I and II errors, power, the central limit theorem, common distributions, regression assumptions and multiple testing. Prepare one-sentence definitions:

• p-value: the probability of a result at least this extreme if there were truly no effect. Not the probability the null is true.
• 95% confidence interval: from a procedure that captures the true value in 95% of repeated samples. Not a 95% chance this interval contains it.
• Power: the probability of detecting an effect of a given size when it is real.

Expect a probability question in data science clothing. Say 1% of transactions are fraud, your model flags 90% of fraud and wrongly flags 5% of legitimate ones. Flagged fraud is 0.01 x 0.90 = 0.009; flagged legitimate is 0.99 x 0.05 = 0.0495; so 0.009 / 0.0585, about 15%, of flags are fraud. With a low base rate, most of a good detector's flags are still wrong.

How do I answer an A/B testing case question?

Answer in a fixed order: goal, metrics, design, size, risks, decision. The formula matters less than picking the right metric and knowing when not to trust a result.

1. Clarify the change, the users it targets and the business goal.
2. Name one primary metric and guardrails such as latency or cancellations.
3. Choose the unit of randomization (user, session, account) and flag network effects.
4. Size the test from the baseline, the minimum detectable effect (MDE), significance and power.
5. Check for sample ratio mismatch, novelty effects and peeking before reading results.
6. Decide in advance what ships, and what happens if the primary metric wins while a guardrail loses.

Worked example: minimum detectable effect and sample size

Say checkout converts at 10% and the team wants to detect a lift to 11%: an absolute MDE of 1 percentage point (10% relative). With two-sided 5% significance (z = 1.96) and 80% power (z = 0.84), the standard approximation is:

n per arm = 2 x (1.96 + 0.84)^2 x p(1 - p) / MDE^2
          = 2 x 7.84 x (0.10 x 0.90) / 0.01^2
          = about 14,100 users per arm (about 28,200 in total)

Calculators that use both arms' variances give about 14,750; either is fine if you say which formula you used. At about 4,000 eligible users a day you reach the sample in just over a week, so you would run two full weeks and see each weekday twice.

Sample size at a 10% baseline, 5% significance, 80% power and 4,000 users a day (illustrative inputs)

The line that earns credit: halving the MDE roughly quadruples the sample, because the MDE is squared in the denominator. If five weeks is too long, accept a larger MDE, choose a more sensitive metric, or reduce variance with a pre-period covariate (CUPED), and say which you would pick.

What is a product sense interview for data scientists?

A product sense interview tests whether you can define success for a product and reason about it with metrics, without code. The two common prompts are "how would you measure success for this feature?" and "this metric dropped; what happened?"

For success, go from goal to funnel to metrics: one north-star metric tied to the goal, supporting metrics along the funnel, and guardrails that catch harm elsewhere. For a drop, rule out boring causes first: a logging change, pipeline delay or new metric definition; then segment by platform, app version, country and new versus returning users; then separate internal causes (your launch or experiment) from external ones (a holiday, outage or competitor). End with what you would check next and what evidence would change your mind.

Structuring a metrics case answer as a non-native speaker

Say your structure before you start, then number each point: "I'll cover three things: the goal, the metrics, and the trade-offs." Signposting gives the interviewer a map, so a grammar slip never costs you the thread, and it buys thinking time. Write metric names in the shared doc as you say them, keep one idea per sentence ("The goal is retention. So the main metric is day-7 retention."), and check in at each transition. Our guide to STAR answers for non-native English speakers has more phrase patterns and a practice routine.

Data science behavioral interview questions

Behavioral rounds probe influence more than technique: an analysis that changed a decision; a time your result was wrong; a disagreement with a product manager about a metric; explaining an uncertain result to a senior leader. Build four or five STAR stories, each ending with the decision your work drove and a number from your own project.

For a take-home, a one-page summary a PM could act on beats a notebook of every chart; our take-home assignment guide covers scoping, the README and the debrief. Ask for the deadline up front: while you are unemployed on OPT, each day without employer information counts toward the 90-day limit (150 days in aggregate with STEM OPT, including days used on post-completion OPT), as Study in the States' unemployment counter page explains.

Do data scientist roles qualify for STEM OPT and H-1B?

Data scientist roles often can, but neither question turns on the job title. STEM OPT depends on your degree, duties that relate directly to it and the employer's compliance; H-1B depends on whether the job is a specialty occupation. Your DSO handles the first; the employer's immigration attorney handles the second.

STEM OPT: your degree and duties, not your title

The USCIS STEM OPT page (updated 01/30/2026) requires a degree on the DHS STEM Designated Degree Program List, an employer "enrolled in E-Verify", and a Form I-983 training plan the employer signs. It adds that the opportunity "must be directly related to the previously obtained STEM degree" and "commensurate with those of similarly situated U.S. workers in duties, hours, and compensation." So the I-983 has to explain how the analysis or modeling work uses your degree. Data science programs carry different CIP codes at different schools, so check yours on Study in the States' eligible CIP codes page and our list of which majors and CIP codes qualify for STEM OPT.

If you file on time and your OPT EAD expires while the extension is pending, USCIS "will automatically extend your employment authorization for 180 days." The 2026-10-14 final rule amended only 8 CFR 274a.13 and does not mention STEM OPT. A STEM I-765 filed after your OPT EAD expires gets no 180 days.

H-1B: the specialty-occupation test for data scientists

USCIS's H-1B specialty occupations page describes a job requiring "theoretical and practical application of a body of highly specialized knowledge" and a bachelor's or higher in a directly related specific specialty, meeting one of four criteria in 8 CFR 214.2(h)(4). For data scientists (SOC 15-2051), the Bureau of Labor Statistics says employers typically want "at least a bachelor's degree in mathematics, statistics, computer science, or a related field," and O*NET puts the occupation in Job Zone Four.

Whether a posting's degree language supports a petition is the attorney's question, not one to argue in the loop. Your part is confirming the duties you are interviewing for are the real ones, because DOL guidance says the occupation code "should not be based solely on the title of the employer's job offer."

What wage level does a "Data Scientist I" title land in?

There is no fixed answer: the level follows the job's requirements, not the "I" in the title. DOL's 2009 Prevailing Wage Determination Policy Guidance says "All prevailing wage determinations start with a Level I determination," then adds points when experience, education, special skills or supervision exceed the occupation's norm. For data scientists, whose O*NET Job Zone expects a bachelor's:

• A Data Scientist I posting asking for a bachelor's and the usual experience starts at Level I and may stay there.
• A master's requirement adds one point, per the guidance's own example ("If the occupation generally requires a Bachelor's degree and the employer's job offer requires a Master's degree, enter a 1"), pointing toward Level II. A PhD adds two.
• "Senior" or "lead" in a title are "indicators that a Level III wage should be considered."

That is the worksheet's logic, not a prediction; the employer and its counsel decide what goes on the Labor Condition Application. Pay also matters for the lottery: under wage-weighted selection (in effect since 2026-02-27), a registration is entered once at Level I and up to four times at Level IV, based generally on the highest OEWS wage level the offered wage meets for the occupation and area, per USCIS's registration page. Level I sits at the 17th percentile of local wages today; DOL's prevailing-wage proposal (RIN 1205-AC30) would raise it to the 34th but is still proposed as of 2026-11-01. See our explainer on H-1B wage levels I through IV.

To see real filings, use DOL's OFLC data. Its FY2026 Q3 Selected Statistics report 426,659 LCA applications received (down 8.8%), with software developers at 29.1% of certified positions; data scientist filings sit in the quarterly disclosure files. Ask about wage level at the offer stage with HR, not mid-loop with the data science lead.

How to answer sponsorship when the interviewer is a data science lead

A data science lead rarely decides sponsorship; they want to know you can start on time and will be there when the roadmap lands. Answer that in two or three sentences, then return to the work. If every part is true for you: "I'm on F-1 OPT, and my degree is STEM-designated, so I plan to apply for the 24-month STEM OPT extension, which needs an E-Verify employer. Longer term I'd need H-1B sponsorship; your recruiting or immigration team usually handles that, and I'm happy to give them details."

Follow-ups a lead might raise:

• "Doesn't H-1B cost $100,000 now?" Not as of 2026-11-01: courts blocked the agency policies, though Proclamation 11069 stands on paper through 2027-09-21, and the USCIS I-129 page warns, "If this order is later lifted, DHS still plans to collect the payment." Changes of status USCIS grants inside the US were never covered; a denied change of status or a consular-notification petition would be if collection resumed. A separate $103,265 fee on cap-subject petitions (RIN 1615-AD20) is only proposed, with no final rule as of 2026-11-01. Suggest looping in the immigration team.
• "Isn't OPT going to cost $70,000?" That is a proposed rule, not law. The school would be the legal payer and could pass the cost on to students or employers: $70,000 before the initial OPT recommendation and $30,000 before each later one, STEM OPT included. It would take effect 60 days after a final rule; OPT already approved or recommended by a DSO before then is not covered, but a later request after that date would be, and the text leaves open whether a student whose first OPT was fee-free would owe $70,000 or $30,000. Ask your DSO how your school reads it.
• "What if you aren't selected in the lottery?" Be honest that selection is never guaranteed and the FY2028 registration period is not announced; USCIS must give at least 30 days' notice (8 CFR 214.2(h)(8)(iii)(A)(3)).

If someone says "we don't sponsor" halfway through, see what to say when a recruiter mentions no sponsorship mid-interview. To find teams that do, start with data science H-1B sponsorship in 2026.

How to check a role's STEM OPT and H-1B fit before the final round

1. Find the CIP code on your Form I-20 and look it up on Study in the States' eligible CIP codes page; ask your DSO if it is unclear.
2. Search the employer's legal name in the E-Verify Employer Search, which matches only the name given at enrollment.
3. Download the latest LCA disclosure file (FY2026 Q3, through 2026-06-30) from DOL's OFLC performance page and filter by employer and SOC 15-2051 for titles, wages and levels.
4. Check the employer in the USCIS H-1B Employer Data Hub, which now says it covers through FY2026 Q4.
5. Ask the recruiter, not the data science lead, whether the team has sponsored data scientists and who handles immigration.
6. Bring the offer to your DSO for the Form I-983 conversation, and send H-1B questions to the employer's immigration attorney.

Common mistakes

• Writing SQL before asking what one row means. Grain errors produce confident wrong answers.
• Calling a p-value "the chance the result is real." It assumes no effect.
• Extending a test after peeking at results, or stopping when it first looks significant. Both inflate false positives.
• Listing ten metrics in a product case. Pick one north star and defend it.
• Saying you will never need sponsorship when you will need H-1B later. Say what is true now and what you will need.
• Assuming a "Data Scientist" title decides STEM OPT eligibility or wage level. The degree, duties and requirements decide.
• Debating immigration law with the hiring manager. Give the short version; route the rest to the immigration team.

Frequently asked questions

What are the rounds in a data scientist interview?

Most product and analytics data scientist loops start with a recruiter screen and a SQL or Python screen, then four or five final interviews covering SQL, statistics and probability, A/B testing and metrics, a product case and behavioral questions. Some teams add a take-home. On OPT or STEM OPT, expect the work-authorization question at the recruiter screen and again with the hiring manager.

How do I answer an A/B testing case question?

Clarify the goal, name one primary metric plus guardrails, choose the unit of randomization, then size the test from the baseline rate, the minimum detectable effect, 5% significance and 80% power. For a 10% baseline and a 1-point lift that is about 14,100 users per arm. Close with checks for sample ratio mismatch and peeking, and a ship rule set in advance.

What is a product sense interview for data scientists?

It tests whether you can define success for a product and reason about it with metrics, without writing code. Typical prompts ask how you would measure a new feature or why a key metric dropped. Strong answers clarify the goal, map the funnel, pick one north-star metric with guardrails, and end with a decision.

Do data scientist roles qualify for STEM OPT and H-1B?

Neither turns on the job title. STEM OPT depends on your degree's CIP code being on the DHS STEM list, duties directly related to that degree, an E-Verify employer and a signed Form I-983, and the 180-day automatic extension for timely filings is still in effect as of 2026-11-01. For H-1B the job must be a specialty occupation; BLS says data scientists typically need at least a bachelor's in mathematics, statistics, computer science or a related field, and the employer's immigration attorney builds the case.

How should a non-native speaker structure a metrics case answer?

Announce your structure first, for example the goal, the metrics and the trade-offs in that order, then number each point as you go. Write metric names in the shared doc, keep to short sentences, and check the interviewer agrees before moving on. A clear structure carries your reasoning even when a sentence comes out imperfectly.

Sources

• USCIS: STEM OPT extension, H-1B Specialty Occupations, Form I-129, H-1B Electronic Registration Process, H-1B Employer Data Hub
• DOL: OFLC Performance Data, Prevailing Wage Determination Policy Guidance (November 2009)
• Federal Register: Weighted Selection (2025-12-29), Improving Wage Protections, proposed, OPT Fees, proposed, Fee for Certain H-1B Petitions, proposed, Removal of Automatic EAD Extensions
• Study in the States: Eligible CIP Codes for STEM OPT, Unemployment Counter
• E-Verify Employer Search, BLS Occupational Outlook Handbook: Data Scientists, O*NET 15-2051.00

How we checked this: we read the USCIS, DOL, DHS, Federal Register, BLS and O*NET pages above on 2026-11-01; the sample-size figures are our arithmetic on illustrative inputs.

This is general information, not legal or tax advice; your own I-20, EAD and USCIS notices, your DSO (for OPT and STEM OPT) and the employer's immigration attorney (for H-1B) govern your situation.

If you want help finding data science teams that sponsor and getting ready for their loops, talk to F1Jobs.