Articles
Data Analyst vs Data Scientist: Which Role Should You Prepare For?
Compare data analyst vs data scientist work through job descriptions, retention examples, and practice questions. Choose the interview skills to prepare first.
Data analyst vs data scientist is best understood through the work you want to own. Do you want to explain a business metric, decide whether a change caused an improvement, or build a prediction used in a product? Those responsibilities overlap, but they lead to different preparation priorities. A job title alone cannot tell you which one an employer needs.
You do not need to prepare for every branch equally.
This guide combines official job descriptions checked on September 12, 2026, with an original retention scenario. The practice recommendations are editorial judgments, not a universal interview sequence. No candidate reports are used to claim how these employers currently interview.
Data Analyst vs Data Scientist: Compare the Deliverable
A useful starting distinction is that an analyst often makes an existing business process measurable and understandable. A scientist may additionally own statistical inference or a predictive system. However, neither boundary is exclusive: analysts can run experiments, and scientists can spend substantial time writing SQL and communicating findings.
Official occupational descriptions support that breadth. O*NET’s Data Scientists profile includes preparing data, evaluating analytical methods, visualizing results, and communicating recommendations. It is a US occupational reference, not a definition of every employer’s title.
The practical difference is the question your work must survive. A reporting result must survive checks of its denominator and data quality. An experimental conclusion must survive challenges to its causal interpretation. A model must survive evaluation on information it could actually have used when making a prediction. All three require technical reasoning and clear explanations.
Avoid treating data science as the inevitable next rung after analytics. Becoming a stronger analyst can mean owning harder decisions, improving measurement across teams, or developing domain expertise. Moving into modeling changes the work you must defend; it does not automatically make the work more valuable to a particular employer.
What Three Official Job Descriptions Actually Show
These are a small sample of official role descriptions, not a survey of the labor market. The Contact Automation posting showed its full description on initial retrieval but displayed an expired notice on a later same-day check; it is a dated responsibility example, not an active vacancy. They are experienced roles, so their requirements should not be copied into an entry-level checklist.
Amazon’s Data Analyst, Global Tax Services, in Seattle, combines data-quality analysis, tax reporting, requirements, and stakeholder recommendations. Its basic qualifications include analytical experience and substantial Excel experience; SQL appears under preferred qualifications. That is a useful warning against assuming that every analyst vacancy is primarily a SQL coding job. This particular position also has domain requirements. Official Amazon posting.
Wise’s Lead Data Scientist — Pricing, in London, spans pricing experiments, causal inference, forecasting, reporting automation, and data pipelines. Its responsibilities cross the tidy categories in many career comparison charts. If you pursued similar work, preparing only model algorithms would leave the pricing-decision and experiment-design responsibilities uncovered. Official Wise pricing posting.
Wise’s Lead Data Scientist — Contact Automation, also in London, emphasizes evaluation of LLM systems, online experiments, failure analysis, product outcomes, and collaboration with engineers. It explicitly asks for production coding experience in Python, TypeScript, or Java. “Data scientist” here includes making an AI system useful and reliable, not merely producing an offline score. Official Wise contact-automation posting.
The preparation implication is an inference from those responsibilities: inspect verbs and outputs before buying another course. None of these job descriptions establishes a fixed assessment format, interview duration, or guaranteed question type.
One Retention Problem, Three Different Deliverables
Consider a fictional language-learning app. Its team asks: “Why are fewer new learners returning, and what should we do?” The following numbers are synthetic teaching inputs, not a company case or an observed experiment.
Define D7 retention as the percentage of signed-up learners with at least one qualifying lesson event on the seventh calendar day after signup, using UTC dates. Count each learner once. Include only cohorts whose entire D7 date has finished by the reporting cutoff.
Suppose an earlier mature cohort contains 1,000 learners, with 420 returning on D7. A later mature cohort contains 1,000 learners, with 350 returning. Retention is 42% versus 35%: a decline of 7 percentage points, or approximately 16.7% relative to the earlier rate. Those are different descriptions of the same change.
Deliverable one: an analyst’s diagnostic memo
The analyst first checks whether the decline is measurable. Did an event name change? Are lesson completions delayed? Did the acquisition mix shift? Are both cohorts mature? A chart built before those checks can turn a logging change into a false product narrative.
The useful output is a reproducible cohort query, a segmented result, and a short recommendation. For example: “The measured decline is concentrated in learners using the new Android onboarding flow. Validate event coverage before testing an onboarding change.” That is a hypothesis and next action, not proof that the flow caused the decline.
A good interview explanation names the grain of each table before joining it. Signup data may contain one row per learner, while lesson data contains many rows per learner. Counting joined rows would overstate retained learners. State how you deduplicate qualifying activity and how you check that every eligible learner remains in the denominator.
Deliverable two: a product scientist’s experiment decision
Now suppose the team proposes a reminder. The product scientist defines who is eligible, what is randomized, and which outcome will determine whether the reminder helps. Randomizing eligible learners before reminder delivery lets the team compare assigned groups; selecting only people who opened the reminder would answer a different question.
Imagine a separate synthetic test with 400 D7 returners among 1,000 assigned control learners and 430 among 1,000 assigned treatment learners. The estimated difference is 3 percentage points. Under a simple independent Bernoulli approximation, the standard error is about 2.2 percentage points, producing an approximate 95% interval from −1.3 to 7.3 percentage points.
That interval is deliberately wide. The point estimate alone does not justify declaring a reliable improvement. The decision memo should discuss uncertainty, the minimum worthwhile effect, delivery failures, and guardrails such as notification opt-outs. This arithmetic illustration is not a complete experiment analysis or a stopping rule.
The scientist’s output is therefore a decision with assumptions and uncertainty. It cannot be replaced by the analyst’s segment chart, even though both use retention data.
Deliverable three: a modeling scientist’s prediction system
A third proposal is to prioritize support for learners likely not to return. Define a scoring time first: for example, the end of the first 24 hours after signup. Features must be available then. A variable recording whether someone completed a lesson on D7 would reveal the outcome and make the evaluation misleading.
The output could be a ranked list plus an evaluation report. Compare it with a simple baseline, evaluate on a later cohort, and ask whether performance remains acceptable across relevant segments. If the team can contact only 100 learners, evaluation at that contact capacity may be more useful than an untargeted overall accuracy number.
Crucially, predicting non-return does not identify who benefits from a reminder. The highest-risk learner might be unreachable or unaffected by the intervention. Choosing whom to help introduces an intervention question, which may require experimental evidence. A model ranking and an estimated treatment effect are different deliverables.
Read a Job Description Beyond Its Title
Use a short worksheet for each vacancy you are seriously considering. Copy three responsibility phrases, write the artifact each implies, then name the uncertainty you still need to resolve. Keep required qualifications separate from preferences.
| Responsibility you notice | Practice output to prepare |
|---|---|
| Define KPIs, reconcile reports, investigate changes | A checked query and a diagnostic recommendation |
| Design experiments, estimate impact, advise launches | A treatment definition and an uncertainty-aware decision |
| Train, validate, score, or monitor models | A baseline comparison and leakage-safe evaluation |
| Build pipelines or maintain production services | A tested implementation and a failure-recovery explanation |
Here are three original annotations you can adapt. “Own weekly retention reporting” suggests practicing cohort definitions and data checks; ask who decides the metric. “Evaluate onboarding experiments” suggests practicing assignment, outcomes, and decision thresholds; ask whether the role designs experiments or analyzes ones designed elsewhere. “Deploy churn scores” suggests practicing feature availability, serving, and monitoring; ask who owns production reliability.
These annotations are not an automated career score. A phrase can be ambiguous. “Build models” could mean financial spreadsheets, statistical estimation, or deployed machine learning. Ask the recruiter: “What would a successful first project produce, and which part will the interview ask me to demonstrate?” A concrete artifact often clarifies the role faster than another discussion of titles.
Also check the level and domain. Experience with a classroom dataset does not establish readiness to lead regulated reporting or own a production system. You can still identify a transferable skill without claiming the full responsibility already belongs on your résumé.
Choose Practice That Reveals the Missing Skill
The same person may be comfortable writing a query and uncomfortable defending its business meaning. Another may explain statistical uncertainty clearly but struggle to construct a reliable dataset. Practice should expose that gap rather than repeat your strongest skill.
Start with the item that matches your target deliverable and explain your assumptions before checking an answer.
| Practice question | What to demonstrate |
|---|---|
| Define and Analyze Early Engagement | Turn a broad product goal into a defensible measure |
| Build cohort 30-day retention from signup date | Preserve cohort eligibility and distinct-user counting |
| Recommend a Decision After a Nonsignificant Experiment | Explain what uncertainty permits you to recommend |
| Apply Leakage-Safe Imputation and Ordinal Encoding | Keep evaluation data out of preprocessing decisions |
| Resolve a Mismatch Between Regression and Classification Metrics | Align the target, model output, and evaluation metric |
For an analytics-heavy role, complete one metric investigation end to end before adding more algorithms. For an experimentation role, explain why a plausible comparison is biased and how you would improve it. For a modeling role, show why the evaluation matches the future decision, including the timing of every feature.
Official interview guidance can refine this allocation when available. CZI’s published Data Science Interview Guide, whose URL is dated February 2025, explicitly prefers SQL for its Product Analytics track and discusses analytical reasoning and communication. It illustrates why “DS interview” does not automatically mean only machine learning. Your own current invitation takes precedence. CZI guide.
Choose a Direction Without Starting Over
If you already have analyst experience, preserve your evidence of data quality, domain judgment, SQL, and communication. Add the missing responsibility deliberately: an experiment decision for product science, or a properly evaluated prediction task for modeling. Renaming a dashboard a “data science project” does not demonstrate that additional work.
If you have a modeling background and prefer analytics roles, show that you can simplify. Explain a metric, detect a misleading denominator, and recommend a useful action without requiring a model. The ability to stop at a sufficient analysis is relevant evidence too.
Choose one target vacancy and write a single sentence: “This team needs someone to produce ___, and my strongest evidence is ___.” Then identify one gap you can demonstrate through a small, reviewable exercise. If the sentence remains vague, return to the responsibilities or ask about the first project.
To build relevant experience while narrowing that direction, you can also explore volunteer data analyst opportunities on Jooble that may let you practice data quality, cohort reporting, and stakeholder communication through real-world work.
For your next practice session, open the Data Scientist question collection, choose the branch that matches that sentence, and explain one complete answer aloud. The aim is a defensible work sample for a specific responsibility, not mastery of every topic carrying a data title.