McKinsey & Company Data Analyst Interview Guide

Most candidates prep for McKinsey's Data Analyst interview the way they'd prep for a tech company loop. Then they hit a round where they're asked to structure a healthcare cost problem from scratch and recommend what to tell a hospital CFO. The candidates who wash out here aren't weak technically. They never practiced framing analytical work the way McKinsey's consulting teams actually consume it.

McKinsey & Company Data Analyst Role

You're not shipping dashboards to a product manager. In McKinsey's Health Systems & Services practice, where DA hiring has been concentrated, you're building data marts from claims extracts in Databricks and packaging cost-driver analyses into exhibits that a partner presents to a payer's VP of Network Strategy. Success after year one looks like owning an analytics workstream end-to-end on a client engagement, from wrangling messy provider taxonomy files through to the steering committee slide that changed a decision.

A Typical Week

What will surprise you is the reactive texture of the analytical blocks. Your calendar might show three hours of focus time, but an associate's Slack message can redirect your afternoon toward an entirely different dataset with a next-morning deadline. The writing-heavy split also catches people off guard, because "writing" at McKinsey means translating model outputs into exhibits that meet the firm's notoriously precise chart standards, not drafting Jira tickets.

Projects & Impact Areas

Claims data reconciliation is the unsexy foundation: deduplicating provider records, resolving NPI mismatches, and standardizing classification files before any real analysis can begin. That pipeline work, often orchestrated in Airflow with Databricks as the compute layer, feeds directly into the client-visible deliverables like propensity models for patient segmentation or enrollment forecasts prototyped with Prophet for scenario planning. McKinsey's published research on AI-powered customer interactions reflects the kind of applied ML you'd contribute to, where model selection and business framing matter more than novel architectures.

Skills & What's Expected

The skill scores show ML rated high, but don't confuse that with research-depth machine learning. McKinsey's version means picking the right regression or decision tree for a client problem and explaining the tradeoff to a partner who doesn't know what regularization is. The underrated dimension is communication: you're producing exhibits for C-suite healthcare executives, and candidates who can write complex Spark jobs but can't storyline a deck plateau fast.

Levels & Career Growth

Most external hires land at Analyst or Senior Analyst. What separates Senior Specialist from Specialist isn't technical chops; it's whether you can independently shape an engagement's analytical direction and manage stakeholder expectations without your engagement manager scaffolding every step. The promotion blocker worth knowing: McKinsey's headcount reductions reported in 2025 have made the path to Expert (Principal-equivalent) less predictable, and reaching that tier from what candidates report requires a sellable niche like healthcare analytics or GenAI strategy that partners can repeatedly staff you against.

Work Culture

McKinsey's "obligation to dissent" culture means you're expected to push back on a partner when your data contradicts their hypothesis, which is a genuinely different dynamic than most corporate DA seats where you confirm what leadership already believes. The intensity is spiky rather than constant: engagement deadlines can push weeks past 55 hours, but bench periods between studies offer real breathing room. On location, the firm's hybrid stance varies by practice and engagement manager, and some engagements may require client-site travel, so if you're assuming fully remote analytical work, recalibrate early in the recruiter screen.

McKinsey & Company Data Analyst Compensation

Because McKinsey is a private partnership, stock grants generally don't exist for data analyst roles. That shifts the math: your bonus becomes the only variable component, and at senior levels it carries real weight. The spread between floor and ceiling TC widens as you climb precisely because individual ratings and firm performance drive that bonus number up or down each year.

The single biggest negotiation lever most candidates overlook is the sign-on bonus. Base salary has some flex within a band (particularly if you're holding a competing offer), but sign-on is where recruiters have the most room to close you. Put your counter in writing, keep it to one or two specific asks, and anchor on scope: the client-facing travel expectations, the SQL/Python/experimentation skill set the role demands, and the fact that you'd be owning data engineering work alongside analysis. Location or office alignment is another lever worth raising if it matters to you.

McKinsey & Company Data Analyst Interview Process

Budget about six weeks from application to offer. Gaps of 1-2 weeks between rounds are common because your interviewers are staffed on live client engagements at hospitals, banks, or insurers, and their availability bends to client timelines, not recruiting timelines. The rejection reason candidates report most often is unstructured problem solving: jumping into the McKinsey Solve or the SQL round without framing the problem first, which in a firm that puts analysts in front of healthcare C-suites is a dealbreaker.

From what candidates describe, no single round can carry you. A strong SQL & Data Modeling session won't offset a flat behavioral showing, because McKinsey's PEI-style final round is specifically designed to test whether you can own a client-facing analytics narrative, not just write correct queries. If you only prep the technical dimensions, you're leaving the rounds that assess consulting judgment completely uncovered.

McKinsey & Company Data Analyst Interview Questions

SQL & Data Manipulation

Expect questions that force you to translate messy payments/product prompts into correct SQL under time pressure. You’ll be evaluated on joins, window functions, cohorting, and debugging logic to produce decision-ready tables.

For each listing, compute the trailing 28-day booking revenue, excluding the current day, and return the top 50 listings by that metric for yesterday. Bookings can be refunded, so use net revenue per booking.

AirbnbMediumWindow Functions and Time Windows

Sample Answer

Compute daily net revenue per listing, then sum it over the prior 28 days using a date-based window that excludes the current day. You avoid double counting by aggregating to listing-day before windowing, then filtering to yesterday at the end. Use $[d-28, d-1]$ as the window, not 28 rows, because missing days exist. Net revenue should incorporate refunds at the booking level before the listing-day rollup.

SQLCopy

1WITH booking_net AS (
2  SELECT
3    b.booking_id,
4    b.listing_id,
5    DATE(b.booking_ts) AS booking_day,
6    COALESCE(b.gross_amount_usd, 0) - COALESCE(b.refund_amount_usd, 0) AS net_amount_usd
7  FROM bookings b
8  WHERE b.status IN ('confirmed', 'completed', 'refunded')
9),
10listing_day AS (
11  SELECT
12    listing_id,
13    booking_day,
14    SUM(net_amount_usd) AS net_revenue_usd
15  FROM booking_net
16  GROUP BY 1, 2
17),
18scored AS (
19  SELECT
20    listing_id,
21    booking_day,
22    SUM(net_revenue_usd) OVER (
23      PARTITION BY listing_id
24      ORDER BY booking_day
25      RANGE BETWEEN INTERVAL '28' DAY PRECEDING AND INTERVAL '1' DAY PRECEDING
26    ) AS trailing_28d_net_revenue_excl_today_usd
27  FROM listing_day
28)
29SELECT
30  listing_id,
31  trailing_28d_net_revenue_excl_today_usd
32FROM scored
33WHERE booking_day = CURRENT_DATE - INTERVAL '1' DAY
34ORDER BY trailing_28d_net_revenue_excl_today_usd DESC NULLS LAST
35LIMIT 50;

You need host-level cancellation rate for the last 90 days, where the numerator is guest-initiated cancellations and the denominator is all bookings that reached confirmed status. Hosts can have multiple listings, and booking status changes are tracked in an events table with one row per status transition.

AirbnbHardEvent Log Deduping and Conditional Aggregation

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more SQL & Data Manipulation questions

Product Sense & Metrics

The bar here isn’t whether you know a metric name—it’s whether you can structure an analysis plan that maps to decisions. You’ll need to define success, identify leading vs lagging indicators, and anticipate confounders and data limitations.

How would you define and choose a North Star metric for a product?

EasyFundamentals

Sample Answer

A North Star metric is the single metric that best captures the core value your product delivers to users. For Spotify it might be minutes listened per user per week; for an e-commerce site it might be purchase frequency. To choose one: (1) identify what "success" means for users, not just the business, (2) make sure it's measurable and movable by the team, (3) confirm it correlates with long-term business outcomes like retention and revenue. Common mistakes: picking revenue directly (it's a lagging indicator), picking something too narrow (e.g., page views instead of engagement), or choosing a metric the team can't influence.

Outbound delivery speed for the company Logistics improved from 2.3 to 2.1 days, but CS contacts per 1,000 orders increased by 12% in the same period. You have order, shipment scan, and contact reason data, propose a metric framework to diagnose whether the speed win is causing the contact increase.

AmazonMediumMetric Decomposition and Customer Contacts

Sample Answer

You could do cohort decomposition by promised speed buckets or you could do contact rate standardization holding mix constant. Cohort decomposition wins here because it exposes which promise and ship-method segments changed behavior, and whether contacts spike specifically where speed improved. Standardization is faster but can hide operational failure modes like more partial shipments or missed promised windows. You want both eventually, but the cohort view answers causality-adjacent questions with fewer assumptions.

A company reduces the guest service fee by 1 percentage point in 5 countries, and Finance wants a metric tree that separates demand lift from margin impact and host behavior changes. Propose the primary success metric, the decomposition you would show (with formulas), and 2 guardrails that prevent gaming or long-run supply damage.

AirbnbHardMetric Tree and Unit Economics

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Product Sense & Metrics questions

A/B Testing & Experiment Design

What is an A/B test and when would you use one?

EasyFundamentals

Sample Answer

An A/B test is a randomized controlled experiment where you split users into two groups: a control group that sees the current experience and a treatment group that sees a change. You use it when you want to measure the causal impact of a specific change on a metric (e.g., does a new checkout button increase conversion?). The key requirements are: a clear hypothesis, a measurable success metric, enough traffic for statistical power, and the ability to randomly assign users. A/B tests are the gold standard for product decisions because they isolate the effect of your change from other factors.

You run an experiment on the guest cancellation flow and randomize by user_id, but a guest can book multiple trips and see both variants across devices. How do you detect and quantify interference, and what changes to the design or analysis would you make?

AirbnbMediumSUTVA violations and unit of randomization

Sample Answer

Reason through it step by step as if thinking out loud. Start by checking exposure logs, does the same guest_id have multiple variant assignments across sessions or devices, and what share of traffic is affected. Next, quantify bias risk by comparing outcome rates for cross-exposed users versus clean users, and by re-estimating the lift on the clean subset to see how sensitive the result is. Then pick a fix: enforce sticky assignment at the guest_id level with consistent identity stitching, or move randomization to a higher level like booking_id if the treatment is per booking. If you cannot fix assignment, analyze with cluster-robust standard errors at the guest_id level, or switch to a design like switchback only if the interference is time-based rather than user-based.

A company runs 8 simultaneous experiments on the host pricing page, and your experiment shows $p = 0.03$ on booking conversion and $p = 0.20$ on contribution margin. How do you decide whether this is a real win, and what correction or validation would you apply?

AirbnbHardMultiple testing and ambiguous readouts

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more A/B Testing & Experiment Design questions

Statistics

Most candidates underestimate how much applied stats shows up in fraud analytics, from thresholding to false-positive tradeoffs. You’ll need to reason clearly about distributions, sampling bias, and how to validate signals with limited labels.

What is a confidence interval and how do you interpret one?

EasyFundamentals

Sample Answer

A 95% confidence interval is a range of values that, if you repeated the experiment many times, would contain the true population parameter 95% of the time. For example, if a survey gives a mean satisfaction score of 7.2 with a 95% CI of [6.8, 7.6], it means you're reasonably confident the true mean lies between 6.8 and 7.6. A common mistake is saying "there's a 95% probability the true value is in this interval" — the true value is fixed, it's the interval that varies across samples. Wider intervals indicate more uncertainty (small sample, high variance); narrower intervals indicate more precision.

A company Logistics changed a routing rule and late deliveries dropped from $2.4\%$ to $2.1\%$ over 14 days, but shipment volume also increased and the mix shifted toward longer-distance lanes. How do you estimate whether the routing change reduced late deliveries, and which statistical model or adjustment would you use?

AmazonMediumConfounding and Adjusted Comparisons

Sample Answer

Walk through the logic step by step as if thinking out loud. Start by defining the outcome as a binary label per shipment (late or not) and list the main confounders, lane distance, carrier, promised speed, weather, day-of-week, and volume. Then fit a logistic regression with a treatment indicator for the routing rule change plus those covariates (and ideally lane fixed effects) so you compare like with like. Validate by checking pre-change trends by lane, then report an adjusted lift with a CI, not just the raw $2.4\%$ to $2.1\%$ delta.

An AWS Console UI experiment shows a $+1.2\%$ lift in weekly active users, but the metric has heavy-tailed session counts and the variance doubled during the test. How do you decide whether to ship, and what statistical technique would you use to make the result decision-ready?

AmazonHardVariance, Heavy Tails, and Robust Inference

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Statistics questions

Data Modeling

When you design tables for analytics, you’re being tested on grain, keys, and how modeling choices impact BI performance and correctness. Expect star schema reasoning, fact/dimension tradeoffs, and how you’d model common product/usage datasets.

An ETL job builds fct_support_interactions from Zendesk tickets, chat transcripts, and on-chain deposit events, and you notice a sudden 12% drop in interactions after a schema change in chat. What data quality checks and pipeline safeguards do you add so this does not silently ship to dashboards again?

CoinbaseMediumETL Monitoring, Data Quality

Sample Answer

Get this wrong in production and your CX dashboards underreport demand, staffing and SLA decisions get made on fake stability. The right call is to add volume and freshness checks (row count deltas by source, max event timestamp lag), completeness checks on required keys (ticket_id, interaction_id, user_id), and distribution checks on critical dimensions (channel, product surface). Gate the publish step with alerting and fail-closed thresholds, plus backfill logic and schema versioning so a renamed field cannot null out a join unnoticed.

A company wants a single "gross bookings" metric used by Finance and Product, but your model has cancellations, modifications, partial refunds, and multiple payment captures per reservation. How do you model facts and keys so that gross bookings, net bookings, and revenue can be computed without double counting across these flows?

AirbnbHardFact Modeling, Double Counting, Payments and Reservations

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Data Modeling questions

Visualization

When dashboards become the source of truth, small choices in charting and narrative can change decisions. You’ll be tested on picking the right visual, communicating insights to non-technical stakeholders, and proposing actionable next steps.

A Tableau dashboard for the company Retail shows conversion rate by store, but the VP wants stores ranked and "actionable" by tomorrow. What is your default chart and sorting approach, and what adjustment do you make to avoid overreacting to small-sample stores?

AppleMediumRanking, Variability, and Visualization Choice

Sample Answer

The standard move is a ranked bar chart of conversion with a reference line for the fleet median, plus a small table for traffic and transactions. But here, sample size matters because $n$ varies wildly by store, so the ranking is mostly noise for low-traffic locations. You either filter to a minimum volume threshold or plot a funnel chart (conversion versus sessions) with confidence bands, then call out only statistically stable outliers for action.

You ship an exec dashboard for iOS crash rate by build, but a new build rollout causes an apparent crash-rate jump. How do you redesign the dashboard so leadership can tell whether the build is worse versus the user mix changing due to staged rollout?

AppleHardCohorting, Segmentation, and Narrative Design

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Visualization questions

Data Pipelines & Engineering

In practice, you’ll be asked how you keep reporting accurate when pipelines break or definitions drift. Strong answers cover validation checks, anomaly detection, backfills, idempotency, and communicating data incidents to stakeholders.

What is the difference between a batch pipeline and a streaming pipeline, and when would you choose each?

EasyFundamentals

Sample Answer

Batch pipelines process data in scheduled chunks (e.g., hourly, daily ETL jobs). Streaming pipelines process data continuously as it arrives (e.g., Kafka + Flink). Choose batch when: latency tolerance is hours or days (daily reports, model retraining), data volumes are large but infrequent, and simplicity matters. Choose streaming when you need real-time or near-real-time results (fraud detection, live dashboards, recommendation updates). Most companies use both: streaming for time-sensitive operations and batch for heavy analytical workloads, model training, and historical backfills.

You need a trustworthy daily metric for App Store subscriptions that powers Finance reporting and product dashboards, and events can arrive up to 72 hours late. How do you design the warehouse tables and the incremental rebuild logic so the metric is both stable and correct?

AppleMediumLate Arriving Data and Incremental Loads

Sample Answer

Start with what the interviewer is really testing: "This question is checking whether you can keep a metric stable for stakeholders while still correcting for late data." You typically separate an immutable raw events layer from a curated fact layer keyed by a stable business grain (for example, subscription, day, storefront), and define a rolling recompute window of the last $N$ days where $N \ge 3$ to absorb late arrivals. You also need a clear definition of when the metric is final (for example, a watermark), plus audit columns like load_timestamp and source_max_event_time so you can explain changes and meet SLAs.

An Airflow DAG builds a daily fact table for payouts to hosts, partitioned by payout_date, and finance reports missing payouts for a two week window after a backfill. How do you design the backfill and data quality safeguards so you avoid double counting, preserve idempotency, and keep downstream Superset dashboards stable?

AirbnbHardBackfills, Idempotency, and Data Quality Gates

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Data Pipelines & Engineering questions

Causal Inference

What is the difference between correlation and causation, and how do you establish causation?

EasyFundamentals

Sample Answer

Correlation means two variables move together; causation means one actually causes the other. Ice cream sales and drowning rates are correlated (both rise in summer) but one doesn't cause the other — temperature is the confounder. To establish causation: (1) run a randomized experiment (A/B test) which eliminates confounders by design, (2) when experiments aren't possible, use quasi-experimental methods like difference-in-differences, regression discontinuity, or instrumental variables, each of which relies on specific assumptions to approximate random assignment. The key question is always: what else could explain this relationship besides a direct causal effect?

Hulu ad load was reduced for a subset of DMAs, but advertisers also shifted budgets toward those same DMAs mid-flight due to a sports schedule. You need the causal effect of ad load reduction on ad revenue per hour, do you use a geo-based diff-in-diff or an instrumental variables approach, and why?

DisneyMediumGeo Experiments and Instrumental Variables

Sample Answer

You could do geo diff-in-diff or instrumental variables (IV). Geo diff-in-diff is simpler, but it loses credibility here because budget reallocation creates time-varying confounding correlated with treatment assignment. IV wins if you have a plausibly exogenous instrument for ad load, for example a policy or capacity constraint that shifts ad load but does not directly shift demand, then you estimate $\text{AdLoad}\rightarrow\text{RevenuePerHour}$ via 2SLS and defend exclusion and relevance.

A company runs a retargeting campaign for the company+ lapsed subscribers, but exposure is highly selective because it targets users with high predicted return probability. How do you design a quasi-experiment to estimate incremental resubscription lift, and what diagnostics convince you the estimate is not driven by selection bias?

DisneyHardSelection Bias and Quasi-Experimental Design

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Causal Inference questions

The weight toward pipeline and query work (Data Engineering/ETL plus SQL Analytics) signals that McKinsey's DA loop is structured to find people who can own the full path from raw claims ingestion in Databricks to polished metrics a partner can present at a client steering committee. Where it gets especially hard is the overlap between Healthcare Data Modeling and those two areas. You'll face questions about designing episode-of-care marts or PMPM cost schemas, then immediately need to write the SQL or Spark logic that makes them work against real payer data with late-arriving corrections. The prep mistake most candidates make is spending the bulk of their time on ML algorithms and statistical theory while underestimating how much of this loop tests whether you can model, build, and query healthcare-specific data end to end.

Practice McKinsey-style questions across all six areas at datainterview.com/questions.

How to Prepare for McKinsey & Company Data Analyst Interviews

McKinsey is betting heavily on embedding AI directly into client engagements, not just advising on it from the sidelines. Their Technology Trends Outlook 2025 maps where generative AI will reshape industries, while their research on AI-powered customer interactions shows the firm already building predictive models that ship inside client deliverables. For DAs, this means your day-to-day isn't exploratory analysis that sits in a notebook. You're constructing the data pipelines and models that partners present to healthcare CEOs and bank CFOs as the analytical backbone of a recommendation.

The single biggest "why McKinsey" mistake is talking about prestige or "working with the best." Partners hear that dozens of times a week. What lands is naming a specific piece of the firm's published research, like the State of Fashion report or the Global Banking Annual Review, and connecting it to analytical work you've actually done. "I noticed your demand-forecasting methodology in the fashion report and I've built similar pipelines from messy retail data" tells the interviewer you understand how DAs feed the firm's intellectual product, not just that you want the name on your LinkedIn.

Try a Real Interview Question

1WITH base AS (
2  SELECT
3    a.market,
4    a.variant,
5    s.session_id,
6    CAST(s.did_book AS DOUBLE) AS did_book
7  FROM experiment_assignments a
8  JOIN sessions s
9    ON s.user_id = a.user_id
10   AND s.session_start >= a.assigned_at
11   AND s.session_start < a.assigned_at + INTERVAL '7' DAY
12  WHERE a.experiment_name = 'search_ranker_v2'
13), agg AS (
14  SELECT
15    market,
16    variant,
17    AVG(did_book) AS conversion_rate
18  FROM base
19  GROUP BY 1, 2
20)
21SELECT
22  market,
23  MAX(CASE WHEN variant = 'control' THEN conversion_rate END) AS conv_control,
24  MAX(CASE WHEN variant = 'treatment' THEN conversion_rate END) AS conv_treatment,
25  (MAX(CASE WHEN variant = 'treatment' THEN conversion_rate END)
26   - MAX(CASE WHEN variant = 'control' THEN conversion_rate END)) AS lift
27FROM agg
28GROUP BY 1
29ORDER BY 1;

McKinsey's DA loop includes a case study round borrowed from their consulting track, which means even technical questions get framed around a client scenario (a hospital system's readmission rates, a payer's claims cost drivers) rather than abstract table joins. Practicing problems that force you to interpret business context before writing a query is the best way to prepare for that blend. Drill those at datainterview.com/coding.

Test Your Readiness

McKinsey's process spans case reasoning, SQL, statistics, and behavioral rounds, so a weak spot in any one area can sink you. Calibrate across all of them at datainterview.com/questions.