Interview guides · Harvey

Harvey Data Scientist Interview Questions & Guide 2026

Harvey Data Scientist interview guide: reported rounds, practice questions with worked solutions and a 7-day prep plan.

The 60-second brief

As a Data Scientist at Harvey, you are joining a mission-critical team at the forefront of transforming legal and professional services. Harvey is not just an AI company; it is an enterprise-grade platform that is redefining how knowledge work is performed globally. You will operate at the intersection of frontier agentic AI, complex user behavior, and high-stakes business strategy, helping to scale a company that already supports hundreds of customers across dozens of countries.

Coding rounds for this role are usually data-manipulation shaped rather than data-structure shaped: group-bys, joins, time windows, ranking within a partition. Confirm the format before spending a week on graph traversal.

Focus areas

Turn a vague request into a measurable question
Size an experiment before anyone launches it
Decompose a metric move by segment and mix

Overview

This role is inherently cross-functional and strategic. Whether you are partnering with the GTM organization to optimize growth or embedding with Marketing to refine acquisition channels, you are expected to build functions from the ground up. You will own the analytical frameworks that define Harvey's success, translating ambiguous business challenges into scalable metrics and actionable models. If you thrive on autonomy, intensity, and the challenge of building a generational company, this role offers an unmatched opportunity to influence the trajectory of the business.

The Harvey interview process

Harvey candidates report 3 rounds · ≈ 3-5 weeks. The stages below are what candidates describe, not a published process.

Recruiter Screen

Data Scientist covers at least four different jobs: experimentation, product analytics, causal work on observational data, and applied modelling that ships into a system. A screening call is the cheapest place to find out which of them is being hired for, and doing that diagnosis openly reads as senior rather than fussy. Ask what the last few pieces of work on the team actually were, and roughly how a week splits between querying, modelling and stakeholder time. Then say which parts of that you have done and which you have not. Claiming the whole range is the fastest way to be caught one round later.

What to demonstrate

Whether you can distinguish the flavours of the role and locate your own experience inside one of them honestly
Whether you name what you have not done instead of stretching to cover every line of the posting
Whether your hard constraints (notice period, location, work authorisation, level) surface now rather than at offer stage

How to prepare

Map the last two years of your time into rough percentages across query writing, experiment design, modelling and stakeholder work, so a question about scope has a real answer
Mark every responsibility in the posting as done, adjacent or new, and prepare one sentence for each adjacent item naming the closest thing you have actually built
Decide which logistics are non-negotiable before the call so you can state them in one sentence rather than negotiating live

Technical Rounds

Much of what gets scored here happens out loud while you type. Nobody can see your reasoning inside a half-written query, so five silent minutes read as being stuck even when they are not. State the plan in plain language first: which tables, what grain you are aggregating to, and the one filter that defines the population. Then write it. The narration doubles as insurance, because a wrong plan gets caught early and cheaply while a wrong query gets caught at the end with no time left to redo it. A timed statistics section, where one exists, is a separate test with its own clock.

What to demonstrate

Whether the query you write matches the plan you just described
What you do with a hint, meaning whether the correction gets absorbed or the first approach gets defended
Whether you can debug your own wrong output by reading the result set and naming which part of the query produced the anomaly

How to prepare

Solve three problems while screen-sharing into a recording, then watch it back and mark every stretch longer than thirty seconds where you said nothing
Practise compressing the plan into one sentence before typing, then check afterwards whether the finished query actually matched it
Time yourself on statistics questions that carry a business reading, such as what a confidence interval does and does not claim, rather than re-reading notes without a clock

Behavioral Interviews

Most of the weight in this round sits on the disagreement questions. Data work routinely produces an answer someone senior did not want, and the interviewer is trying to learn what you do in that hour. Both failure modes are common: folding as soon as a director pushes back, and treating the pushback as ignorance to be corrected with a better chart. A strong answer usually contains a specific thing the other person knew that you did not, and describes how you found out whether it changed the conclusion.

What to demonstrate

Whether you can state the other side's argument accurately before you explain why you disagreed
What you treated as evidence during the disagreement, such as a rerun under their assumption or a holdout check, rather than persuasion technique
Whether you distinguish being overruled from being wrong, and can give an example of each

How to prepare

Write out one disagreement where you turned out to be wrong, and say what in the data misled you. Candidates prepare the story where they were right, and the follow-up asks for the other one.
For your main disagreement story, be ready to say what result would have made you drop your position. If no such result exists, you were not arguing from the data.
Practise stating the opposing position out loud in one sentence the stakeholder would accept, then continue the story.

Where candidates lose points

Counting on an identity key that changes underneath the metric

visitor_id is per browser and per device, and it resets on cookie clearance, private browsing and platform privacy changes, so the distinct-visitor count drifts upward for reasons unrelated to reach. Any rate with visitors in the denominator therefore decays over time even when behaviour is constant, and any rate with visitors in the numerator inflates. The stitching at signup makes it worse in both directions: a user who signed up on mobile and returns on desktop is two visitors and one user, while a shared device is one visitor and several users. Decide which key each metric is counted on, write it into the definition, and when comparing a period before and after a platform privacy change, expect a level shift in every visitor-keyed metric and do not attribute it to the product.

Watching an experiment daily and stopping when it crosses significance

A fixed-sample test controls type I error at one pre-declared look. Checking repeatedly and stopping at the first p < 0.05 inflates the false positive rate to roughly 0.15 to 0.20 for ten looks, and it rises further with more frequent checks, because the p-value takes a random walk that will eventually dip below the threshold under the null. The usual defences are a fixed horizon declared before launch, group-sequential boundaries such as O'Brien-Fleming that spend alpha across a planned number of looks, or always-valid confidence sequences that are correct under continuous monitoring. Compounding it, the effect size reported conditional on having crossed the threshold is biased away from zero, and the bias is larger the lower the power was, so an underpowered test that 'won' typically overstates the lift it found.

SQL that silently fans out on a one-to-many join

State the grain of each table and the grain you want in the result before writing the join. Pre-aggregate the many side to the join key, or use EXISTS or a window function, and verify with a row count against COUNT(DISTINCT id) rather than trusting that the numbers look plausible.

Reaching for a model before the target metric exists

Before naming an algorithm, write down the label, the prediction time, and the action that changes when the score crosses a threshold. If you cannot say what decision the output drives, any modelling choice is guesswork dressed up as method.

Questions & practice

Simulate the false positive cost of repeated peeking

Medium · simulation · peeking · type i error

Quantify the cost of peeking. Simulate a two-arm experiment with no true effect: each arm accumulates Bernoulli conversions at a base rate of 0.10 up to 40,000 units per arm. Run a two-sided two-proportion z-test at alpha 0.05 at ten equally spaced interim points, and record whether the test ever crossed. Report the false positive rate over at least 10,000 replications, alongside the rate for a single look at the final sample only. Use a fixed seed and report a Monte Carlo standard error on both figures.

Approach

Generate each replication as two cumulative sums of Bernoulli draws, then read the interim points off the cumulative arrays. Regenerating data at each look would make the looks independent, which destroys exactly the dependence the exercise is about: later looks share data with earlier ones.
Use the pooled-variance two-proportion z: p_pool = (x1+x2)/(n1+n2), z = (p1-p2) / sqrt(p_pool*(1-p_pool)*(1/n1 + 1/n2)), reject when |z| > 1.96. State that the normal approximation is fine here because the smallest look has roughly 400 expected conversions per arm.
Vectorise across replications rather than looping: draw a (reps, n) array of uniforms, threshold at 0.10, cumsum along axis 1 and slice the ten look indices. A per-replication loop at 10,000 by 40,000 is unnecessarily slow.
Record the any-cross indicator per replication, take the mean, and compute the Monte Carlo standard error as sqrt(p*(1-p)/reps) so the reported figure comes with its own precision.
Report the single-look rate in the same run as a control. If it does not land near 0.05, the bug is in the test statistic and not in the peeking argument.

Follow-up

Re-run with 40 looks instead of 10. Why does the curve flatten rather than continue rising linearly?
Among the replications that crossed, what is the mean observed lift, and why is it not zero?
What does an O'Brien-Fleming boundary or an always-valid confidence sequence change about this simulation, and what does each cost in power?

Permutation test for a difference in conversion rates

Medium · permutation test · resampling · inference

Write a two-sided permutation test from scratch for a difference in conversion rates, using no scipy hypothesis function. Input: a DataFrame with unit_id, variant in {control, treatment} and converted in {0,1}, one row per randomisation unit. Compute the observed difference in proportions, then build the null distribution by reshuffling the variant labels while holding each arm's size fixed. Report the p-value as (1 + the count of permuted statistics at least as extreme in absolute value) / (B + 1) with B at least 10,000, and return the permutation distribution.

Approach

Name the null being tested: the sharp null that each unit's outcome is the same under either label. That is what licenses permuting labels, and it is stronger than the null of equal means, which matters when someone asks whether the test is valid under unequal variances.
Extract converted to a single numpy array of 0s and 1s and record n_treatment. Every permutation is then just a reshuffle of one array, and the treatment mean is the mean of the first n_treatment entries of the shuffled array.
Vectorise the B permutations with rng.permuted on a tiled 2-D array, or with argsort of a (B, n) random matrix. A Python loop calling np.random.shuffle B times is correct but roughly an order of magnitude slower and often runs past the time limit.
Use the +1 correction in both numerator and denominator. Without it a p-value of exactly 0 is reportable, which is false: the observed labelling is itself one of the permutations, so the smallest attainable p-value is 1/(B+1).
Compare the resulting p-value against a two-proportion z-test as a sanity check. At these sample sizes they should agree closely; a large divergence means the statistic or the shuffle is wrong, not that the permutation test found something subtle.

Follow-up

The arms are 200 and 20,000 units. Does the permutation test stay valid, and what happens to its resolution at B = 10,000?
Give a 95 percent confidence interval for the difference. Can you get it from this permutation distribution, and if not, what would you run instead?
The randomisation unit is user_id but the outcome is per session. What breaks, and what is the fix?

Split a pooled conversion drop into rate and mix

Medium · mix shift · decomposition · pandas

You have weekly visit-to-signup counts by segment: a DataFrame with week, device_type, referrer_channel, visitors and signups. The pooled rate fell 0.84 percentage points between two consecutive weeks while several individual segments rose. Write a function that, for a caller-supplied list of segment columns, splits the pooled change into a rate effect, a mix effect and an interaction term that sum exactly to the observed change. Return those three scalars plus a per-segment contribution table sorted by absolute contribution, so the largest single driver can be named.

Approach

State the algebra before coding: the pooled rate is r = sum over segments of w_s * r_s, with w_s the segment's share of the denominator. Then r1 - r0 decomposes exactly into sum(w_s0 * (r_s1 - r_s0)) for rate, sum((w_s1 - w_s0) * r_s0) for mix, and sum((w_s1 - w_s0) * (r_s1 - r_s0)) for interaction. The identity is per-segment, so it holds for any numbers you put in the four slots.
Pivot both weeks onto a common segment index with an outer join so a segment that appeared or vanished is kept rather than dropped, then decide what rate to give a segment with no visitors in one of the weeks, and document the choice. The identity stays exact either way because the missing week's weight is 0, but the attribution does not. Filling the missing rate with 0 sends an appearing segment's entire w_s1 * r_s1 into the interaction term, since w_s0 = 0 makes both the rate term and the mix term (w_s1 - w_s0) * r_s0 identically zero; a vanishing segment then splits as -w_s0 * r_s0 in rate, -w_s0 * r_s0 in mix and +w_s0 * r_s0 in interaction.
The convention used below instead imputes the missing week's rate as that week's pooled rate. A vanishing segment then lands wholly in mix at -w_s0 * r_s0, with rate and interaction cancelling; an appearing segment puts w_s1 * r_pooled0 in mix (volume arriving at the average rate) and only w_s1 * (r_s1 - r_pooled0) in interaction (its rate differing from that average). Impute by which week the segment is missing from, never by argument order, or the swap identities below stop holding.
Guard the division where visitors is 0 so no NaN enters the vectors, because a single NaN poisons every sum. A segment with zero visitors in both weeks contributes exactly 0 and can be dropped; a segment missing from only one week does not contribute 0, and where its contribution lands is settled by the convention above, not by the guard.
Compute the three components as vectors over segments, then sum. Keep the vectors, because the per-segment contribution table is what turns the decomposition into an explanation.
Assert that the three components sum to the observed pooled change within floating-point tolerance. This identity is exact, so a mismatch means an implementation bug, not a modelling judgement.

Worked solution (25 min)

Aggregate to one row per (week, segment tuple) with summed visitors and signups, then split into w0 and w1 frames and align with an outer join, filling missing visitors and signups with 0.
Compute w_s = visitors / visitors.sum() within each week, and r_s = signups / visitors only where visitors > 0. Where a week's visitors are 0, set that week's r_s to that week's pooled rate, the stated convention; never leave it NaN.
rate_effect = (w0 * (r1 - r0)).sum(); mix_effect = ((w1 - w0) * r0).sum(); interaction = ((w1 - w0) * (r1 - r0)).sum().
contribution = w0*(r1-r0) + (w1-w0)r0 + (w1-w0)(r1-r0) per segment, which reduces to w1r1 - w0r0; sort by abs and return the head.
assert abs(rate + mix + interaction - (pooled1 - pooled0)) < 1e-12.

Expected result: Three scalars summing exactly to the observed pooled change, plus a segment table whose contribution column also sums to that change. In the stated case rate and mix carry opposite signs: several segments converted better while volume moved toward a weaker-converting one, which is what makes the pooled move look inexplicable until it is split.

Verify it yourself

On synthetic data where every segment appears in both weeks, every segment rate is held constant and only the weights move, rate_effect and interaction are exactly 0.
On synthetic data where every segment appears in both weeks, weights are held constant and only rates move, mix_effect and interaction are exactly 0.
The per-segment contribution column sums to the pooled change, matching the sum of the three scalars.
Swapping the two weeks negates the total, as it must, but not the parts. Assert the three identities that actually hold: the interaction term is exactly invariant, because both of its factors change sign; the swapped rate effect equals -(rate_effect + interaction); and the swapped mix effect equals -(mix_effect + interaction). They still sum to the negated pooled change. Expecting all three to flip sign is the common wrong assertion.
Plant a segment present only in week 1, holding 3 percent of week-1 visitors at a rate well below pooled. Under the documented pooled-rate convention its mix contribution is w_s1 * r_pooled0 and its interaction contribution is w_s1 * (r_s1 - r_pooled0), with a rate contribution of exactly 0; under a zero-fill convention its mix contribution is exactly 0 and the whole w_s1 * r_s1 lands in interaction. Assert whichever one the function documents, and fail if mix is asserted non-zero while the code zero-fills.

Follow-up

The mix effect accounts for 0.71 of the 0.84 point drop, driven by paid_social volume. What is your recommendation, and what would change it?
Why is a two-way split into a counterfactual rate and a residual also exact, and when would you prefer it to the three-way version?
Segmenting on device and channel leaves a large interaction term. What does that tell you about the choice of segments?

Describe your approach to building a reliable data pipeline when source data is messy or inconsistent.

Medium · sql and data manipulation

Approach

State the window function and its partition and ordering out loud before writing it.
Compute rates by summing numerator and denominator separately, never by averaging rates.
Handle the rows that do not match: a LEFT JOIN with a NULL check is usually the question.

Follow-up

What breaks if events arrive late or out of order?
How does the query change if the join becomes one-to-many?

Write a query to calculate the rolling 30-day retention rate for our top-tier legal firm customers.

Medium · sql and data manipulation

Approach

Check whether any join is one-to-many before aggregating, or the sums inflate.
Say which table is the grain you start from, and join outward from it.
Compute rates by summing numerator and denominator separately, never by averaging rates.

Follow-up

How would you verify this result without re-running the same query?
How does the query change if the join becomes one-to-many?

Using SQL window functions, identify the first and last interaction dates for users within a specific cohort.

Medium · sql and data manipulation

Approach

Check whether any join is one-to-many before aggregating, or the sums inflate.
Handle the rows that do not match: a LEFT JOIN with a NULL check is usually the question.
Compute rates by summing numerator and denominator separately, never by averaging rates.

Follow-up

What breaks if events arrive late or out of order?
How would you verify this result without re-running the same query?

Rebuild sessions from raw events with a thirty-minute gap

Medium · window-functions · sessionisation · gaps-and-islands

From fct_event (event_id, visitor_id, occurred_at_utc) alone, rebuild sessions: a new session begins when the gap from that visitor's previous event exceeds 30 minutes, and every session is force-closed at UTC midnight so none spans two calendar dates. Return one row per session with visitor_id, session_start, session_end, event_count and session_date. Do not read fct_session; the point is to reproduce it. Assume duplicate occurred_at_utc values exist for the same visitor.

Approach

Get the previous timestamp per visitor with LAG(occurred_at_utc) OVER (PARTITION BY visitor_id ORDER BY occurred_at_utc, event_id). The event_id tiebreaker is required, not stylistic: with duplicate timestamps an unstable ordering makes the boundary flags non-deterministic between runs.
Set a boundary flag when prev IS NULL, or occurred_at_utc - prev > interval '30 minutes', or occurred_at_utc::date <> prev::date. The third disjunct is the midnight rule, expressed as a date change rather than a clock comparison so it holds across any gap length.
Number the islands with SUM(flag::int) OVER (PARTITION BY visitor_id ORDER BY occurred_at_utc, event_id ROWS UNBOUNDED PRECEDING). Because event_id is unique the ordering is total, so no two rows are peers and RANGE UNBOUNDED PRECEDING would compute exactly the same numbers here. Write ROWS anyway: it is the half of the guard that survives someone later simplifying the ORDER BY back to occurred_at_utc alone, at which point the default RANGE frame gives every row sharing a timestamp one shared running total.
GROUP BY visitor_id and the island number, then MIN(occurred_at_utc) AS session_start, MAX(...) AS session_end, COUNT(*) AS event_count, session_start::date AS session_date.
Reconcile against fct_session on one sample day. The counts should agree except for server-emitted events carrying no client session, so a systematic difference beyond those is a bug in the gap rule or in the ordering.

Worked solution (30 min)

Pick one high-volume visitor and dump their ordered event timestamps for a day to trace by hand.
Add LAG with the tiebreaker and eyeball the computed gaps.
Add the three-part boundary flag and confirm the first event of each date is flagged.
Add the running SUM with an explicit ROWS frame, then group and aggregate.
Compare total event_count against the raw input count, and compare session counts to fct_session for the sample day.

Expected result: One row per (visitor_id, island). SUM(event_count) equals the input row count exactly, with no events lost or duplicated. No session_end falls on a different date from its session_start.

Verify it yourself

Total events in equals total events out.
MAX(session_end::date - session_start::date) is 0 across all rows.
For a visitor ordered 09:00, 09:10, 09:55, 23:58, 00:02, the result is four sessions, not three.
No two sessions for the same visitor have overlapping [session_start, session_end] ranges.
Swapping ROWS for RANGE while event_id stays in the ORDER BY changes nothing; dropping event_id from both ORDER BY clauses does change the boundaries for a visitor with duplicate timestamps, which locates the guard in the tiebreaker rather than in the frame keyword.

Follow-up

Why 30 minutes? What does a 5-minute rule do to sessions-per-visitor and to any per-session conversion rate?
The same person uses phone then laptop. Two visitor_ids, two sessions. What breaks if you sessionise on user_id instead?
The midnight rule splits an overnight session. Which metrics does that bias, and in which direction?

How would you measure the success of a new AI-driven feature for legal research?

Medium · product sense and metrics

Approach

Restate the decision this analysis has to support, and who acts on the answer.
Fix the population and the time window before naming any metric.
Name one primary metric, then the guardrail that stops it being gamed.

Follow-up

What would you do if the primary metric and the guardrail moved in opposite directions?
Which segment would you cut first, and what would that rule out?

If our core engagement metric drops by 10% overnight, what is your systematic process for diagnosing the root cause?

Medium · product sense and metrics

Approach

State what result would change your recommendation, so the answer is falsifiable.
Decompose the metric into the rates that drive it, and say which one you would check first.
Name one primary metric, then the guardrail that stops it being gamed.

Follow-up

What would you do if the primary metric and the guardrail moved in opposite directions?
How would you detect that the metric is being gamed rather than genuinely improving?

How would you design a metric to measure "agentic quality" in our platform?

Medium · product sense and metrics

Approach

State what result would change your recommendation, so the answer is falsifiable.
Decompose the metric into the rates that drive it, and say which one you would check first.
Restate the decision this analysis has to support, and who acts on the answer.

Follow-up

How would you detect that the metric is being gamed rather than genuinely improving?
What would you do if the primary metric and the guardrail moved in opposite directions?

What are the most common experimentation pitfalls you have encountered, and how do you mitigate them?

Medium · a/b testing and experimentation

Approach

Name the guardrails that would stop a launch even on a positive primary result.
State the primary metric and the minimum effect worth shipping, then size the test.
Decide the analysis before seeing data, including how long it runs and when you look.

Follow-up

What would you conclude if the result is positive but the test is underpowered?
How would you handle interference between treated and control units?

Explain the difference between statistical significance and practical significance in an A/B test.

Medium · a/b testing and experimentation

Approach

State the primary metric and the minimum effect worth shipping, then size the test.
Name the guardrails that would stop a launch even on a positive primary result.
Name the randomisation unit first; it decides the variance and what the test can detect.

Follow-up

What would you conclude if the result is positive but the test is underpowered?
What would you do if you could not randomise at all?

Measure a paid channel without a user-level experiment

Hard · difference-in-differences · geo holdout · incrementality · clustered inference

Finance needs the incremental contribution of paid social before next year's budget. It carries 18% of last-touch-attributed signups in dim_user.last_touch_channel, but the ad platform will not randomise users and unexposed users are unobservable. The ad platform targets at market level, your warehouse resolves a market label onto fct_session, and 40 markets are eligible with 12 months of weekly signups derivable from dim_user. Design the read, size it on total signups rather than attributed ones, and state the identifying assumption and the check that would falsify it.

Approach

Say why the attributed number cannot answer the question. Last-touch assigns full credit to whichever touch was recorded nearest the conversion, which systematically rewards channels that intercept demand that already existed, and the counterfactual of not running the channel is nowhere in the touchpoint data.
Choose a geo holdout: randomise markets into a paused arm and a spend-as-usual arm, and measure total signups per market. The metric has to be the total, because the question is whether demand is created or merely relabelled, and attributed signups fall to zero in a paused market by construction.
Size on markets, since the market is the randomisation unit and therefore the variance unit. Forty markets give at most twenty per arm, so form matched pairs on twelve months of pre-period weekly signups and randomise within pair, which removes the between-market level differences that would otherwise dominate the residual variance.
Fix the identifying assumption in plain words: absent the pause, treated and control markets' signup trends would have moved in parallel. Falsify it by running the same estimator on placebo cut dates inside the pre-period; a non-zero placebo estimate or visibly diverging pre-trends kills the design as specified.
Estimate with market and week fixed effects and cluster standard errors on market. Below about forty clusters the cluster-robust error is biased downward, so use a wild cluster bootstrap or randomisation inference over the pair assignment. If only one or two markets can be paused, switch to synthetic control: weight untreated markets to match the treated market's pre-period path and use in-space placebos for inference.

Worked solution (40 min)

Aggregate weekly signups per market for twelve months from dim_user with is_internal = FALSE, and drop markets whose history is too short or too volatile to pair.
Rank markets on pre-period mean and trend, form twenty matched pairs, and randomise one of each pair into the paused arm.
Derive the MDE from the between-pair residual standard deviation of weekly signups, not the within-market one, because power is set by the number of pairs; express it as a percentage of total signups and compare it with the effect finance cares about.
Run placebo estimates at three pre-period cut dates before committing; re-pair or abandon if any is significant.
After the pause, estimate difference-in-differences with market and week fixed effects, report a wild cluster bootstrap interval, and convert the point estimate into incremental cost per signup against the spend withheld.

Expected result: A matched-pair geo holdout across twenty pairs, read on total signups with market and week fixed effects and market-clustered inference, yielding an incrementality ratio of incremental signups to attribution-credited signups. The identifying assumption is parallel counterfactual trends, checked with pre-period placebo cuts. The MDE is set by twenty pairs rather than by millions of sessions, and is usually far coarser than the requester expects.

Verify it yourself

Pre-period placebo estimates should centre on zero with intervals covering it. A systematically non-zero placebo means the pairing failed and the main estimate inherits that bias.
Treated plus control market signups in the pre-period must reconcile with the total the finance dashboard reports, otherwise the market label is silently dropping sessions.
An incrementality ratio above 1 implies the channel drives conversions it never touched. That is possible through view-through effects the model never sees, but it demands a stated mechanism rather than a shrug.

Follow-up

Campaign commitments mean the pause starts in different weeks in different markets. What does staggered timing do to a two-way fixed effects estimate when effects vary across cohorts?
How long do you pause, given that paid social has a delayed conversion tail, and how do you keep that tail out of the pre-period?
The estimate says the channel is 40% incremental. What happens to the last-touch number on the dashboard after that?

Tell a pipeline outage from a collapse in usage

Medium · pipeline-failure · instrumentation · event-volume

Weekly active accounts completing a core action fell 9%, and almost the entire fall sits in accounts whose events carry surface = 'ios'. App crash rates and store reviews are unchanged. You have fct_event with occurred_at_utc, received_at_utc, event_name, is_core_action, app_version and surface, plus fct_session and the ingestion job run log. Establish within the hour whether iOS engagement fell or iOS events stopped arriving, name the evidence that distinguishes them, and say what you would publish on the dashboard in the meantime.

Approach

Compare the event-name composition inside surface = 'ios' against the prior four weeks as shares, not counts. A behaviour collapse scales most event names together; a dropped event definition or a broken downstream filter hits specific event_name values while page_view and session-opening events hold steady. That shape difference is the fastest discriminator available.
Profile the received_at_utc minus occurred_at_utc distribution per day for surface = 'ios'. A stalled-then-backfilling pipeline shows a fat upper tail and a recovering p99; a silently dropped stream shows an unchanged lag distribution over a smaller volume. The two failure modes have different remedies and different histories.
Cut by app_version. A logging SDK change arrives with one build and ramps with its adoption curve; an infrastructure fault arrives across every build within the same hour. Checking this costs one group-by and rules out half the hypothesis space.
Cross-check against a signal that does not travel the suspect path: server-emitted events with session_id NULL, and subscription or billing activity for the same accounts. If those accounts are still transacting, the users did not leave.
Publish an ex-iOS total with an explicit annotated break rather than a blended total. A blended number during a known ingestion gap is wrong in a direction you can already name, and republishing it daily spreads the artefact into every downstream report.

Follow-up

Suppose the events do eventually backfill. What is your policy for restating the published weekly numbers, and who needs to be told?
What monitor would have caught this before a human noticed the weekly metric, and what would it alert on?
If is_core_action is maintained in the tracking plan, what governance would stop a change to that list from silently moving a north-star metric?

Suggested practice checklist

For a candidate whose interviews will centre on A/B testing, metric movement and causal claims. Design comes before arithmetic, arithmetic before analysis, and the week ends by rehearsing the readout rather than the derivation.

Day 1 — Design one test end to end on paper

Take a single feature change and write the full design: randomization unit, the exact point of exposure, the primary metric with its grain, guardrails, allocation, planned duration, and the decision rule committed before any data exists.
Write why the randomization unit must sit at or above the level where treatment can spill over, and give one case where user-level randomization is still contaminated (shared accounts or devices, or two participants in the same marketplace).
State in advance what you will do if the primary metric is flat while a secondary metric is significant.

Deliverable: A one-page test design with a decision rule written before launch.

Day 2 — Power arithmetic until it is automatic

Compute required sample size per arm for a binary metric with the normal approximation, n is approximately 2 times (z for alpha/2 plus z for power) squared times p(1 minus p) divided by delta squared, for baselines of 2, 10 and 40 percent at a 5 percent relative lift, and note that for a fixed relative lift the requirement falls as the baseline rises because delta grows proportionally with p.
Redo the calculation for a continuous metric using variance in place of p(1 minus p), and show why a heavy-tailed quantity such as revenue per user needs either far more traffic or a capped version with a stated cap.
Convert one of the results into weeks given a weekly eligible traffic figure, then list the two honest ways to shorten it (accept a larger detectable effect, or reduce variance) and write why quietly lowering the power target is a decision to miss more real wins, not a speedup.

Deliverable: A small script or sheet that maps baseline, minimum detectable effect, alpha and power to sample size and weeks, cross-checked against a published calculator.

Day 3 — Variance and the unit-of-analysis problem

Take a ratio metric whose denominator is not the randomization unit (clicks per session, randomized by user) and compute the standard error twice, once naively at session level and once by the delta method or a user-level bootstrap, then record how much the naive version understates it.
Implement CUPED on simulated data: choose a pre-period covariate X measured before assignment, estimate theta as Cov(Y, X) divided by Var(X), and analyse Y minus theta times (X minus its mean) in place of Y. Confirm the variance of the adjusted outcome equals the raw variance multiplied by one minus the squared correlation between Y and X, so a correlation of 0.45 removes about 20 percent of the variance and not 80.
Now run that simulation a few hundred times and confirm the adjusted effect estimate is unbiased for the same effect rather than numerically identical to the raw one. Within any single run the two differ, sometimes by a large fraction of the true effect, because the two arms' pre-period covariate means never coincide exactly in a finite sample; they agree in expectation, which is the property that matters and the one to state out loud.

Deliverable: A notebook showing the adjusted estimator with a measurably smaller variance than the raw one, plus a repeated-simulation table showing the two estimators agreeing on average while differing run by run.

Day 4 — Validity threats you can actually test for

Run a sample ratio mismatch check as a chi-square goodness-of-fit test against the intended allocation, and write the three causes you would chase first (assignment logged before exposure, an arm-specific redirect or load failure, bot filtering applied asymmetrically).
Simulate peeking: generate A/A data, test daily at alpha 0.05 across 14 looks, record the inflated false positive rate, then apply an alpha-spending boundary or commit to a fixed horizon and confirm the rate returns to nominal.
Write how you would separate a novelty effect from a durable lift using the treatment effect plotted against days since first exposure, and what shape would change your recommendation.

Deliverable: One table showing the peeking false positive rate before and after correction, plus a written SRM triage list.

Day 5 — When randomization is not available

Write the identifying assumption for difference-in-differences (parallel trends in the absence of treatment), then plot pre-period trends for two candidate control groups and justify rejecting one of them.
Design a switchback test for a change where user-level randomization would leak across participants, choosing a time-block length against the carryover you expect and saying how you would detect carryover in the data.
List what an interrupted time series or a synthetic control buys you and the one thing neither can rule out: an unobserved shock that coincides with the launch.

Deliverable: A one-page memo recommending a single quasi-experimental design and naming its weakest assumption explicitly.

Day 6 — The readout query

Write the assignment-to-exposure join that returns exactly one row per unit per experiment, and handle units appearing in both arms by excluding and counting them rather than silently keeping one.
Compute the per-arm metric, its variance and the relative lift with a confidence interval in SQL, then reproduce the identical numbers in a notebook as a cross-check.
Add a segment breakdown and write the sentence that keeps it from being p-hacking: segments declared in advance, everything else reported as exploratory and corrected for multiplicity.

Deliverable: A single query that outputs the full readout table, matched to a notebook recomputation.

Day 7 — Present it to someone who will not read the appendix

Give a 10-minute readout of a real or simulated experiment in the order decision, number, uncertainty, caveat.
Have your listener ask "can we ship it" in the case where the primary is flat and a guardrail moved, and answer with a recommendation rather than a request for more data.
Rewrite your opening line so the recommendation lands before any methodology.

Deliverable: A one-page readout whose first line is the recommendation.

Your part, not the team's

Most data work is done by groups, so an interviewer has to work out which piece was yours. An answer that runs on 'we' for several minutes gets interrupted with a question about what you personally did, and by then the answer sounds defensive even when it is true. Mark your own contribution as you go, and name the parts that belonged to someone else instead of leaving them ambiguous. Keep a few specifics back as well, like the name of the metric or who actually objected, so a probe can be answered with something you had not already said.

If an experiment shows a positive result but stakeholders are skeptical, how do you validate the findings?

Medium · behavioural and stakeholder questions

Approach

Close with what you would do differently, concretely.
Quantify the outcome, including what you would not claim credit for.
State the situation in two sentences and spend the rest on your reasoning.

Follow-up

How did you know the outcome was caused by your change?
What would you do differently if you ran that project again?

Walk through an analysis you got wrong and what changed

Easy · postmortem · data quality · self-assessment

Describe an analysis of yours that turned out to be wrong after somebody had already acted on it. You have four minutes. The account must name the defect mechanically, the join, the filter, the window or the identity key, rather than describing it as a communication problem. It must also say who did what because of the wrong number, how the error surfaced, how long it stood, and what control you put in place so that class of error cannot reach a decision again. Do not pick an error nobody acted on.

Approach

Recognise what is being probed: whether you can be specific about your own failure without minimising it or performing contrition. The discriminator is whether the defect has a mechanism the listener could reproduce in their own warehouse.
Choose the case by blast radius rather than by comfort. An error nobody acted on tests nothing, and picking one signals that you are managing the interview instead of answering it.
Structure the account in six beats: the number, the decision it drove, the defect, the detection, the correction, the control. Keep the defect to one reproducible sentence, for example an inner join to fct_subscription_period that dropped accounts with no subscription row and so computed retention over payers only.
State the direction of the bias, not only its existence. A filter or join that removes rows usually moves a metric predictably, and knowing which way shows you diagnosed the mechanism rather than patched the symptom.
Be exact about detection and elapsed time. 'A colleague noticed' and 'the row-count assertion failed before publication' are different answers about the same organisation, and the second one is the one your control is supposed to produce next time.
End on the control, its cost, whether it has fired since, and one thing it does not cover.

Follow-up

What did the control cost, and has it fired since? If it never has, how do you know it works?
How long did the wrong number stand before anyone questioned it, and what does that say about the review path it went through?
What is the equivalent mistake you are most likely to make in this role, given the tables you would be working in?

Quantify your own impact without claiming the topline you touched

Hard · self-assessment · attribution · communication

You are writing the impact section of your own review. Over the year you ran four experiments, one of which shipped and three of which were flat; you corrected the definition of gross monthly revenue churn so that cancellation is recognised at period_end_utc; and you built a self-serve funnel dashboard. Weekly active accounts rose 14% over the same period. Your reviewer knows the data well. Write the three impact claims you would defend, stating for each what you contributed, what evidence supports it, and what portion of the outcome you are not claiming.

Approach

Recognise what is being probed: whether you apply to your own work the causal standard you would apply to somebody else's roadmap claim. Nearly everyone who would reject 'accounts that do Y retain better' will write 'I drove a 14% increase' without noticing it is the same error with a friendlier subject.
Sort the work by the kind of evidence it can carry. The shipped experiment is the only item with a randomised estimate, so it is the only one where an effect size is defensible, and you claim the interval rather than the point estimate.
Claim the three flat experiments as decisions prevented and price them. Features not built, or built differently, on evidence, with the engineering weeks reallocated as the number somebody else can verify. A defensible null is a delivered decision and should be written as one.
Claim the definition fix as correctness, not as improvement. The old figure was overstated by a specific percentage and appeared in a specific set of recurring documents; the impact is the change it produced in the forecast built on top of it, not a change in churn itself.
Claim the dashboard on usage and displacement: distinct weekly users of it, and the ad-hoc request count for six months before against six months after. If the request log does not exist, record the claim as unverified rather than estimating it upward.
Disclaim the 14% explicitly and once. State that it cannot be separated from seasonality, other teams' launches and a pricing change, and bound your own contribution from above using the shipped experiment's interval converted into headline units.

Follow-up

Your shipped experiment's interval was +0.2pp to +1.4pp on activation. How much of the 14% can that account for, and how do you say so without undercutting yourself?
A peer in the same cycle claims the full 14%. What, if anything, do you do about it?
If you could only keep two of your three claims, which do you drop, and why that one?