Interview guides · Intercom

Intercom Data Scientist Interview Questions & Guide 2026

Timed data scientist practice for Intercom: short SQL and probability sets run end to end, then scored against a rubric.

The 60-second brief

As a Data Scientist at Intercom, you sit at the intersection of product innovation, customer experience, and rigorous data analysis. Intercom is a leader in customer messaging, and the Data Science team is instrumental in shaping how the platform evolves. You will work on high-impact initiatives, ranging from product analytics that inform feature development to the optimization of AI-driven tooling that powers modern customer support.

Ask early whether the loop includes an asynchronous take-home or a timed live case, because the two are graded on different things. A take-home is read as an artifact: the question you decided to answer, what you did about missing or malformed records, and a conclusion stated plainly enough for someone to act on. A reviewer who cannot rerun your notebook discounts the result whatever score is printed in it. Hold to the stated time box and write down what you would have done with more of it, since the follow-up round is usually a live defence of the same work.

Focus areas

Separate novelty effects from durable behaviour change
Turn a vague request into a measurable question
Define numerator, denominator and window precisely

Overview

Your role is to transform ambiguous business problems into clear, data-backed strategies. You will be expected to design metrics, evaluate the success of new product launches through A/B testing, and diagnose performance shifts in real-time. Whether you are working on Product Analytics or AI Tooling, your work directly influences how thousands of businesses interact with their customers, making this a role where your technical expertise has a tangible, visible impact on the bottom line.

The Intercom interview process

Intercom 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 Assessment

Before anything else, this round is a reading test. You are given a small schema and a question phrased in business language, and most of the difficulty sits in the gap between them. Who counts as an active user, does a refunded order still count as an order, is that date column an event time or a load time. Weak answers start typing immediately and compute something precise about the wrong population. Strong ones pin the definition in one sentence, name the column that encodes it, then write the query. On a timed assessment with nobody to tell, write the definition in a comment anyway.

What to demonstrate

Whether an ambiguous term becomes a specific column and filter before any computation happens
Whether you read the schema for keys and cardinality rather than only for column names
Whether the result answers the question at the grain it was asked at, per user or per session or per day

How to prepare

Take three metrics you already use and write down the exact filter and exact grain behind each, then practise stating one of them in a single sentence out loud
On a schema you have never seen, spend the first minute writing what one row of each table means and which key it is unique on, then predict which joins can duplicate rows
Rehearse a version where the definition changes halfway through, and edit the query you have instead of starting over

Virtual or Onsite Panel

A day of back-to-back interviews samples your floor, not your ceiling. Four hours in, the habits that carry a good answer are the first to go: restating the question before solving it, asking what the data would have to look like, checking a number before quoting it. What the day decides is whether the tired version of you is still someone to leave alone with an ambiguous problem. The round that sinks a candidate is usually not the hardest one. It is the one immediately after the round that went badly.

What to demonstrate

Whether the late rounds get the same clarifying questions as the first one, or whether you start answering immediately to save effort
Whether a weak answer stays in the room it happened in, instead of following you into the next conversation as apology or distraction
Whether the quality of your questions holds up, since fatigue removes curiosity about the problem before it removes knowledge of the method

How to prepare

Rehearse the length, not just the content: book four mock interviews of different types in one afternoon with short gaps, because the one you need to observe is the fourth
Put the two or three questions you ask at the start of any problem on a card in front of you, so that under fatigue it is a habit you run rather than a decision you make
Decide in advance what the gap between rooms is for: water, one line of notes on anything you promised to follow up, and an explicit close on the round that just ended so it does not travel
Prepare a different closing question for each interviewer, so the end of a long day does not produce the same one four times

Where candidates lose points

Comparing cohort retention curves of different maturities, or building the curve from users who are still present

A cohort four weeks old has no week-8 value, so an average taken across cohorts silently drops young cohorts from the later columns and keeps them in the earlier ones. The curve then bends upward at the tail, and the reading that 'retention is improving over time' is an artefact of which cohorts survived to be measured. The same error appears in the denominator when retention is computed over users active in the current period rather than over the full original cohort, which conditions on survival and guarantees a flattering number. The fix is a triangle: fix the cohort at signup, bound every window on both sides, and only compare cells where every cohort has had the full elapsed time, publishing the rest as blank rather than as a partial average.

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.

Reading a dozen metrics with no multiplicity control

Nominate one primary metric before launch and treat the rest as guardrails or exploratory, with Bonferroni or Benjamini-Hochberg applied when you intend to make claims from them. Twenty independent tests at 0.05 under the null produce at least one false positive about 64 percent of the time.

Optimising accuracy on a heavily imbalanced target

State the base rate first, then choose the metric from the relative cost of a false positive against a false negative: precision and recall at the operating threshold, PR-AUC, or expected cost. At a 1 percent positive rate, predicting the majority class for everyone scores 99 percent accuracy and is worthless.

Questions & practice

Cluster bootstrap for a per-session rate randomised on users

Hard · cluster bootstrap · randomisation unit · variance

An experiment randomised on user_id reports a per-session conversion rate, so sessions inside a user are correlated. Input: one row per session with user_id, variant in {control, treatment} and converted in {0,1}. Write a cluster bootstrap from scratch: resample users with replacement within each arm, keep every session of a drawn user, recompute each arm's ratio of converted sessions to sessions, and take the difference. Return the point estimate, a 95 percent percentile interval from at least 2,000 resamples, the naive session-level interval that ignores clustering, and the ratio of their widths.

Approach

Name the estimand precisely: it is a ratio of sums, sum(converted) over sum(sessions) within an arm, not the mean of per-user rates. Those differ whenever session counts vary across users, and the ratio is what the reported metric is.
Resample the cluster, not the row. Draw n_users user ids with replacement inside each arm and take every session belonging to each draw, including duplicate draws of the same user. Keeping the user count fixed per arm rather than the session count is what preserves the sampling design.
Precompute per-user (converted_sum, session_count) once, so each resample is two vector lookups and a division rather than a repeated filter over the session frame. That turns 2,000 resamples from minutes into under a second.
Take the 2.5th and 97.5th percentiles of the 2,000 differences for the interval, and report the point estimate from the full data rather than from the bootstrap mean, since the bootstrap mean carries the resampling bias.
Compute the naive interval from the session-level binomial standard error and compare widths. The expected inflation is roughly sqrt(1 + (m-1)*rho), with m the mean sessions per user and rho the intraclass correlation of converted within users, so a computed ratio far from that value points at a bug in one of the two intervals.

Worked solution (35 min)

per_user = df.groupby(['variant','user_id'])['converted'].agg(['sum','size']); split into two arrays per arm.
point = (t_sum.sum() / t_n.sum()) - (c_sum.sum() / c_n.sum()).
For b in range(B): idx = rng.integers(0, len(t_sum), len(t_sum)); ratio_t = t_sum[idx].sum() / t_n[idx].sum(); same for control; store the difference. Vectorise by drawing a (B, n) index matrix if memory allows.
ci = np.percentile(diffs, [2.5, 97.5]); naive_se = sqrt(p_t*(1-p_t)/n_sessions_t + p_c*(1-p_c)/n_sessions_c); naive_ci = point +/- 1.96*naive_se.
width_ratio = (ci[1]-ci[0]) / (naive_ci[1]-naive_ci[0]).

Expected result: A point estimate identical to the full-data ratio difference, a percentile interval containing that point estimate, and a width ratio greater than 1 whose value approximates sqrt(1 + (m-1)*rho) for the data's mean cluster size m and intraclass correlation rho. A ratio of approximately 1.0 means sessions were resampled instead of users.

Verify it yourself

Construct data where every user has exactly one session: the cluster and naive intervals then agree to within Monte Carlo error, since m = 1 makes the design effect 1.
Construct data where converted is constant within each user: rho is 1, and the width ratio should approach sqrt(m).
The bootstrap distribution's mean sits close to the point estimate; a large offset means the resampled arms have drifted in size.
Doubling B narrows the Monte Carlo noise in the percentile endpoints without shifting the interval's centre.

Follow-up

Users average 3.4 sessions and the intraclass correlation is 0.12. What width ratio do you predict before running it, and does your bootstrap land there?
Give the delta-method standard error for this ratio and say when you would prefer it to the bootstrap.
Half the users in the treatment arm have exactly one session. What does that do to the cluster bootstrap's coverage, and how would you check it?

Implement seven-day activation from its written definition

Medium · metric definition · cohorts · pandas · joins

Implement the seven-day activation rate. Inputs: dim_user with user_id, account_created_at_utc and is_internal; fct_event with user_id, occurred_at_utc and is_core_action. A user activates when core-action events carrying a non-NULL user_id fall on at least two distinct UTC dates inside [account_created_at_utc, account_created_at_utc + 7 days). The denominator is every non-internal user whose account_created_at_utc lands in the cohort week, including users with no events at all. Return one row per cohort week with numerator, denominator and rate, publishing only weeks whose last signup is at least eight days old.

Approach

Build the denominator first, from dim_user alone, filtered on is_internal = False. Deriving it from the join is the standard way to lose every user who never fired an event, which is exactly the population the metric is about.
Join events to users on user_id with a left join from the user side, then apply the window as a half-open interval: occurred_at >= created AND occurred_at < created + 7 days. The right bound is exclusive, so an event at exactly created + 7 days does not count.
Count distinct UTC dates per user, not distinct events. Floor occurred_at_utc to date before the nunique, and do it in UTC rather than local time so the threshold does not move with the user's country.
Apply the >= 2 threshold, aggregate to cohort week, and compute the rate by re-summing numerator and denominator per week rather than averaging any per-user or per-day rate. Fix the week anchor explicitly: cohort_week is the Monday of the signup week in UTC, which is what Postgres DATE_TRUNC('week') returns and what any SQL version of this metric will produce. In pandas, subtract dt.weekday days from the floored timestamp. If you reach for periods instead, the anchor that matches is to_period('W') (equivalently 'W-SUN'), whose weeks end Sunday and therefore start Monday; to_period('W-MON') labels weeks that end on Monday, so it runs Tuesday through Monday and its start_time is a Tuesday. Mixing the two shifts every cohort label by one day and silently moves Mondays into the previous week.
Suppress immature weeks: drop any cohort week whose maximum account_created_at_utc is within 8 days of the data cut, and return them as absent rather than as a partial number.

Follow-up

The threshold is 2 distinct days. What changes in the reported history if someone moves it to 3, and how would you publish that change?
Invited seats and SSO-provisioned users have no pre-signup session. Should they be in this denominator at all, and what does including them do to the rate for sales-assisted accounts?
How would you produce the same metric at account grain, and which of the two would you put on the dashboard?

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.

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?

Write a query using SQL window functions to calculate the running total of messages sent per user over a 30-day period.

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.
Check whether any join is one-to-many before aggregating, or the sums inflate.

Follow-up

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

Given two tables (users and events), how would you calculate the retention rate for a cohort of users?

Medium · sql and data manipulation

Approach

Compute rates by summing numerator and denominator separately, never by averaging rates.
Say which table is the grain you start from, and join outward from it.
Handle the rows that do not match: a LEFT JOIN with a NULL check is usually the question.

Follow-up

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

Read an experiment from first exposure, not assignment

Hard · experiments · window-functions · srm

fct_experiment_exposure holds experiment_id, unit_type, unit_id, variant, user_id, assigned_at_utc, first_exposed_at_utc, is_in_analysis_population and planned_end_utc. fct_event holds user_id, occurred_at_utc, is_core_action, and carries events up to a known data cut, :data_cut_utc. For one experiment randomised on unit_type = 'user', return per variant: exposed units, units with at least one core action in the seven days after that unit's own first exposure, the rate, and the variant share of exposed units. Only units whose seven-day window has fully elapsed as of the data cut belong in the readout. Units appearing under more than one variant are excluded from both arms and counted separately.

Approach

Run the contamination pass as an aggregate, not a window: SELECT unit_id FROM fct_experiment_exposure WHERE experiment_id = :exp GROUP BY unit_id HAVING COUNT(DISTINCT variant) > 1, then anti-join it away. PostgreSQL rejects COUNT(DISTINCT variant) OVER (PARTITION BY unit_id) outright, since DISTINCT is not implemented for window functions; if you want the test inline, MIN(variant) OVER (PARTITION BY unit_id) <> MAX(variant) OVER (PARTITION BY unit_id) is the equivalent that does run.
Do not resolve contamination by keeping the earliest variant. A unit that saw both arms carries treatment from both, so assigning it to either one biases that arm.
Define the population as is_in_analysis_population = TRUE AND unit_type = 'user' AND first_exposed_at_utc < planned_end_utc AND first_exposed_at_utc + interval '7 days' <= :data_cut_utc. The horizon filter is what makes the readout reproducible next week instead of drifting with every re-run; the data-cut filter is the one that actually buys seven days of follow-up, since a unit exposed an hour before the horizon otherwise contributes an hour of observation to a seven-day rate.
Measure the outcome on a per-unit relative window: LEFT JOIN fct_event on user_id with is_core_action = TRUE and occurred_at_utc in [first_exposed_at_utc, first_exposed_at_utc + interval '7 days'). LEFT JOIN so units with no outcome stay in the denominator at zero rather than being deleted by an inner join.
Check the sample ratio before reading the effect: variant share of exposed units against the intended split, tested as a binomial. Run it on the truncated population as well as on the full exposed set, because if one arm exposes later on average the data-cut filter removes more of that arm and can manufacture a ratio mismatch the randomisation did not have. A mismatch on the full set means the exposure data is not a valid randomisation and invalidates the readout rather than being a footnote under it.
Report the per-variant rate, the absolute difference, and the fact that the variance unit is unit_id. That is straightforward here only because the grain is already one row per user; a per-session outcome under user randomisation would need a delta-method or bootstrap standard error instead.

Worked solution (35 min)

Count units per variant and units with more than one variant; report the contaminated count before anything else.
Build the exposed population CTE with both the horizon filter and the seven-days-before-data-cut filter, assert one row per unit_id, and record per variant how many units the second filter removed.
LEFT JOIN the bounded outcome events and flag conversion with a boolean aggregate.
Aggregate per variant to exposed, converted, rate and share, and run the binomial check on the share, both before and after the truncation.
Re-run with a fixed calendar window from experiment_start_utc and record the difference as the size of the dilution you avoided.

Expected result: One row per variant with exposed, converted, rate and share, plus a scalar count of excluded contaminated units. Every retained unit was first exposed at least seven days before the data cut, so each one has a full seven days of observation and the rates are directly comparable.

Verify it yourself

Exposed counts across variants sum to the population count minus the contaminated units.
MAX(first_exposed_at_utc) among retained units is at least seven days before :data_cut_utc.
Variant shares sum to 1.
No unit_id contributes to two variant rows.
The truncation drops a similar share of each arm; if it does not, note the imbalance alongside the SRM result rather than reading the effect.
Swapping the relative window for a calendar one changes the rates; if it does not, exposure timing is balanced and you can say so.

Follow-up

Some units were assigned days before they were exposed. What does analysing the assigned set instead do to the estimated effect, and in which direction?
The split is 51/49 on 400,000 exposed units. Do you read the result?
The treatment arm exposes on average two days later than control. What does that do to a fixed calendar outcome window, and which arm does it favour?

How do you prioritize your work when you have multiple competing requests from different product teams?

Medium · product sense and metrics

Approach

Name one primary metric, then the guardrail that stops it being gamed.
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

Which segment would you cut first, and what would that rule out?
How would you detect that the metric is being gamed rather than genuinely improving?

How would you measure the success of a new feature in the Intercom messenger?

Medium · product sense and metrics

Approach

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.
State what result would change your recommendation, so the answer is falsifiable.

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?

How would you design a metric to track the quality of customer support conversations?

Medium · product sense and metrics

Approach

Name one primary metric, then the guardrail that stops it being gamed.
Restate the decision this analysis has to support, and who acts on the answer.
State what result would change your recommendation, so the answer is falsifiable.

Follow-up

How would you detect that the metric is being gamed rather than genuinely improving?
Which segment would you cut first, and what would that rule out?

What are the primary trade-offs when choosing between a long-term engagement metric and a short-term conversion metric?

Medium · product sense and metrics

Approach

Fix the population and the time window before naming any metric.
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

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 common experimentation pitfalls that could lead to false positives?

Medium · a/b testing and experimentation

Approach

Decide the analysis before seeing data, including how long it runs and when you look.
Name the guardrails that would stop a launch even on a positive primary result.
Say whether units interfere with each other, and switch design if they do.

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?

Walk me through the steps to set up an A/B test for a new pricing page.

Medium · a/b testing and experimentation

Approach

Say whether units interfere with each other, and switch design if they do.
Decide the analysis before seeing data, including how long it runs and when you look.
Name the guardrails that would stop a launch even on a positive primary result.

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?

If we notice a 5% drop in daily active users on our dashboard, how would you go about diagnosing the cause?

Medium

Approach

Work from the decision backwards to the evidence you would need.
Clarify what is being asked and what a complete answer would contain.
State your assumptions explicitly before working the problem.

Follow-up

What assumption would you test first?
How would you know your answer was wrong?

Estimate a threshold-triggered programme with regression discontinuity

Hard · regression discontinuity · instrumental variables · identification

Accounts reaching seats_licensed >= 25 in dim_account are automatically assigned a dedicated onboarding specialist; below 25 they are not. Leadership wants the programme's effect on 12-month net revenue retention and will not randomise coverage away from any account. Three years of dim_account and fct_subscription_period rows are available, including mrr_cents_constant_fx. Specify the design, the estimand it identifies, two threats that would invalidate it, and what changes when 8% of accounts below the threshold received a specialist anyway.

Approach

Set up a regression discontinuity on the running variable seats_licensed with a cutoff at 25, comparing accounts just below with accounts just above. The identifying assumption is continuity: absent the programme, expected 12-month NRR would be a continuous function of seat count through 25.
Name the estimand honestly and early. This is a local average treatment effect at 25 seats. It says nothing about a 5-seat or a 200-seat account, and that belongs in the first line of the answer rather than a footnote.
Estimate with local linear regression on each side, a triangular kernel, an MSE-optimal bandwidth and robust bias-corrected confidence intervals. Do not fit a high-order global polynomial; it imports weight from observations far from the cutoff and is known to manufacture discontinuities.
Test the two threats that actually apply. Manipulation: an account that wants the specialist can buy a 25th seat, which piles density just above the cutoff, so run a density test on seats_licensed around 25 and look for a spike at exactly 25. Bundling: if a price break, a plan tier or a support SLA also switches at 25 seats, the discontinuity measures the whole bundle, so check current_plan_tier and the price schedule at the cutoff.
Treat the 8% crossover as a fuzzy design. Treatment probability jumps at the cutoff without going from 0 to 1, so divide the jump in NRR by the jump in the probability of receiving a specialist. That is a Wald instrumental-variables estimator with the cutoff indicator as the instrument; it needs exclusion, which is exactly what the bundling check is about, plus monotonicity, and it narrows the estimand further to compliers at the cutoff.

Worked solution (45 min)

Fix the running variable as seats_licensed at the moment the assignment rule was evaluated, and fix the outcome as 12-month NRR computed from fct_subscription_period on mrr_cents_constant_fx for the account's cohort.
Plot mean NRR in one-seat bins on each side of 25 with a local linear fit. The picture comes before the estimate, because a discontinuity invisible in the binned plot is rarely real.
Run the density test at 25 and a continuity check on pre-cutoff characteristics such as billing_country, account_type and pre-programme MRR; these must be smooth through the cutoff.
Estimate the sharp RD with an MSE-optimal bandwidth and robust bias-corrected intervals, then repeat at half and double the bandwidth as a sensitivity check.
Estimate the first stage, the jump in specialist assignment at 25, and report the fuzzy estimate as the ratio with the estimand stated as the complier effect at the cutoff.

Expected result: A fuzzy regression discontinuity at 25 seats, estimated as the ratio of the NRR jump to the specialist-probability jump, identifying the local complier average effect at the cutoff under continuity, exclusion and monotonicity. With 8% crossover below the cutoff and near-full compliance above it, the first stage is about 0.92, so the fuzzy estimate is roughly 1.09 times the reduced-form jump and its standard error inflates by the same factor.

Verify it yourself

Characteristics fixed before the cutoff is crossed must not jump at 25. If account_type mix or billing_country jumps, the local comparison is not as-good-as-random.
The estimate should be stable across half and double the MSE-optimal bandwidth. Strong bandwidth sensitivity means the functional form is doing the work, not the discontinuity.
The reduced-form jump divided by the first-stage jump must reproduce the fuzzy point estimate exactly; if it does not, the two are being computed on different samples.
The effective sample inside the bandwidth, not the full account count, drives the standard error. Local linear RD typically needs on the order of three to four times the sample a randomised comparison would for the same power.

Follow-up

The density test shows a spike at exactly 25 seats. Is the design dead, and what would you do next?
You have 40,000 accounts but the bandwidth keeps 900. How does the detectable effect compare with a randomised comparison of the same nominal size?
Seat counts change over time. Which value of seats_licensed is the running variable, and what breaks if you pick the wrong one?

Separate a definition edit from a behaviour change

Easy · definition-change · metric-versioning · cohorts

Seven-day activation rate is documented as core actions on at least 2 distinct UTC dates inside the first seven days after account_created_at_utc, and every cohort published through last week was computed on that definition. The dashboard now reads 38% for the signup cohort that started Monday, against 31% for the cohort before it. Two things landed in that week: an onboarding change, and a commit to the metric's SQL that you have not read yet and that came with no documentation update. You have dim_user, fct_event and the metric's version history. Deliver two numbers: how many of the 7 points are behaviour and how many are the definition. Then say which published cohorts are no longer comparable.

Approach

Diff the metric SQL across the commit boundary and enumerate every changed predicate individually: the distinct-day threshold, the is_internal exclusion, the is_core_action source, the window bounds, and the identity key. A single commit routinely changes more than one, and each has its own sign.
Dual-run both definitions over the same 26 cohort weeks. The definition effect is the level shift between the two series computed on identical data, which isolates it from anything that happened in the world.
Read the behaviour effect only from the old-definition series, comparing the new cohort to its own trailing cohorts. The new-definition series cannot answer the behaviour question because it has no pre-period.
Explain the mechanism of the changed predicate rather than only its size. Lowering a distinct-day threshold from 2 to 1 admits every user who acted once in a single session, which is a large and low-intent population, so a several-point jump is the expected magnitude and not a surprise.
State the restatement plan: pick one definition, backfill the full history on it, and annotate the break date. A series carrying two definitions is not a time series.

Follow-up

The onboarding team wants the 1-day threshold kept because it is easier to move. What is your argument, and what would change your mind?
How would you make a definition change visible to every consumer of this metric without relying on people reading a changelog?

Suggested practice checklist

For someone who has spent the last year in notebooks, dashboards or modelling work and has not written raw SQL under time pressure. The first four days rebuild query fluency against a fixture you control and can verify by hand; the last three attach that fluency to the rest of the loop.

Day 1 — Build a fixture you can check answers against

Create a local Postgres or SQLite database with four tables (users, sessions, events, orders) holding roughly 200 rows you generated yourself, so you know the contents well enough to predict every result.
Deliberately seed the cases that break queries: a user with no sessions, a session with no events, two orders sharing a timestamp, a NULL in one join key, and one duplicated user row.
Before writing any SQL, hand-compute five answers on paper (how many users placed at least one order, median orders per ordering user, and three others) and save them as the ground truth for the week.

Deliverable: A one-command seed script plus a text file of five hand-computed answers to grade every later query against.

Day 2 — Joins, filters and NULL semantics

Answer "which users have no orders" three ways (LEFT JOIN with IS NULL, NOT EXISTS, NOT IN) and confirm that the NOT IN version returns zero rows once the subquery contains a NULL, because the comparison is never TRUE.
Reproduce the LEFT JOIN that silently collapses to an inner join by putting a right-table predicate in WHERE, then fix it by moving the predicate into the ON clause, and record both row counts.
Create a fan-out bug on purpose by joining orders to order_items and summing the order total, then correct it with a pre-aggregated subquery and explain in one line which table changed the grain.

Deliverable: One annotated .sql file holding the three join traps, each with the wrong result and the corrected result side by side.

Day 3 — Window functions and frames

Write three window queries against the fixture: a running order total per user, the rank of each order within its user by value, and the day gap to that user's previous order, then check each against the day-one ground truth.
Run ROW_NUMBER, RANK and DENSE_RANK over a column containing ties, print all three side by side, and write one sentence on when each is the correct choice.
Switch one query from the default frame (RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW, which is what you get when ORDER BY is present and no frame is written) to ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW, and explain why the output differs only when the ORDER BY column has duplicates.

Deliverable: Three verified window queries plus a short note explaining the RANGE versus ROWS difference in your own words.

Day 4 — The four analytical query patterns

Write a monthly retention grid: first order month per user, then months-since-first as the column, and verify that month zero equals the cohort size exactly.
Sessionize the events table under a 30-minute inactivity rule using LAG plus a cumulative sum over a new-session flag.
Build a four-step funnel that counts distinct users rather than events at each step, and state the rule you applied to a user who reaches step three without ever logging step two.

Deliverable: One file with the retention, sessionization and funnel patterns, each carrying a one-line note on the assumption it bakes in.

Day 5 — Write SQL the way you will have to write it live

Set a 12-minute timer and solve three medium prompts in a plain editor with no execution and no autocomplete, then run them and tally syntax errors separately from logic errors.
Narrate one solution aloud while writing it, stating the grain of each intermediate result (one row per user, one row per user-day) before you type its body.
Rewrite your slowest solution as a CTE chain where every CTE name states its grain, and time yourself re-solving it from blank.

Deliverable: A recording of one narrated solution plus an error tally that separates syntax from logic.

Day 6 — One day for everything that is not SQL

Write the preconditions of the two-sample t-test from memory, then check them: independent observations, and a difference in means whose sampling distribution is approximately normal, which at large sample sizes follows from the central limit theorem rather than from normality of the raw values.
Write the difference between an odds ratio from logistic regression and a relative risk, and state the condition under which the two are close (low outcome prevalence).
Prepare a 90-second answer to "how would you know this model is any good" that names the metric, the baseline you would beat, and the cost of the errors you care about.

Deliverable: One page of notes covering test preconditions, the odds-ratio caveat and the model-quality answer.

Day 7 — Full loop rehearsal

Run a 45-minute mock with someone willing to interrupt: 20 minutes of SQL, 15 minutes defining a metric, 10 minutes on a past project.
Re-solve from blank the two queries you were slowest on this week and compare the times against day five.
Write a five-line answer to "walk me through a project" that puts a number in the first sentence and names the decision the work changed.

Deliverable: Mock feedback notes plus a timed project narrative you can deliver without reading it.

Rigour against a deadline

Nearly every data role forces a trade between the analysis you want and the one that fits the decision window. Prepare a case where you deliberately shipped something less rigorous, named the weakness to the person relying on it, and said what would change your answer. The naming is the part interviewers listen for.

Tell me about a time you had to explain a complex technical insight to a non-technical stakeholder.

Medium · behavioural and stakeholder questions

Approach

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

Follow-up

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

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?

Handle a request for numbers supporting a decision already made

Hard · integrity · framing · stakeholder

A senior leader has already decided to sunset a plan tier and asks you for the analysis showing it is the right call. Accounts on that tier carry 6% of MRR at constant FX and have the highest licensed-seat utilisation in the book. The leader's support matters to your next review cycle, and the decision is being presented in four days. Deliver what you produce, what you decline to produce, and the exact sentence you will say in the meeting where the number appears on a slide.

Approach

Recognise what is being probed: whether you can find the legitimate request inside an illegitimate framing instead of either complying or refusing on principle. The generic answer promises to push back; the strong one produces something genuinely useful and states its limits in the room, without ambushing anybody.
Separate the decision from the justification. Sunsetting the tier may be correct for reasons the data does not hold, such as support cost, roadmap surface area or sales motion. What you decline is a one-sided document. What you produce is the case read both ways, which also happens to be more useful to the leader.
Build the symmetric analysis: MRR at risk at constant FX, the share of affected accounts with a plausible migration path given seats_licensed and billing_term, the recovery rate assumed for that migration and where it came from, and the downside case in which high-utilisation accounts treat the sunset as a reason to re-evaluate the vendor entirely.
Surface the inconvenient fact privately and early. The highest seat utilisation in the book is a retention signal, and the leader should hold it before the room does, so they can incorporate it rather than be caught by it.
Agree the meeting sentence in advance with the leader, so that nobody is surprised. Something to the effect that the tier is 6% of MRR and its accounts are the most heavily used in the book, and that the case for sunsetting rests on cost and focus rather than on revenue. That is true, it supports the decision on its real grounds, and it stops the deck claiming the numbers endorse it.
Decide your own line before you need it: what you will not put your name to, and that the route if asked anyway is your own manager rather than a confrontation in the meeting.

Follow-up

The deck circulates with your analysis included and the downside case removed. What do you do, and by when?
What changes if the honest analysis says the sunset is clearly the wrong call?
How do you write the same memo when the leader is your skip-level and the meeting is tomorrow?