Data Analyst Interview Questions

First I confirm it's real, not an instrumentation break — check whether the tracking or a pipeline changed. Then I segment: by platform, geo, new vs returning, and time-of-day, to localize where the drop concentrates. I look for a coincident release or campaign change at that boundary. I form one hypothesis, validate it against a holdout or a prior period, then quantify the impact before recommending a fix. The discipline is: rule out measurement first, then segment to localize, then test one cause.

I pre-register the primary metric and the minimum detectable effect, size the test for power before it launches, and I don't peek-and-stop. I check the p-value alongside the confidence interval and the absolute effect size — statistical significance on a tiny effect isn't worth shipping. I once killed four 'winning' tests that were just under-powered noise once I enforced this.

S: Marketing's CAC didn't match finance's. T: Resolve without a turf war. A: I traced both definitions to source — different attribution windows and one included refunds. I proposed one canonical definition, documented it, and built the query both teams now use. R: Ended months of 'whose number is right' and the metric became trusted.

Bucket customers by their first-purchase month (the cohort), then aggregate revenue per cohort per subsequent month — a self-join or window function on a date-truncated order table. I'd use DATE_TRUNC for the cohort and activity months, COUNT/SUM grouped by both, and a window function (LAG) to compute the MoM delta. Then pivot cohort-month against activity-month for the retention triangle.

I start from the decision it's meant to support, not the data I have. One headline metric, a small number of drill-downs, clear definitions on hover, and no vanity charts. I sit with the users for the first week and cut anything they don't look at. My revenue dashboard got adopted because it answered finance's actual monthly question, not because it had more charts.

Correlation is what a query shows; causation needs a design — an experiment, a natural holdout, or at least a difference-in-differences with a credible control. When I can't run a true A/B, I'm explicit that the finding is directional and I name the confounders, rather than presenting a correlation as a proven driver.

S: Leadership assumed discounting drove our growth. T: Test it. A: I ran a cohort analysis isolating discounted vs full-price acquisition and tracked 6-month retention and LTV. Discounted cohorts churned far faster. R: We cut blanket discounts, reallocated to a referral program, and blended LTV:CAC improved within two quarters.

I triage by decision value and reversibility — what decision does this unblock, how big is it, and when's it actually made? A board number due Friday beats a 'nice to know' exploration. I publish a short queue so requesters see the trade-offs, and I templatize the recurring asks so they self-serve.