Interview Questions
Data Science Interview Questions and Answers
Data science interviews blend statistics fundamentals, SQL/data manipulation, and business judgment -- being able to explain a p-value clearly matters as much as being able to compute one.
Example: Reading an A/B test result
A new checkout page shows a 4% conversion rate vs. the old page's 3.6%, from a test with 10,000 users per variant, and the difference is statistically significant (p < 0.05). That means: if there were truly no real difference between the pages, seeing a gap this large by pure chance would happen less than 5% of the time -- so it's reasonable to conclude the new page genuinely performs better, not just that this particular sample got lucky.
Frequently Asked Questions
The probability of seeing a result at least as extreme as what you observed, assuming there's actually no real effect (the "null hypothesis" is true). A small p-value (conventionally < 0.05) suggests the observed result is unlikely to be pure chance -- but it's a probability statement about the data given the null hypothesis, not a direct statement about the probability the null hypothesis itself is true, which is a very common misinterpretation.
Two variables can move together (correlation) without one causing the other -- both could be driven by a third factor, or the relationship could be coincidental. Ice cream sales and drowning incidents both rise in summer (correlated), but ice cream doesn't cause drowning -- warm weather drives both. Establishing causation typically needs a controlled experiment (like an A/B test), not just observed correlation.
Define a single clear success metric upfront, randomly assign users to control and variant groups (randomization is what makes the comparison valid), calculate the sample size needed to detect a meaningful effect with reasonable statistical power before starting, run the test for a predetermined duration, and only then check statistical significance -- peeking at results early and stopping as soon as you see significance inflates the false-positive rate.
A Type I error is a false positive -- concluding there's an effect when there really isn't one. A Type II error is a false negative -- failing to detect a real effect that actually exists. There's generally a tradeoff between the two; reducing one (e.g. by requiring stronger evidence) tends to increase the other.
Depends on why it's missing and how much. Options include dropping rows/columns with missing values (fine if it's a small fraction and missingness is random), imputing a reasonable value (mean/median, or a value predicted from other features), or -- if the missingness itself is informative (e.g. a survey question people skip specifically when the true answer is unfavorable) -- treating "missing" as its own meaningful category rather than filling it in.
Considerable overlap, but generally: a data analyst focuses more on describing what already happened -- dashboards, reporting, exploratory analysis to answer specific business questions. A data scientist more often builds predictive models and does deeper statistical/ML work to answer "what will happen" or "what should we do," though the exact line varies a lot by company.
Aggregations (GROUP BY, HAVING), window functions (RANK, LAG/LEAD for period-over-period comparisons), joins across multiple tables, and writing correct, readable queries under time pressure using real business scenarios ("find the top 10% of customers by spend last quarter") rather than purely syntactic trivia.