Data science interviews test three things at once: whether you actually understand statistics rather than reciting definitions, whether you can pull and shape data yourself with SQL and pandas, and whether you can turn a model into a decision someone will act on. Below are the questions asked most often, each with a model answer. Junior: core theory, definitions, and simple practical cases.
A population statistic describes everyone you care about; a sample statistic estimates it from a subset, and that estimation adds uncertainty a population parameter doesn't have. It shows up concretely in variance: dividing by n-1 instead of n for sample variance (Bessel's correction) corrects a bias that would otherwise systematically underestimate the true population variance.
Correlation means two variables move together; causation means one produces a change in the other. Ice cream sales and drowning deaths correlate because both rise in summer heat — neither causes the other. Treating a correlation as causal without a controlled comparison or a plausible mechanism is the single most common analytical mistake junior analysts make.
INNER JOIN keeps only rows that match on the join key in both tables; LEFT JOIN keeps every row from the left table and fills in NULLs where there is no match on the right. Picking INNER JOIN by default when you actually want LEFT JOIN silently drops rows — customers with no orders disappear entirely instead of showing up with a zero, which quietly biases any aggregate built on top.
A normal distribution is symmetric around its mean, with about 68% of values within one standard deviation and 95% within two — it underlies most classical statistical tests. Real business data is often not normal: revenue and session duration are right-skewed with a long tail, and counts of rare events follow a Poisson shape instead, which is why you check the distribution before assuming a t-test is valid.
.loc selects by label — the actual index or column name — while .iloc selects by integer position regardless of what the labels are. They diverge painfully after filtering or sorting a DataFrame, because the positional order no longer matches the original index; using .iloc[0] there gives you the first row of the current view, not "index 0."
P-hacking is running many analyses — different metrics, different cuts of the data, different time windows — until one crosses p < 0.05, then reporting only that one as if it were the single test you planned. It happens by accident when you check a dashboard daily and stop the moment a metric looks significant. The fix is pre-registering what you'll measure and when you'll look before the data comes in, not after.
In a real interview you speak under pressure. Cam asks these same questions, scores every answer, and shows exactly what to fix.
Practice a Junior Data Scientist interview →