Data Science interview questions and answers

Bias is systematic error of the expected estimate versus the truth

MLE chooses parameters that maximize the probability of the observed data under a specified model.

Heteroskedasticity means residual variance changes with the level of a predictor or group, violating the constant-variance assumption.

Robust statistics remain stable when a fraction of points are contaminated, such as the median, MAD, or Huber estimators.

The joint distribution describes two or more variables together.

The base-rate fallacy ignores prior prevalence when translating a sensitive test into a posterior.

Probability treats parameters as given and describes data

A Bernoulli trial is one success-or-failure draw

Use rank-based tests when the mean is not the estimand you care about, tails are heavy, or normality of the sampling distribution is doubtful in small samples.

Bonferroni splits alpha across m tests by using alpha/m for each, controlling the family-wise error rate under weak assumptions.

FDR is the expected share of rejected nulls that are false discoveries.

Paired tests use within-unit differences, such as before/after on the same user or matched markets, which removes stable unit-level noise and often raises power.

Prefer them when the estimand is a median or quantile, the metric is skewed with small n, or you distrust the analytic variance.

Classical p-values assume the sample size or stopping rule was fixed in advance.

They let you look at results continuously while still controlling Type I error, using martingale or mixture sequential probability ratio ideas.

A p-value conditions on the null and does not use a prior over effect sizes.

For a continuous test with a correctly specified null, p-values are uniform on 0 to 1.

CUPED is a variance-reduction technique that adjusts the outcome using a pre-experiment covariate, often the same metric before launch.

SRM is a statistically large deviation from the planned traffic split, detected with a goodness-of-fit test on assignment counts.

When treatment spills over, such as marketplace pricing, shared inventory, or social feed effects, user-level randomization is contaminated.