roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Prepare for statistics questions in data science and analytics interviews, from probability to experiment design. Covers inference and A/B testing, regression, Bayesian and causal methods and ML math.
750 of 750 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 750 new. 0 cut, outside your plan.
In most loops statistics is spread across rounds rather than confined to one. Data science and product analytics loops carry the most of it: a screen with a probability puzzle, a fundamentals conversation on inference and modeling, and a product case built around an experiment or a metric that moved. For machine learning roles the questions turn mathematical: why least squares has a closed-form solution, what an L2 penalty does to the coefficients.
The product case is where many loops are decided. Signups fell overnight: broken tracking, traffic shifting toward a weaker platform, or a real change in behavior? A test reads significant on day three: what does stopping it there do to the false positive rate? The treatment arm holds more users than the 50/50 split promised: what can you still trust? The fundamentals conversation tests recited definitions: what a p-value is a probability of, and why a positive result on a test for a rare condition usually still means you do not have it.
Seniority shows in what a candidate checks before believing a number: whether the unit of randomization matches the unit of analysis, whether users in one arm can influence the other, whether the lift will outlast its novelty. Where nobody could randomize, the stronger candidate names the design that stands in for it and the assumption it cannot test. Early in the level range, a correctly run test explained in plain words usually carries the round.
Descriptive statistics comes first because the product case usually starts there, with a biased sample or a metric that moved because the mix of users changed rather than their behavior. Probability follows with the distributions and the limit theorems the next stretch relies on every time it calls a statistic approximately normal.
Drawing Conclusions is the center of gravity. Inference builds the machinery, from sampling distributions and intervals to errors and power, and A/B testing then puts it to work, since sizing an experiment is a power calculation with a business decision attached. The experiment module continues past the readout into threats to validity and designs that need fewer users. Causal inference closes the stretch by asking what can still be said when nobody randomized, down to quasi-experimental designs such as difference-in-differences.
Modeling turns from deciding to estimating. Regression comes after causal inference so that reading a coefficient can lean on what confounding has already taught, and generalized linear models carry it to counts and yes-or-no outcomes. Bayesian statistics revisits the earlier questions with a prior and a posterior, giving the experiment readout a second vocabulary. Time series ends the stretch by dropping the independence earlier modules took for granted: seasonality, forecasts judged on data they never saw, and anomalies and changepoints in a metric watched over time.
Math for ML comes last because it serves a different round. Linear algebra explains least squares as a projection and principal components through the SVD; calculus and optimization explains how a model is fitted when no closed form exists, by following the gradient downhill.
When a product analytics or data science loop is close, give most of the hours to Data & Chance and Drawing Conclusions, read whole, since the case round and the screen both draw on them. The controls earn their place in the second half: importance cut to must know leaves the modeling and math modules their core and drops the specialist corners. With a few months, switch study order to by topics, because each idea here is built from the one before: intervals need sampling distributions, power needs both, and sizing an experiment needs power. Rehearse the case questions aloud, because a sizing calculation or the [diagnosis of a dropped metric](/topics/found-math-statistics-descriptive-metric-change-diagnosis) is judged on the reasoning you speak, not on the final number. Candidates trained in statistics usually hold the theory and lack the product framing, so their time goes to metric diagnosis and validity threats. Machine learning engineering candidates can begin at Math for ML and work back toward regression, since their round asks about gradients more often than p-values. The plan teaches the statistics and stops short of the learning algorithms built on it. Decision trees, ensembles and neural networks, and the SQL and Python a data science loop also tests, belong to the [AI & Data Scientist roadmap](/roadmaps/ai-data-scientist).
When a product analytics or data science loop is close, give most of the hours to Data & Chance and Drawing Conclusions, read whole, since the case round and the screen both draw on them. The controls earn their place in the second half: importance cut to must know leaves the modeling and math modules their core and drops the specialist corners.
With a few months, switch study order to by topics, because each idea here is built from the one before: intervals need sampling distributions, power needs both, and sizing an experiment needs power. Rehearse the case questions aloud, because a sizing calculation or the diagnosis of a dropped metric is judged on the reasoning you speak, not on the final number.
Candidates trained in statistics usually hold the theory and lack the product framing, so their time goes to metric diagnosis and validity threats. Machine learning engineering candidates can begin at Math for ML and work back toward regression, since their round asks about gradients more often than p-values.
The plan teaches the statistics and stops short of the learning algorithms built on it. Decision trees, ensembles and neural networks, and the SQL and Python a data science loop also tests, belong to the AI & Data Scientist roadmap.