Statisticians exist because the world speaks in noise and we want the signal.\nEvery measurement is contaminated — by sampling, by error, by how it was collected\n— yet decisions must be made: does this drug work, did the policy change anything? A\nstatistician quantifies how much we can trust a conclusion drawn from limited,\nimperfect data, and designs the collection so trust is warranted: the formal study\nof variation.
\n","wordCount":68},{"heading":"Core Mission","id":"core-mission","markdown":"Turn data into calibrated belief — estimates with honest uncertainty, and claims\nthat survive \"compared to what, and how could this be fooling me?\"","html":"Turn data into calibrated belief — estimates with honest uncertainty, and claims\nthat survive "compared to what, and how could this be fooling me?"
\n","wordCount":23},{"heading":"Primary Responsibilities","id":"primary-responsibilities","markdown":"The visible output is a number with an interval; the work determines whether it\nmeans anything. A statistician frames the question as an estimand before touching\ndata; designs studies so the comparison is fair and the effect identifiable;\nchooses models matching the data-generating process, not the ones that fit best;\nchecks assumptions and residuals; quantifies uncertainty through standard errors,\nintervals, or posteriors; and translates it for those who act. Much of the job is\ndefensive: spotting confounding and selection effects. Most value lands in design,\nwhere good randomization makes analysis trivial and a bad one impossible.","html":"The visible output is a number with an interval; the work determines whether it\nmeans anything. A statistician frames the question as an estimand before touching\ndata; designs studies so the comparison is fair and the effect identifiable;\nchooses models matching the data-generating process, not the ones that fit best;\nchecks assumptions and residuals; quantifies uncertainty through standard errors,\nintervals, or posteriors; and translates it for those who act. Much of the job is\ndefensive: spotting confounding and selection effects. Most value lands in design,\nwhere good randomization makes analysis trivial and a bad one impossible.
\n","wordCount":97},{"heading":"Guiding Principles","id":"guiding-principles","markdown":"- **The design is the analysis.** Most validity is fixed before collection; a\n confounded design cannot be rescued by a model. Randomization buys causal claims\n no regression can.\n- **Quantify uncertainty, always.** A point estimate without an interval is a guess\n in a lab coat. Deliver the estimate *and* its error.\n- **All models are wrong, some are useful.** (Box.) The model serves a purpose, not\n truth — ask which decision it serves.\n- **Distrust the data you didn't collect.** Every dataset came from some process —\n who was included, who dropped out, what got recorded. Reconstruct it.\n- **Significance is not importance.** A p-value answers \"would this surprise me\n under the null?\" — not \"does this matter?\" A tiny effect is significant with\n enough n, a large one nonsignificant with too few.\n- **Look at the data before you model it.** (Tukey.) Plot it, tabulate it, find the\n impossible values. Surprises live in the margins.\n- **Pre-specify, then explore — label which is which.** The garden of forking paths\n turns curiosity into false discovery; confirmatory and exploratory analyses claim\n under different licenses.","html":"Statisticians are almost always embedded in someone else's problem — a clinician's\ntrial, an economist's policy question, a product team's experiment. The most\nvaluable conversation happens before data collection, preventing an unanswerable\ndesign rather than apologizing for one. The recurring friction is the collaborator\nwho arrives with data collected and a conclusion desired. Good statisticians say\n"this design cannot answer that," then help reshape it. They translate both ways: a\nvague aim into an estimand, an interval into a layperson's decision.
\n","wordCount":80},{"heading":"Ethics","id":"ethics","markdown":"The statistician sits at the gate between data and belief, a position of easy abuse.\nThe duties: report the uncertainty, not just the estimate; disclose every analysis\nattempted, not only the one that worked; refuse to p-hack on request; resist letting\na desired conclusion drive the method. Confidentiality and re-identification risk are\nlive whenever the data are about people. There is a duty against false precision —\nfour decimals on a noisy number mislead as surely as a lie. The replication crisis\nwas an ethical failure: incentives rewarded surprising over true findings.","html":"The statistician sits at the gate between data and belief, a position of easy abuse.\nThe duties: report the uncertainty, not just the estimate; disclose every analysis\nattempted, not only the one that worked; refuse to p-hack on request; resist letting\na desired conclusion drive the method. Confidentiality and re-identification risk are\nlive whenever the data are about people. There is a duty against false precision —\nfour decimals on a noisy number mislead as surely as a lie. The replication crisis\nwas an ethical failure: incentives rewarded surprising over true findings.
\n","wordCount":93},{"heading":"Scenarios","id":"scenarios","markdown":"**A vaccine trial readout.** The team reports the vaccine \"significantly reduced\ninfections, p = 0.03.\" The statistician asks for the effect size and its interval:\nrelative risk reduction 12%, 95% interval 1% to 22% — barely excluding zero. Then\nthe harder questions: how many endpoints were tested, was this the pre-registered\nprimary, and did the arms differ at baseline because randomization was imperfect in\na small trial? The writeup reports \"a modest effect, imprecisely estimated,\nwarranting a confirmatory trial\" — not the certainty the p-value implies.\n\n**The \"training program works\" claim.** HR shows that employees who took an optional\nleadership course were promoted more often, and wants to mandate it. The statistician\nsketches the DAG: ambition drives both enrolling *and* getting promoted — a textbook\nconfounder, so the association is partly, maybe entirely, selection. The proposal: a\nrandomized pilot offering the course to a random half. If impossible, match on\npre-course performance plus a sensitivity analysis for how strong an unmeasured\nconfounder must be to erase the effect. Either way, the association is a hypothesis,\nnot a finding.\n\n**Surprising A/B test winner.** A product experiment shows a new checkout flow\nlifting conversion by 8%, hugely significant. The statistician distrusts it: was the\nrandomization unit correct (user, not session), did the test run a full weekly cycle\nto avoid day-of-week confounding, was this the only metric among forty? They find\nthe team peeked daily and stopped the moment it crossed significance — optional\nstopping that inflates the false-positive rate. The fix: rerun with pre-committed\nsample size and one look.","html":"A vaccine trial readout. The team reports the vaccine "significantly reduced\ninfections, p = 0.03." The statistician asks for the effect size and its interval:\nrelative risk reduction 12%, 95% interval 1% to 22% — barely excluding zero. Then\nthe harder questions: how many endpoints were tested, was this the pre-registered\nprimary, and did the arms differ at baseline because randomization was imperfect in\na small trial? The writeup reports "a modest effect, imprecisely estimated,\nwarranting a confirmatory trial" — not the certainty the p-value implies.
\nThe "training program works" claim. HR shows that employees who took an optional\nleadership course were promoted more often, and wants to mandate it. The statistician\nsketches the DAG: ambition drives both enrolling and getting promoted — a textbook\nconfounder, so the association is partly, maybe entirely, selection. The proposal: a\nrandomized pilot offering the course to a random half. If impossible, match on\npre-course performance plus a sensitivity analysis for how strong an unmeasured\nconfounder must be to erase the effect. Either way, the association is a hypothesis,\nnot a finding.
\nSurprising A/B test winner. A product experiment shows a new checkout flow\nlifting conversion by 8%, hugely significant. The statistician distrusts it: was the\nrandomization unit correct (user, not session), did the test run a full weekly cycle\nto avoid day-of-week confounding, was this the only metric among forty? They find\nthe team peeked daily and stopped the moment it crossed significance — optional\nstopping that inflates the false-positive rate. The fix: rerun with pre-committed\nsample size and one look.
\n","wordCount":262},{"heading":"Related Occupations","id":"related-occupations","markdown":"The statistician shares deep tooling and a love of uncertainty with several roles\nbut is defined by the formal study of variation itself. Data scientists apply the\nsame methods at scale, leaning toward prediction over inference. Mathematicians\nsupply the probability theory statistics rests on, caring about proof where\nstatisticians care about data. Actuaries are statisticians of risk in finance and\ninsurance. Epidemiologists are domain statisticians of disease, fluent in study\ndesign and causal inference. Machine learning engineers optimize prediction and\ntreat the bias–variance tradeoff operationally. Economists wield the same\ncausal-inference toolkit on policy.","html":"The statistician shares deep tooling and a love of uncertainty with several roles\nbut is defined by the formal study of variation itself. Data scientists apply the\nsame methods at scale, leaning toward prediction over inference. Mathematicians\nsupply the probability theory statistics rests on, caring about proof where\nstatisticians care about data. Actuaries are statisticians of risk in finance and\ninsurance. Epidemiologists are domain statisticians of disease, fluent in study\ndesign and causal inference. Machine learning engineers optimize prediction and\ntreat the bias–variance tradeoff operationally. Economists wield the same\ncausal-inference toolkit on policy.
\n","wordCount":95},{"heading":"References","id":"references","markdown":"- *Statistical Inference* — Casella & Berger\n- *Exploratory Data Analysis* — John Tukey\n- *Bayesian Data Analysis* — Gelman, Carlin, Stern, Dunson, Vehtari & Rubin\n- *All of Statistics* — Larry Wasserman\n- *Causality* — Judea Pearl\n- *The Elements of Statistical Learning* — Hastie, Tibshirani & Friedman","html":"