Learning epidemiology from scratch: the statistics you need first

What epidemiology actually is

Epidemiology asks three questions of populations: who has the condition, why do they have it, and what would happen if we intervened? Its answers steer public-health policy, clinical guidelines, and regulatory decisions. Where clinical medicine treats one patient at a time, epidemiology treats the pattern — and it has developed a specialised vocabulary (incidence, prevalence, confounding, effect modification) to talk about that pattern precisely.

The two big challenges of epi are that (a) you usually can't randomise the exposure of interest (you can't randomly assign someone to be a smoker or to live in a particular ZIP code), and (b) real-world data are noisy, biased, and confounded. The discipline's methods exist to squeeze credible causal claims out of that reality.

A staged learning path

Epidemiology teaches best in three layers: the statistics, then the methodology, then the causal-inference machinery. Every layer needs the one below to make sense.

Layer 1 — statistical foundations

Before any epi method, be fluent in:

• Descriptive statistics — the essential first pass on any dataset. Every good epi paper opens with a Table 1.
• 2×2 tables — the atomic unit of epi. Chi-square and Fisher's exact for the test; the four cells give you the odds ratio, the risk ratio, the risk difference, and their CIs.
• Choosing the right test — start here for the general framework; every epi analysis is a specific instance of it.
• Logistic regression — the epi workhorse for binary outcomes with covariate adjustment. If you learn one regression model deeply, learn this one.
• Survival analysis — Kaplan–Meier for the curve, log-rank for the group comparison, Cox regression for hazard-ratio estimation with covariates. Essential for cohort studies with follow-up time.
• Confidence intervals and effect sizes — epi is an estimation discipline, not a p-value discipline. Report the OR/RR/HR and its 95% CI; the p-value is almost never the point.

Layer 2 — epidemiological methods

The vocabulary that turns statistics into epi:

• Study designs — cross-sectional, case-control, cohort (prospective or retrospective), ecological, RCT. Each answers a different question and has its own statistical footprint.
• Measures of frequency — incidence rate, cumulative incidence, prevalence. Knowing which one a paper reports (and why) is half the battle when reading epi literature.
• Measures of association — odds ratio (OR), risk ratio (RR), hazard ratio (HR), prevalence ratio, and their absolute counterparts (risk difference, attributable risk). Know when to use which — and why RR isn't OR unless the outcome is rare.
• Confounding and effect modification — the two most-confused concepts in epi. Confounding is a nuisance you adjust away; effect modification is a finding you report.
• Bias — selection bias, information bias (recall, interviewer, misclassification). Each has textbook examples and its own way of biasing the estimate toward or away from the null.
• Standardisation — direct and indirect age-standardisation, so mortality rates from populations with different age structures can be compared without the older one always looking worse.
• Screening and diagnostic-test evaluation — sensitivity, specificity, PPV, NPV, likelihood ratios, ROC curves. Foundational for any biomarker or diagnostic paper.

Layer 3 — causal inference and specialist tools

The modern epi frontier is causal inference — the toolkit for making credible causal claims from observational data:

• Directed Acyclic Graphs (DAGs) — the visual language for reasoning about which variables to adjust for and which not to (adjusting for a collider is a common way to introduce bias into a "controlled" analysis).
• Propensity scores — matching, weighting (IPTW), stratification. The workhorse for mimicking a randomised trial in observational data when a randomised trial is impossible.
• Target trial emulation — the Hernán-and-Robins framework that reframes every observational analysis as "what randomised trial would answer this question, and how do we imitate its design in the data we have?"
• Instrumental variables and difference-in-differences — econometric methods increasingly cross-pollinating with epi for policy evaluation.
• Meta-analysis and evidence synthesis — the tool for aggregating findings across studies (see also our note on learning meta-analysis).

The tools working epidemiologists actually use:

• R with epitools, epiR, survival, and MatchIt / WeightIt for propensity work. The open-source pillar.
• dagitty — free browser-based tool (and R package) for drawing DAGs and identifying the correct adjustment set.
• Stata — dominant in academic epi and health-services research; has the best complex-survey support (svy family).
• SAS — still common in large national surveillance systems (CDC, WHO) and pharmacoepi.

Where thericerca can help — and where it can't

Epi has a large "everyday" layer that any general stats platform handles well: 2×2 tables, logistic regression, Cox regression, group comparisons, sample size. It also has a specialist causal-inference layer that lives in dedicated tools.

Resources worth your time

• Books. Rothman, Greenland, Lash, Modern Epidemiology, is the field's bible — comprehensive and dense. Szklo & Nieto, Epidemiology: Beyond the Basics, is a friendlier second read. And the causal-inference book by Hernán & Robins, Causal Inference: What If, is free online (Harvard) and is the standard reference for the causal layer.
• Tutorials. The Society for Epidemiologic Research (SER) publishes free short-course materials. UCLA's OARC statistical consulting notes have concrete R walkthroughs for most standard epi analyses.
• Practice datasets. NHANES (freely downloadable US health survey), the Framingham Heart Study public-use file, and the Rossi recidivism dataset (in the lifelines R/Python packages; used in our recidivism sample study) are the standard teaching sets for cohort and survival analysis.
• Community. The r-epi mailing list, the American Journal of Epidemiology (AJE), and causalinf.org gather the working methodologists in one place.

A concrete first month

If you're starting with a working stats background, spend Week 1 on the 2×2 table until every measure (RR, OR, RD, prevalence ratio, PAR%) reads instantly. Week 2, work through logistic regression and confounding / effect-modification adjustment. Week 3, learn survival analysis end-to-end. Week 4, read Chapters 1–4 of Hernán & Robins and try one causal-inference exercise (a simple IPTW analysis on a real dataset).

thericerca can carry most of the applied work in that first month. Use the guided study pipeline for the 2×2 tables, logistic regressions, and survival analyses; use sample size for planning; and use the test-picker to double-check that (say) a paired McNemar is what you want for a matched case-control question, not a chi-square.

For the causal-inference layer — DAGs, propensity scoring, target trial emulation — you'll graduate to R's dedicated packages or specialised software. That's the way in, and there's no shortcut. But the two-by-two-table intuition and the regression fluency you build first are what keep you from being lost when you get there.