Communicating Inferences from Epidemiologic Data

Core Concepts

Statistical inference: separating signal from chance

Studies observe a sample and make statements about the population it came from; that leap is statistical inference, and it is never certain. Two tools quantify the uncertainty. A p-value is the probability of observing a result at least as extreme as the one found, if the true effect were zero (the null hypothesis). A small p-value (conventionally below 0.05) says the result is unlikely under the null — it does not say the effect is large, important, or caused. A confidence interval (CI), usually 95%, is a range of plausible values for the true effect built from the data — roughly, if the study were repeated many times, 95% of such intervals would contain the true value. The CI is more informative than the p-value because it shows precision: a wide CI means an imprecise estimate, whatever the p-value says. For a relative risk, a CI that includes 1.0 means the data are compatible with no association. The critical judgment — statistical significance versus clinical or public-health significance — is human: a tiny effect can be "significant" in a huge study, and a meaningful effect can be "not significant" in a small one.

Causal inference: from association to causation

Association is not causation. An observed association may be real but still not causal. The classic framework for judging causality is the set of considerations associated with Austin Bradford Hill (often called Hill's criteria or viewpoints): temporality (exposure precedes outcome — the only truly essential one), strength (larger associations are harder to explain away), dose–response (more exposure, more effect), consistency (reproduced across studies and settings), specificity (one exposure, one outcome — helpful when present but weak when absent), plausibility (a believable mechanism), coherence (fits what else is known), experiment (evidence from interventions), and analogy (similar exposures behave similarly). These are considerations for judgment, not a checklist where checking boxes proves causation — no combination of them can replace a well-designed study.

Threats to valid inference: bias and confounding

When a study's result does not reflect the truth in the target population, the culprits are usually bias or confounding. Selection bias arises when the people studied differ systematically from the population of interest — e.g., a survey that only reaches people with phones, or a case-control study whose controls come from a different population than its cases. Information bias arises when data are measured or recalled inaccurately — e.g., recall bias in case-control studies, or misclassification of exposure or outcome; non-differential misclassification usually dilutes an effect, while differential misclassification can create or exaggerate one. Confounding occurs when a third variable is associated with both the exposure and the outcome and is not on the causal pathway — the classic example being a study of coffee and heart disease where smoking is more common among coffee drinkers; the smoking, not the coffee, may drive the association. Designs handle confounding through randomization, restriction, or matching; analysis handles it through stratification and adjustment. Confounding is distinct from effect modification (interaction), where the exposure's effect genuinely differs between subgroups (e.g., a drug works in younger people but not older people) — a real finding to report, not an error to remove.

Communicating numbers honestly: relative, absolute, and plain

The same finding can be told two honest ways that feel very different. Relative measures (RR, OR, "50% higher risk") express the change in risk between groups; absolute measures (risk difference, "from 2% to 3%") express the actual change in probability. Reporting only the relative number inflates perceived danger when the baseline risk is small, and reporting only the absolute number can bury a large effect when the baseline is high — good communication gives both, with the time frame and population the numbers apply to. Public-health messages should also respect the audience: use plain language and concrete comparisons ("about 1 extra case per 100 people over 10 years"), state uncertainty honestly without hedging into mush, and present numbers and graphs (epidemic curves, trend lines) with honest scales — a truncated y-axis can turn a flat line into a cliff. When the evidence is incomplete or conflicting, say so; trust is built on candor about limits.

The nurse as communicator: the message, the audience, the moment

Whether the setting is a clinic visit, a community meeting, or a news interview, the same discipline applies: know what the study can and cannot claim; lead with the bottom line; give the absolute and relative picture; name the uncertainty; and connect the numbers to what the audience can do. Write materials at the audience's literacy level, avoid jargon, and be ready to say "our understanding is still developing" — an honest statement, not an evasion.

Common Confusions