Epidemiologic studies are grouped by whether the investigator assigns the exposure and whether the study compares groups.

Prospective vs. retrospective cohort. Both start with exposure and look for later disease. In a prospective cohort, the investigator follows people into the future; in a retrospective (historical) cohort, existing records define exposure in the past and outcomes up to the present (for example, occupational records of workers exposed to a solvent).

Strengths and limitations

Measures of association (2 × 2 layout)

• Relative risk (risk ratio, RR) = [a ÷ (a + b)] ÷ [c ÷ (c + d)] = incidence in exposed ÷ incidence in unexposed (cohort studies, RCTs).
• Odds ratio (OR) = (a × d) ÷ (b × c) = odds of exposure in cases ÷ odds of exposure in controls (case-control studies). When the disease is rare, OR approximates RR.
• Interpretation of RR or OR: 1 = no association; greater than 1 = exposure associated with higher risk (possible risk factor); less than 1 = lower risk (possible protective factor). If the 95% confidence interval includes 1, the association is not statistically significant.
• Attributable risk (AR, risk difference) = incidence in exposed − incidence in unexposed: the excess risk due to the exposure among the exposed; the amount of disease prevented if the exposure were removed.
• Attributable risk percent (AR%) = AR ÷ incidence in exposed × 100 = (RR − 1) ÷ RR × 100.
• Vaccine efficacy = (attack rate in unvaccinated − attack rate in vaccinated) ÷ attack rate in unvaccinated × 100 = (1 − RR) × 100. Measured in an RCT it is called efficacy; the same formula applied to real-world observational data is called vaccine effectiveness.
• Number needed to treat (NNT) = 1 ÷ absolute risk reduction.

Error and bias

• Random error: chance variation; reduced by larger sample size (improves precision, not validity).
• Selection bias: study groups differ systematically in how they were chosen (for example, hospital controls).
• Information (measurement) bias: includes recall bias (cases remember exposures more thoroughly than controls) and interviewer bias.
• Confounding: a third factor linked to both the exposure and the disease distorts the association (for example, smoking confounds the link between coffee drinking and lung cancer).
• Controlling confounding: randomization, restriction, matching (design stage); stratification and multivariable analysis (analysis stage). Increasing the sample size does not control confounding.
• Blinding (single, double) reduces information bias in trials.

Hierarchy of evidence (highest to lowest): systematic reviews and meta-analyses of RCTs → individual RCTs → cohort studies → case-control studies → cross-sectional and ecological studies → case reports, expert opinion.

Each calculation below was done twice.

Example 1 — Cohort study (RR, AR, AR%). 2,000 smokers and 4,000 nonsmokers, all free of the disease, are followed for 10 years.

• Incidence (exposed) = 60 ÷ 2,000 = 0.03 = 30 per 1,000
• Incidence (unexposed) = 20 ÷ 4,000 = 0.005 = 5 per 1,000
• RR = 0.03 ÷ 0.005 = 6.0 → smokers are 6 times as likely to develop the disease.
• AR = 30 − 5 = 25 per 1,000 smokers over 10 years.
• AR% = 25 ÷ 30 × 100 = 83.3%; check: (6 − 1) ÷ 6 × 100 = 83.3%. About 83% of the disease among smokers is attributable to smoking.

Example 2 — RR from rates. Lung cancer incidence is 150 per 100,000 among smokers and 10 per 100,000 among nonsmokers.

RR = 150 ÷ 10 = 15. AR = 150 − 10 = 140 per 100,000.

Example 3 — Case-control study (OR). 100 people with the disease (cases) and 200 without (controls) are asked about a past exposure.

• Odds of exposure in cases = 60 ÷ 40 = 1.5
• Odds of exposure in controls = 50 ÷ 150 = 0.333
• OR = 1.5 ÷ 0.333 = 4.5; check with cross-product: (60 × 150) ÷ (40 × 50) = 9,000 ÷ 2,000 = 4.5.

Interpretation: cases had 4.5 times the odds of past exposure compared with controls. Incidence and RR cannot be calculated, because the investigator chose how many cases and controls to include.

Example 4 — Vaccine trial (RCT). 2,000 people receive the vaccine and 2,000 receive placebo. Disease develops in 20 vaccinated and 100 placebo recipients.

• Attack rate, placebo = 100 ÷ 2,000 = 5%; vaccinated = 20 ÷ 2,000 = 1%
• RR = 1 ÷ 5 = 0.2 (protective)
• Vaccine efficacy = (5 − 1) ÷ 5 × 100 = 80%; check: (1 − 0.2) × 100 = 80%
• Absolute risk reduction = 5% − 1% = 4% = 0.04; NNT (number needed to vaccinate) = 1 ÷ 0.04 = 25

Choosing a design — quick guide

• Estimate how common a condition is now → cross-sectional.
• Rare disease, or an outbreak whose exposed population cannot be fully listed (community-wide) → case-control. Outbreak in a small, well-defined group with a complete list (wedding, banquet) → retrospective cohort (food-specific attack rates, RR).
• Rare exposure (an industrial chemical) or need for incidence and temporality → cohort.
• Test whether an intervention works → RCT (or community trial for community-wide interventions).
• Compare national cigarette consumption with lung cancer death rates → ecological.

• Read research critically: identify the design, the measure of association, and possible bias or confounding before applying findings.
• Use the right level of evidence for practice and policy (evidence-based public health nursing).
• Participate in data collection (surveys, interviews during outbreak case-control studies) using standardized questions to reduce bias.
• Protect participants: explain voluntary participation, confidentiality, and the right to withdraw.
• Translate findings for communities and decision makers in plain language (for example, "six times the risk").

• Federal Common Rule (45 CFR 46) governs federally funded human-subjects research: Institutional Review Board (IRB) review and informed consent. Public health surveillance activities conducted by a public health authority are excluded from the definition of research under the revised rule, but analytic studies that go beyond routine practice need review.
• Belmont Report principles: respect for persons (consent, protection of people with diminished autonomy), beneficence (maximize benefit, minimize harm), and justice (fair selection of participants).
• The US Public Health Service syphilis study at Tuskegee (1932–1972) is the classic example of unethical observational research; it shaped modern protections and continues to affect trust in some communities.
• RCTs require equipoise — genuine uncertainty about which arm is better; assigning a known harmful exposure is unethical, which is why smoking was studied with cohort and case-control designs.
• HIPAA allows research use of protected health information with authorization or an IRB/privacy board waiver.

Case 1 — Outbreak study. After a wedding with a complete guest list, 40 guests become ill. The nurse helps find the food responsible.

Correct design: a retrospective cohort study of all guests — classify everyone by what they ate and whether they became ill, then compare food-specific attack rates and RR (as in the communicable disease topic). Use a case-control study when the exposed population cannot be fully listed (for example, a community-wide outbreak linked to a grocery product).

Why: in a small, well-defined group, everyone can be classified by exposure and outcome, so incidence and RR can be calculated directly.

Case 2 — Interpreting OR. A study reports OR = 0.6 (95% CI 0.4–0.9) for daily physical activity and depression.

Correct interpretation: activity is associated with lower odds of depression, and the result is statistically significant because the interval does not include 1.

Why: OR less than 1 suggests a protective association.

Case 3 — Temporality question. A county survey finds that residents with diabetes exercise less. A board member says inactivity caused the diabetes.

Correct response: explain that a cross-sectional study cannot show which came first; a cohort study is needed.

Why: temporality is essential for causation.

• Calculating RR in a case-control study — use the OR.
• Calling a case-control study "prospective" — it starts with disease and looks back.
• Believing a bigger sample removes confounding; it only reduces random error.
• Treating ecological associations as individual-level facts (ecological fallacy).
• Assuming cross-sectional data prove cause.
• Misreading RR less than 1 as "no effect" — it indicates reduced risk.
• Confusing AR (difference) with RR (ratio).
• Ignoring confidence intervals: an RR of 2.0 with a CI of 0.8–5.0 is not significant.

• Cross-sectional → prevalence, no temporality; ecological → group data, ecological fallacy.
• Case-control → starts with disease, looks back; OR = (a × d) ÷ (b × c); best for rare diseases; recall bias.
• Cohort → starts with exposure, follows forward; incidence, RR, AR; clearest temporality among observational designs; best for rare exposures.
• RCT → random assignment; strongest evidence; controls known and unknown confounders.
• RR or OR = 1 no association; greater than 1 risk; less than 1 protective; CI including 1 = not significant.
• AR = Ie − Iu; AR% = (RR − 1) ÷ RR × 100; vaccine efficacy = (1 − RR) × 100; NNT = 1 ÷ ARR.
• Confounding is controlled by randomization, restriction, matching, stratification, multivariable analysis — not by sample size.
• IRB review and informed consent under the Common Rule; Belmont principles: respect for persons, beneficence, justice.