Epidemiology and Population Health
Epidemiology and Population Health tests your command of disease frequency, association, causation, and bias at the population level, plus the policy-facing tools: screening criteria, surveillance logic, and health equity measurement. Expect computations in a two-by-two table and conceptual questions about what a measure can and cannot support.
Measuring frequency and association
- Prevalence counts existing cases; incidence counts new cases per person-time and is the only sound basis for risk.
- Risk ratio comes from cohort data; odds ratio from case-control data or logistic regression.
- Attributable risk in the exposed equals incidence in exposed minus incidence in unexposed; the population attributable fraction estimates the share of disease removable by eliminating exposure.
- Crude rates mislead when age structures differ; direct standardization applies a common standard population.
Bias and causation
Selection bias, information bias (recall, interviewer, surveillance), and confounding are distinct problems requiring distinct remedies: design (restriction, matching, randomization) or analysis (stratification, multivariable adjustment). Effect modification is not confounding; it is a real difference in effect across strata and should be reported, not adjusted away. Bradford Hill viewpoints (temporality, strength, consistency, plausibility, dose-response, experiment, analogy) guide judgement, and only temporality approaches a necessary condition.
Worked example
In a cohort, 1,000 exposed workers have 50 cases and 1,000 unexposed have 10. Risk in exposed is 5 percent, in unexposed 1 percent, so RR = 5.0; attributable risk is 4 percentage points and the attributable fraction in the exposed is 80 percent. The odds ratio here is (50/950)/(10/990), about 5.2, close to the RR because the outcome is rare. Had the outcome been 40 percent, the OR would have overstated the RR substantially.
Screening and population health
Wilson-Jungner criteria require an important condition, accepted treatment, adequate facilities, a recognizable latent stage, a suitable test, and acceptable cost. Lead-time bias makes survival look longer without extending life; length bias favours detection of slow-growing disease; overdiagnosis inflates incidence without benefit. Herd immunity thresholds depend on the basic reproduction number: threshold = 1 - 1/R0.
Traps that cost marks
- Using prevalence as if it measured the risk of becoming ill.
- Adjusting for a mediator or a collider and thereby creating bias.
- Reporting a crude comparison across populations with different age structures.
- Confusing lead-time bias with length bias.
- Treating an ecological association as an individual-level effect.