Epidemiology: principles, measures and study designs

In short

Epidemiology is the study of the distribution and determinants of health-related states in defined populations, and the application of that study to control health problems. It supplies the numbers behind every public health decision — who is getting ill, how fast, why, and whether an intervention works. It is the core paper of BPH/B.Sc Public Health degrees and is examined in MPH entrance tests, ICMR/NIE and AIIMS-CMC research exams, and the UPSC/state public-health service papers. Employers test it in interviews for roles in NHM programme units, ICMR institutes, clinical research organisations (CROs), health insurers, Gates/PATH/Jhpiego-type NGOs, and surveillance teams such as IDSP/IHIP and the National Centre for Disease Control.

Explained

Principles, concepts and scope

• Definition (John Last/IEA): "the study of the distribution and determinants of health-related states or events in specified populations, and the application of this study to the control of health problems." Three elements: distribution (descriptive), determinants (analytical), application (control).
• Aims: describe disease burden; identify causes and risk factors; study natural history; evaluate interventions and health services; provide the basis for policy and planning.
• Population, not patient: the clinician asks "what has this patient got"; the epidemiologist asks "how much of it is there, in whom, why, and is it rising".
• Epidemiological triad: agent, host, environment, held in balance; disease results when the balance shifts. Extended into the web of causation for multifactorial (chronic) disease and the wheel model (host genome at the centre, biological/physical/social environment around it).
• Natural history of disease: pre-pathogenesis (agent, host, environment interacting) → pathogenesis: subclinical (incubation/induction) → clinical → outcome (recovery, disability, death). Iceberg phenomenon: clinically evident cases are the tip; carriers, subclinical and latent infections lie submerged.
• Levels of prevention: primordial (prevent emergence of risk factors — e.g. preventing tobacco uptake in adolescents); primary (health promotion, specific protection — immunisation, iodised salt); secondary (early diagnosis and treatment — screening for cervical cancer, HbA1c in NCD clinics); tertiary (disability limitation and rehabilitation). Levels of prevention map to the natural history stage at which you intervene.
• Spectrum of disease frequency: sporadic, endemic (constant usual presence), hyperendemic, epidemic/outbreak (occurrence clearly in excess of expectation), pandemic. (Outbreak field methods belong to communicable disease epidemiology.)
• Disease surveillance: continuous scrutiny of occurrence and spread. In India: IDSP, now delivered through the Integrated Health Information Platform (IHIP), plus programme-specific systems (Ni-kshay for TB, NVBDCP/NCVBDC for vector-borne disease, HMIS and NFHS for population indicators).
• Legal and ethical frame: notification of specified diseases is a state function under state public health Acts and the Epidemic Diseases Act, 1897; the National Health Policy 2017 is the current policy document. Human research must follow the ICMR National Ethical Guidelines for Biomedical and Health Research Involving Human Participants (2017), the New Drugs and Clinical Trials Rules, 2019 (CDSCO) and registration on the Clinical Trials Registry – India (CTRI). Personal health data: until 13 May 2027 the IT Act s.43A and the SPDI Rules apply; the Data Fiduciary duties under the DPDP Rules, 2025 (notice, consent, breach intimation, erasure) take effect from that date.

Measures of disease frequency

• Ratio (a/b, unrelated numerator and denominator), proportion (a/(a+b), numerator part of denominator), rate (proportion with time in the denominator).
• Incidence rate (cumulative incidence / risk): new cases in a period ÷ population at risk at the start. Dimensionless over a stated period; measures risk.
• Incidence density (person-time rate): new cases ÷ total person-time at risk. Units: per person-year (per 1,000 PY). Used when follow-up is unequal or people enter and leave.
• Attack rate: a cumulative incidence used for a short, sharply defined epidemic, expressed as a percentage; secondary attack rate = cases among contacts of a primary case ÷ susceptible contacts.
• Point prevalence: existing cases (old + new) at a point in time ÷ population examined. Period prevalence uses a span.
• Relationship: Prevalence ≈ Incidence × mean Duration (P = I × D) when both are stable and prevalence is low. A disease can have high prevalence either because it is common or because it lasts long — so prevalence rises if treatment prolongs life without curing (HIV on ART, diabetes).
• Use: incidence for aetiology, risk and evaluating prevention; prevalence for planning beds, drugs, staff and workload.
• Mortality measures: crude death rate, age-specific death rate, cause-specific mortality rate, case fatality rate (deaths among diagnosed cases — a measure of severity/virulence), proportional mortality ratio, infant mortality rate, maternal mortality ratio. (Detailed vital rates and life tables belong to demography and vital statistics.)
• Standardisation: removes the effect of differing age structures when comparing populations. Direct (apply the study population's age-specific rates to a standard population → age-adjusted rate) and indirect (apply a standard population's rates to the study population → expected deaths → SMR = observed/expected × 100).
• Composite burden measures: DALY = YLL + YLD; QALY; used in the India State-Level Disease Burden Initiative (ICMR–PHFI–IHME) and in health technology assessment (HTAIn).

Measures of association, effect and impact

• Relative risk (risk ratio, RR): incidence in exposed ÷ incidence in unexposed. Estimable only where incidence is measurable — cohort studies, RCTs. RR = 1 no association; >1 risk factor; <1 protective.
• Odds ratio (OR): ad/bc from a 2×2 table; the measure of a case-control study. Approximates RR when the disease is rare (<~10%); otherwise it exaggerates the RR away from 1.
• Attributable risk (risk difference, AR): Ie − Iu — the excess incidence caused by exposure; basis for absolute benefit.
• Attributable risk percent (AR%) = (Ie − Iu)/Ie × 100 — the proportion of disease in the exposed attributable to exposure (aetiologic fraction in the exposed).
• Population attributable risk (PAR) = Ipop − Iu; PAR% = Pe(RR − 1)/[Pe(RR − 1) + 1] × 100 — depends on both strength of association and prevalence of exposure. A weak risk factor that is very common (e.g. ambient air pollution) can carry a larger population burden than a strong rare one.
• Relative risk reduction = 1 − RR; absolute risk reduction (ARR) = Iu − Ie in a trial; number needed to treat (NNT) = 1/ARR; number needed to harm = 1/attributable risk.
• Hazard ratio from survival (Cox) analysis — the ratio of instantaneous event rates; reported in cohort studies and trials with time-to-event outcomes.
• Precision and inference: a 95% confidence interval for RR or OR that includes 1 is not statistically significant; interpret width (precision) along with the point estimate. (Significance testing itself belongs to biostatistics.)

Study designs

• Classification: observational (descriptive and analytical) versus experimental (interventional); and by unit of observation — individual versus population (ecological).
• Descriptive studies: describe disease by time (secular trend, cyclic, seasonal, short-term fluctuation), place (international, national, rural–urban, local clustering) and person (age, sex, occupation, SES, marital status, migration). Outputs: hypotheses, spot maps, burden estimates. Include case reports, case series and cross-sectional surveys.
• Cross-sectional (prevalence) study: exposure and outcome measured at the same time in a sample. Cheap, fast, gives prevalence and health service needs; cannot establish temporal sequence, and misses short-duration and rapidly fatal cases. Indian examples: NFHS-5, National NCD Monitoring Survey, GATS-2.
• Ecological study: the unit is a group (district, state); correlates aggregate exposure with aggregate outcome. Cheap, uses routine data; prone to the ecological fallacy — a group-level association need not hold for individuals.
• Case-control (retrospective) study: start with cases and suitable controls, look back at exposure. Steps: define and ascertain cases → select controls (hospital, neighbourhood, or matched) → measure past exposure identically in both → build the 2×2 table → compute OR with CI. Good for rare diseases, cheap, quick, multiple exposures; limits — recall bias, selection of controls, no incidence or RR, poor temporality. Variants: nested case-control (a case-control study within a cohort), case-cohort, incidence-density (risk-set) sampling, case-crossover.
• Cohort (prospective/longitudinal/incidence/follow-up) study: start with exposed and unexposed free of disease, follow for incidence. Types: prospective, retrospective (historical) using past records, and ambispective. Yields incidence, RR, AR, multiple outcomes, clear temporality; limits — large n, long time, costly, loss to follow-up, unsuited to rare diseases.
• Randomised controlled trial (RCT): the design closest to an experiment. Steps: protocol → define reference and experimental population → eligibility and written informed consent → randomisation (simple, block, stratified, cluster) → intervention vs control (placebo or standard care) → blinding (single, double, triple) → follow-up → analysis by intention to treat (preserves randomisation) with per-protocol as secondary. Randomisation controls both known and unknown confounders; blinding controls information bias. Phases I–IV of drug trials; CONSORT for reporting.
• Other experimental designs: field trials (healthy individuals, e.g. a vaccine trial), community trials (whole communities, e.g. salt-fluoridation or IEC campaigns), cluster-randomised and stepped-wedge designs, crossover trials; quasi-experimental where randomisation is impossible.
• Hierarchy of evidence (ascending): case report → case series/ecological → cross-sectional → case-control → cohort → RCT → systematic review and meta-analysis of RCTs. Reporting guidelines: STROBE (observational), CONSORT (trials), PRISMA (reviews).

Error, bias, confounding and causation

• Random error reduces precision (fixed by larger sample size); systematic error (bias) reduces validity (not fixed by sample size).
• Selection bias: the way participants get in distorts the association — Berkson's (hospital admission) bias, healthy worker effect, non-response bias, loss to follow-up, prevalence–incidence (Neyman) bias.
• Information/measurement bias: recall bias (cases remember exposure better — classic in case-control studies), interviewer bias, misclassification (differential vs non-differential), observer bias, lead-time and length bias in screening.
• Confounding: a third factor associated with the exposure, an independent risk factor for the outcome, and not an intermediate step. Control at design stage — randomisation, matching, restriction; at analysis stage — stratification (Mantel–Haenszel), standardisation, multivariable regression, propensity scores.
• Effect modification (interaction) is not bias — the effect genuinely differs across strata and must be reported stratum-wise.
• Validity: internal (the result is correct for the study population) before external (generalisability).
• Bradford Hill considerations for causation: temporality (the only essential one), strength of association, dose–response, consistency, specificity, biological plausibility, coherence, analogy, experimental evidence. Association ≠ causation.
• Screening test validity: sensitivity (ability to detect disease), specificity (ability to exclude), PPV/NPV (depend on prevalence), accuracy, Youden's index, ROC curve and area under it, kappa for agreement/reliability. Screening criteria follow the Wilson–Jungner principles; India screens for hypertension, diabetes, and oral, breast and cervical cancer for those aged 30+ under the NP-NCD (formerly NPCDCS) through Ayushman Arogya Mandirs.

Key points

• Epidemiology - distribution, determinants and control of health-related states in populations.
• Incidence - new cases per population at risk per time; the measure of risk and of aetiology.
• Prevalence - existing cases at a point; the measure of burden for service planning. P = I × D.
• Attack rate - cumulative incidence in a short, well-defined epidemic, given as a percentage.
• Case fatality rate - deaths ÷ diagnosed cases; measures severity, not population risk.
• Relative risk - Ie/Iu; from cohort studies and trials only.
• Odds ratio - ad/bc; the case-control measure; approximates RR only for rare disease.
• Attributable risk - Ie − Iu, the excess due to exposure; PAR% also depends on exposure prevalence.
• NNT - 1/ARR; how many must be treated for one extra good outcome.
• Standardisation - removes age-structure differences; indirect method yields the SMR.
• Cross-sectional study - prevalence and needs; no temporality, so no causal inference.
• Case-control - best for rare disease and quick answers; vulnerable to recall and selection bias.
• Cohort - gives incidence, RR and temporality; costly, slow, loses people to follow-up.
• RCT - randomisation balances unknown confounders; analyse by intention to treat.
• Ecological fallacy - group-level associations need not apply to individuals.
• Confounding - controlled by randomisation, matching, stratification or regression; effect modification is reported, not removed.
• Bradford Hill - temporality is essential; strength, dose–response and consistency strengthen a causal claim.
• Sensitivity and specificity - properties of the test; PPV and NPV move with prevalence.

Formulas and examples

Example 1 — incidence, prevalence and person-time. A village has 5,000 people. On 1 January 2026, 150 have hypertension. During 2026, 90 new cases are detected among the 4,850 at risk; no one dies or leaves.

• Cumulative incidence = 90 ÷ 4,850 = 0.01856 → 18.6 per 1,000 per year.
• Point prevalence on 31 December = (150 + 90) ÷ 5,000 = 0.048 → 48 per 1,000.
• If instead only 3,000 person-years of follow-up were accumulated for those 90 cases, incidence density = 90 ÷ 3,000 = 30 per 1,000 PY.
• Check with P = I × D: if mean duration is 20 years and incidence 0.0186/yr, P ≈ 0.372 — far above the observed 0.048, which signals that detection is incomplete (the iceberg), a reasonable exam comment.

Example 2 — cohort study: RR, AR, PAR%. 10,000 smokers and 20,000 non-smokers followed 10 years for lung cancer: 120 cases among smokers, 40 among non-smokers.

• Ie = 120/10,000 = 12 per 1,000; Iu = 40/20,000 = 2 per 1,000.
• RR = 12 ÷ 2 = 6.0 — smokers have six times the risk.
• AR = 12 − 2 = 10 per 1,000 over 10 years.
• AR% = (10 ÷ 12) × 100 = 83.3% of lung cancer in smokers is attributable to smoking.
• Pe = 10,000/30,000 = 0.3333. PAR% = [0.3333 × 5] ÷ [0.3333 × 5 + 1] × 100 = 1.6667/2.6667 × 100 = 62.5%.
• Cross-check from rates: Ipop = 160/30,000 = 5.333 per 1,000; PAR = 5.333 − 2 = 3.333; PAR% = 3.333/5.333 = 62.5% ✓.

Example 3 — case-control study and the rare-disease assumption. Oral cancer cases 200 (160 tobacco chewers), controls 400 (120 chewers).

• Following the 2×2 convention above (a exposed-diseased, b exposed-healthy, c unexposed-diseased, d unexposed-healthy): a = 160, b = 120, c = 40, d = 280. OR = (160 × 280) ÷ (120 × 40) = 44,800 ÷ 4,800 = 9.33.
• Attempting RR on these numbers gives (160/280) ÷ (40/320) = 0.571/0.125 = 4.57 — wrong, because in a case-control study the number of controls is fixed by the investigator, so a/(a+b) is not an incidence.
• Exposure attributable fraction in cases ≈ (OR − 1)/OR = 8.33/9.33 = 89.3%.

Example 4 — RCT: ARR, RRR, NNT. In a trial of a new antihypertensive, strokes occur in 40/2,000 on the drug and 70/2,000 on placebo over 5 years.

• Ie = 0.020; Iu = 0.035. ARR = 0.035 − 0.020 = 0.015 (1.5 percentage points).
• RR = 0.020/0.035 = 0.571; RRR = 1 − 0.571 = 42.9%.
• NNT = 1 ÷ 0.015 = 66.7 → 67 people treated for 5 years to prevent one stroke.

Example 5 — screening test validity. A rapid diabetes screen is applied to 1,000 people; 100 truly have diabetes. It flags 90 of them positive and also flags 135 of the 900 non-diabetics.

• TP 90, FN 10, FP 135, TN 765. Sensitivity = 90/100 = 90%; Specificity = 765/900 = 85%.
• PPV = 90/(90 + 135) = 90/225 = 40%; NPV = 765/(765 + 10) = 765/775 = 98.7%.
• Same test in a low-prevalence population (1% of 10,000 = 100 cases among 10,000): TP 90, FP = 0.15 × 9,900 = 1,485 → PPV = 90/1,575 = 5.7% — the test did not change, prevalence did.

Where it is used

• National health programmes: NP-NCD, NTEP (TB), NVBDCP/NCVBDC and the immunisation programme use incidence, prevalence, coverage and attack rates for target setting; district and state programme officers and NHM consultants read these tables daily.
• Surveillance and response: IDSP/IHIP units, NCDC, state and district surveillance officers, and India EIS/field epidemiology training officers compute attack rates and test hypotheses on outbreak data.
• Research institutes and surveys: ICMR institutes (NIE Chennai, NIRT, NIN), PHFI, IIPS Mumbai (NFHS), AIIMS and St John's — designing cohorts, case-control studies and prevalence surveys.
• Clinical research industry: CROs and pharma (Syneos, IQVIA, Parexel, Sun, Cipla, Serum Institute) hire for protocol design, randomisation, biostatistics support and CTRI/CDSCO regulatory submissions under the New Drugs and Clinical Trials Rules, 2019.
• Insurance, HTA and hospital quality: risk scoring and claims analytics in health insurers; HTAIn and state HTA cells use DALYs and cost-effectiveness; hospital infection-control committees compute device-associated infection rates per 1,000 device-days.
• Development sector and global health: Gates Foundation, PATH, Jhpiego, WHO India and state Technical Support Units — monitoring and evaluation, baseline–endline surveys and quasi-experimental programme evaluations.

Interview questions

• Q: Differentiate incidence and prevalence, and say when you would use each. A: Incidence counts new cases per population at risk per unit time and measures risk — used for aetiology and evaluating prevention; prevalence counts all existing cases at a point and measures burden — used for planning staff, drugs and beds.
• Q: Why can't you calculate relative risk from a case-control study? A: The investigator fixes how many cases and controls are enrolled, so the denominators are not populations at risk and no incidence can be derived; the odds ratio is used instead.
• Q: When does the odds ratio approximate the relative risk? A: When the disease is rare in the population (roughly under 10% incidence), so that b ≈ a + b and d ≈ c + d; for common outcomes the OR overstates the effect.
• Q: What is confounding and how do you handle it? A: A third variable associated with the exposure and independently causing the outcome without being an intermediate step; controlled by randomisation, matching or restriction at design, and by stratification, standardisation or multivariable regression at analysis.
• Q: How is effect modification different from confounding? A: Confounding is a distortion to be removed; effect modification is a real difference in effect between strata and must be reported stratum-wise rather than adjusted away.
• Q: Why is intention-to-treat analysis preferred in a trial? A: It analyses everyone in the group they were randomised to, preserving the comparability randomisation created and giving an unbiased, pragmatic estimate of effect; per-protocol analysis reintroduces selection by compliance.
• Q: A screening test has 95% sensitivity and 95% specificity but most positives turn out healthy — explain. A: Positive predictive value depends on prevalence; in a low-prevalence population the small number of true positives is swamped by false positives drawn from a very large healthy group.
• Q: Which Bradford Hill criterion is indispensable? A: Temporality — the exposure must precede the outcome; every other consideration strengthens or weakens a causal argument but none is by itself necessary or sufficient.
• Q: What is the ecological fallacy? A: Inferring an individual-level association from group-level averages — for example concluding that fat-eating individuals get heart disease because countries with high mean fat intake have high mean mortality.
• Q: Explain attributable risk percent versus population attributable risk percent. A: AR% is the fraction of disease among the exposed caused by the exposure; PAR% is the fraction in the whole population that would disappear if the exposure were removed, so it also depends on how common the exposure is.
• Q: What is NNT and why do clinicians like it? A: The number needed to treat, 1/absolute risk reduction, is the number of patients who must receive the intervention for one additional good outcome — it expresses benefit in absolute, decision-relevant terms, unlike relative risk reduction.
• Q: Which study design would you choose for a rare cancer with a long latency, and why? A: A case-control study — a cohort would need an enormous sample and decades of follow-up, whereas starting from existing cases gives an answer quickly and cheaply, accepting the risk of recall bias.
• Q: Why does prevalence rise when a disease becomes better treated but not cured? A: Because P = I × D: survival lengthens the duration of illness, so cases accumulate even if incidence is unchanged or falling.

Common mistakes

• Using prevalence data to claim causation, or reporting a prevalence odds ratio from a cross-sectional study as if it were a relative risk.
• Computing RR from a case-control table, or calling ad/bc a "relative risk".
• Confusing case fatality rate with cause-specific mortality rate — the denominators differ (diagnosed cases vs total population).
• Omitting the time period or the "per 1,000" multiplier, and mixing cumulative incidence with person-time incidence in the same comparison.
• Quoting sensitivity and specificity as if they predicted an individual's disease status, and ignoring prevalence when interpreting PPV.
• Thinking a larger sample size corrects bias — it reduces random error only; and treating matching as a cure-all when over-matching destroys the ability to study the matched variable.
• Reporting a statistically significant association as clinically important without the absolute risk difference, NNT or confidence interval.
• Dropping non-compliant participants from a trial analysis, or forgetting that loss to follow-up in a cohort is a selection bias, not just a smaller n.

Related topics

• Biostatistics for health sciences · Communicable disease epidemiology and outbreak investigation · Non-communicable diseases: epidemiology and prevention · Demography and vital statistics · Screening and health surveillance systems · Research methodology and health programme evaluation