Epidemiological Study Designs, Bias, Confounding and Causation: Cohort vs Case-Control vs RCT (NEET PG PSM)
Start Here — Pick the Right Design
🎯 Exam priority: High-yield (Tier 1). Questions here are short and conceptual: which design for a rare disease, which measure comes out of which design, what a bias or confounder is, and which is the essential criterion for causation. Learn the grid below until you can rebuild it.
CBME mapping: CM 7. Facts checked against WHO's Basic Epidemiology and the CDC's Principles of Epidemiology.
Observational vs Experimental
Observational studies — the investigator only observes exposure and disease. The common types are cohort, case-control and cross-sectional studies; ecological studies use groups as the unit.
Experimental (intervention) studies — the investigator decides the exposure for each person (clinical trial) or community (community trial) and follows them over time. A study that merely assesses exposure and follows people forward is a cohort, not a trial, however long it runs.
Design | Unit / direction | Best for | Key limitation |
Ecological (correlational) | Groups, not individuals | Generating hypotheses cheaply from routine data | Ecological fallacy: a group-level link may not hold for individuals |
Cross-sectional | Individuals; exposure and outcome measured at the same time | Prevalence; planning services | Cannot tell whether exposure preceded the effect |
Case-control | Starts from disease (cases vs controls), looks back at exposure | Rare diseases; many exposures; quick and cheap | Recall and selection bias; gives OR, not incidence or RR |
Cohort | Starts from exposure, follows forward to disease | Best observational evidence on causation; direct measure of risk; rare exposures | Long, costly; loss to follow-up |
Randomized controlled trial | Investigator assigns the intervention at random to patients | Treatment efficacy; strongest evidence | Ethics, cost, generalisability |
Field trial | Healthy people presumed to be at risk, studied 'in the field' | Preventive interventions (vaccines) | Very large numbers, expensive |
Community trial | Whole communities receive the intervention | Interventions that can only be delivered to a whole community | Few units; harder to randomize |
Case-Control and Cohort, Side by Side
Case-control: start with cases (people with the disease) and controls (without it); compare how often each group was exposed. Efficient for rare diseases. Measure: odds ratio. Called "retrospective" because inquiry runs from disease back to cause — though data collection itself may continue forward.
Cohort: start with exposed and unexposed people free of the disease; follow them to see who develops it. Gives incidence, relative risk and attributable risk, and WHO calls it the best observational source of information on causation. Good for rare exposures (an occupational chemical, the Bhopal gas leak).
Cross-sectional studies measure exposure and effect at the same moment — useful for prevalence, weak on causation because the time order is unknown.
Errors: Random, Systematic and Confounding
WHO's comparison of the four observational designs:
Probability of… | Ecological | Cross-sectional | Case-control | Cohort |
Selection bias | not applicable | medium | high | low |
Recall bias | not applicable | high | high | low |
Loss to follow-up | not applicable | not applicable | low | high |
Confounding | high | medium | medium | medium |
Time required | low | medium | medium | high |
Cost | low | medium | medium | high |
Error | What it is | How to reduce it |
Random error | A sample value differs from the true value by chance alone (biological variation, sampling error, measurement error) | Larger sample size; more precise measurement |
Selection bias (systematic) | People in the study differ systematically from those not in it (e.g. self-selection, 'healthy worker' effect) | Careful choice of study population and controls |
Measurement (information) bias | Exposure or disease is measured or classified inaccurately; includes recall bias | Standardised tools, validation, blinding |
Confounding | A third factor linked to both the exposure and the disease distorts the association | Randomization, restriction, matching; stratification or statistical modelling at analysis |
Blinding: in a blind study the investigators do not know how participants are classified; in a double-blind study neither the investigators nor the participants know. Blinding reduces measurement bias.
Validity is the degree to which a study or test measures what it is meant to — no systematic error, small random error.
Controlling confounding: at the design stage by randomization (WHO calls it the ideal method in experimental studies, as it spreads potential confounders evenly between groups), restriction (e.g. study only non-smokers) and matching (e.g. match each case to a control of the same age and sex); at the analysis stage by stratification and statistical modelling.
Is It Cause or Mere Association?
Consideration (WHO, after Bradford Hill) | Question to ask |
Temporal relationship | Does the cause precede the effect? (essential) |
Plausibility | Is the association consistent with other knowledge (mechanism, animal evidence)? |
Consistency | Have similar results been shown in other studies? |
Strength | How strong is the association (relative risk)? An RR above 2 can be considered strong. |
Dose–response relationship | Does more exposure bring more effect? |
Reversibility | Does removing the cause reduce the risk? |
Study design | Is the evidence based on a strong design? |
Judging the evidence | Weigh all the above together |
Only the temporal relationship is essential: a cause must come before its effect. The other considerations add weight but none is obligatory.
Note for the exam: many textbooks teach Bradford Hill's original nine "viewpoints" (1965). WHO's list above is a condensed version of the same ideas; we have not reproduced Hill's own wording because the paper is not openly licensed.
Traps the Examiner Sets
Rare disease → case-control; rare exposure → cohort.
Relative risk and incidence need a cohort or trial; a case-control study gives only the odds ratio.
The ecological fallacy belongs to ecological studies — group-level association applied to individuals.
Recall bias is a form of measurement bias typical of case-control studies.
Field trials are on healthy people (vaccines); clinical trials are on patients.
Temporality is the one essential criterion for causation.
Rapid-Revision Cards
Best design for a rare disease
Case-control study.
Best design for a rare exposure
Cohort study.
Measure from a case-control study
Odds ratio (not incidence or RR).
Unit of analysis in an ecological study
Groups of people, not individuals.
Ecological fallacy
Applying a group-level association to individuals.
Weakness of cross-sectional studies for causation
Exposure and effect measured at the same time.
Field trial vs clinical trial
Field = healthy people at risk; clinical = patients.
Random error is reduced by
A larger sample size.
Two main sources of systematic error
Selection bias and measurement bias.
Confounder
A factor associated with both exposure and disease that distorts their association.
Design-stage control of confounding
Randomization, restriction, matching.
Analysis-stage control of confounding
Stratification and statistical modelling.
Double-blind
Neither investigators nor participants know the allocation.
Essential causal criterion
Temporal relationship — cause precedes effect.
Strong association (WHO)
Relative risk greater than 2.
Sources
Sources
Facts in this article were checked against these documents.
Basic Epidemiology, 2nd ed. — Chapters 3 (Types of studies) and 5 (Causation in epidemiology) — World Health Organization, 2006 (link)
Principles of Epidemiology in Public Health Practice, 3rd ed., Lesson 1 — U.S. Centers for Disease Control and Prevention (public domain), 2006, updated 2012 (link)
CBME Curriculum 2024, Vol. II — Community Medicine (CM 7) — National Medical Commission, 2024
Test Yourself: Practice Questions
Fifteen conceptual questions in NEET PG style.
Practice Questions (NEET PG pattern)
Original questions written in the style of the exam: +4 for a correct answer, −1 for a wrong one. Every answer was checked against the sources listed above. NEET PG papers are not published by NBEMS, so these are not past-paper questions.
Q1. To investigate the risk factors for a rare cancer, the most efficient study design is
(a) cross-sectional survey
(b) community trial
(c) cohort study
(d) case-control study
Show answer
Answer: (d) — Case-control studies are a relatively simple way to investigate the causes of diseases, especially rare diseases.
Q2. Which observational design provides the most direct measurement of the risk of developing a disease?
(a) Ecological study
(b) Case-control study
(c) Cross-sectional study
(d) Cohort study
Show answer
Answer: (d) — WHO notes that cohort studies give the best information about causation and the most direct measurement of risk.
Q3. A study finds that countries with higher average fat intake have higher breast-cancer rates, and concludes that women who eat more fat get breast cancer. The error is
(a) ecological fallacy
(b) Berkson's bias
(c) lead-time bias
(d) recall bias
Show answer
Answer: (a) — An ecological study uses groups as units; applying a group-level association to individuals is the ecological fallacy.
Q4. In which design are exposure and outcome measured at the same point in time, making it hard to know which came first?
(a) Randomized controlled trial
(b) Cross-sectional study
(c) Cohort study
(d) Field trial
Show answer
Answer: (b) — In a cross-sectional study exposure and effect are measured at the same time.
Q5. A trial of a new vaccine is carried out among healthy schoolchildren in their villages. This is a
(a) clinical trial
(b) ecological study
(c) case-control study
(d) field trial
Show answer
Answer: (d) — Field trials involve healthy people presumed to be at risk, with data collected in the field.
Q6. Error that arises purely by chance and is reduced by increasing the sample size is
(a) random error
(b) confounding
(c) measurement bias
(d) selection bias
Show answer
Answer: (a) — Random error is divergence of a sample value from the true value due to chance alone; larger samples reduce it.
Q7. Volunteers who enrol in a screening study are healthier than the general population. This is an example of
(a) random error
(b) measurement bias
(c) observer variation
(d) selection bias
Show answer
Answer: (d) — Selection bias is a systematic difference between the people selected for a study and those not selected.
Q8. In a study of coffee and coronary heart disease, smoking is associated with both coffee drinking and heart disease. Smoking is acting as a
(a) outcome
(b) confounder
(c) effect modifier only
(d) random error
Show answer
Answer: (b) — Confounding occurs when another exposure in the population is associated both with the disease and with the exposure studied.
Q9. Limiting the coffee–heart disease study to non-smokers in order to remove confounding by smoking is called
(a) randomization
(b) restriction
(c) stratification
(d) matching
Show answer
Answer: (b) — Restriction limits the study to people with particular characteristics, removing the confounder's effect.
Q10. Selecting a control of the same age and sex for each case to make confounders evenly distributed is
(a) blinding
(b) restriction
(c) matching
(d) stratification
Show answer
Answer: (c) — Matching selects participants so that potential confounders are evenly distributed between the groups compared.
Q11. A study in which neither the investigators nor the participants know who receives the drug is
(a) open-label
(b) crossover
(c) double-blind
(d) single-blind
Show answer
Answer: (c) — A double-blind study means neither investigators nor participants know how the participants are classified.
Q12. Among the considerations for judging whether an association is causal, the one regarded as essential is
(a) consistency
(b) temporal relationship
(c) dose–response relationship
(d) strength of association
Show answer
Answer: (b) — WHO lists 'Does the cause precede the effect?' as essential; the others add weight.
Q13. Prevalence of hearing loss rising with both noise level and years of exposure illustrates which causal consideration?
(a) Consistency
(b) Reversibility
(c) Dose–response relationship
(d) Plausibility
Show answer
Answer: (c) — A graded increase in effect with increasing exposure is a dose–response relationship.
Q14. Lung-cancer risk falls among people who stop smoking. This supports causation through
(a) temporality
(b) consistency
(c) specificity of the test
(d) reversibility
Show answer
Answer: (d) — Reversibility asks whether removing a possible cause reduces disease risk.
Q15. Loss to follow-up is most likely to be a problem in which observational design?
(a) Ecological study
(b) Cohort study
(c) Cross-sectional study
(d) Case-control study
Show answer
Answer: (b) — WHO rates the probability of loss to follow-up as high for cohort studies, which follow people over time; it does not arise in cross-sectional or ecological designs.
Q16. Cases of a congenital malformation remember drugs taken in pregnancy more thoroughly than mothers of normal babies. This is
(a) confounding
(b) selection bias
(c) recall bias
(d) random error
Show answer
Answer: (c) — Recall bias is differential recall of exposure by cases and controls — a form of measurement bias important in case-control studies.
Q17. According to WHO, a relative risk of what size can be considered a strong association?
(a) Less than 0.5 only
(b) Greater than 1.1
(c) Greater than 2
(d) Exactly 1
Show answer
Answer: (c) — WHO notes that relative risks greater than 2 can be considered strong.
Active Recall
Draw the study-design grid from memory: for each design state the starting point, the measure it gives, what it is best for and its main weakness. Then list three ways to control confounding and the one essential criterion for causation.