Cohort Study Design and Biases

Understanding Cohort Study Design and Key Epidemiologic Measures

In public health research, cohort studies are a cornerstone for investigating the relationship between exposures and disease outcomes over time. This course explains the most important concepts tested in a typical quiz on cohort design, bias, and epidemiologic calculations, providing clear definitions, examples, and practical tips for interpretation.

1. Measuring Association in Prospective Cohort Studies

Risk Ratio (Relative Risk)

The primary measure of association in a prospective cohort is the Risk Ratio (RR), also called the relative risk. It quantifies how much more (or less) likely the disease is to occur in the exposed group compared with the unexposed group.

• Formula: RR = [Incidence in Exposed] / [Incidence in Unexposed]
• Interpretation: RR = 1 indicates no association; RR > 1 suggests increased risk; RR < 1 suggests a protective effect.
• When to use: When the study follows participants forward in time and can directly observe incident cases.

Other measures such as the Prevalence Ratio (PR), Odds Ratio (OR), and Attributable Risk (AR) are useful in different designs, but the RR remains the gold standard for prospective cohorts.

2. Common Sources of Bias in Epidemiologic Research

Selection Bias – Berkson Bias

When cases and controls are drawn from a hospital setting, the control group may not represent the source population. This specific form of selection bias is known as Berkson bias. It occurs because hospitalized patients often have multiple comorbidities, altering the exposure distribution compared with the general community.

• Impact: It can either exaggerate or mask true associations, leading to misleading conclusions.
• Prevention: Use population‑based controls, or adjust analytically for the factors that make hospital patients atypical.

Other Bias Types (Brief Overview)

• Recall bias: Differential accuracy of memory between cases and controls, common in retrospective studies.
• Information bias: Systematic errors in measuring exposure or outcome.
• Confounding bias: A third variable associated with both exposure and outcome that distorts the true relationship (see Section 5).

3. Standardized Mortality Ratio (SMR) – What You Need to Know

The Standardized Mortality Ratio (SMR) compares the observed number of deaths in a study population with the number expected if the population had the same age‑specific mortality rates as a reference (standard) population.

• Required components:
  • Population size of the study area broken down by age groups.
  • Age‑specific mortality rates from the reference population.
  • Total number of deaths observed in the study area.
• Not required: The incidence of the disease under study. SMR uses mortality data, not incidence data.

Formula: SMR = Observed Deaths / Expected Deaths, where Expected Deaths = Σ (Standard Rate × Study Population in each age group).

4. Odds Ratio (OR) in Case‑Control Studies

Why OR Approximates RR When Disease Is Rare

In a case‑control design, researchers cannot directly compute incidence, so they rely on the Odds Ratio (OR). When the disease outcome is rare (<5% prevalence), the odds of exposure among cases approximates the probability of exposure, making the OR mathematically close to the Risk Ratio (RR).

• Rare‑disease assumption: OR ≈ RR because the number of cases is a small fraction of the total population.
• Practical implication: Researchers can interpret the OR as a measure of relative risk, simplifying communication of findings.

Remember, this approximation breaks down when the disease is common; in those situations, the OR may overstate the strength of association.

5. Confounding Factors – Definition and Management

A confounding factor is a variable that is associated with both the exposure and the outcome, but is not on the causal pathway. If not controlled, it can distort the observed relationship, leading to either an over‑ or under‑estimation of the true effect.

• Example: Age can confound the link between smoking (exposure) and lung cancer (outcome) because age influences both smoking prevalence and cancer risk.
• Control strategies:
  • Design phase: Randomization, restriction, or matching.
  • Analysis phase: Stratification, multivariable regression, or propensity‑score methods.

Identifying potential confounders early in study planning is essential for valid inference.

6. Direct Standardization – Applying the Correct Weights

When researchers wish to compare mortality (or any rate) between populations with different age structures, they often use direct standardization. The age‑specific rates of the study population are weighted by the age distribution of a standard population, not by the study population itself.

• Weight used: The proportion of each age group in the standard population.
• Result: An age‑adjusted rate that reflects what the study population’s rate would be if it had the same age composition as the standard.

This method enables fair comparisons across regions, time periods, or demographic groups.

7. Cross‑Sectional Studies and the Prevalence Ratio (PR)

Why PR Is Not a Measure of Risk Over Time

Cross‑sectional designs capture a snapshot of disease status and exposure at a single point. The Prevalence Ratio (PR) therefore reflects the proportion of individuals with disease at that moment, which is influenced by both the incidence of new cases and the duration of existing cases.

• Key point: Because prevalence = incidence × average duration, PR cannot be interpreted as a risk that accumulates over time.
• Implication for researchers: Use PR to describe burden, but rely on cohort or case‑control designs for true risk estimation.

8. Diagnostic Test Performance – Sensitivity (Se)

Sensitivity measures a test’s ability to correctly identify individuals who truly have the disease. It is defined as the proportion of true positives among all diseased persons.

• Formula: Se = True Positives / (True Positives + False Negatives)
• Common misconceptions cleared:
  • Se is not the proportion of true negatives (that is specificity).
  • Se does not change with disease prevalence; it is an intrinsic property of the test.
  • Se is evaluated against a gold‑standard reference, so the choice of gold standard directly influences its estimate.

High sensitivity is crucial for screening tools where missing a case (false negative) could have serious public‑health consequences.

9. Integrating Concepts – A Practical Example

Imagine a prospective cohort of 10,000 workers followed for 5 years to assess the effect of a new chemical exposure on respiratory disease. After follow‑up, 150 exposed workers develop the disease, while 80 unexposed workers do. The incidence rates are:

• Exposed: 150 / 5,000 = 0.03 (3%)
• Unexposed: 80 / 5,000 = 0.016 (1.6%)

The Risk Ratio is 0.03 / 0.016 ≈ 1.88, indicating an 88% higher risk among the exposed. If a case‑control study were conducted instead, and the disease were rare, the calculated OR would be close to 1.88, allowing researchers to infer a similar magnitude of risk.

During analysis, investigators notice that older workers are more likely to be both exposed and to develop disease. Age is therefore a potential confounder. By stratifying the data by age groups or using multivariable regression, the adjusted RR might change, revealing the true exposure effect.

Finally, to compare mortality in this workforce with the national population, the team would compute an SMR. They would need the age‑specific mortality rates from the national reference and the observed deaths in each age group of the cohort. The incidence of respiratory disease is irrelevant for SMR calculation.

10. Key Take‑aways for Public‑Health Professionals

• Prospective cohorts provide direct incidence data; use the Risk Ratio to express association.
• Selection bias, especially Berkson bias, can arise from hospital‑based sampling; choose population‑based controls whenever possible.
• SMR requires age‑specific mortality rates and observed deaths, not disease incidence.
• Odds Ratio approximates the Risk Ratio under the rare‑disease assumption, making it a useful surrogate in case‑control studies.
• Confounding distorts true exposure‑outcome relationships; control it through design or analysis.
• Direct standardization applies the age distribution of a standard population as weights.
• Prevalence Ratio reflects disease burden at a point in time, not cumulative risk.
• Sensitivity is the proportion of true positives; it is independent of disease prevalence.

Mastering these concepts equips epidemiologists and public‑health practitioners to design robust studies, interpret findings accurately, and communicate results effectively to policymakers and the public.