Cohort Studies: Design, Bias, and Critical Appraisal
Learning Outcomes
Describe the key design features of cohort studies.
Define the main differences between prospective and retrospective cohort study designs.
Identify important types of bias affecting cohort studies and strategies to address the risk of bias (selection, confounding, and measurement bias).
Understand key elements in how to assess and critically appraise cohort studies.
Cohort Study Design Overview
Classic Cohort Design: - Start with a population of interest. - Identify a cohort that is initially free of disease. - Measure exposure and classify participants into two groups: exposed to a risk factor or not exposed. - Follow participants over time to observe whether each group develops the disease or not. - Key feature: The cohort must be disease-free before exposure is measured and classified.
The CHAMP Study (Local Example): - Name: Concord Health and Ageing in Men Project. - Population: Men aged and over listed on the electoral roll in three local government areas surrounding Concord Hospital in Metropolitan Sydney. - Baseline Assessment: Participants completed a questionnaire at home and attended a -hour clinic appointment. - Follow-up: Conducted over several years through phone calls every months (to track falls or hospitalizations), repeated clinic assessments, and linkage to death and hospitalization records. - Example Application: A study within CHAMP investigated the association between total cholesterol levels at baseline and longer-term major adverse cardiovascular events in men without pre-existing ischemic heart disease.
Prognosis Study Design: - Participants all have a specific disease at the start of the study. - Assessment focuses on a prognostic factor (a specialized type of risk factor) and its relationship to an outcome related to that specific disease. - Difference from Classic Cohort: In classic studies, the population starts healthy; in prognosis studies, they start with the disease. - Key distinction: Factors associated with the risk of developing a disease are not automatically associated with the risk of a poor prognosis once the disease is present.
Design Comparison: Prospective vs. Retrospective
Shared Features: - In both designs, the exposure occurs and is measured prior to the outcome occurring.
Prospective Cohort Studies: - Conducted as the study moves forward in time. - Advantages: Better control over what is measured; provides the highest level of evidence among observational designs. - Disadvantages: More time-consuming and expensive.
Retrospective Cohort Studies: - Conducted after follow-up has happened and outcomes have already occurred. - The study looks back in time to assemble the cohort and identify historical exposure measurements. - Advantages: No need to wait for follow-up to occur. - Disadvantages: Less control over measurements; risk of inaccuracies or use of outdated exposure measures.
Advantages and Disadvantages of Cohort Studies
Advantages: - Directionality: Exposure occurs before the outcome, establishing a clear temporal relationship. - Evidence Level: Provides the highest level of evidence within observational studies and is second only to Randomized Controlled Trials (RCTs) regarding causality. - Breadth: Capacity to examine multiple exposures and multiple outcomes simultaneously. - Descriptive Power: Facilitates the study of the incidence and natural history of diseases.
Disadvantages: - Confounding: Because there is no randomization, groups are likely to differ in ways that influence outcomes. While adjustment is possible, it is rarely complete, leading to residual confounding. - Logistical Barriers: They are often large, expensive, time-consuming endeavors, and securing funding can be difficult.
Types of Bias in Cohort Studies
Bias Definition: Systematic errors producing results that differ systematically from the truth.
Selection Bias: Occurs when groups compared are different in factors that determine the outcome other than the factor of interest.
Measurement Bias: Occurs when methods of measurement differ between the groups being compared.
Confounding Bias: Occurs when the effect of the primary factor of interest is mixed with or distorted by a second factor. - Criteria for a Confounder: 1. Must be associated with the exposure. 2. Must be associated with the outcome. 3. Must not be on the causal pathway between the exposure and the outcome (acts independently). - Example (Smoking as a Confounder): If studying alcohol consumption and lung cancer, smoking is a confounder because: - It is associated with alcohol consumption. - It is a known cause (associated) of lung cancer. - It is not on the causal pathway (alcohol does not biological cause cancer via smoking; smoking has an independent pathway).
Strategies to Minimize Bias
Dealing with Confounding in the Design Stage: - Restriction: Enrolling only specific participants (e.g., only non-smokers to remove smoking as a confounder). - Matching: Enrolling a participant in the unexposed group with matching characteristics to a participant in the exposed group.
Dealing with Confounding in the Analysis Stage: - Multivariate Analysis / Regression Modeling: adjusting comparisons for other factors. - Stratification: Analyzing results by subgroups of the potential confounding factor. - Note: We can only adjust for known confounders that have been specifically measured.
Minimizing Loss to Follow-up: - Collecting contact details for participants and their friends/relatives at baseline. - Maintaining regular contact (newsletters, holiday cards). - Obtaining permission for passive follow-up via data linkage (hospital records, registries). - Using statistical techniques during the analysis phase.
Critical Appraisal Principles
Question 1: Representative Sample: - Was a defined sample assembled at a common point in time? - For risk factors: Before disease onset. - For prognosis: At a common time point in the disease course, known as an inception cohort (e.g., onset of symptoms, time of diagnosis, or start of treatment like surgery). - Was a representative sampling frame used? (e.g., the electoral roll in the CHAMP study). - What were the inclusion and exclusion criteria?
Question 2: Response Rates: - Defined as the proportion of eligible people approached who agree to participate. - Higher rates are desirable, but older population cohorts often average . - Indirect evidence can prove representativeness: In CHAMP ( response rate), comparison to the Australian census and a national telephone survey showed identical age distributions and similar prevalence of conditions.
Question 3: Adjustments for Confounders: - Were adjustments made for differences between subgroups? - Example: Crow et al. (2020) study on frailty and low-mileage driving adjusted for age, gender, vision, and cognitive health.
Question 4: Follow-up Completeness: - Was follow-up long enough for the outcome to occur? - Was there loss to follow-up? Ideally, this should be low. - Differential Loss to Follow-up: If loss is higher in one group than another and related to the outcome (e.g., $RR$ changes from to ), it produces a biased estimate. Non-differential loss (equal in both groups) generally preserves the relative risk estimate (e.g., $RR$ moves slightly from to ).
Question 5: Outcome Measurement: - Were outcomes measured the same way in all participants? - Were measures objective and reliable (e.g., death registry)? - Was blinding used? Blinding can involve participants, interviewers, measurers, or data analysts to prevent knowledge of exposure status from influencing outcome recording.