ARCHE - Cohort Studies

Introduction

A cohort study follows a group of people (a “cohort”) over time to see how certain exposures (like smoking or a new drug) affect outcomes (like developing lung cancer or symptom relief). There are two main types: prospective (starting now and following forward) and retrospective (using past data to look back). Cohorts are powerful for establishing temporal sequences – you can see if exposure happened before the outcome, which helps infer causation. However, they can be expensive and time-consuming, especially prospective ones. They’re often used to study rare exposures or multiple outcomes from one exposure.

Key Elements to Appraise

Study Question & Design: Ensure the research question is clear and suitable for a cohort design. Cohorts are ideal for studying incidence, natural history, or risk factors over time. Check if it’s prospective or retrospective – prospective is stronger for minimizing recall bias, but retrospective is faster and cheaper.

Cohort Selection: Look at how the cohort was assembled. It should be representative of the target population, with clear inclusion/exclusion criteria. For example, in a study of factory workers exposed to a chemical, the cohort should include both exposed and unexposed workers. Watch for selection bias – if only healthy people are included at baseline, it’s “healthy worker bias.”

Exposure Measurement: Verify that exposures are clearly defined and measured accurately at baseline (and sometimes repeatedly). Good studies use objective measures (e.g. blood tests) over self-reports to reduce misclassification. In prospective cohorts, exposures should be assessed before outcomes to establish temporality.

Outcome Measurement: Outcomes should be clearly defined, objective, and measured the same way for everyone (blinded if possible). Check the follow-up duration – it needs to be long enough for outcomes to occur. Loss to follow-up should be minimal (<20% is often a threshold); high dropout can bias results if dropouts differ from completers.

Confounding Control: Cohorts can have imbalances in confounders (e.g. age, gender). Good studies measure these at baseline and adjust in analysis (e.g. using Cox proportional hazards models or stratification). Matching exposed and unexposed groups can also help.

Statistical Analysis: Common measures are relative risks (RR) or hazard ratios (HR), with confidence intervals. Check for appropriate handling of time-to-event data and censoring. Absolute risks (incidence rates) should also be reported for context.

Reporting: Following STROBE guidelines, the study should detail cohort size, follow-up time, number of events, and handling of missing data. A flow diagram showing participant flow is helpful.

Common Mistakes / Red Flags

High Loss to Follow-Up: If more than 20% of the cohort is lost, and no analysis shows they’re similar to completers, bias is likely. For example, sicker people dropping out could underestimate risks.

Immortal Time Bias: In retrospective cohorts, if time before exposure is counted as “exposed” time, it can falsely lower risk estimates. Good studies start the clock at exposure.

Inadequate Confounding Adjustment: Failing to measure or adjust for key confounders (e.g. not accounting for smoking in a diet-heart disease study) can lead to spurious associations.

Misclassification: If exposures change over time but are only measured once, or if outcomes are based on unreliable data (e.g. self-reported without validation), results are weakened.

Overinterpretation: Even in cohorts, association isn’t always causation – residual confounding can remain. Watch for authors claiming definitive cause without discussing limitations.

Statistical Issues: Small cohorts with few events can lead to imprecise estimates (wide CIs). Also, multiple comparisons without adjustment can produce false positives.

Example

Consider the famous Framingham Heart Study, a prospective cohort following residents of Framingham, Massachusetts, since 1948 to identify risk factors for cardiovascular disease. In appraising a similar study, check cohort selection – was it a random sample? In Framingham, it was, reducing selection bias. Exposures like blood pressure and cholesterol were measured objectively at baseline and follow-ups. Outcomes (heart attacks) were verified by medical records, with low loss to follow-up over decades. Confounders like age and smoking were adjusted in analyses, yielding reliable relative risks (e.g. high cholesterol increases heart disease risk). A red flag would be if a smaller cohort had high dropout or didn’t adjust for key variables, potentially overestimating or underestimating risks.

Quick Checklist

Clear question; appropriate use of prospective/retrospective design.
Representative cohort with defined inclusion/exclusion; minimal selection bias.
Accurate, baseline exposure measurement; temporality established.
Objective outcome assessment; adequate follow-up length and completeness (<20% loss).
Confounders measured and adjusted; appropriate stats (RR/HR with CIs).
Transparent reporting of methods, results, and limitations.

Take-Home Points