Introduction and Overview

Section 1 established what a cohort study is and named the design choices its investigators have to make. This section drills into the most consequential of those choices: whether to count events per person (risk) or events per unit of person-time (rate). The two designs share a 2×2 layout but differ in what they assume about the population and what they let you say about disease frequency. Sample-size planning, surprisingly, is a useful place to start, because the calculation is the same for both designs even when the analysis ends up being different.

Sample Size

Initial sample-size estimates are usually performed assuming an equal number of exposed and non-exposed subjects, and assuming the disease is measured by risk (Section 8.2.2). This approach is often sufficient for initial planning even if the population is open and a rate-based design must ultimately be used.

Risk-Based (Cumulative Incidence) Designs

This is the simplest form of cohort study, but several assumptions must hold:

• Exposure groups are defined at the start of the study and remain unchanged (fixed cohorts).
• The study groups are closed — all subjects must be observed for the full risk period.
• There should be few or no losses (some authors use >10% losses as a cut-point that casts doubt on validity).

2×2 Table: Risk-Based Cohort Design

We select n₁ exposed and n₀ non-exposed individuals (free of disease) from the source population, follow them for the full follow-up period, and observe a₁ exposed cases and a₀ non-exposed cases. The two risks of interest are:

Risk-based designs are conceptually clean but operationally fragile — their assumptions break the moment people leave the cohort or the risk period extends beyond a few months. The rate-based alternative was developed precisely to handle the populations where those assumptions do not hold.

Rate-Based (Incidence Density) Designs

In many cohort studies, not every subject is under observation for the full risk period — especially when:

• The source population is dynamic (subjects enter and leave).
• The follow-up period is long.
• Subjects are added part-way through the biological risk period.
• A significant proportion of subjects withdraw from the study.
• Exposure status itself changes during the study.

In these situations, we cannot just count exposed and non-exposed subjects. Instead, we accumulate the amount of ‘at-risk time’ contributed by each subject in each exposure category. The denominator becomes person-time, not persons.

2×2 Table: Rate-Based Cohort Design

Each subject contributes ‘at-risk’ time until they develop the disease, are lost to follow-up, or the study ends. The two rates of interest are:

You have now seen both designs from the inside. The next subsection puts them side by side; read it as a decision aid for matching design to research situation, not as a statement that one design is generally better than the other.

Comparing the Two Designs

Key Examples

The four examples below illustrate the design choice with real published studies. The first two are risk-based; the last two are rate-based. As you expand each, ask yourself why the investigators chose what they chose — in every case the source population's behaviour and the follow-up window's length will be doing most of the work.

The reflection below asks you to apply the choice from this section to a specific long-running occupational cohort. After working through it, Section 3 turns to a problem that stays mostly hidden in textbook treatments: how do you actually measure exposure when it can change over years or decades of follow-up?

Reflection

Introduction and Overview

Sections 1 and 2 set up the architecture of the cohort study and the choice of risk-vs.-rate design. Both treated “exposed” as if it were a fixed property of each subject. In real cohorts that is rarely true. People start smoking, quit, switch jobs, change diets; doses accumulate. This section is about how exposure is actually measured and handled when it can vary across years or decades of follow-up — the technical problem that distinguishes a working cohort study from a textbook one.

Why Exposure Measurement Matters

In cohort studies, the objective is to identify the consequences of a specific exposure factor. Exposures can range from study-subject characteristics (sex, age) to infectious or noxious agents, environmental exposures, or food-related factors. Exposures that can be manipulated are of special interest because they lead more directly to disease control.

Scales of Exposure Measurement

Exposure status can be measured on different scales, each with implications for design and analysis. The four flip cards below run from the simplest binary contrast up to compound measures that combine intensity and duration. The pattern to take away: more granular measurement makes the design more powerful for detecting dose-response, but only if the underlying biology really has a graded effect.

The choice of measurement scale is independent of a second question: does the exposure stay fixed for the rest of the subject's life, or can it change?

Permanent vs. Non-Permanent Exposures

Both permanent and non-permanent exposures share a complication that becomes visible only when you start counting person-time: the gap between when exposure happens and when disease can plausibly result.

The Induction Period & Time-at-Risk

An important concept in cohort design is the induction period — the time after exposure is completed before disease can reasonably arise.

Changing Exposure Status

If exposure status changes during follow-up, an individual subject can contribute time-at-risk to both exposed and non-exposed categories:

• Previously non-exposed subjects contribute to the exposed category after the exposure threshold is reached.
• Previously exposed subjects contribute to the non-exposed category after any lag effects have ended.
• If a subject develops the disease, the exposure category assigned is the one they were in at the time the outcome occurred.

One last twist on the meaning of “exposure” is worth flagging before we move on, because it shows up surprisingly often in chronic-disease cohorts.

Disease as Exposure

Disease itself can serve as an exposure for other outcomes such as additional diseases, mortality, or quality-of-life measures. Lazo et al (2011) followed 11,000+ adults for 18 years using non-alcoholic fatty-liver disease (NAFLD) as the exposure. Those with NAFLD — whether or not they had elevated liver enzymes — had a similar mortality hazard ratio as those without NAFLD, an interesting null result.

Key Examples

The three published cohorts below illustrate the full range of exposure handling we just covered: a continuous diet exposure validated against multiple recalls, a multi-exposure design with biospecimens, and a multinational study that pushes from estimated relative risk to population-attributable fraction. Expand each one and notice which exposure-measurement decisions shape the rest of the design.

The reflection below puts the measurement-scale decision into a real research scenario. After working through it and the knowledge check, Section 4 closes the lesson by addressing the four practical questions that any cohort investigator faces once exposure is settled: keeping the groups comparable, managing the follow-up period, ascertaining outcomes, and analysing the data.

Reflection

Introduction and Overview

Sections 1–3 walked through the design choices: cohort architecture, risk-vs.-rate handling, and exposure measurement. With those locked in, the remaining work is operational. Four practical questions structure this section in turn: how do you make sure the exposed and non-exposed groups are comparable on the variables that aren't your exposure of interest, how long should you follow them, how do you confirm the outcome happened, and how do you analyse what you collect?

Ensuring Exposed & Non-Exposed Groups Are Comparable

If exposed and non-exposed groups are not comparable (i.e., not exchangeable) with respect to factors related to both exposure and outcome, a biased (confounded) assessment results (Klein-Geltink et al, 2007).

Three approaches help ensure comparability:

Comparability addresses the cross-sectional baseline. The next question is what happens over time — specifically, how long the follow-up should last and what counts as time-at-risk for each subject.

Follow-Up Period

To enhance validity, the follow-up process must be as complete as possible and unbiased with respect to exposure. Achieving unbiased follow-up may require some form of blinding to exposure status:

• In prospective studies: assign follow-up tasks to researchers unaware of exposure status.
• In retrospective studies: keep record reviewers unaware of exposure status when possible.

Losses to follow-up are a perennial concern. Chang et al (2009) describe shared parameter models to reduce bias when losses are not randomly distributed.

Even careful follow-up only matters if the outcome itself is measured well. Cohort studies tend to be ambitious about exposure characterization but surprisingly casual about outcome ascertainment, and the next subsection is the corrective.

Measuring the Outcome

Although the most frequent outcome in cohort studies is the occurrence of a specific disease (measured as risk or rate), outcomes can also be:

• Ordinal: e.g., none, mild, moderate, severe.
• Continuous: e.g., birth weight, blood pressure, BMI, quality of life index (Example 8.2).

Harley et al (2011), for instance, examined polybrominated diphenyl ethers (PBDEs — a flame retardant) and infant birth weight, length, and head circumference — both exposure and outcome on continuous scales.

Diagnostic Criteria

Each study needs explicit protocols for determining outcome occurrence and timing. Clear diagnostic criteria minimise diagnostic errors. In retrospective studies, this can be challenging when only summary diagnostic information is available.

Incidence Requires Two Examinations

Strictly, measuring incidence requires:

1. An examination at the start of follow-up to ensure subjects do not yet have the disease.
2. A second examination to determine whether (and when) the disease developed.

If clinical diagnostic data are used, the incident date is usually based on the time of diagnosis, not the time of underlying disease occurrence — an important caveat for diseases with long subclinical phases. If subjects are screened at regular intervals, the time of disease occurrence is conventionally placed at the midpoint between examinations.

Multiple Outcomes

Comparability, follow-up, and outcome ascertainment are all design-stage decisions. The analysis stage is where those decisions translate into a quantitative result — and where pre-specified analysis plans (the EGAP-style discipline from Lesson 1) earn their keep, because cohort data tempt almost limitless slicing.

Analysis

Time-Dependent Confounders

Gran et al (2010) describe a sequential Cox technique for data with time-dependent confounders — covariates that are affected by past exposure and predict future exposure and outcome (e.g., CD4 count when assessing HIV treatment effects).

Repeated Measurements

Xue et al (2010) explain how marginal and mixed-effect models can handle cohort data with repeated measurements of exposure and covariates, contrasting these with logistic and proportional hazards approaches.

The last operational concern, as in Lessons 3 and 4, is reporting. STROBE returns one more time, now with the items that matter specifically for cohort designs.

Reporting of Cohort Studies (STROBE)

The STROBE statement (von Elm et al, 2007) provides reporting guidelines for observational studies. Tooth et al (2005) elaborate criteria specific to cohort studies. These should be used both to plan and report studies, and to assess the validity of published work.

The reflection below is the section's exit ticket — a realistic review-the-paper prompt that requires you to use everything from this section. After working through it and the knowledge check, the final assessment integrates the four content sections into one 15-question exam.

Reflection

Bringing It All Together

Where Lesson 4 looked backward from outcome to exposure, this lesson followed the arrow forward. Section 1 framed the cohort study as something close to a controlled trial without randomisation: pick a source population, classify exposure, follow people, and watch incidence accumulate. Section 2 then forced the central design choice — risk-based (cumulative incidence) for closed populations followed for a fixed time, or rate-based (incidence density) for open populations where person-time is the natural denominator — and connected each to its analytic toolkit (binomial / log-binomial vs. Poisson / survival).

Sections 3 and 4 zoomed in on the operational details that decide whether a cohort study survives appraisal. How exposure is scaled (dichotomous, ordinal, continuous, compound), whether it changes over time, how the induction period is handled, and how comparability is engineered through restriction, matching, and analytic control all determine whether the rate ratio you eventually report is estimating what you think it is. Blinded outcome ascertainment, careful handling of loss to follow-up, and STROBE-aligned reporting then turn a defensible design into a study other researchers can actually use.

The final reflection asks you to put the entire arc to work as a brief cohort proposal of your own; the 15-question assessment then checks the conceptual content directly. Lesson 6 will pull the camera back further to look at ecological and group-level designs — where the unit of analysis stops being the person — and the comparability and inference logic you just built will keep paying off there.

# 1000 exposed and 1000 unexposed individuals followed for up to 5 years.
#   exposed:  80 events in 4500 person-years
# unexposed:  30 events in 4900 person-years

events <- c(exposed = 80,   unexposed = 30)
n      <- c(exposed = 1000, unexposed = 1000)
py     <- c(exposed = 4500, unexposed = 4900)

risk <- events / n
rate <- events / py * 1000          # per 1000 person-years

RR  <- risk["exposed"] / risk["unexposed"]   # risk ratio
IRR <- rate["exposed"] / rate["unexposed"]   # incidence rate ratio

round(data.frame(risk, rate, RR = RR, IRR = IRR), 3)

## -----------------------------------------------------------------------------
## Stretch: 95% CI for the rate ratio (Wald approximation on the log scale)
## -----------------------------------------------------------------------------
log_irr   <- log(IRR)
se_logirr <- sqrt(1/events["exposed"] + 1/events["unexposed"])
ci_irr    <- exp(log_irr + c(-1, 1) * 1.96 * se_logirr)
round(c(IRR = IRR, lower = ci_irr[1], upper = ci_irr[2]), 3)

Final Assessment

This assessment covers all sections of this lesson. You must score 100% to complete the lesson. Review the feedback after each attempt.