Differences between cohort studies of aetiology and prognosis
Many medical and scientific questions are the sort that cannot be answered using randomised study designs. For example, Does cigarette smoking cause lung cancer?, What is the risk hip fracture in adults?, and Do rats indirectly destroy coral reefs? Scientists often use observational cohort studies to answer these types of questions. In cohort studies, (1) a sample of subjects or a sample of observational units is identified, (2) data on individual subjects’ exposures and outcomes are collected, and (3) associations between exposures and outcomes are determined.
Cohort studies can be used to identify aetiology (causes) of outcomes, or prognosis (progression over time) of outcomes. Cohort studies of aetiology and prognosis have different aims and need to be conducted differently. These concepts are summarized by Herbert (2014). A table of these differences is shown below:
|Aim||To determine causes of health outcomes. The main interest is in effects.||To predict health outcomes. The main interest is in outcomes.|
|Choice of exposures||Only exposures that can be manipulated are of interest because causal effects can only be defined with these.||Any exposure can be a predictor.|
|Independent variables in statistical model||Models must include all non-ignorable variables that determine the outcome including all confounders.||Confounding is irrelevant. Simple models are preferred because these are useful in clinical practice.|
|Nature of analysis||Analysis should be theory driven.||Analysis can be data driven so prevent “over-fitting” the model.|
|Accuracy of measurement||Exposures must be measured with little or no error.||Exposures should be easily measured.|
|Specifying statistical models||Model must be correctly specified or estimates of causal effects may be biased. Eg. may need to consider non-linear effects, interactions, mediators.||Predictions can be accurate even if the model is not correctly specified.|
|Form of analysis||Analysis involves contrasting outcomes with exposures.||Analysis involves estimating expected outcomes.|
|Overall purpose||Findings can be used to develop new health interventions.||Findings can be used to inform people of prognosis and identify those at high risk for whom intervention is most likely to be effective.|
In understanding how and why the thinking behind aetiologic and prognostic cohort studies is different, we see that “associations” between variables are not so interesting unless we can talk about causation or prognosis.
Herbert (2014) Cohort studies of aetiology and prognosis: they’re different. Journal of Physiotherapy 60:241-244.