Cluster Unit Randomization Trials

16. Cohort vs. Cross-sectional Designs

The earlier discussion on analysis at the individual level is most pertinent to cohort designs, where each individual in the study is followed up over time. However in studies enrolling very large clusters, such as entire communities, such detailed follow-up may not be possible. Considerable discussion has therefore arisen in the community intervention trial literature as to the relative advantages of this design to a cross-sectional design, in which different groups of individuals are independently sampled and assessed at each of several time periods.

It is acknowledged that the cohort design is theoretically more powerful from a statistical perspective, since it allows an analysis that controls for individual baseline values, thus allowing the effect of intervention to be estimated with more precision.

However, as shown by Feldman and McKinlay (1994), this advantage must be weighed against the risk of loss to follow-up that arises in any longitudinal study. The “worst-case scenario” arises when the loss to follow-up is differential across intervention groups, since then the final estimate of intervention effect may be subject to substantial bias.

Even when subject attrition is unrelated to treatment assignment, a large loss to follow-up rate may result in reduced efficiency relative to a cross-sectional design. Other disadvantages of the cohort design, as reviewed by Atienza and King (2002), include:

• A loss of representativeness of the target population related to the aging of the cohort; and
• "Learning effects” that may result from repeated assessments on the same individual.

These considerations suggest that a cohort design is most effective when:

• Participating clusters are of relatively small size,
• The study population is relatively stable and compliant; and
• Follow-up times are not lengthy. 

It follows that for studies enrolling large communities, where complete follow-up is rarely feasible, cross-sectional designs have often been preferred, as in the early trials of cardiovascular health referred to in this chapter.They also may be the inevitable choice for any intervention that is evaluated at the cluster level only.

To avoid the analytic limitations of cross-sectional designs, an approach adopted by some investigators has been to augment this design by subsampling a cohort consisting of a relatively small number of subjects in each community.  For example the COMMIT investigators used randomly selected cohorts of heavy and light-to-moderate smokers, respectively, as one means of evaluating the effect of their community-based smoking cessation intervention.

From a conceptual perspective, the choice of design must also be considered in light of how the primary question of scientific interest is posed.  Thus if interest focuses mainly on change at the broader community level, cross-sectional designs may be the more natural choice while cohort designs may be more natural if change at the individual level is of most interest.  Methods of analysis that are particularly suited to cross-sectional designs have been discussed by Nixon and Thompson (2003) and Ukoumunne and Thompson (2001) for the case of binary outcomes, and by Koepsell et al., (1991) and Murray (2001) for the case of continuous outcomes.