Pseudo-cluster randomisation and the contamination problem
Individual randomisation, assigning each participant to an arm independently, is the design most researchers default to, and for good reason: it maximises statistical power for a given sample size. But it has a specific weakness in studies where participants at the same site or under the same care provider can be assigned to different arms: contamination. A care provider delivering both a new intervention and standard care to different patients can end up letting elements of one bleed into the other, consciously or not, diluting the measured difference between arms.
Cluster randomisation, assigning whole sites or providers to a single arm rather than randomising within them, solves the contamination problem directly. It introduces a different one: referral bias. If a provider who prefers a particular treatment approach knows which arm their site has been assigned to, referral patterns into the study can shift in ways that undermine the comparison just as much as contamination would.
Why neither approach solves the whole problem
This is a genuine trade-off, not a solved problem with one obviously correct answer. Individual randomisation risks contamination; cluster randomisation risks referral bias; both risks are real and can meaningfully distort a trial's results if left unaddressed. Work applying pseudo-cluster randomisation in practice looked at exactly this tension in a psychiatric care setting, where both risks were plausible: providers delivering care directly, and referral decisions that could plausibly be shaped by provider preference.
What pseudo-cluster randomisation actually does
The method sits deliberately between the two extremes. Rather than assigning a whole site to one arm, or randomising every individual independently within a site, pseudo-cluster randomisation assigns most participants at a site to one arm, with a smaller, randomised minority assigned to the other. The exact ratio is a design choice, but the logic is the same: enough clustering to reduce contamination meaningfully, enough individual-level randomisation retained to reduce the referral-bias risk that a fully clustered design would carry.
It's a genuinely clever compromise, and like most compromises, it's not free. The statistical analysis has to explicitly account for the resulting clustering structure, and the design is more complex to implement and explain to sites than either pure approach.
When this design is actually worth the added complexity
Pseudo-cluster randomisation isn't a good default for every study. It earns its complexity specifically when both of these are true:
- Contamination is a real, plausible risk. If a single care provider or site could reasonably be expected to deliver aspects of both arms to different participants, individual randomisation without a clustering mechanism is likely to understate the true effect of the intervention.
- Referral bias is also a real, plausible risk. If providers or sites have visible influence over which patients are referred into the study, and could plausibly steer referrals based on their own treatment preferences, a fully cluster-randomised design introduces a comparable threat in the other direction.
If only one of these risks applies, a simpler individual or cluster design, with the appropriate safeguards for the risk that does apply, is usually the better choice. Pseudo-cluster randomisation is a tool for the specific situation where both risks are live at once, not a generally superior default.
The broader lesson for study design
The value of a case like this isn't the specific ratio the study settled on. It's the reminder that randomisation strategy is a genuine design decision with real trade-offs, not a default setting to be applied uniformly regardless of context. A study that reaches automatically for individual randomisation without asking whether contamination is plausible, or automatically for cluster randomisation without asking whether referral bias is plausible, is skipping a step that pseudo-cluster randomisation exists precisely to force: naming both risks explicitly, and designing the randomisation strategy around whichever combination of them is actually present in the specific research setting.