Since the EEF began, we’ve commissioned more than 200 randomised controlled trials to understand what works to improve teaching and learning. Designing trials with sufficient statistical power is essential, and one important choice is at which level to randomise: should treatment and control groups be assigned to schools or classes, or to individual pupils?

This choice matters because it affects both statistical power and how precise the results are.

Individual-level randomisation (where pupils are assigned to treatment or control condition individually) usually produces more precise results for the same total number of participants.

However, this design might not be the most appropriate if, for example, there are high risks that the intervention can be shared with the control group.

Choosing the right level of randomisation means weighing up the benefits of more precise estimates against other risks, including contamination. So how can we help evaluators navigate this trade-off?

What do we mean by ​“level of randomisation”?

Most of our trials use cluster-level randomisation, where a cluster – typically a school, setting, or class – is randomly allocated to the intervention or control group. Everyone in that cluster (all pupils and teachers) gets the same condition.

In individual-level randomisation, pupils within the same school or class are allocated individually to either treatment or control group. This means some children will receive the intervention while others in the same class will be in the control group.

How does randomisation level affect power and precision?

People within the same ​‘cluster’ (school or class) tend to be more alike than people across different schools or classes. This similarity is measured by the intra-cluster correlation coefficient (ICC). Because of this, cluster-randomised trials are statistically less efficient.

As a result:

• Individual-level randomisation gives more precise estimates.
  
• Cluster trials often need more participants to reach the same level of precision.

How to decide on the right level of randomisation

Even though individual randomisation is statistically more efficient, it isn’t always practical or appropriate. Three key factors help guide the decision:

1. Contamination risk

   Contamination happens when the control group is exposed to the intervention, for example, through shared teaching practices. This can bias the results and reduce the measured impact of the intervention. 

2. Recruitment and sample size

   Cluster trials need larger samples, which can be harder and more expensive to recruit and retain. 

3. Perceptions and practicalities

   Attitudes toward randomisation vary. In cluster trials, some schools may drop out if they end up in the control group and none of their pupils receive the intervention. Whereas, in individual trials, schools may find it difficult to explain why some pupils are chosen and others are not. 

How to assess contamination risk

Some contamination is always possible, but its likelihood and impact depend on:

• the design of the intervention
• who delivers it
• how easily aspects of the intervention could be shared with the control group

Our contamination checklist helps evaluators think through these issues systematically and assess contamination risks on a hypothetical individual-level scenario.

Choosing the level of randomisation

Individual-level randomisation isn’t suitable for every intervention. For example:

• Whole-school programmes can only be tested at the school level.
  
• Some interventions are hard for teachers to apply inconsistently. For example, a programme encouraging teachers to use more praise may be difficult to restrict only to pupils in the treatment group.

The EEF encourages evaluators to:

• Use individual randomisation when appropriate and contamination risks are manageable.
• Measure and monitor contamination in trials that randomise individuals.
• Consider alternative levels of randomisation – such as class, year group, or teacher – when school-level or individual-level designs aren’t suitable. Examples include:
  • Class-level randomisation (ReflectED efficacy trial)
    
  • Teacher-level randomisation (Mastering Maths effectiveness trial)
  • Year-group randomisation (Stop & Think effectiveness trial)

Looking ahead

We’ll continue to refine and innovate trial designs to improve statistical power and produce robust, practical evidence that can support schools, colleges, and early years settings to improve teaching and learning outcomes.