Table 2 Authors’ recommendations for managing sources of heterogeneity in pragmatic trial design, conduct and analysis

• Ensure that centres where results are intended to be applied are clearly defined.

• Attempt to recruit centres typical of target population (rather than convenience sample).

• Increase number of centres (possibly at cost of decreasing sample size per centre).

• Ensure that patient populations to whom results are intended to apply are clearly defined and reflected in selection criteria, which should not be too restrictive.

• Stratify by centre in randomisation, whenever possible.

• Consider stratifying randomisation by other important prognostic factors.

• Sample size may need to be increased to reflect pragmatic features, including potentially attenuated intervention effect and larger variance estimate.

• In cluster randomised trials, ensure that intracluster correlation and cluster size variation are taken into account.

• Inclusion criteria should be applied consistently across all centres and investigators.

• Include stratification factors in analysis.

• Limit number of planned subgroup analyses to those relevant to usual clinical decision-making.

• Implement the trial such that, aside from the intervention, the trial has the lowest possible impact on patients and caregivers so as not to distort usual care conditions.

• Consider likely extent of non-adherence, contamination, and co-interventions when specifying effect size for sample size calculation.

• Allow tailoring but maintain core intervention features that define the intervention being assessed.

• Avoid interventions to monitor and promote compliance that may change patient behaviour unless they can be incorporated into future scale-up of intervention.

• Assessment of compliance (by centres, providers, and patients) should be incorporated into intervention or assessed as a secondary outcome but measured in an unobtrusive way that does not interfere with compliance that would be expected outside the trial.

• Conduct analysis of superiority trials by intention-to-treat.

• Ancillary process analyses can shed light on understanding mechanisms of observed treatment effects as long as data to facilitate such analyses are collected unobtrusively.

• Ideally, outcomes should be selected to be relevant to patients or trial stakeholders and assessed in usual care.

• Avoid blinding outcome assessors except if there is a high risk of detection bias; select an outcome as objective as possible.

• If needed, standardise how the outcome is assessed and adjudicate it.

• Sensitise data monitoring committees to the pragmatic nature of the trial to prevent undue intrusion through excessive data collection.

• Plans for recruitment and informed consent must adhere to internationally accepted ethical principles even though this may affect pragmatism and bias.

• Inclusion of vulnerable participants is appropriate provided plans are in place to identify and protect those who cannot provide informed consent and who may be at risk because of co-morbidities.

• “Clinical-style” or integrated consent may be appropriate for recruitment in clinical settings when interventions involve usual care.

• Waiver of consent may be requested from a research ethics committee when a cluster-level intervention poses only minimal risk.