Visualising harms in Randomised Controlled Trial publications: a consensus and recommendations

Objective: To improve communication of harm in RCT publications we identified researchers’ recommendations for visualising harm outcomes. Design: Consensus study evaluating visualisation methods. Setting: 15 UKCRC registered CTUs, an academic population health department, Roche Product Ltd and the BMJ. Participants: Experts in clinical trials: 20 academic statisticians, one industry statistician, one academic health economist, a data graphics designer and two clinicians. Data sources: Visualisations were primarily identified via a methodological review of statistical methods developed specifically to analyse harm outcomes, these were considered alongside visualisations recommended by consensus group members. Interventions: None Main outcomes measured: Consensus for visualisations to recommend achieved over a series of three meetings with participants. Participants reviewed and critically appraised candidate visualisations against an agreed framework. Appraisals were summarised and presented back to participants to inform discussions. After discussions participants voted on whether to endorse each visualisation. Eligibility criteria: Visualisation receiving at least 60% of the available votes were endorsed. Scores marginally below this threshold (50-60%) were revisited for further discussions and votes retaken until a consensus was reached. Results: Twenty-eight visualisations were considered, of which ten are recommended to researchers to consider in publications of main research findings. The choice of visualisations to present will depend on outcome type e.g., binary, count, time-to-event or continuous and the scenario e.g., summarising multiple emerging events or one event of interest. A decision tree to assist trialists decide which visualisations to use is presented. Examples of each endorsed visualisation, along with example interpretation, potential limitations and signposting to code for implementation across a range of standard statistical software are provided. Clinician feedback was incorporated into the explanatory information provided in the recommendations to aid understanding and interpretation. Conclusions: Visualisations provide a powerful tool to communicate harms in clinical trials, offering an alternative perspective to the traditional frequency tables. Increasing the use of visualisations for harm outcomes in clinical trial manuscripts and reports will provide clearer presentation of harm information and thus enable informative interpretation, especially valuable for assessing the profile of harm. Whilst we endorse each of the visualisations presented, we also note their limitations and provide examples of where their use would be inappropriate. Though the decision tree aids the choice of visualisation the statistician and clinical trial team must ultimately decide the most appropriate visualisations for their data and objectives. We recommend trialists continue to examine crude numbers alongside visualisations to fully understand harm profiles.