Patterns may arise across multiple studies that are difficult to discern in individual results. Aggregating participant-level data across multiple studies enables researchers to understand how analytical methods selection, trial conduct, and design characteristics can impact research outcomes. These learnings can uncover new insights and improve future research, extending the value of existing data and patient contributions.

Vivli provides the secure access that makes this kind of inquiry possible. Researchers can draw on anonymized participant-level data from completed studies to investigate questions beyond the aims of the original work, extending the value of the data and participants’ contributions.

When those data span thousands of participants, researchers can pursue queries like:

Why do participants leave a study early? Can control data collected earlier be used again without weakening the analysis? Which statistical methods give the clearest picture of treatment effects?

The studies below show how Vivli supports efforts to answer these questions.

When participants leave a study early, the resulting gaps can weaken the evidence and complicate its interpretation. Ryan McChrystal, David McAllister, and colleagues at the University of Glasgow analyzed participant-level data from 90 randomized studies involving 86,107 people across 10 conditions. By tracing when participants left and the reasons reported for their departure, the team found that attrition was generally highest near the beginning of follow-up and was most often associated with adverse events or voluntary withdrawal. The scale of the analysis made it possible to show that attrition changes over time and differs by cause, giving teams a more precise basis for anticipating where losses are most likely to occur. Those patterns can inform enrollment targets, the timing of retention efforts, and how incomplete follow-up is handled in the final analysis.

 

Control groups raise a different methodological question. In studies that add treatments over time, control participants enrolled earlier may differ from those enrolled alongside a later treatment group. Using the earlier data could reduce the number of new control participants needed, but changes in who enrolls or how care is delivered can skew the comparison. Junichi Asano, Akihiro Hirakawa, and colleagues at the Institute of Science Tokyo, developed a Bayesian method that determines how much influence the earlier control data should have. They tested it through simulations and applied it to COVID-19 data accessed through Vivli. When results from the earlier and concurrent control groups were similar, the method gave the earlier data more weight. When they differed, it gave them less. The approach can make better use of existing controls while limiting the chance that changes over time are mistaken for a treatment effect.

The same need for careful judgment extends to the analysis of patient-reported outcomes. Jiajun Yan, Feng Xie, and colleagues at McMaster University examined whether the statistical method used to analyze EQ-5D, a widely used measure of health and quality of life, could change the estimated effect of a treatment. Previous guidance had rested largely on simulation studies and expert recommendations. Using participant-level data from 13 completed studies, the team applied four common models to each dataset separately and compared the results. The models generally reached the same conclusion about whether a treatment made a meaningful difference. This suggests that, in many cases, the overall interpretation did not depend heavily on the method chosen. However, the comparison also showed where the methods began to diverge when many responses were missing or clustered at the top of the scale. In those cases, the choice of analysis could shape the result. These findings give researchers a firmer basis for analyzing quality-of-life outcomes in future work.

 

Shared data makes it possible to examine the decisions that shape clinical research with the same rigor used to study treatments and outcomes. By revisiting data from completed studies, investigators can trace how decisions made during design and analysis influence the findings. Vivli gives this data a continuing life, allowing knowledge to build from one study to the next.

Recent methodology and prediction publications and public disclosures using Vivli-hosted data

Modeling rates of trial attrition: an analysis of individual participant data from 90 randomized controlled trials of pharmacological interventions for multiple conditions Journal of Clinical Epidemiology| PI: David McAllister, University of Glasgow | Data Request 9492

Bayesian Power Prior in Platform Trials With Non-Concurrent Control for Binary Outcomes Statistics in Medicine| PI: Akihiro Hirakawa, Tokyo Medical and Dental University | Data Request 9907

An Empirical Comparison of Statistical Methods for Estimating Treatment Effects on EQ-5D in Randomized Clinical Trials Value in Health| PI: Jiajun Yan, McMaster University | Data Request 9253

Estimating hypothetical estimands with causal inference and missing data estimators in a diabetes trial case study Biometrics| PI: Johnathan Bartlett, London School of Hygiene & Tropical Medicine | Data Request 6764

Statistical Modeling to Adjust for Time Trends in Adaptive Platform Trials Utilizing Non-Concurrent Controls Biometrical Journal| PI: Günter Höglinger, Hannover Medical School | Data Request 6495

Assurance Methods for Adaptive Clinical Trials with a Delayed Treatment Effect White Rose eTheses Online| PI: James Salsbury, University of Sheffield | Data Request 9422