Predictive Probability of Success (PPOS) in Longitudinal Endpoint to Predict Trial Outcome: What can we learn from a retrospective interim analysis?
Niclas Jonsson1, Richard Anziano1, Andrea Henrich1, Steven Kern2, Yasir Shafiq3, Niclas Jonsson1
1Pharmetheus , 2Bill & Melinda Gates Foundation, 3Center of Excellence for Trauma and Emergencies and Department of Community Health Sciences, Aga Khan University
Background: Clinical trials require significant resources (subjects, time and money) and often fail in achieving their objectives. Predicting the final outcome based on interim data can enable adaptations (futility, randomization ratios etc.) to maximize the information gathered in the trial. We used a pharmacometrics model at interim time points to allow prediction of likely outcomes for each individual subject, regardless of how much data had been gathered. Allowing for variability in response and model uncertainty offered the opportunity to assess final outcomes by using clinical trial simulation to predict the probability of success at the end of the trial. We had planned to implement this analysis as the trial was accruing but the Data Monitoring Committee (DMC) did not agree to allow early unblinded data, hence we performed a retrospective interim analysis at various points in the presence of full knowledge of the trail outcome. Objectives: To provide an example of PPOS using a longitudinal endpoint at various stages of recruitment and show how the convergence to the eventual outcome can be model and time dependent. Data: Data used in this analysis is from a randomized controlled trial , sponsored by the Bill & Melinda Gates Foundation and conducted by the VITAL Pakistan Trust. The trial, Mumta Lactating Women (LW), focused on assessing nutritional support for lactating women and Azithromycin for infants to improve growth outcomes in the peri-urban slums of Karachi, Pakistan. Methods: Individual predictions were done in two ways based on whether or not subjects had data at the interim analysis. If a subject had data, their individual posterior was used to predict the missing outcomes through the final observation period. For subject having no data at the interim analysis, the population posterior was used to predict the unobserved outcomes. In this way individuals with observed data would drive their outcomes while the assumption was that those who had not been observed would behave as the typical subject. Comparisons are made to what we would have known if we ran this exercise live rather than retrospectively to inform considerations for implementation. Results: The ability of the interim data to predict the outcome in the final analysis could be done very early on but the predictions were somewhat dependent on the model chosen, where better models predicted eventual outcomes earlier. In all cases, longitudinal models outperformed landmark analyses. Conclusions: PPOS is a viable means of predicting trial outcome with interim data. The degree and timing of how early adequate confidence in the predictions were dependent on the quality of the model used. This approach could be considered for applications where much is known about the longitudinal trajectory of the outcomes. Acknowledgements: This analysis was supported by the Gates Foundation. The contents are the sole responsibility of the authors and may not necessarily represent the official views of the Gates Foundation or other agencies that may have supported the primary data study used in the present analysis. The authors wish to recognize the members of the Ki community, including the principal investigators and their study team members for their generous contribution of the data that made this analysis possible and the members of the Ki team who directly or indirectly contributed to the study.