|
Abstract:
|
Meta-analysis allows the combination of evidence from multiple studies, making it a natural and powerful tool to understand the toxicological profile of newly developed medical interventions and identify risk factors for serious adverse events, which can be life-threatening. However, these events are often rare or subject to left-censoring, since low-incidence events may be omitted from official trial reports. An additional challenge is that risk factors may interact with one another, but including interaction terms greatly increases the number of predictors and makes accurate effect estimation difficult. We address this challenge by imposing sparsity on the interaction terms, through the use of horseshoe priors in a Bayesian variable selection framework. We demonstrate through simulation studies that our method enables the identification of relevant interactions, and improves the accuracy of risk prediction over existing methods. Finally, we illustrate our proposed method with an application to meta-analysis of rare adverse events in cancer immunotherapy, identifying factors that elevate risk for specific classes of serious adverse events.
|
