← Back to Event List

Doctoral Dissertation Defense: Yun-Ju Cheng

Advisor: Dr. Yi "Yvonne" Huang


Mathematics/Psychology : 102

Date & Time

October 29, 2018, 9:00 am11:00 am


Title: Marginal meta-analysis combining randomized clinical trials with rare events

The adverse outcomes in safety studies are commonly rare in randomized clinical trials (RCT). Rare events challenge the standard Meta-analysis methodology in three aspects when combining those RCT.  First, whether and how to include trials with zero event in one or both arms are debated in academic communities, and there are increasing concerns on various ad-hoc continuation corrections methods.  Even though those ad-hoc methods may improve numerical performance of those meta-analysis combining RCTs with rare events, they may bias the result in various ways.  Second, the inferences and conclusion made from the analysis may change dramatically in various standard MA estimations with or without continuation corrections using the same data and same effect measures.  Third, the Cochran’s heterogeneity test, often used to determine fixed vs.  random effects MA methods, has low power in rare events setting.

The first part of this dissertation relaxed the typical “effect at random” assumption under standard MA framework to a more flexible set of “studies at random” assumptions by removing their association to a specific effect measure. Under the assumption studies completely at random (SCAR), a marginal MA estimator was proposed to estimate the average treatment effect combining RCTs with rare evetns. It provides not only a consistent treatment effect estimate for marginal causal effect, but also addresses the challenges stated above. Simulation study showed the proposed estimator performs reasonably well under various setting. Avandia study is our real application where the proposed marginal MA and standard MA methods are compared.

To relax the SCAR setting to allow study selection bias indicated by individual patients characteristics in RCTs, the second part of this dissertation focus on a patient-level marginal meta-analysis incorporating inverse-probability-sampling-weight (IPSW) as a tool to deal with individual level stochastic sampling error within the framework of study at random (SAR). When study selection bias can be correctly modeled using observed patient level data across RCTs, this IPSW MA estimator is a consistent estimator. Simulation studies showed that the proposed IPSW method's performance is non-inferior to currently existing approaches, and could beat the standard approaches under rare event settings.