Accommodating missingness when assessing surrogacy via principal stratification.

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Causal Inference

Background:

• Surrogate markers are crucial in clinical trials for outcomes that are late-occurring or hard to measure.
• Understanding the causal effect (CE) of treatment on both the surrogate and the outcome is key to assessing a surrogate's value.
• Traditional regression methods for evaluating surrogate markers have limitations, including instability and inability to handle unmeasured confounding.

Purpose of the Study:

• To extend the principal surrogacy approach to handle missing outcome data in clinical trials.
• To address potential unmeasured confounding in the surrogate-outcome relationship.
• To incorporate counterfactual components of missingness, a novel aspect not previously addressed.

Main Methods:

• Utilized a principal surrogacy framework based on counterfactuals within prerandomization principal strata.
• Extended existing Bayesian methods to accommodate missing outcome data under ignorable and nonignorable missingness mechanisms.
• Allowed for counterfactual aspects of missing data, enhancing robustness.

Main Results:

• Applied the developed methods to a glaucoma clinical trial comparing surgery and medication, using intraocular pressure (IOP) control as a surrogate.
• Simulations explored the impact of nonignorability, prior sensitivity, and model selection using the decision information criterion (DIC).
• Demonstrated that even with non-identifiable parameters, the approach can yield meaningful results.

Conclusions:

• Model parameters for surrogate marker evaluation are often not fully identifiable from data.
• Informative priors can introduce bias, while noninformative priors may lead to wide credible intervals.
• The proposed methods, despite identifiability challenges, offer a robust way to assess surrogate marker validity by exploring sensitivity and identifiability.