Assessing the predictive value of a binary surrogate for a binary true endpoint based on the minimum probability of a...

Area of Science:

• Biostatistics
• Clinical Trials
• Causal Inference

Background:

• The individual causal association (ICA) is a recent metric for surrogate endpoint evaluation within a causal inference framework.
• ICA quantifies the association between individual causal effects on surrogate (ΔS) and true (ΔT) endpoints, ranging from 0 to 1.
• Interpreting ICA becomes challenging beyond deterministic or independent relationships between ΔT and ΔS.

Purpose of the Study:

• Introduce novel surrogacy metrics: minimum probability of prediction error (PPE) and reduction in prediction error (RPE).
• Provide metrics with more straightforward interpretations for surrogate endpoint validation.
• Develop an R package for practical application of these surrogacy metrics.

Main Methods:

• Defined minimum probability of prediction error (PPE) for binary endpoints, measuring erroneous prediction of ΔT using ΔS.
• Introduced reduction in prediction error (RPE) to address PPE's upper bound dependency on the true endpoint.
• Illustrated methodology with data from two clinical trials.

Main Results:

• PPE offers a more interpretable measure of surrogate predictive value compared to ICA.
• RPE provides a normalized measure (0 to 1) of how well a surrogate predicts the true endpoint, with 1 indicating perfect prediction.
• The RPE metric effectively quantifies the information ΔS conveys about ΔT.

Conclusions:

• PPE and RPE offer improved and more interpretable metrics for surrogate endpoint validation in causal inference.
• The RPE metric provides a robust assessment of surrogate utility, ranging from no information to perfect prediction.
• A user-friendly R package 'Surrogate' is available for implementing these validation methods.