On the individual surrogate paradox.

Area of Science:

• Biostatistics
• Clinical Trial Design
• Epidemiology

Background:

• Surrogate endpoints are crucial when primary outcomes are difficult to measure in clinical trials.
• Existing surrogate paradox definitions do not fully capture individual-level treatment effect heterogeneity.
• The potential for a treatment to positively impact a surrogate but negatively impact a primary outcome in specific individuals is a critical concern.

Purpose of the Study:

• To define and investigate the "individual surrogate paradox," a phenomenon where treatment effects on surrogate and primary outcomes diverge at the individual level.
• To evaluate existing surrogate endpoint criteria against the manifestation of the individual surrogate paradox.
• To develop novel criteria for identifying and excluding surrogate endpoints susceptible to the individual surrogate paradox.

Main Methods:

• Analysis of existing surrogate criteria (e.g., strong binary surrogate) concerning the individual surrogate paradox.
• Application of sharp bounds to derive new criteria for paradox exclusion.
• Illustrative case study on intensive glycemia treatment and diabetic retinopathy progression.

Main Results:

• The individual surrogate paradox can occur even when average or distributional causal effects suggest a valid surrogate.
• Only the strong binary surrogate criterion inherently avoids this paradox without further assumptions.
• Novel criteria using sharp bounds were developed to identify and mitigate the individual surrogate paradox.

Conclusions:

• The individual surrogate paradox represents a significant challenge in surrogate endpoint validation, particularly for capturing treatment effect heterogeneity.
• The strong binary surrogate offers protection against this paradox, but novel criteria are needed for broader applicability.
• The proposed methods provide a framework for selecting reliable surrogate endpoints that avoid paradoxical individual-level treatment effects, as demonstrated in the diabetic retinopathy example.