One in 10 Americans is living with diabetes, a chronic disease that occurs when the pancreas is no longer able to make insulin, or when the body cannot make good use of the insulin it produces. Diabetes impacts an individual’s quality of life, decreases their life expectancy, and increases their risk of heart disease, stroke, kidney failure, and dementia. These consequences of diabetes highlight the importance of identifying effective strategies for prevention of diabetes, and treatment and management of diabetes after diagnosis. In studies that aim to formally test such interventions, it is often the case that long term follow-up of patients is needed in order to observe the primary outcome. The identification and use of surrogate markers in such intervention studies have the potential to support more timely decision-making about the intervention’s effectiveness. While incredible progress has been made in the development of statistical methods to identify surrogate markers, existing methods for evaluating surrogate markers and appropriately using surrogate markers to test for a treatment effect in a future study still face key challenges. Importantly, there are no robust methods to identify when a surrogate may be useful for only certain subgroups of patients, where such subgroups are defined by multiple patient characteristics. Existing methods are limited to a single patient characteristic, but in practice, heterogeneity of a surrogate is more complex and is likely a function of multiple covariates. For example, the surrogacy of changes in hormone levels likely varies by not just sex but also age and body mass index. The impact of such complex heterogeneity is especially important to understand with respect to its potential to lead to violations of surrogacy assumptions that existing methods commonly require to ensure that future testing on the surrogate would accurately reflect testing on the primary outcome. When such testing is not in alignment, termed a surrogate paradox, incorrect conclusions about a treatment may be made. In this study, we aim to develop and apply robust statistical methods to address these challenges. Our methods, software, and results have the potential to inform and improve the design and analysis of future studies aimed at diabetes prevention by identifying when and how a surrogate marker can be used in future studies.