Propagating Ambiguity into Decision Analyses – Center for Health Decision Science

Conventional decision analytic models typically use probability distributions to represent parameter uncertainty. Thomas Trikalinos has developed a novel approach for representing uncertainty when available information is too limited to support meaningful probability distributions, which he presented in a recent CHDS seminar. Trikalinos is Professor of Health Services, Policy, and Practice and Director of the Center for Evidence Synthesis in Health at Brown University.

Ambiguity, also known as “deep uncertainty” or “pervasive uncertainty,” is defined as uncertainty that a given analyst is unwilling or unable to describe with a probability measure model. Alternatively, ambiguous parameters can be represented by an uncertainty set wherein each parameter is characterized only by their lower and upper bounds. Uncertainty sets are compatible with probability measure models, allowing analysts to mix both representations of uncertainty within a given analysis based on available evidence and computational power.

As an example, Trikalinos described a decision analysis comparing test-and-treat strategies for latent tuberculosis infection among immigrants in the United States. The analysis was extended to incorporate deep uncertainty about screening test performance. When using uncertainty sets, outcomes like net monetary benefit (NMB) are not expressed as expected values but rather as ranges (outcome set). Trikalinos compared NMB ranges for each test-and-treat strategy across different optimziation criteria (e.g., maximax, maximin, Hurwicz) to identify possible optimal strategy sets under deep uncertainty.

This framework gives decision modelers greater flexibility in how they represent uncertainty and holds particular use when evidence is limited and distributional assumptions are not robust for some parameters.

Read the case, Cost-Effectiveness of Testing and Treatment for Latent Tuberculosis Infection in Residents Born Outside the United States With and Without Medical Comorbidities in a Simulation Model

Join the CHDS mailing list to be notified about future CHDS seminars.