SHELF (Sheffield Elicitation Framework)
SHELF is a structured method that helps experts share their best guesses about uncertain quantities, especially when there isn't enough data. It combines expert opinions into a probability distribution used in health technology assessments.
At a glance
Use when
Empirical data are sparse or unavailable, and uncertainty must be quantified for decision-analytic models
Avoid when
Sufficient high-quality data exist to inform parameters, or when experts lack familiarity with the topic or probabilistic reasoning
Inputs
Expert judgments (e.g., quantiles, central estimates, confidence levels), facilitator guidance, predefined elicitation questions
Outputs
Subjective probability distributions for model parameters, often in the form of fitted distributions suitable for PSA
How it works
SHELF is a formal expert elicitation protocol developed at the University of Sheffield by Oakley and O'Hagan. It provides structured templates and guidance for facilitating group elicitations, aiming to quantify uncertainty in model parameters through subjective probability distributions. It supports probabilistic sensitivity analysis (PSA) in decision models and is implemented with a companion CRAN R package for analysis and aggregation.
- HTA domains
- Clinical Effectiveness, Costs & Economic Evaluation, Patient and Social Aspects
- Assumptions
- Experts are knowledgeable, willing to express uncertainty probabilistically, and can be effectively facilitated; group processes reduce individual bias
- Strengths
- Provides a rigorous, transparent, and reproducible approach to expert elicitation; includes tools and templates to support facilitation and analysis; integrates with statistical models and PSA frameworks
- Limitations
- Relies on availability and quality of expert input; potential for cognitive biases despite structured process; requires skilled facilitation and time commitment
- Also known as
- Sheffield Elicitation Framework
Questions this answers
- › What is the most likely value of an uncertain parameter when empirical data are limited?
- › How can expert opinion be systematically combined to represent uncertainty?
- › What probability distribution best reflects group expert judgment for a model input?
- › How confident are experts in their estimates, and how should this uncertainty be quantified?
- › Which sources of uncertainty in a model can be addressed through expert judgment?
- › How can bias in expert opinion be minimized during elicitation?
References & sources
Similar by meaning
Beta record. Based on the original catalogue summary; primary-source enrichment pending.