SENSITIVITY ANALYSIS AND VALUE OF INFORMATION ANALYSIS USING REGRESSION METAMODELING IN R

Overview: In this course, participants will learn the theory and application of metamodeling in sensitivity analysis and value of information analysis using a dataset of probabilistic sensitivity analyses.

Background: Sensitivity analysis and value of information analysis (VOI) are key concepts in decision analyses. VOI informs study design and resource allocation for further research. However, traditional methods of conducting sensitivity analyses, and VOI are often computationally demanding. Metamodels are statistical models that can reveal various characteristics of the original simulation model, including sensitivity analysis. In this course, participants will learn the basics of regression metamodeling and its use to efficiently conduct sensitivity analysis and VOI using a single dataset of probabilistic sensitivity analysis (PSA) in R.

Format Requirements: This course consists of lectures explaining sensitivity analyses and the theory of VOI interspaced with “hands-on” experience conducting these analyses using linear regression metamodeling in R. Participants will work through structured examples using their own computers. Data sets and files needed for the course will be distributed during the course session. A basic level experience with PSA, regression and VOI are preferred, but not required.

Description and Objectives:

The purpose of this course is to familiarize participants with the concepts and application of linear regression metamodeling in sensitivity analysis and VOI. By the end of this course, participants will gain hands-on-experience using linear regression metamodeling as a tool to produce a panel of sensitivity analyses as recommended by the SMDM-ISPOR Task Force and calculating various measures of VOI from a PSA dataset, including:

• One- and two-way parameter sensitivity analyses, and threshold analyses.
• Expected value of perfect information (EVPI): The value of eliminating all sources of parameter uncertainties in a model.
• Expected value of partial perfect information (EVPPI): The value of eliminating uncertainty for one or more parameters accounting for possible correlation. 
• Expected value of sample information (EVSI): The value of collecting additional information of all parameters of the model for a given sample size (n).
• Expected value of partial sample information (EVPSI): The value of collecting additional information on one or more parameters accounting for possible correlation for a given n.
• Expected net benefit of sampling (ENBS): The benefit of collecting additional information after accounting for the cost of research for a range of sample sizes and alternative research study designs.
• Optimal sample size (n*):  The maximum number of patients in a research study design that provides the highest ENBS.

In addition, we will provide R code to produce publication quality figures and tables for these measures.