MODELLING SURVIVAL DATA FOR DECISION ANALYSIS

Format Requirements: This is not a statistics course and so statistical knowledge / skills are not a pre-requisite. However, we will be discussing statistical modelling and so familiarity with regression-type modelling will be assumed. The focus on the use of survival analysis for decision modelling will assume that participants are familiar with the basic concepts of model building and understand the distinction between decision trees, Markov models, discrete event simulation, etc. The course will be a mixture of informal lectures and practical sessions. The practicals will use Microsoft Excel for calculations so no knowledge of statistical software is required. We will finish the course with a question and answer session which will allow participants to seek clarification on any of the content and encourage debate about how best to utilise survival models for decision analysis purposes.

Survival data are ubiquitous in medical research and medical decision modelling often requires modelling the full lifespan of patients in order to calculate outcomes for comparative effectiveness such as life-years and Quality Adjusted Life Years (QALYs). A standard medical statistics course on survival analysis would focus on Kaplan-Meier and Cox regression methods with median survival as the summary statistic of interest and effect sizes measured by hazard ratios. However, decision modellers typically need parametric survival models to obtain estimates of mean survival and effect sizes expressed as differences in life expectancy between intervention arms. In this course, we will introduce and review the use of survival data from the perspective of the decision modeller and provide practical demonstrations. In particular, the objectives are:

• To introduce the Kaplan-Meier method and demonstrate how it can be used to calculate (usually restricted) life expectancy and transition probabilities
• To introduce common parametric survival models used in decision analysis and illustrate how they can be used for extrapolating (and therefore obtain unrestricted life expectancy estimates) and partitioning survival curves
• To demonstrate how parametric survival models can be used to adjust differences in life expectancy for covariate imbalance and/or obtaining life expectancy for sub-groups of interest
• To demonstrate how outputs from survival models can be used as inputs in Markov models, discrete event simulations and cumulative incidence implementations