ADVANCED TOPICS IN DECISION ANALYTIC MODELING

Overview: This course will review different types of decision analytic models and recommendations of the ISPOR-SMDM Modeling Good Research Practices Task Force. We will then highlight a few typical mistakes that can lead to biases in the outcomes of interest and provide advanced methods to avoid such mistakes, which could be unavoidable with commonly used software.

Background: Decision analytic models are commonly used in medical decision making. The choice of modeling technique—cohort versus individual-level (microsimulation, discrete-event simulation and agent-based modeling), static versus dynamic, and continuous- versus discrete-time—can influence the results. In addition, incorrect modeling assumptions can also provide misleading outcomes. This course will review different types of decision analytic models. We will then highlight a few typical mistakes that can lead to biases in the outcomes of interest and provide advanced methods to avoid such mistakes, which could be unavoidable with commonly used software.

Format Requirements: The course will include a combination of didactic lectures, discussions and hands-on exercises. This course is designed for those with basic understanding of decision analytic modeling (e.g. comfortable with introductory modeling courses offered at SMDM) and interested in building relatively complex models. It is recommended that participants bring a Windows based computer.

Description and Objectives:

This course will cover advanced topics in the following three modeling approaches: cohort-based models, patient-level models, and population-based models. We will provide several examples illustrating the mistakes and bias, and practical solutions to avoid such mistakes. Participants will acquire the following skills from this short course.

1.    Discrete-time Markov model (DTMM)

The choice of DTMM with commonly used cycle lengths (e.g., a year) can lead to biased outcomes.

a)    Half-cycle correction: Understand biases in modeling outcomes (costs, life-years, etc.) by ignoring half-cycle correction in DTMM and learn different methods for half-cycle correction to reduce bias.

b)   Competing risk in Markov models: Understand issues with changing cycle lengths in Markov models without accounting for competing risk, and learn methods to incorporate competing risks to avoid bias in results.

2.    Continuous-time Markov model (CTTM)

Since in most biological and healthcare systems state transitions can occur at any time, use of DTMM instead of CTMM can result in biased outcomes. Using a series of examples, we illustrate the formulation and numerical simulation of DTMM and CTMM, and examine the differences between the two approaches.

3.    Cohort- versus individual-level models

Understand biases with cohort-based approach because of non-linear relationship between the model inputs and outcomes, and how individual-level models (Markov simulation models, discrete event simulations, agent-based models), can incorporate heterogeneity in data.

4.    Static versus dynamic models

Understand differences between "static" (e.g. Markov) versus dynamic population models and learn about incorporating indirect effects into static models of infectious diseases using analytical formulae.

5.    Probabilistic sensitivity analysis (PSA)

Learn different sampling techniques (e.g. Latin Hypercube) to efficiently perform PSA in individual-level models, which can be computationally challenging and impractical with a commonly used Monte Carlo sampling.