OBJECTIVES: For Health Technology Assessment (HTA) bodies, the use of probabilistic modeling increasingly informs decision making. However, running probabilistic patient-level microsimulations in HTA-accepted software like Excel requires substantial computation time. The objective of this analysis is to determine a practical method for optimizing computational time for multifactorial microsimulations.

METHODS: Previous work suggests that computation time can be optimized using equations involving precision required (c₂), and k, a ratio of mean patient-level variance and probabilistic run variance of incremental net benefit (INB). To calculate these parameters, a set of 3^d probabilistic runs is required, where d is the number of probabilistic inputs. However, this is infeasible in microsimulation models that have more than 5 inputs. Instead, use of one-way sensitivity analysis (OWSA) to select the two or three most impactful variables is proposed. The model used to test this approach was a diabetes microsimulation model comparing two therapies A and B.

RESULTS: A OWSA revealed that two input parameters that most impacted the estimate of INB were the difference in efficacy between A and B and the number of cost units of B. Creating a 3X3 matrix with microsimulation runs of n = 100,000 patients for each variable run as mean or mean+/- 1.5 SD allowed for calculation of k. This resulted in an optimal n of 18,556 and an optimal number of probabilistic simulations (N) of 3,200 for c₂ = 0.05. However, the run-time for these parameter values would be over two months in Excel using a four-core machine with 16Gb of RAM.

CONCLUSIONS: Selecting the top variables is a practical approach to determine optimal n and N for probabilistic microsimulations. However, the use of Excel to run these analyses is impractical, and use of languages such as R may be necessary.