THE NEED FOR PROBABILISTIC SENSITIVITY ANALYSIS IS NOT A REASON TO PREFER COHORT MODELS TO MICROSIMULATION

Purpose: To dispel a persistent myth. Background: The literature has for some years contained examples of probabilistic sensitivity analysis (PSA) carried out upon microsimulation models (also known as individual sampling models or ISMs). Despite this, one still hears the claim that cohort models are preferable, and that PSA is a reason for preferring cohort models. This is usually accompanied by the assertion that cohort models run more quickly than microsimulations. While it is true that a simple cohort model will take longer to run if converted to a microsimulation, models using complex patient pathways are modeled as microsimulations because they would take much longer to run as cohort models preserving the richness of structure. 

Method: Comparison of use of the Birmingham Rheumatoid Arthritis Model (BRAM) under deterministic and probabilistic sensitivity analysis. The BRAM is a microsimulation model that has been used for successive appraisals of rheumatoid arthritis drugs by the UK National Institute for Health and Clinical Excellence.

Result: When run with a fixed set of parameters, it was often necessary to run as many as 100,000 simulated patients to obtain sufficiently accurate estimates of the population results using those parameters. On the other hand, when run under PSA, 5000 patients per parameter set were sufficient. Thus a PSA with 1000 parameter sets needed only a total of 5 million simulated patients, that is 50 times the computational effort of an analysis with fixed parameters. An equivalent PSA for a cohort model necessarily requires 1000 times the computational effort of the analysis with fixed parameters. Assuming that the cohort model takes the same time to run for a single parameter set as the microsimulation with 100,000 patients, this means that the running time for a PSA with the cohort model is equivalent to running the microsimulation for 100 million patients, 20 times as many as the microsimulation.

Conclusion: Given equivalent complexity of models, a PSA under microsimulation can easily be 20 times quicker than the PSA for a cohort model. The need for PSA is a good reason to prefer microsimulation models when the complexity of the patient pathway warrants their use in deterministic analysis.