Presentation (CTI)

OBJECTIVES: Patient-level simulations (PLS) offer flexibility for complex health economic decision problems but are intrinsically probabilistic. Historically, computational constraints in Excel and Visual Basic for Applications (VBA) have suppressed N, introducing under-recognised decision risk. We investigated metrics to quantify convergence in an HTA microsimulation of resistant hypertension, and explored the relative benefit of one large cohort versus many smaller ones.
METHODS: A VBA microsimulation was replicated and performance-optimised in R, enabling convergence to be examined up to N=1,000,000. Monte-Carlo sampling error was quantified by holding second-order parameters fixed and varying random seeds across 50 iterations, summarised by coefficient of variation (CV) on costs, quality-adjusted life years (QALYs) and incremental cost-effectiveness ratios (ICERs). Single-cohort runs (N=20,000; 100,000; 1,000,000) were compared against many-small-cohort designs (50 sets of 100 simulations at N=1,000).
RESULTS: Baseline rates, survival and costs stabilised at N values between 8,000 and 15,000; QALYs at N values between 15,000 and 20,000. CVs fell non-linearly with N: costs 2.75%, 1.00%, 0.26%; QALYs 3.74%, 1.70%, 0.51%; ICERs 4.00%, 1.90%, 0.48% for N=20,000, 100,000 and 1,000,000 respectively. Aggregating 100 small cohorts of N=1,000 produced markedly worse stability (CVs 1.68%, 3.27%, 3.46%) than a single cohort of equivalent total size, demonstrating that variance does not pool linearly across replications. R implementation delivered a 62-fold runtime reduction versus VBA (24 seconds vs 25 minutes, N=100,000), enabling this simulation.
CONCLUSIONS: PLS should deploy a single cohort of at least 20,000 patients, with diminishing marginal benefit beyond this threshold. Many small cohorts offer less stability than one large cohort of equivalent total size, a non-intuitive result with direct implications for HTA reproducibility and payer decision risk.