OBJECTIVES: Probabilistic uncertainty analysis (PUA) is common in economic evaluations, but generating enough random samples to achieve stability can be computationally intensive. This study evaluates the relative efficiency of alternative random number sampling methods to stably estimate uncertainty within a cost-effectiveness analysis.

METHODS: We modified a published cost-effectiveness model of kidney disease including 18 random parameters to test the number of samples required for the average of the PUA results to converge to the deterministic value (i.e., “efficiency”). Two sampling methods were compared to the traditional random draw: full factorial (FF) and Latin hypercube (LH). Both divide the parameter space into discrete sections, sampling one value from each. The FF method samples one value from every section in the space, requiring n_sections^n_parameters total samples. The LH method samples n_sections values such that each section for each parameter is sampled exactly once.

RESULTS: The number of samples required for the FF method functionally limited our experiment to 3 random parameters. The FF method was as efficient as random draws of an equivalent number of samples when each parameter space was divided into 5 (n=125 samples), 10 (n=1000), or 25 (n=15,625) sections. With 18 parameters, the LH method was as efficient as random draws with 10 divisions of space (n=10 samples), 50 (n=50), 500 (n=500), and 5000 (n=5000). One notable difference was that less variability was associated with multiple sets of LH samples than with multiple sets of an equivalent number of random draw samples.

CONCLUSIONS: As an alternative to purely random sampling, the FF method is impractical. In our test model, the LH sampling method achieved results in individual sample sets with less variability than an equivalent number of random samples, suggesting that the LH method could potentially achieve PUA stability faster in models with higher overall uncertainty.