THE BENEFITS OF EMPLOYING COMMON RANDOM NUMBERS IN MICROSIMULATION MODELING: AN EXAMPLE OF A CERVICAL CANCER MODEL

Purpose: The use of Common Random Numbers (CRN) in Monte Carlo simulation modeling can reduce stochastic noise between model runs. We estimated the time and variance reductions of CRN using a previously developed individual-based model of cervical cancer.

Method: Two versions of the model were used with one version utilizing CRN and the other using a seeded random number generator (RNG).  For each version, simulations ranging in size from 1,000 to 1,000,000 women were run estimating the benefits of cervical cancer screening.  For each cohort size, 50 pairs of runs were conducted, simulating two strategies:  (1) no screening and (2) HPV screening 3 times in a lifetime at ages 35, 40, and 45.  Both sets in the pair had the same random number starting seed, but each of the 50 pairs had distinct seeds in order to generate the mean and range of results for each cohort size. Standard deviations in model outcomes were calculated for each set of results to estimate variance reduction when using CRN.

Result: Percent reduction in cancer deaths from HPV screening compared to no screening was calculated for each pair. With CRN, the mean percent reduction in cancer deaths ranged from 44.3% to 44.8% when using cohort sizes ranging from 1,000 to 1,000,000; standard deviations decreased from 10.0% (n=1,000) to 0.33% (n=1,000,000).  With RNG, the mean in reductions ranged from 39.9% to 45.1%, and standard deviations were roughly twice as large as the corresponding result in the CRN set.  The ranges of results within each set of 50 were also twice as wide when comparing RNG to CRN runs.  Furthermore, our results showed that to generate a standard deviation in a RNG run comparable to that of a CRN run, the sample size must be increased at least threefold. When considering outputs besides cancer deaths, the trend of decreased noise with CRN was consistent.  A drawback to CRN was the slight increase in computation time per run.

Conclusion: Utilizing CRN in a model of cervical cancer resulted in lower variance than when using the more conventional RNG.  The moderate increase in computation speed necessary for CRN implementation was compensated by the advantage of being able to simulate smaller sample sizes without a large variance which decreased overall processing time.