PRM62 - Variance Reduction Through Antithetic Variates as a Means of Developing Complex VBA Models with Reasonable Computation Times

Abstract

Modelling chronic conditions such as diabetes necessitates the development of complex models, raising issues of computational intensity and execution time. While the use of compiled languages such as C++ is more computationally efficient, concerns may exist regarding their transparency compared to commonly used VBA in Microsoft Excel. This study investigates the application of antithetic variates to the pseudo-random numbers of Monte Carlo simulation, to provide reductions in stochastic uncertainty, through the introduction of negatively correlated pairs of simulation replicates, as a means of developing complex VBA models with reasonable computation times. A simulation model of type 2 diabetes, based on the UKPDS 68 outcomes equations, was executed with and without the application of antithetic variates. The impact of the technique was evaluated through comparison of total cost and benefit estimates, predicted over a long-term horizon of 40 years. An approximate four-fold reduction was observed in the Monte Carlo Error (MCE) associated with estimated mean incremental costs and benefits, when antithetic variates were applied over 1,000 simulations of 1,000 individuals. For a fixed number of runs (1,000), the number of replicated individuals required to achieve 99% accuracy (MCE/mean<1%) in incremental cost and benefit estimates fell from approximately 500 and 550 respectively, to fewer than 50 with antithetic variates. Similarly, for a fixed cohort size (1,000) the same level of precision was produced with fewer than 10% of the simulation runs required otherwise. The use of antithetic variates can improve the precision of modelling output; reducing the number of simulation runs and thus computation time required to perform analyses. The use of such variance reduction techniques should be pursued in the simulation of chronic conditions, as a means of achieving manageable run times and facilitating the extensive scenario and sensitivity analyses required as part of economic evaluations.