Combining bootstrap-based stroke incidence models with discrete event modeling of travel-time and stroke treatment: Non-normal input and non-linear output

Abstract

Incidence rates in simulation models are often assumed to stem from Poisson processes, with rates based on analyses of real-life data. In cases where the record of data is limited, or observed rates are low, the stochastic process involved in sampling from modeled distributions may not adequately reflect the uncertainty around the estimated input parameters. We present a conceptually simple, but computationally demanding, method for generating variance in incidence through the use of bootstrapping; for each subsample, a regression model is fitted, and the simulation model is run repeatedly sampling from the fitted model. Stochasticity is introduced at two levels; data for fitting the regression, and sampling from the fitted model. We illustrate this hybrid approach using Norwegian stroke records to generate stroke incidences with age, sex, and location, in a simulation model made to analyze travel time, queuing, and time to treatment in regional stroke units.