Fig. 1

Behavioral sensitivity analysis process. The outputs of Monte Carlo simulations of mathematical models are classified into behavior pattern modes and pattern measures are defined for each of the modes. Standardized input parameter values from Monte Carlo simulations are used to build regression models for each pattern measure from each of the behavior pattern modes. Smoothing spline curves are fit to the simulation outputs if necessary to eliminate noise and enable calculation of the behavior pattern measures. Variable selection and model fit evaluation methods are used to find each best-fit regression model. Validity of assumptions for the best-fit regression model is evaluated; dependent (simulation outputs) and independent (parameters) variable transformations or other appropriate approaches such as time-dependent coefficients are used to meet the regression model assumptions if necessary. To obtain a most parsimonious regression model, parameters with relatively small coefficients in the best-fit model are eliminated, starting with the smallest, if there is no substantial change in model fit or the other parameter coefficients. Validity of assumptions is re-evaluated for the most parsimonious regression model