Expanding behavior pattern sensitivity analysis with model selection and survival analysis.


Interventions aimed at the enteric bacterial population ecology, such as diet changes, may be effective at reducing the prevalence of tetracycline-resistant enteric bacteria in beef cattle. Behavior pattern sensitivity analysis is a useful and flexible tool for conducting a sensitivity analysis on models with varied output behavior, enabling prioritization of input parameters via regression model selection techniques. Cox proportional hazard models are an alternative to linear regression when behavior pattern measures are censored or linear regression assumptions cannot be met.

We have expanded the behavior pattern sensitivity analysis procedure by incorporating model selection techniques to produce parsimonious linear regression models that efficiently prioritize input parameters. We also demonstrate how to address common violations of linear regression model assumptions. Finally, we explore the semi-parametric Cox proportional hazards model as an alternative to linear regression for situations with censored data. In the example mathematical model, the resistant bacteria exhibited three behaviors during the simulation period: (1) increasing, (2) decreasing, and (3) increasing during antimicrobial therapy and decreasing after therapy ceases. The behavior pattern sensitivity analysis identified bacterial population parameters as high importance in determining the trajectory of the resistant bacteria population.

Sensitivity analysis is an essential step in mathematical modeling because it identifies parameters with a strong influence on model output, due to natural variation or uncertainty in the parameter values. Recently behavior pattern sensitivity analysis has been suggested as a method for sensitivity analyses on models with more than one mode of output behavior. The model output is classified by behavior mode and several behavior pattern measures, defined by the researcher, are calculated for each behavior mode. Significant associations between model inputs and outputs are identified by building linear regression models with the model parameters as independent variables and the behavior pattern measures as the dependent variables. We applied the behavior pattern sensitivity analysis to a mathematical model of tetracycline-resistant enteric bacteria in beef cattle administered chlortetracycline orally. The model included 29 parameters related to bacterial population dynamics, chlortetracycline pharmacokinetics and pharmacodynamics. The prevalence of enteric resistance during and after chlortetracycline administration was the model output. Cox proportional hazard models were used when linear regression assumptions were not met.

MIDAS Network Members