Latin hypercube sampling and the sensitivity analysis of a Monte Carlo epidemic model

Area of Science:

• Computational Biology
• Epidemiology
• Statistical Modeling

Background:

• Discrete, algorithmic, and Monte Carlo simulations are widely used across various scientific disciplines.
• Assessing model sensitivity to feature variations is crucial but challenging, as traditional methods for differential equations are inapplicable.
• Current ad hoc methods for sensitivity analysis are often limited in scope and computational efficiency.

Purpose of the Study:

• To compare the effectiveness of Latin hypercube (LH) sampling with full factorial (FF) sampling for sensitivity analysis in discrete simulations.
• To evaluate the trade-offs between computational cost and information gain in sensitivity studies.
• To demonstrate the application of these methods using a Monte Carlo model of influenza virus epidemics.

Main Methods:

• Implementation of a sensitivity analysis using Latin hypercube (LH) sampling.
• Comparison with a sensitivity analysis employing a full factorial (FF) sampling design.
• Utilizing a discrete, Monte Carlo simulation model of human community influenza epidemics as a case study.

Main Results:

• Both LH and FF sampling designs yielded comparable predictive abilities for the simulation model.
• Both methods provided similar levels of information regarding simulation sensitivity to model features.
• The LH sampling design required substantially fewer samples (over 14 times fewer) than the FF design, indicating greater computational efficiency.

Conclusions:

• Latin hypercube (LH) sampling is a viable and computationally efficient alternative to full factorial (FF) sampling for sensitivity analysis in discrete simulation models.
• LH sampling can achieve comparable results to FF sampling with significantly reduced computational burden.
• This study highlights the utility of LH sampling for complex models in fields like population biology and epidemiology.