Simulation studies are virtually all motivated by decision making. Because simulation output is stochastic and input models are never perfect, all decisions should include an accounting for risks. Input model risk refers to the exposure due to imperfect simulation input models that are estimated from real-world data, involving both the level of uncertainty about the input itself and the propagation of uncertainty in the input to the output. This dissertation addresses both issues: reducing and measuring input model risk, with a focus on the latter. For reducing the input model risk, we propose model averaging, which is a weighted average of the candidate distributions in a given set with the weights tuned by cross-validation, and extend the implementation in the probability space to the quantile space that emphasizes the tail behavior. For measuring the input model risk, we propose a family of solutions to measure the local sensitivity of an output property to an input property, focusing on the output mean or variance with respect to the input mean or variance. We extend existing stochastic gradient methods to identify the point and error estimators for any member of the family from the nominal simulation experiment only. Based on this basic framework, we create a local sensitivity analysis technique for the clinical trial enrollment simulation at SAS Institute and demonstrate it on a realistic case for the U.S.

Creator

• JIANG, XI

DOI

• https://doi.org/10.21985/n2-t10t-pv39

Subject

• Operations research

Language

• en

Alternate Identifier

• etdadmin_upload_763007
• http://dissertations.umi.com/northwestern:15233

Keyword

• model averaging
• input modeling
• input model risk
• local sensitivity measures
• sensitivity analysis
• input uncertainty

Date created

• 2020-01-01

Resource type

• Dissertation

Rights statement

• In Copyright