Sensitivity analysis of censored output through polynomial, logistic, and tobit regression meta-models: theory and case study

Abstract

The paper focuses on simulation output that may be censored; that is, the output has a limited range (examples are simulations that have as output the time to occurrence of a specific event, such as a 'rare' event, within a fixed time horizon). For sensitivity analysis of such simulations we discuss three alternatives: (i) traditional polynomial regression models, (ii) logistic or logit regression, and (iii) tobit analysis. The case study concerns the control of a specific animal disease (namely, IBR) in The Netherlands. The simulation experiment has 31 environmental factors or inputs, combined into 64 scenarios, each replicated twice. Traditional polynomial regression gives some estimated main effects with wrong signs. Logit regression correctly predicts whether simulation output is censored or not for 92% of the scenarios. Tobit analysis does not give effects with wrong signs; it correctly predicts censoring for 89% of the scenarios.