GFDI Student Seminar: Ahmed Elshall
Theory and Application of Bayesian Multimodel Analysis with Application to Microbial Soil Respiration Models
Abstract: Models in biogeoscience are subject to parametric and conceptual uncertainties. To accommodate different sources of uncertainty, multimodel analysis such as model selection and model averaging are becoming popular. This talk will present the theoretical and practical challenges of Bayesian multimodel analysis, using a microbial soil respiration modeling example. We are interested in these models because global soil respiration releases about ten times more carbon dioxide to the atmosphere than all anthropogenic emissions. Improving our understanding of microbial soil respiration is essential for reducing the uncertainties of earth system models. This study focuses on a poorly understood phenomena, which is the soil microbial respiration pulses in response to episodic rainfall pulses, the “Birch effect”. The hypothesis is that the “Birch effect” is generated by three mechanisms that will be discussed during the talk. To test the hypothesis, five microbial-enzyme models were developed and assessed against field measurements from a semiarid Savannah that is characterized by pulsed precipitation. These five models evolve stepwise such that the first model includes none of the three mechanism, while the fifth model includes the three mechanisms. The first part of the talk will illustrate Bayesian model selection for the five models. Bayesian inference, which involves updating a prior parameter disruption to a posterior parameter distribution using a likelihood function, will be illustrated as well as the estimation of Bayesian model evidence for model selection purpose. The second part will discuss an important theoretical and practical challenge, which is the effect of likelihood function selection on both Bayesian model selection and model averaging. The talk will show that making valid inference from scientific data is not a trivial task, since we are not only uncertain about the candidate models, but also about the statistical techniques that are used to appraise these models.