Bayesian Dynamic Network Modeling with Censored Flow Data

There is an abundance of applied and theoretical statistical research focused on the analysis of network data. However, few applications have the flexibility to account for the inherently limited flow that results from constrained capacity at destination nodes and, thus, may provide an incomplete picture of the underlying data generating process. This thesis works to address this shortcoming by applying Bayesian Dynamic Flow Modeling in a context where the capacity at the destination node is limited. To that end, it develops a methodology for updating beliefs about flow rates when the flow is censored. These methods are applied to a publicly available bike sharing dataset that exhibits censoring during high-volume times of the day. The results show a comparison of network characterization from a model built under the assumption of censored flows and a model without that assumption. This analysis highlights specific circumstances in which the estimates of underlying demand from both models are most at odds with one another and provides a framework for guiding the analysis of datasets that can be similarly represented.