PhD Dissertation Proposal Defense: Pracheta Amaranath, Simulation for Causal Inference and Causal Inference for Simulation: A Generative Modeling Approach
Content
Speaker
Abstract
This thesis explores the interplay between simulation and causal inference through the lens of generative modeling. In the first part, we construct metamodels for discrete-event simulations of queueing networks using modular dynamic Bayesian networks (MDBNs), which serve as compact, causal-preserving surrogates for the underlying simulators. Our preliminary work demonstrates that these models can approximate system behavior across a range of interventional settings, particularly in Markovian systems.
Building on this foundation, we propose to improve the efficiency and scalability of inference in MDBNs by integrating advances in Bayesian inference—specifically, using probability generating functions for tractable closed-form solutions or exploiting GPU-accelerated matrix operations for large-scale model evaluation. Furthermore, we aim to extend the applicability of MDBNs to semi-Markovian queueing systems, which pose new challenges for dynamic causal modeling.
In the second part, we use simulation as a tool to systematically evaluate and improve the evaluation of causal inference estimators. Our current investigations highlight inconsistencies that arise from the sensitivity of generative models to user-defined parameters and architectural biases. To address this, we develop a principled Bayesian framework, named simulation-based inference for causal evaluation (SBICE) that incorporates priors over these parameters, enabling the generation of synthetic datasets that better reflect the underlying data-generating process.
We propose to investigate ensemble-based generative models that incorporate priors over model classes themselves, providing a robust method for benchmarking causal estimators across diverse scenarios. Through these proposed directions, the thesis aspires to deepen the methodological connections between simulation and causal inference and offer new tools for efficient modeling, scalable inference, and principled evaluation of causal inference estimators.
Advisors
Peter Haas and David Jensen