LIDS Tea Talk | Mansi Sood & Phevos Paschalidis

Phevos Paschalidis

"Causal Inference in Sequential Settings under Interference and Latent Confounding"

We study causal inference under outcome interference for sequential, observational settings.  Specifically, we consider settings where the binary outcomes over N units are Markovian across T time steps. At each time step, the outcomes of N units have dependencies captured through an Ising model;  each outcome is also impacted through an external field capturing the effects of its treatment as well as latent confounders. Similar to panel data literature, these latent confounders are modeled to have a low-rank factor structure. Our data is a single sample from this high-dimensional distribution.  To estimate causal quantities of interest, we provide a computationally efficient method based on Maximum Pseudo-Likelihood Estimation (MPLE) for learning the model parameters. Under mild assumptions, we establish non-asymptotic consistency for parameter estimation and show this translates to faithful estimation of causal quantities of interest after sampling from the learned model. We demonstrate the efficacy of the method through synthetic experiments as well as a real-world case-study investigating causal effects of vaccine rates on COVID-19 death rates within US counties nationwide.

Phevos Paschalidis is a second-year Ph.D. student in EECS advised by Professors Constantinos Daskalakis and Devavrat Shah. He is broadly interested in theoretical guarantees for learning methods, including for causal inference and diffusion models. He obtained his B.A. in Computer Science at Harvard in 2025 where he worked with Professors Na Li and Salil Vadhan. He is supported by an NSF Graduate Research Fellowship.

‍

‍

Mansi Sood

"Learning and Intervening under Heterogeneity: From Mixtures to Causal Inference"

Latent heterogeneity is pervasive in observational data, where samples may arise from distinct populations, hidden states, or operating regimes. To capture latent heterogeneity in observational settings, we consider the task of learning mixtures of high-dimensional exponential-family distributions, where observations exhibit within-sample dependencies through a shared interaction parameter. We propose an estimator that first minimizes a computationally tractable convex loss, avoiding the log-partition function. We conclude by discussing the implications of this approach for learning and intervening under latent heterogeneity in complex systems, with applications ranging from causal inference to physical and analog computing systems.

Mansi Sood is a Schmidt Science Fellow at the Massachusetts Institute of Technology Laboratory for Information and Decision Systems hosted by Devavrat Shah. Her research focuses on foundations for learning and inference in complex networked systems. She previously completed her PhD at Carnegie Mellon University (CMU), where she received the A.G. Jordan Award for Outstanding Thesis ('25), an Information Theory and Applications (ITA) Graduation Day Award ('24), and the Best Paper Award at the IEEE International Conference on Communications (ICC) ('21). She has been twice recognized as an Electrical Engineering and Computer Science (EECS) Rising Star by MIT ('23) and Georgia Tech ('21).