Statistical prediction models are often trained on data that is drawn from different probability distributions than their eventual use cases. My work uses insights from causal inference to develop new methods for building machine learning models that are robust to environmental changes.

Combining multiple populations or contexts induces universal confounding on a structural causal model (SCM). Assuming a bound on the cardinality of a discrete universal confounder turns the problem into a mixture model, allowing identification of within-source probability distributions. This perspective expands the notion of causal identifiability, as many graphically unidentifiable relationships can be identified.

Many medical settings with privacy concerns deny direct access to data, requiring us to synthesize conclusions at a higher level. This setting is riddled with paradoxes - namely that conclusions are not necessarily transitive. We use “expert graphs” to define a new notion of consistency in networks of conclusions from differing contexts.

We process non-coding regions of the genome which contain duplication and mutation signatures. These mutation profiles have been shown to be predictive of various forms of cancer.