Research
Bayesian methodology for learning from complex, evolving data.
My research is broadly about Bayesian learning for complex data that arrive or evolve over time. I am especially interested in methods that preserve sequential updating, causal questions involving longitudinal measurements, and nonparametric models that provide flexibility without giving up principled uncertainty quantification.
01
A recurring goal in my work is to make Bayesian updating genuinely sequential: when new data arrive, inference should be updated from what has already been learned rather than recomputed from the beginning. This is particularly important in time-series and streaming settings, where monitoring, forecasting, and decision-making occur continuously.
My master’s work developed computational methods for k-parametric dynamic generalized linear models (k-DGLMs). These models extend dynamic generalized linear models to response distributions with multiple evolving parameters. The methodology combines conjugate updating with information-geometric approximations so that filtering, smoothing, and forecasting can be carried out efficiently for non-Gaussian dynamic models. I have also worked on extensions for multiple related time series while preserving the sequential structure of the analysis.
This work is implemented in the kDGLM R
package.
Related work: k-parametric DGLMs · kDGLM package paper
02
A second line of my research concerns surrogate markers: intermediate measurements that may capture part of a treatment’s effect on a primary outcome. The central causal question is not simply whether a surrogate predicts the outcome, but how much of the treatment effect on the outcome is explained through the surrogate.
I developed a framework for settings in which both the surrogate and the primary outcome are longitudinal. The method defines treatment-effect and proportion-of-treatment-effect-explained quantities over time, allowing surrogate strength to be studied locally and cumulatively along the trajectory. State-space models provide an efficient estimation strategy that can accommodate temporal evolution and subject-level variability, while bootstrap inference and a temporal-homogeneity test address uncertainty and changes in surrogate strength over time.
The methodology is implemented in the OnlineSurr
R package.
Related work: A Causal Framework for Evaluating Jointly Longitudinal Outcomes and Surrogate Markers
03
My current work uses Bayesian nonparametric mixture models to make surrogate evaluation more flexible and more naturally compatible with streaming data. The motivation is to reduce sensitivity to a prespecified working mean model while retaining a full probabilistic description of uncertainty.
I am studying Dirichlet-process mixture formulations in which the joint behavior of the outcome, surrogate, treatment, and covariates can adapt to heterogeneous or multimodal data. A major computational question is how to carry the posterior forward as new batches arrive without retaining all previously observed subject-level data. This creates an intersection between Bayesian nonparametrics and sequential computation: the model should remain expressive, but its posterior must also admit a practical finite representation for online updating.
Current directions include sequential posterior approximations for mixture models, extensions to longitudinal settings, and using flexible posterior models to study robustness and heterogeneity in causal surrogate evaluation.
