Biography

I am a PhD student in the Methodology and Statistics department at Utrecht University, supervised by Sara van Erp, Gerko Vink, Stef van Buuren and Ellen Hamaker. I am part of the missing data (MICE) and Bayesian statistics research groups.

My research is about uncertainty quantification for prediction models on tabular data: how much can we trust a prediction, and how much can we trust our estimate of how well a model performs? I approach this from four angles:

• (Approximate) Bayesian inference. When MCMC is too computationally demanding, can we take some shortcuts? I compare MCMC with faster alternatives, such as variational inference, for Bayesian penalized regression.
• Prior-data fitted networks (PFNs). PFNs such as TabPFN are trained on synthetic datasets drawn from a prior (with exceptions such as TabDPT and RealTabPFN), and approximate the posterior predictive distribution in a single forward pass. I study what this prior contributes when data is scarce.
• Uncertainty induced by missing data. Missing values make predictions less certain, and this should be reflected in the uncertainty of a predicted value. We show that this is not yet perfect in PFNs, and that conformal prediction can be used to adjust the estimates. The method is implemented in TabPFN Extensions.
• Uncertainty in model evaluation. A performance estimate is uncertain too, especially when a model moves to a new setting. We showed that the AUC of clinical prediction models varies substantially between settings, and that we need prior information to obtain good estimates of the uncertainty around this performance metric in a new setting.

Previously, I worked at the Leiden University Medical Center under the supervision of Ewout Steyerberg and Erik van Zwet, where I also contributed to projects on traumatic brain injury. My full CV can be found here.

Feel free to reach out if you are interested in some of my work or want to collaborate.