I aim to increase the number of data sets from which we can draw trustworthy causal conclusions. To achieve this goal, I develop new causal inference methods on a strong mathematical foundation.
To facilitate the deployment of the new methods in practice, I maintain several open source R software packages, see Software.
Regression discontinuity designs
Extensions of the regression discontinuity design, which is one of the most credible ways to draw causal conclusions. With application to arthroplasty data.
Sensitivity analysis for missing outcomes
Bernstein-von Mises results for a Bayesian sensitivity analysis method for data where some outcomes are missing, possibly not at random.
More info coming soon
Posterior contraction theorems, results on uncertainty quantification, variable selection and more for shrinkage priors, in particular for the popular horseshoe prior.
Bayesian survival analysis
Bernstein-von Mises results in the supremum norm for various survival objects, justifying the use of credible bands to quantify the uncertainty in the survival function.
Cumulative sum charts
Developing a new monitoring tool for survival outcomes, with applications to arthroplasty data. In addition, development of cumulative sum charts for liver procurement data.
Bayesian community detection
We introduce a Bayesian estimator of the underlying class structure in the stochastic block model, when the number of classes is known, and show strong consistency.
Combining data from a randomised clinical trial and an observational study as part of the orthopedic APOLLO trial.
We derive asymptotic properties of caliper matching, a form of propensity score matching.
More info coming soon
Bayesian Additive Regression Trees
Posterior contraction results for Bayesian regression trees and forests.
Competing risks in orthopedic data
Guidelines for working with arthroplasty data, taking into account the competing risks structure and the dependence between various joints.
Merged block randomisation
A novel restricted randomisation method designed for small clinical trials (at most 100 subjects) or trials with small strata, for example in multicenter trials. It can be used for more than two groups and unequal randomisation ratios.
Detecting syntactic differences
Using the minimum description length principle, as well as other approaches, we developed tools to automatically detect syntactic differences between languages based on parallel corpora.