They show that, unlike for the corresponding model under just marginalization (mDAGs), there are subtle differences between the ‘single-world’ intervention scheme, and weaker probing schemes (Ansanelli et al., 2024). They also provide a separation criterion that is at least as powerful as m-separation.

• it is frugal: it makes use of variation independent quantities, so these can be switched without compromising the interpretation;

• it is flexible: in our case it uses neural networks, but in principle it can make use of any method;

• it is marginal: the estimand is marginally causal, i.e. it is of the form $Y(t)$ not $Y(t) | Z=z$;

• it is faithful: the estimand can be exchanged for any other, while still retaining the structure of the original data;

• it is longitudinal: longitudinal and survival data can be estimated/simulated, as well as the static case;

• it is extrapolative: continuous treatments can be extended under an additive noise model;

• it is privacy preserving: it supports differential privacy.

We show that it is competitive with a range of other, more specific, generative methods over a suite of these settings.

The work is joint with Xinwei Shen (UW) and Robin Evans.

Daniel Manela and Linying Yang have been presenting their paper Testing Generalizability in Causal Inference at Uncertainty in Artificial Intelligence 2025. The paper provides a new method for assessing whether or not a dataset is transportable.

Today we welcome a new researcher, Jonathan Rollings. He joins us from the University of Surrey for eight weeks, and he’ll be working with Daniel Manela on the link between vine copula independence models and graphical models.

Daniel Manela and Linying Yang’s paper on a new framework for testing generalizabilty has been accepted to Uncertainty in Artificial Intelligence 2025. Congratulations, and have fun in Rio!

In particular, it can filter out instruments and spurious variables, and allows for inference that out-performs the original marginal ratio.

The work is joint with Robin Evans.

Xi Lin’s paper on a power likelihood method for combining randomized controlled rials and observational data has been accepted to Biometrics. Co-authors are Jens Magelund Tarp (Novo Nordisk) and Robin Evans.

Welcome to Jakob Zeitler, who joins us as a Postdoctoral Researcher following a PhD at UCL with Ricardo Silva. He is funded by the Danish Foundations’ SMARTbiomed Pioneer Centre, and co-supervised by Prof. Erin Gabriel.

Jakob’s research interests include partial identification, active learning and Bayesian optimization.

Dan Manela and Laura Battaglia presented a poster associated with de Vassimon Manela et al. (Marginal Causal Flows for Validation and Inference) at the Thirty-Eighth Annual Conference on Neural Information Processing Systems, in Vancouver. The paper provides a generative method using normalizing flows to simulate from causal models under the frugal parameterization.