Safe Causal Inference consortium

Making observational causal inference safer to use, by bridging the gap between conclusions based on expert knowledge and data-driven methods.

Who we are

We are researchers interested in causal inference, and in making sure that the claims we make are substantiated with evidence, so that we can make better-informed decisions.

Safe Causal Inference is an NWO project running from 2025 to 2031. Many of us also take part in the Causal Inference for AI meetings, an informal research community in the Netherlands.

  • Delft University of Technology
  • Erasmus MC
  • Erasmus University Rotterdam
  • Leiden University Medical Center
  • Vrije Universiteit Amsterdam

What we do

We develop tools and methodologies that help other researchers and practitioners come to causal conclusions safely.

People

Supervisors

PhD candidates

  • Photo of Matej Havelka

    Matej Havelka

    TU Delft

    Project 1.1: Conditional Exchangeability

  • Another PhD Candidate

    University B

    Short description of the PhD project

Other consortium members

  • Other Member

    Institute C

Projects

Photo of Matej Havelka

Matej Havelka

Project 1.1: Detecting and Quantifying Conditional Exchangeability Violations

Many causal analyses assume that there is no hidden confounding or selection bias beyond what has been adjusted for, which cannot be tested from a single dataset without further assumptions. Researchers instead rely on sensitivity analysis, to provide a way to include uncertainty of hidden confounding within their findings. This project studies existing sensitivity analysis methods and extends them into various different settings, while trying to provide a more data-informed methods to find sufficient sensitivity parameter.

Supervisors

PhD Candidate B

Project 1.2: Detecting and Mitigating Positivity Violations

Positivity requires that every treatment option has a real chance of occurring for every kind of individual we want to draw conclusions about. This project develops methods to detect positivity violations in high-dimensional and time-varying treatment settings, such as intensive care medicine. It then identifies the subgroups and treatment policies for which the data still support trustworthy causal claims.

Supervisors

PhD Candidate C

Project 1.3: Incorporating Assumption Uncertainty into Inference

Detecting assumption violations is not enough: the remaining uncertainty about the assumptions should also be reflected in the causal conclusions. This project uses nonparametric Bayesian methods to combine statistical uncertainty, modelling uncertainty and expert knowledge into the uncertainty about causal estimates. It also develops sensitivity analyses for proximal inference, which relies on negative control variables.

Supervisors

PhD Candidate D

Project 2.1: Formalization of Triangulation

Causal triangulation compares estimates from methods that rely on different sets of causal assumptions, where agreement between them strengthens the evidence. This project formalizes what counts as successful triangulation, including how to account for random error and for methods that target slightly different causal questions. The resulting guidance also makes it possible to pre-register triangulation studies.

Supervisors

  • Supervisor B
  • Supervisor C

PhD Candidate E

Project 2.2: Integration of Triangulated Estimates

So far, triangulation has only compared estimates from different methods; this project investigates when and how their evidence can be combined. It develops ways to pool estimates from methods that rely on different causal assumptions, and to integrate methods that produce bounds rather than a single estimate, such as partial identification. It also addresses what to conclude when triangulation fails.

Supervisors

PhD Candidate F

Project 3.1: Performance Assessment for Conditional Average Treatment Effects

Conditional average treatment effects (CATEs) describe how the effect of an intervention differs between subgroups, but different estimators can give very different answers and the true effects are never observed. This project studies how reliable current methods for evaluating CATE estimators are when causal assumptions are violated, and makes them more robust. It also develops evaluation metrics that target the decisions the estimates will be used for.

Supervisors

  • Photo of Jesse Krijthe Jesse Krijthe
  • Supervisor A
  • Supervisor B
  • Supervisor C
  • Supervisor D
  • Supervisor E

PhD Candidate G

Project 3.2: Performance Assessment for Predictions under Interventions

Sometimes the question is not the size of an effect, but the expected outcome under a specific intervention, such as a patient’s prognosis without treatment. This project develops doubly robust performance measures for such predictions under interventions, including for time-to-event outcomes, that work for internal as well as external validation. It also studies how sensitive these measures are to unmeasured confounding and positivity violations.

Supervisors

PhD Candidate H

Project 3.3: Triangulation of Performance Assessment Methods for Conditional Causal Quantities

Evaluating a conditional causal model itself relies on causal assumptions about the evaluation data. This project investigates whether combining evaluations made under different assumptions, datasets and approaches, including comparisons with trial results, gives a more complete and safer view of a model’s performance. It builds on the triangulation work in Work Package 2.

Supervisors

  • Supervisor A
  • Supervisor D

Publications

2026

  1. An example paper title

    A. Author, B. Author, C. Author

    Conference on Example Research

    Code

2025

  1. Another example paper

    D. Author, A. Author

    Journal of Examples

    PDF

Causal book club

In the causal book club we read and discuss a paper or book chapter on causal inference together. Everyone is welcome: come prepared and join the discussion.
2 Sep 2026

Kick-off for the new academic year

When
Wednesday 2 September 2026, 10:00–11:00
Where
28 Informatica 6.W950 Fawcett
16 Sep 2026

Fundamentals of Causal Inference with R, chapter 3

Babette A. Brumback

When
Wednesday 16 September 2026, 10:00–11:00
Where
28 Informatica 6.W950 Fawcett
Led by
Isak Aslund (Presentation)
Prepare
Everyone reads chapter 3.
23 Sep 2026

Fundamentals of Causal Inference with R, chapter 5

Babette A. Brumback

When
Wednesday 23 September 2026, 10:00–11:00
Where
28 Informatica 6.W950 Fawcett
Led by
Matej Havelka (Presentation about causal discovery)
Prepare
Everyone reads chapter 5.
30 Sep 2026

Detecting Hidden Confounding in Observational Data Using Multiple Environments

Rickard Karlsson

When
Wednesday 30 September 2026, 10:00–11:00
Where
28 Informatica 6.W950 Fawcett
Led by
Matej Havelka (Discussion)
Prepare
Read the paper.
7 Oct 2026

Two-phase experimentation

When
Wednesday 7 October 2026, 10:00–11:00
Where
28 Informatica 6.W950 Fawcett
Led by
Francesco Di Giuseppe (Presentation)
Prepare
No required reading.