High-dimensional nonparametric Bayesian causal inference

Informations projet

BayCause

N° de convention de subvention: 101074802

DOI

10.3030/101074802

Date de signature de la CE 13 Decembre 2022

Date de début 1 Septembre 2023

Date de fin 31 Août 2028

Financé au titre de

• European Research Council (ERC)

Coût total

€ 1 596 136,25

Contribution de l’UE

€ 1 499 770,00

Coordonné par

STICHTING VU
Netherlands

Periodic Reporting for period 1 - BayCause (High-dimensional nonparametric Bayesian causal inference)

Période du rapport: 2023-09-01 au 2026-02-28

Résumé du contexte et des objectifs généraux du projet

Causal questions are a key part of life. They take many shapes, impacting decisions ranging from perhaps more minor matters like "What is the effect of drinking orange juice on my cold?" to questions with major impact like "Will I live longer if I undergo chemotherapy?".

Causal inference methods can provide answers to questions for which experiments are impossible due to ethical, financial or practical constraints. For example, adopting a vegetarian diet is a choice that cannot be ethically assigned randomly, and research into its effects necessarily relies on causal inference. Causality is complex, and potential causal links can often be explained away by other factors. Continuing the example, perhaps vegetarians live longer because they tend to smoke less or exercise more and not because of the diet.

An obvious solution would be to measure every single thing that could possibly explain away a causal effect, but there's a catch. Contrary to many people's intuition, when aiming to establish the cause of something, it is usually best not to include every factor. However, when you have already measured many things, it's not so easy to decide what should be included. That's the focus of this project. We develop new mathematical tools and theory to select the right information out of a huge amount of data, so that reliable causal relationships can be established. This is done for a variety of settings relevant to real-life.

The expected impact is that more causal questions can be answered using already existing data, so that people can make better informed choices for themselves.

Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent

We have developed theory for the use of possibly invalid adjustment sets in causal inference. This unconventional approach is very promising, and our methods greatly reduce the search space for suitable sets. This is especially relevant to the high-dimensional setting, because it shows that even if an invalid adjustment set is selected, this may still lead to excellent results in terms of mean squared error.

We have developed theory and a method for finding valid adjustment sets when the causal DAG is only partially known. This is relevant for the high-dimensional setting, where it's not realistic to assume that a DAG is fully known.

We have developed a method for causal inference with high-dimensional survival data, based on Bayesian additive regression trees. A software package to carry out the method is freely available as well.

Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)

A new avenue towards high-dimensional causal inference has been opened with the focus on invalid adjustment sets. A way to have credible causal conclusions despite not fully knowing the DAG has been created. And a new approach towards survival data in high-dimensional causal inference is now available. Together, the impact of these results is that more causal questions can be reliably answered.