Frugal, Flexible, Faithful: Causal Data Simulation via Frengression

TL;DR

This paper introduces frengression, a deep generative model that accurately simulates multivariate, time-varying causal data and supports interventional sampling, enhancing benchmarking and evaluation in causal inference.

Contribution

It presents frengression, a novel generative approach for modeling causal data with guarantees on consistency and extrapolation, enabling faithful simulation and intervention analysis.

Findings

Accurately models joint distribution of covariates, treatments, outcomes

Enables direct sampling from interventional distributions

Validated on real-world clinical trial data

Abstract

Machine learning has revitalized causal inference by combining flexible models and principled estimators, yet robust benchmarking and evaluation remain challenging with real-world data. In this work, we introduce frengression, a deep generative realization of the frugal parameterization that models the joint distribution of covariates, treatments and outcomes around the causal margin of interest. Frengression provides accurate estimation and flexible, faithful simulation of multivariate, time-varying data; it also enables direct sampling from user-specified interventional distributions. Model consistency and extrapolation guarantees are established, with validation on real-world clinical trial data demonstrating frengression's practical utility. We envision this framework sparking new research into generative approaches for causal margin modelling.