And here are my posters:

Poster 1 - Falsification of Unconfoundedness by Testing Independence of Causal Mechanisms.
Thursday 11:00, E-2212

Poster 2 - Robust estimation of heterogeneous treatment effects in randomized trials leveraging external data.
Friday, Scaling Up Interventions Model workshop.

Thanks Aleksander!! And appreciate the suggestion, didn't know about this hashtag but will definitely use it from now on :)

If you're interested in this work, please feel free to reach out. I'd be happy to chat!

Also, this new paper extends on the work that we presented at NeurIPS 2023, so please also read it if you're interested in this topic in general. arxiv.org/abs/2205.13935

Our main idea is that, in the absence of unmeasured confounders, testable implications may exist in this type of data. We provide theoretical guarantees for when such implications arise and demonstrate the effectiveness of testing them empirically using both simulated data and semi-synthetic data.

Our paper addresses this challenge by proposing a method to falsify the assumption of no unmeasured confounding. Specifically, we introduce a new strategy that leverages data from multiple sources, such as different hospitals or regions, and can be implemented via a simple two-stage algorithm.

In many real-world settings, we estimate the effect of interventions using observational data. These analyses typically assume that all relevant confounders been measured. But if this assumption is violated, the resulting conclusions from such data can be seriously misleading.

Curious to learn more? Check out the full paper here:

arxiv.org/abs/2502.06231

And feel free to reach out with any questions!

To implement this, we design an efficient two-stage algorithm that maintains valid Type 1 error rates. Compared to our previous method (HGIC) and another strong baseline, our new method achieves higher power—meaning it’s more effective at falsification.

By testing this null hypothesis, we get a direct way to potentially falsify unconfoundedness. We also show theoretically and empirically that the null hypothesis is violated when unmeasured confounding exists.

Mathematically, we prove that under no unmeasured confounding and independent causal mechanisms, a specific testable null hypothesis holds. This null hypothesis translates to an independence condition between the parameters of the outcome and treatment models (e.g. propensity score).

Our key insight: If unmeasured confounding is present under environmental distribution shifts, the parameters of the causal mechanisms we try to estimate will appear dependent. So, if we assume these mechanisms should be independent—but observe otherwise—confounding is a likely explanation.

Observational data often comes from different sources—e.g., hospitals, schools, or time periods—which we refer to as environments. We show how to leverage such data to test for unmeasured confounding, as distribution shifts between environments can expose hidden information about confounders.

I often refer people to pcalg if they want to do causal discovery with R

Will do!

Happy to hear you like it and thanks for sharing it with others! There is a follow-up paper in the pipeline further exploring the possibilities of detecting hidden confounding in the multi-environment setting. Hope to be able to share it soon 👀

At the same time, this looks like a place where everyone is welcome. Differences-in-differences are left outside the door.