Can causal_forest() be used to explore moderation of an observational relationship with unmeasured confounding?

[jonasdora]

Hello, thank you for this great package.
I am planning an exploratory observational study on the relationship between W and Y using intensive longitudinal data. Each participant contributes multiple observations over time, with both W and Y measured repeatedly (data are clustered within participants). We also have multiple situational variables (X) measured at each timepoint.
Our research goal is to understand how the W-Y relationship varies across different situations. We do not have any concrete hypotheses for individual moderators, we are at a quite exploratory point in this line of research. Specifically, we want to:

Robustly identify which situational variables (if any) most strongly moderate the W-Y relationship
Visualize how the strength of the W-Y relationship varies across these situational variables
Account for the nested data structure (observations within participants)

The causal_forest() approach is appealing because it:

Focuses on how one key relationship varies across other variables
Provides variable importance measures for moderators
Handles clustering through the clusters parameter
Offers useful visualization and analysis tools

However, my key concern is that we cannot assume unconfoundedness. There are almost certainly unmeasured variables (e.g., personality traits) that affect both W and Y.

I am wondering if causal_forest() can be validly used to study patterns of moderation in our case? We do not need to make causal claims, we just want to explore the heterogeneity of the W-Y relationship across X.
If not, are there alternative approaches you would recommend that could provide similar insights about moderation while accounting for our clustered data structure?

[ricor07]

Hello, I'd be glad to work on this. Can i get the issue assigned? Thanks

[swager]

@jonasdora this is an interesting question; unfortunately we don't have any out-of-the box workflows for this task.

@ricor07 thanks for volunteering. To me, this looks more like an open research question (i.e., a topic for a stand-alone research paper) rather than just a feature to be added. If/when a standard solution to this overall problem is adopted by the community, we could talk about whether it makes sense to add an implementation to GRF (or instead have the implementation live in a separate package, like we did with policytree).