Introduction
This paper introduces a framework for causal inference in settings where confounders are deterministic functions of the observed treatment variables—a scenario that technically violates the fundamental "positivity" assumption. The authors propose Level-set Orthogonal Descent Estimation (LODE), an algorithm that uses gradient flow to find "surrogate interventions" that allow for accurate causal effect estimation even when traditional methods fail.
The Problem: Functional Confounders and Positivity
In classical causal inference, two pillars hold up the house: ignorability (we've observed all confounders) and positivity (for any given set of confounders, there’s a non-zero chance of receiving any treatment).
However, in fields like genomics (e.g., GWAS), we encounter Estimation with Functional Confounders (EFC). Here, the confounder $z$ is a function of the pre-outcome variables $t$ (e.g., $z = h(t)$). For example, in a genetic study, the "population structure" (a confounder) is calculated directly from the same genetic markers (SNPs) we are testing as treatments.
This creates a mathematical paradox: if $z$ is a deterministic function of $t$, then for a specific $t$, we can never observe a different $z$. This violates positivity, making the regression of the outcome on both $t$ and $z$ unidentifiable. Essentially, there are infinitely many ways to split the influence between the treatment and the confounder because you cannot tell if the treatment caused the outcome or if the confounder (which is perfectly determined by the treatment) did.
Proposed Methods
The authors address this unidentifiability by identifying conditions under which causal effects are still estimable:
- Functional Positivity: If we only intervene on a part of the treatment (a "functional intervention"), we might still have enough variation to estimate effects.
- C-REDUNDANCY (Causal Redundancy): This is a novel condition where the outcome function is "redundant" in a specific way—it uses the value of the confounder from its second argument rather than re-computing it from the treatment. Mathematically, this involves the gradient of the outcome being orthogonal to the gradient of the confounder function.
- LODE (Level-set Orthogonal Descent Estimation): This is the core algorithm. Since we can't observe the effect of an arbitrary intervention $t^*$ at a specific confounder level $z$, LODE searches for a "surrogate intervention" $t'$ that:
- Produces the same causal effect as $t^*$.
-
Actually exists in the observed data for that confounder level.
LODE finds this surrogate by solving a gradient flow equation (using Euler integration) to traverse the "level-sets" of the outcome function. If the C-REDUNDANCY condition holds, we can move along a path in the treatment space that changes the confounder value without changing the causal effect. This allows us to "transport" an unobservable intervention to an observable one.
Datasets and Key Results
The paper evaluates the approach on two main fronts:
- Simulated Data: The authors used two EFC causal models with varying strengths of confounding and outcome smoothness to compare estimates against a known ground truth. LODE significantly outperformed baseline models in Root Mean Squared Error (RMSE) across various confounding strengths. While baselines saw error increase as confounding got stronger, LODE remained stable.
- Real-World GWAS Data: The Wellcome Trust Celiac disease dataset consists of ~12,000 individuals and over 337,000 SNPs. They focused on 1,050 SNPs to see if LODE could correctly identify relevant genetic markers while adjusting for population structure. In this study, LODE identified 31 relevant SNPs. Crucially, 13 of these were previously reported in the literature as being associated with the disease. Furthermore, LODE's estimates did not simply mirror Lasso coefficients; it identified significant effects for SNPs where the Lasso coefficient was zero, demonstrating its ability to adjust for confounding factors that the outcome model alone ignored.
Limitations and Future Work
- Euler Integration Errors: The accuracy of LODE depends on the step size of the Euler integrator. Large steps can lead to "confounder mismatch," though the authors provide error bounds to quantify this.
- Known Confounder Function: The current framework assumes the function $h(t)$ that defines the confounder is known or specified by a practitioner.
- Search Strategy: The current search relies on gradient descent, which can get stuck in local minima if the level-sets are complex. Future work could explore more robust optimization techniques like momentum or second-order methods.
- Ignorability Assumption: The paper assumes all confounders are captured by $h(t)$. If there are additional unobserved confounders, the estimates would still be biased.
Summary updated