Abstract
Causal inference relies on two fundamental assumptions: ignorability and positivity. We study causal inference when the true confounder value can be expressed as a function of the observed data; we call this setting estimation with functional confounders (EFC). In this setting, ignorability is satisfied, however positivity is violated, and causal inference is impossible in general. We consider two scenarios where causal effects are estimable. First, we discuss interventions on a part of the treatment called functional interventions and a sufficient condition for effect estimation of these interventions called functional positivity. Second, we develop conditions for nonparametric effect estimation based on the gradient fields of the functional confounder and the true outcome function. To estimate effects under these conditions, we develop Level-set Orthogonal Descent Estimation (LODE). Further, we prove error bounds on LODE's effect estimates, evaluate our methods on simulated and real data, and empirically demonstrate the value of EFC.
Introduction
Traditional causal inference frameworks typically assume ignorability (all confounders are observed) and positivity (every treatment has a non-zero probability for any given confounder value). However, these assumptions break down in scenarios where the confounder is a deterministic function of the treatment variables, a setting defined as Estimation with Functional Confounders (EFC).
A prominent example of EFC is found in Genome-Wide Association Studies (GWAS). While researchers seek the effect of specific genetic variations (SNPs) on a disease, population structure (ancestry) acts as a confounder. In practice, ancestry is often computed directly from the SNPs themselves (e.g., via PCA). Because the confounder $h(t)$ is a deterministic function of the treatment $t$, positivity is fundamentally violated: one cannot observe different confounder values for the same treatment set, making standard non-parametric estimation impossible.
Methodology
To address the positivity violation, this work introduces theoretical frameworks and a novel algorithm:
Functional Interventions and F-Positivity
Rather than intervening on the full treatment vector, we can estimate effects for functions of the treatment (e.g., cumulative dosage). This is feasible if there remains sufficient randomness in the data, a condition termed Functional Positivity (F-Positivity).
Causal Redundancy (C-REDUNDANCY)
For full interventions, the authors propose the C-REDUNDANCY assumption. This posits that the underlying data-generating mechanism depends on the explicit value of the confounder rather than its re-computation from the treatment variables.
Level-set Orthogonal Descent Estimation (LODE)
LODE is the core algorithmic contribution designed to find surrogate interventions. When a desired treatment-confounder pair is unobserved, LODE employs Euler integration to solve a gradient flow equation. It traverses the "level-set" of the conditional effect, adjusting treatment variables along a path that maintains a constant outcome until the computed confounder matches the target value.
Evaluation and Results
The proposed methods were evaluated using both synthetic and real-world datasets.
Simulated Experiments
Using synthetic datasets with 20 pre-outcome variables, the authors tested LODE against linear and non-linear functional confounders. LODE successfully corrected for confounding across varying strengths ($\gamma$), maintaining low Root Mean Squared Error (RMSE). In contrast, standard baseline regressions exhibited significant bias as confounding strength increased.
GWAS Application: WTCCC Dataset
The method was applied to the Wellcome Trust Case Control Consortium (WTCCC) dataset, comprising ~12,000 individuals and ~337,000 SNPs related to Celiac disease. LODE identified 31 relevant SNPs, 13 of which are corroborated by existing medical literature as associated with Celiac or related inflammatory bowel diseases.
Notably, LODE's estimates decoupled from base Lasso regression coefficients. For instance, it assigned identical causal effects (0.17) to two SNPs (rs13151961 and rs2237236) despite their disparate Lasso coefficients (0.32 and 0.00, respectively). This demonstrates LODE's ability to actively adjust for functional confounding (population structure) that standard models overlook.
Discussion and Key Insights
The research highlights that strict positivity is not always a prerequisite for causal estimation. If the outcome function possesses sufficient structure (C-REDUNDANCY), one can utilize the outcome's level sets to identify surrogate treatments within the observed distribution.
However, Effect Connectivity remains a critical requirement. Non-parametric estimation in EFC requires that every level set of the conditional effect contains an intervention for every possible confounder value. Without this connectivity, the true causal effect remains unidentifiable as multiple models could explain the same observational data.
Limitations and Future Directions
While LODE provides a robust framework, it currently relies on simple Euler integration. Future improvements could involve more sophisticated ODE solvers (e.g., Runge-Kutta) or adaptive optimization techniques to reduce numerical integration error. Additionally, extending the framework to handle "partial confounding"—where some confounders are not functional—remains an open area for research.
Summary updated