Robust Weighted Triangulation of Causal Effects Under Model Uncertainty

1. Problem Statement

Causal inference with observational data often faces model uncertainty, where multiple plausible causal models exist, each relying on different (and potentially untestable) identifying assumptions.

• The Challenge: Analysts typically face a dilemma: rely on a single model (risking misspecification bias) or combine estimates from multiple models (triangulation).
• Limitations of Existing Methods:
  • Voting/Selection: Traditional triangulation often involves selecting the "best" model or requiring a plurality of models to be correct. This introduces post-selection inference problems and bias.
  • Evidence Factors: Existing methods (e.g., Rosenbaum) focus on hypothesis testing (combining p-values) rather than effect estimation and rely on strong assumptions about the independence of bias sources.
  • Bayesian Averaging: Often relies on parametric assumptions (e.g., linearity) or specific model classes.
• Goal: Develop a framework to aggregate causal effect estimates from multiple candidate models into a single robust estimator without explicit model selection, while providing valid statistical inference.

2. Methodology

The authors propose a Smooth Weighted Triangulation Functional that combines causal discovery (testable implications) with semiparametric theory.

A. Core Framework

Let $\theta$ be the true causal effect and $\{M_1, \dots, M_K\}$ be $K$ candidate causal models. Each model $M_k$ has an identifying functional $\psi_k$ (e.g., backdoor formula, frontdoor formula, IV estimator).

• Testable Implications ($\beta_k$): For each model, the authors define an observed data parameter $\beta_k$ such that $\beta_k = 0$ if and only if the identifying assumptions of $M_k$ hold (e.g., conditional independence, Verma constraints, or generalized equality constraints).
• The Naive Functional: A naive approach would average only those $\psi_k$ where $\beta_k = 0$. However, since $\beta_k$ is unknown and estimated from finite data, a hard threshold (indicator function) is unstable and leads to post-selection bias.

B. The Triangulation Functional

To avoid hard selection, the authors propose a smooth aggregation using Gaussian kernels:
$$\psi = \sum_{k=1}^K w_k \psi_k$$
where the weights $w_k$ are defined as:
$$w_k = \frac{\delta_a(\beta_k)}{\sum_{j=1}^K \delta_a(\beta_j)}$$
Here, $\delta_a(\beta_k) = \frac{1}{|a|\sqrt{\pi}} e^{-(\beta_k/a)^2}$ is a Gaussian kernel centered at 0 with bandwidth $a$.

• Mechanism: If a model is correct ($\beta_k \approx 0$), its weight is high. If a model is incorrect ($\beta_k$ is large), its weight decays exponentially.
• Parameter $a$: Controls the sharpness of the approximation. As $a \to 0$, the method approaches the naive selector; in practice, a small fixed $a$ (e.g., 0.1) balances robustness and numerical stability.

C. Theoretical Guarantees (Theorem 1)

The paper establishes a bound on the bias of the triangulation functional:
$$|\psi - \theta| \leq \frac{\max_{k \in \text{Incorrect}} |\psi_k - \theta|}{1 + D_a}$$
where $D_a$ is a discrimination factor dependent on the ratio of weights assigned to correct models versus incorrect models.

• Robustness: If at least one model is correct and testable (i.e., its $\beta_k=0$ while incorrect models have $\beta_k \neq 0$), the discrimination factor $D_a$ becomes large as sample size increases (or $a \to 0$), driving the bias of $\psi$ toward zero.
• Key Insight: The estimator is robust to misspecification in some models as long as at least one model is correct and its violations are detectable in the data.

D. Inference Procedure

The authors derive valid statistical inference for $\psi$ using asymptotic linearity:

1. Estimators: Assume estimators $\hat{\psi}_k$ and $\hat{\beta}_k$ are asymptotically linear with influence functions $\phi_{\psi_k}$ and $\phi_{\beta_k}$.
2. Delta Method: Since $\psi$ is a smooth function of these estimators, the Delta method yields the asymptotic distribution:
   $$\sqrt{n}(\hat{\psi}_n - \psi) \xrightarrow{d} N(0, \gamma^T \Sigma \gamma)$$
   where $\gamma$ is the gradient vector of partial derivatives (derived in Lemma 1) and $\Sigma$ is the covariance matrix of the influence functions.
3. Implementation Strategies:
   • Preferred: Influence function-based estimation (using one-step or estimating equations) allows for machine learning nuisance estimators.
   • Alternative: If influence functions are unavailable, use parametric bootstrap (for parametric nuisance models) or subsampling (for machine learning-based plug-in estimators to avoid bias issues).

3. Key Contributions

1. Principled Triangulation: Provides the first quantitative framework for triangulating causal effects that avoids explicit model selection and post-selection inference bias.
2. Robustness to Misspecification: Demonstrates that the estimator remains consistent for a weighted functional that converges to the true causal effect if at least one candidate model is correct and testable.
3. Integration of Causal Discovery: Unifies causal discovery (testing structural assumptions via $\beta_k$) with semiparametric estimation (estimating $\psi_k$).
4. Valid Inference: Derives the asymptotic distribution and provides practical variance estimation strategies (Delta method, bootstrap, subsampling) for the triangulated estimator.

4. Results

The authors validate the method through simulations and an empirical application:

• Scenario 1 (M-Bias):

  • Setup: Multiple adjustment sets where some induce M-bias (collider bias) and others are valid.
  • Result: The triangulated estimator remained robust even when the majority of candidate models were incorrect. It successfully down-weighted models with M-bias based on the testable implications ($\beta_k$).
  • Coverage: Achieved nominal 95% coverage for both the functional $\psi$ and the true parameter $\theta$ at $n=5000$.

• Scenario 2 (Backdoor vs. Frontdoor vs. IV):

  • Setup: Triangulating between Backdoor, Frontdoor, and Instrumental Variable (IV) models, which rely on qualitatively distinct assumptions.
  • Result: The estimator maintained robustness when only one model (e.g., Frontdoor) was correct. It also showed lower variance than relying solely on the IV model when multiple models were correct.
  • Coverage: Achieved coverage close to nominal (93-96%) across different truth scenarios.

• Empirical Application (Framingham Heart Study):

  • Task: Estimate the effect of blood glucose on coronary heart disease.
  • Models: Backdoor (adjusting for sex), Frontdoor (via hypertension), and IV (using education).
  • Outcome: The IV model received near-zero weight due to diagnostic failures. The Frontdoor model received the highest weight. The triangulated estimate (4.7% increase) provided a robust conclusion that balanced the Backdoor and Frontdoor estimates, effectively filtering out the unreliable IV estimate.

5. Significance

This work addresses a critical gap in causal inference: how to combine evidence from conflicting models without arbitrarily picking a winner.

• Practical Impact: It offers a "safe" default for analysts dealing with model uncertainty, allowing them to leverage all available structural assumptions while penalizing those that fail data-driven tests.
• Theoretical Advance: It moves beyond the "all-or-nothing" nature of model selection, providing a continuous, differentiable path to robustness that is amenable to standard statistical inference.
• Flexibility: The framework is general and can accommodate various causal structures (backdoor, frontdoor, IV, proxy variables) and estimation techniques (parametric, semi-parametric, machine learning).

In summary, the paper formalizes causal pluralism by creating a weighted average of causal estimates where the weights are dynamically determined by the empirical validity of each model's assumptions, ensuring robustness against misspecification while maintaining statistical rigor.