Abstract
Instrumental variable methods have been widely used to identify causal effects in the presence of unmeasured confounding. A key identification condition known as the exclusion restriction states that the instrument cannot have a direct effect on the outcome which is not mediated by the exposure in view. In the health and social sciences, such an assumption is often not credible. To address this concern, we consider identification conditions of the population average treatment effect with an invalid instrumental variable which does not satisfy the exclusion restriction, and derive the efficient influence function targeting the identifying functional under a nonparametric observed data model. We propose a novel multiply robust locally efficient estimator of the average treatment effect that is consistent in the union of multiple parametric nuisance models, as well as a multiply debiased machine learning estimator for which the nuisance parameters are estimated using generic machine learning methods, that effectively exploit various forms of linear or nonlinear structured sparsity in the nuisance parameter space. When one cannot be confident that any of these machine learners is consistent at sufficiently fast rates to ensure $\surd{n}$-consistency for the average treatment effect, we introduce a new criteria for selective machine learning which leverages the multiple robustness property in order to ensure small bias. The proposed methods are illustrated through extensive simulations and a data analysis evaluating the causal effect of 401(k) participation on savings.
Introduction: The Challenge of Invalid Instruments
In causal inference, researchers often rely on Instrumental Variables (IV) to estimate effects in the presence of unmeasured confounding. A valid instrument must satisfy the exclusion restriction, meaning it affects the outcome only through the treatment. However, in real-world applications like the social sciences, this assumption is frequently violated. An instrument may have a direct effect on the outcome, rendering it "invalid" and leading to biased results in standard IV analyses.
As shown in the figure above, the direct arrow from $Z$ to $Y$ represents a violation of the exclusion restriction. For instance, eligibility for a 401(k) plan (the instrument) might change a person's financial literacy or general savings behavior (the outcome) even if they do not ultimately participate in the plan (the treatment). This paper addresses how to identify the Average Treatment Effect (ATE) in such scenarios while leveraging modern Machine Learning (ML) techniques.
Methodology
The proposed framework introduces a three-pronged solution to the problem of invalid instruments:
- Nonparametric Identification: The authors prove that the ATE remains identifiable without the exclusion restriction, provided there is sufficient heteroscedasticity—specifically, the variance of the treatment must vary with the instrument. This variance provides the mathematical leverage needed to separate the instrument's direct effect from the treatment's effect.
- Multiple Robustness: By deriving the Efficient Influence Function (EIF), the authors propose an estimator that is consistent if any one of three distinct sets of underlying "nuisance" models (modeling the instrument, treatment, or outcome residuals) is correctly specified. This provides a significant safety net compared to traditional doubly robust methods.
- Selective Machine Learning (SML): Standard Debiased Machine Learning (DML) requires high accuracy from ML models to ensure unbiased treatment effect estimates. The authors introduce a cross-validated quadratic pseudo-risk metric. SML uses this metric to select the combination of learners that minimizes overall bias, rather than simply picking the model with the lowest prediction error.
Empirical Evaluation
Synthetic Simulations
The framework was tested using simulated data with five baseline covariates. The multiply robust estimator ($\hat{\gamma}_{mr}$) significantly outperformed standard G-estimators and plug-in estimators, maintaining high coverage (near 95%) and low bias even when parts of the model were intentionally misspecified.
Real-World Application: 401(k) Participation
The authors re-analyzed the 1991 Survey of Income and Program Participation (SIPP) dataset ($n = 9,915$) to evaluate the effect of 401(k) participation on net financial assets. Using 401(k) eligibility as an instrument, the SML approach found that participation increases net financial assets by approximately $13,000 to $14,000. Crucially, the method detected that eligibility had a direct effect on non-401(k) savings (asset substitution), a nuance that standard IV methods would have missed or been biased by.
Discussion and Future Work
The "aha moment" of this research is that heteroscedasticity can serve as a lifeline when the exclusion restriction fails. Furthermore, the SML approach shifts the focus of ML selection from prediction accuracy to bias reduction, which is critical for causal inference.
While the estimators are locally efficient, future work could improve finite-sample performance by incorporating prior knowledge about the degree of exclusion restriction violation. Additionally, extending this framework to handle "weak" invalid instruments and refining post-selection inference for valid p-values remain important areas for development.
Summary updated