Abstract
Instrumental variables (IVs) are a popular and powerful tool for estimating causal effects in the presence of unobserved confounding. However, classical methods rely on strong assumptions such as the exclusion criterion, which states that instrumental effects must be entirely mediated by treatments. In the so-called "leaky" IV setting, candidate instruments are allowed to have some direct influence on outcomes, rendering the average treatment effect (ATE) unidentifiable. But with limits on the amount of information leakage, we may still recover sharp bounds on the ATE, providing partial identification. This package implements methods for ATE bounding in the leaky IV setting with linear structural equations. For details, see Watson et al. (2024) <a href="https://doi.org/10.48550%2FarXiv.2404.04446" target="_top">doi:10.48550/arXiv.2404.04446</a>.
Introduction
In causal inference, researchers often seek to estimate the effect of a treatment ($X$) on an outcome ($Y$) when unobserved confounders influence both. The Instrumental Variable (IV) approach is the standard solution, relying on an instrument ($Z$) that affects the outcome only through the treatment—a requirement known as the exclusion criterion.
In practice, this assumption is frequently violated. An instrument might have a "leaky" direct influence on the outcome, making the exact treatment effect unidentifiable. Traditional approaches often force a choice between biased point estimates or abandoning the IV method entirely. The leakyIV package offers a middle ground through partial identification, calculating the tightest possible range (bounds) for the true average treatment effect (ATE) given a bounded amount of leakage.
Methodology
The leakyIV package operates within the framework of linear structural equations. It addresses two main tasks: bounding the ATE and testing the validity of the exclusion assumption.
Bounding the ATE
The core function, leakyIV, calculates the lower and upper bounds of the ATE by relaxing the exclusion assumption. It allows the direct effect of instruments on the outcome (represented by the vector $\gamma$) to be non-zero, provided it remains within a user-defined threshold ($\tau$). This threshold can be specified using different $p$-norms, allowing researchers to limit the total magnitude of leakage across all instruments.
Testing Exclusion
The exclusion_test function performs a Monte Carlo test against the null hypothesis that the minimum leakage is zero. This test utilizes "tetrad constraints"—geometric properties of the covariance matrix. If the exclusion criterion holds, there must be a specific linear dependence between the covariance of $(Z, X)$ and $(Z, Y)$. The test statistic measures the deviation from this geometric requirement to determine if the data is consistent with a non-leaky instrument.
Software Implementation
The package is designed to be robust and flexible, supporting various estimation methods for the covariance matrix, which is crucial for high-dimensional or noisy data.
As shown in Figure 4, the leakyIV function accepts parameters for data input, leakage thresholds (tau), and the norm power (p). It also supports bootstrapping (classical or Bayesian) for uncertainty quantification.
Figure 3 details the parameters for the exclusion_test, including options for the covariance estimator (method) such as Maximum Likelihood (MLE), Empirical Bayes Shrinkage, or Graphical Lasso.
Key Insights
- Partial Identification: The package shifts the focus from potentially biased point estimates to scientifically honest intervals. By acknowledging leakage, the results become more robust to model misspecification.
- Geometric Constraints: The use of tetrad constraints provides a rigorous mathematical basis for testing the exclusion assumption, measuring how far the data deviates from the geometric requirements of a valid instrument.
- Robustness: By incorporating Higham’s (2002) algorithm for nearest positive definite matrices, the package ensures stability even when encountering singular covariance matrices in high-dimensional or small-sample settings.
Conclusion
The leakyIV package provides a robust framework for causal inference when the strict exclusion criterion is violated. By allowing for bounded direct effects, it enables researchers to perform more nuanced and honest analyses in the presence of "leaky" instruments, moving from potentially biased point identification to sharp partial identification.
Summary updated