TL;DR
This paper introduces a novel "instrumental inequality" that provides a testable condition for determining if a variable can be considered a valid "instrument" in causal models with unobserved confounders. This inequality offers a way to empirically falsify causal models that were previously untestable, moving beyond purely subjective assumptions.
What Problem Does This Paper Address?
In causal inference, we often want to understand the effect of a variable X on an outcome Y. However, unobserved factors (latent variables) can confound this relationship, making it difficult to isolate the true causal effect. Instrumental variables (IVs) are a technique to address this, where a third variable Z influences X but not Y directly (except through X), and is also independent of the unobserved confounders. The challenge has been that models with latent variables and IVs are often untestable; even if the data violates expected conditional independencies, it's hard to pinpoint whether the model structure or the assumptions about the latent variables are at fault. This paper tackles the problem of empirically testing the validity of an instrument Z in such models, particularly when latent variables are present.
Methods
The core contribution is the derivation of an "instrumental inequality." This inequality is a mathematical condition that must be satisfied by the observed joint probability distribution of X, Y, and Z if Z is a valid instrument for the causal relationship between X and Y, even in the presence of latent variables. The proof involves decomposing the joint probability distribution and deriving bounds based on the structure of the causal model. The paper also explores the implications of this inequality for both discrete and continuous variables, highlighting differences in testability. For discrete variables, the inequality provides a direct test. For continuous variables, the testability is more nuanced, and the paper suggests that in many continuous cases, the instrumental inequality might not offer a constraint, making such models inherently harder to falsify.
Datasets
Not discussed in the paper. The paper focuses on theoretical derivations and conditions for testability rather than empirical evaluation on specific datasets.
Key Insights and Ideas
- Empirical Testability of Instruments: The most significant insight is that the concept of an "instrumental variable" is not just a theoretical construct but can be empirically tested. This moves beyond relying solely on assumptions about the unobserved confounders.
- Instrumental Inequality: The derived inequality provides a concrete mathematical condition to check. If this inequality is violated, the model (or the assumption that Z is a valid instrument) is falsified.
- Discrete vs. Continuous Variables: The paper highlights a crucial difference in testability between discrete and continuous variables. For discrete variables, the instrumental inequality offers a powerful falsification tool. For continuous variables, the same inequality may not impose constraints, suggesting that models with continuous variables and latent confounders are generally harder to test.
- Connection to Bell's Inequality: The paper draws an interesting parallel between the instrumental inequality and Bell's inequality in quantum physics, noting that both delineate classes of observed correlations that cannot be explained by latent common causes.
Key Results
The paper derives the instrumental inequality: $$ \max_z [\max_y P(x, y|z)] \le 1 $$ This inequality serves as a necessary condition for a variable Z to be an instrument relative to (X, Y) in models with latent variables. For binary variables, this inequality reduces to four specific conditions that can be directly checked against data. The paper also suggests that for continuous variables, this inequality might not provide a testable constraint in many cases, meaning that a variable Z could appear to be a valid instrument even if it's not, due to the flexibility of continuous distributions.
Limitations and Future Work
- Sufficiency: The paper notes that the derived instrumental inequality is a necessary condition, but not always sufficient, except in specific cases (e.g., when X is binary). This means that satisfying the inequality doesn't guarantee Z is a valid instrument, but violating it definitively proves it's not.
- Continuous Variables: The testability of instrumental variables in continuous settings remains a significant challenge. The paper conjectures that for continuous variables, any joint density might be generatable by an instrumental process, implying a lack of testability in many such scenarios. Future work is needed to fully prove this conjecture and explore conditions under which continuous instrumental variables might still be testable.
- Monotonicity Assumptions: The paper briefly touches upon tightening the inequalities with monotonicity assumptions (e.g., no contrarians). Further exploration of how such assumptions impact testability and the derivation of stronger inequalities would be valuable.
- General Proofs: The paper mentions that a general proof for Conjecture 1 (regarding continuous variables) is still underway, indicating this as a key area for future research.
Summary updated