Abstract
Instrumental variables (IVs) are widely used to estimate causal effects in the presence of unobserved confounding between exposure and outcome. An IV must affect the outcome exclusively through the exposure and be unconfounded with the outcome. We present a framework for relaxing either or both of these strong assumptions with tuneable and interpretable budget constraints. Our algorithm returns a feasible set of causal effects that can be identified exactly given relevant covariance parameters. The feasible set may be disconnected but is a finite union of convex subsets. We discuss conditions under which this set is sharp, i.e., contains all and only effects consistent with the background assumptions and the joint distribution of observable variables. Our method applies to a wide class of semiparametric models, and we demonstrate how its ability to select specific subsets of instruments confers an advantage over convex relaxations in both linear and nonlinear settings. We also adapt our algorithm to form confidence sets that are asymptotically valid under a common statistical assumption from the Mendelian randomization literature.
Introduction
In causal inference, Instrumental Variables (IVs) are often considered the "gold standard" for estimating the effect of a treatment ($X$) on an outcome ($Y$) when unobserved confounding is present. However, standard IV methods rely on strict, untestable assumptions that are frequently violated in practice—particularly in fields like genetics where biological pathways are complex and interconnected.
BudgetIV is a new framework designed to handle "invalid" instruments. While traditional robust methods often fail if more than 50% of instruments are invalid (the "majority rule"), BudgetIV provides a way to identify the tightest possible range of causal effects even when the vast majority of candidate instruments do not satisfy standard IV assumptions.
Problem Statement: The Challenge of Invalid Instruments
A valid IV must satisfy three core criteria: relevance (it must be associated with the treatment), unconfoundedness (it must not share common causes with the outcome), and the exclusion restriction (it must affect the outcome only through the treatment).
In real-world scenarios such as Mendelian Randomization, these assumptions are often violated. For instance, a genetic variant might influence the outcome through biological pathways independent of the treatment (instrumental leakage) or be correlated with unobserved confounders. Most existing "robust" methods break down when more than 50% of instruments are invalid. Furthermore, existing partial identification methods often use "convex relaxations" that result in overly conservative, wide bounds that lack practical utility.
Methodology: Budget Constraints and Sharp Identification
BudgetIV shifts the focus from point estimation to finding the set of all plausible causal values consistent with user-defined "budget" constraints on violations.
Key Innovations
- Budget Constraints: Instead of assuming a majority of valid instruments, users specify thresholds ($\tau$) for violation magnitudes and a "budget" ($b$) for the number of instruments allowed to exceed these thresholds.
- Star Domains and Non-Convexity: The framework formalizes the search space as a "star domain," allowing for solution sets that are non-convex or even disconnected. This flexibility is crucial for capturing the true complexity of the identification problem.
- Sharpness: The algorithm is proven to be "sharp," meaning it returns the smallest possible set of causal effects consistent with the data and assumptions.
- Algorithmic Implementation: The framework includes
budgetIV_scalarfor single treatment effects (polynomial-time) and a generalbudgetIVcombinatorial search for multi-dimensional treatments.
Key Results
The paper evaluates BudgetIV through simulations mimicking challenging real-world conditions, including linear models and large-scale benchmarks with 100 candidate instruments where 70% are invalid.
In linear simulations, BudgetIV produced significantly narrower bounds than traditional $L_1$ or $L_2$ norm relaxations, which often "blow up" as uncertainty increases. In high-violation scenarios (70% invalid instruments), standard Mendelian Randomization benchmarks (e.g., MR-Egger, IVW) failed to include the true causal effect in their confidence intervals. BudgetIV was the only method, aside from an "Oracle" with perfect knowledge, to correctly capture the true effect within its identified set.
Discussion and Future Work
BudgetIV demonstrates that meaningful causal information can be extracted even when most data sources are "lying," provided one accepts partial identification. The ability to produce disconnected sets is a significant feature, as separate "islands" of possible effects can represent distinct, plausible biological hypotheses.
However, challenges remain. For multi-dimensional treatments, the problem is NP-hard, limiting its use in very high-dimensional settings without further optimization. Future work may involve relaxing the "No Measurement Error" (NOME) assumption and exploring "instrument discovery"—using the framework to identify valid instruments within large datasets.
Summary updated