A Novel Approach to Instrumental Variable Estimation: TEAM-IV

Technical Summary: TEAM-IV

Problem Statement
Instrumental variable (IV) analysis is a standard approach for estimating causal effects in the presence of unmeasured confounding. However, the validity of IV estimates relies on the exclusion restriction, which posits that instruments affect the outcome only through the exposure. In practice, particularly in Mendelian randomization, instruments often violate this restriction via direct effects (horizontal pleiotropy). Existing robust IV methods, such as sisVIVE and CIIV, rely on "majority-validity" or "plurality-validity" conditions, requiring that valid instruments constitute more than half (or the largest group) of the candidate set. These methods can fail or produce biased estimates when invalid instruments are prevalent, when direct effects are large, or when the majority condition is violated.

Methodology: TEAM-IV
The authors propose TEAM-IV (Team-based Instrumental Variable estimation), a novel approach designed to identify and aggregate sets of instruments that appear valid together ("teams"), allowing for reliable estimation even when only a small number of candidate instruments are valid (potentially as few as two). The method proceeds through four key stages:

1. Ridge-Based Ordering: To mitigate heterogeneity in instrument strengths which can distort direct-effect estimates, the authors first construct an adjusted exposure vector ($d_{adj}$) such that its projection onto the instrument space corresponds to homogeneous instrument strengths. They then fit a ridge-penalized model to estimate direct effects ($\hat{\alpha}$) and order the instruments based on these estimates. Theoretical results suggest that under symmetric population regimes, this ordering places valid instruments into a contiguous block.
2. Team Construction via Windowed MCP: The ordered instruments are analyzed using overlapping contiguous windows. Within each window, a Minimax Concave Penalty (MCP) regression is fitted to estimate direct effects. A "teamness" matrix is constructed by aggregating dyadic evidence across windows; instruments that frequently have estimated direct effects shrunk to zero together (indicating concordant validity status) are grouped into candidate "teams."
3. Cross-Validation Evaluation: Each candidate team is evaluated as a potential set of valid instruments. The method employs a sisVIVE-style cross-validation criterion: for a candidate valid set $G$, the remaining instruments ($G^c$) are treated as covariates in a regression of the outcome on the projected exposure. The out-of-sample prediction risk is estimated via K-fold cross-validation. Teams containing invalid instruments (omitted from the covariate set) are expected to have higher prediction risk.
4. Union and Estimation: The final set of valid instruments is defined as the union of "near-best" teams (those with cross-validation risk within one standard error of the minimum). The causal effect is then estimated using two-stage least squares (2SLS), where instruments in the selected union serve as excluded instruments, and instruments classified as invalid are included as exogenous covariates to adjust for their direct effects.

Key Contributions

• New Selection Procedure: The paper introduces a team-based selection-and-aggregation procedure that constructs multiple candidate valid instrument sets rather than relying on a single penalized fit.
• Theoretical Results: The authors establish two theoretical properties:
  1. Under a simplified symmetric population regime with compound-symmetric covariance, ridge-based ordering asymptotically places valid instruments in a contiguous block.
  2. The window-specific MCP estimator asymptotically assigns a common sign-with-tolerance classification to valid instruments within a window, even when their estimated direct effects are non-zero.
• Robustness to Invalidity: The method is designed to function when invalid instruments constitute a majority of the candidate set, a scenario where sisVIVE and CIIV often struggle.

Simulation and Empirical Results

• Simulation Study: Across simulations varying the number, magnitude, and signs of invalid direct effects, instrument correlation, and exposure endogeneity, TEAM-IV generally attained lower absolute estimation error than sisVIVE and CIIV.
  • The relative gains were largest when invalid instruments were prevalent (violating the majority-valid condition) and when invalid direct effects were large.
  • TEAM-IV remained robust under correlated instruments (autoregressive and compound-symmetric structures) and mixed-sign direct effects, though the performance gap with sisVIVE narrowed in mixed-sign settings.
• MESA Application: The method was applied to the Multi-Ethnic Study of Atherosclerosis (MESA) to estimate the effect of LDL cholesterol on carotid intima–media thickness (IMT). In the empirical analysis, both TEAM-IV and sisVIVE retained all eight candidate SNPs and produced identical point estimates ($\hat{\beta} = 2.19$).
• Validation Studies: Two data-anchored validation studies highlighted differences between the methods:
  1. Adversarial Stress Test: Using outcome-informed pseudo-instruments (deliberately invalid), TEAM-IV correctly identified 100% of the pseudo-instruments as invalid, whereas sisVIVE identified only 1.2%. TEAM-IV's estimates remained closer to the empirical reference than sisVIVE's.
  2. Semi-Synthetic Outcome: In a setting with known ground truth where 62.5% of instruments were invalid, TEAM-IV achieved perfect instrument classification (100% sensitivity and specificity) and lower bias, MAE, and RMSE compared to sisVIVE.

Significance and Claims
The paper claims that TEAM-IV offers a robust alternative to existing penalized IV methods, particularly in challenging configurations where invalid instruments are numerous or strong. The authors emphasize that TEAM-IV is "generally" more accurate than sisVIVE and CIIV in simulations with high proportions of invalid instruments.

However, the authors maintain a modest tone regarding their claims:

• They state that the results support the robustness of TEAM-IV "under the data-generating mechanisms examined" but do not imply "uniform superiority across all instrumental-variable settings."
• They note that in the empirical MESA analysis, the agreement between methods does not prove instrument validity, especially given the weak joint first-stage F-statistic.
• They acknowledge limitations, including the use of a fixed sample size ($n=2000$) and a small number of instruments ($L=10$) in simulations, and call for future work on settings with large $L$ relative to $n$, theoretical guarantees for selection, and valid uncertainty quantification.

The paper concludes that penalized instrument-selection methods, combined with post-selection 2SLS, can provide substantial robustness to direct instrument effects, with TEAM-IV being particularly promising when invalid instruments violate the majority-valid condition required by existing methods.