When a hidden variable confounds a treatment and an outcome, one fix is to find "proxies": measured variables that carry information about the hidden confounder in a specific, checkable way. Recent work automates this search with covariance rank tests. These tests check whether a block of the observed covariance matrix is rank-deficient, which is the algebraic trace a hidden bottleneck leaves behind. In realistic data, though, the search usually returns several proxy sets that all pass the test. The estimators built on them differ sharply in accuracy, because each one inverts a different cross-covariance matrix, and some of those matrices are nearly singular. Wireless engineers met the same issue decades ago. When a channel estimate is later inverted, an error along a weak direction is amplified catastrophically, while an equal-sized error along a strong direction does no harm. They responded by designing estimation for the downstream task rather than for generic accuracy.
We propose to carry that lesson into proximal causal inference. Among the proxy sets certified by rank tests, we select the one, or the weighting of several, that minimizes a predicted finite-sample error of the causal effect itself. The prediction is a directional sensitivity criterion: it measures how much of the inverse cross-covariance acts on the effect coordinate. This differs from the matrix's overall condition number, which is the obvious heuristic. The proposal also tests a sharp prediction imported from estimation theory: shrinking (ridge-regularizing) the proxy covariance before inverting it should hurt rather than help.
Core Claim
Automating the Selection of Proxy Variables of Unmeasured Confounders introduced Proxy-Rank. It searches for negative-control exposures $Z$ (proxies that affect the outcome only through the confounder) and negative-control outcomes $W$ (proxies unaffected by the treatment) whose cross-covariance submatrices have rank deficiency equal to the number $q$ of hidden confounders. It then estimates effects with a matrix-determinant generalization of the proximal estimator. Its target is consistency, so any valid set is as good as any other.
Our claim is that in finite samples the choice among valid sets dominates the error, and that the right criterion is directional. In the linear case the proximal estimate solves a moment system $\hat\theta = \hat M^{-1}\hat m$. Here $M$ is the cross-covariance between the stacked treatment and NCO variables $(A, W)$ and the NCEs $Z$, $m$ is the cross-covariance between the outcome $Y$ and $Z$, and the causal effect is the first coordinate $\theta_1$. A first-order expansion gives
$$ \operatorname{Var}(\hat\theta_1) \approx \frac{1}{n}\, e_1^\top M^{-1}\, \Omega\, M^{-\top} e_1 , $$
where $n$ is the sample size, $e_1$ selects the effect coordinate, and $\Omega$ is the covariance of the per-sample moment residual $Z\,(Y - (A,W)^\top\theta)$.
Only the first row of $M^{-1}$ matters. A proxy set can therefore have a huge condition number while giving a stable effect estimate, as long as its weak direction lives in nuisance coordinates (the confounder loadings). Conversely, a well-conditioned set can be fragile if the weak direction aligns with $e_1$. This is exactly the directional-sensitivity argument of Training sequence design for MIMO channels: an application-oriented approach, which allocates estimation precision to the subspaces that matter after zero-forcing inversion, now transplanted from a physical channel to a proxy covariance.
Position Relative to Prior Work
The motivation comes from three papers. Condition number bounds for causal inference shows that algebraically equivalent identification formulas can have wildly different numerical sensitivity, but it works with discrete do-calculus expressions and offers no selection rule. We make sensitivity a selection objective, in continuous linear models with latent confounders. MSE-optimal adjustment sets in linear Gaussian causal models with finite sample size shows that the asymptotically best adjustment set loses to a sample-size-dependent choice when $n$ is small. It reports MSE 0.2440 versus 0.2789 at $n=10$, with convergence by $n=1000$ (0.0016 versus 0.0015). That work assumes no hidden confounders; we target exactly the hidden-confounder case, where the candidates come from rank tests rather than graphical criteria.
The method draws on two sources. Proxy-Rank supplies the candidate generator. Estimator Selection: End-Performance Metric Aspects and Rojas et al. (2015) supply two things. First, they show that when an estimate is later inverted, the unbiased ML estimator zeroes the gradient of downstream excess MSE while MMSE shrinkage does not. Second, they supply zeroth-order Taylor and trimming devices for handling inverses whose moments do not exist.
We distinguish ourselves from neighboring "task-oriented causal" work. Reward-oriented Causal Representation Learning also argues for learning only what the task needs, but it uses active interventions and precision-matrix differences; we stay purely observational. Fast Proxy Experiment Design for Causal Effect Identification chooses proxies to make an effect identifiable at minimum cost, which is a yes/no target, not estimation accuracy.
Research Program
The first component is the selection rule. Proxy-Rank produces a candidate pool of rank-certified $(Z, W)$ sets. For each candidate we estimate the directional variance above with a plug-in $\hat M$ and $\hat\Omega$, and we use sample splitting so that selection does not reuse the estimation data. Near-singular candidates have heavy-tailed estimators, so we score them with a trimmed version of the criterion that follows the regularized-trimming construction of Rojas et al. We also flag any candidate whose estimated smallest singular value is not separated from sampling noise; those candidates are discarded rather than scored.
The second component extends selection to weighting. Valid proxy sets can be stacked, and the efficient combination of stacked moment conditions is a generalized method of moments (GMM) problem. We derive when picking a single set beats pooling. Pooling adds weak moments whose estimated optimal weights are themselves noisy, so we expect single-set selection to win at small $n$ and pooling to win at large $n$, mirroring the Rutsch et al. crossover.
The third component tests estimator choice. We will prove, for the single-confounder determinant estimator, the analogue of the Rojas et al. result: ridge or Ledoit–Wolf-style shrinkage of $\hat M$ introduces bias aligned with the weak direction, and that bias enters the effect at first order. We will then check this numerically for $q > 1$.
Baselines
The naive reference point is OLS of $Y$ on $A$ ignoring confounding. Xie et al. report it as heavily biased on both their synthetic data and the mouse obesity data. We include it to bound the bias scale, not as a serious competitor.
The strongest existing method is Proxy-Rank itself, with its matrix-determinant estimator applied to whichever valid set its search returns. We run it on identical data with the same rank-test significance level used to build our candidate pool, so that the only difference is the selection step. Xie et al. report no finite-sample effect MSE for it (no reported result found on Lacuna for Proxy-Rank effect-estimation MSE), so we measure it ourselves. Beating it means a lower effect MSE across the $n$ sweep, with the gap concentrated at small $n$ and vanishing asymptotically.
The heuristic competitor selects the candidate with the smallest condition number of $\hat M$. It is cheap and is the first thing a reviewer would try.
The pooling competitor is two-step GMM over all certified proxies, using the same candidate pool and no extra tuning.
The shrinkage competitor uses Proxy-Rank's set with a ridge-regularized $\hat M$. The ridge strength is tuned by cross-validated moment fit on held-out data, with the same tuning budget we give our trimming parameter.
The key ablation replaces the directional criterion with an isotropic one: the Frobenius norm of $\hat M^{-1}\hat\Omega\hat M^{-\top}$, with everything else unchanged. This isolates whether the gains come from directionality specifically.
Why not just…?
Why not just pick the proxy set with the best condition number? It is the obvious move, and it will match our rule whenever the weak direction of $M$ happens to fall on the effect coordinate. The variance formula shows that the two rules diverge when the weak direction lies in nuisance coordinates, and it predicts which way the divergence goes. Our synthetic design deliberately includes both regimes. If condition-number selection matches the directional rule even in the regime where they disagree, the directional claim is falsified, and we would report the cheaper rule.
Why not just bootstrap each candidate's estimator and keep the one with the lowest variance? Near-singular inverses have heavy tails, and moments may not exist. Estimator Selection: End-Performance Metric Aspects needed trimming or Taylor approximations precisely for this reason, so naive bootstrap variances are unstable selectors. We include the bootstrap selector as an additional baseline and compare its selection regret against ours as $n$ shrinks.
Why not just pool every valid proxy with efficient GMM? This is a real threat. At moderate $n$ it may dominate any single-set choice. We do not argue it away. The second component of the program exists to locate the crossover. If pooling wins at every $n$ we test, the contribution reduces to choosing the GMM weighting by directional sensitivity, and we would reframe the work that way.
Why not just shrink the covariance, the standard fix for noisy inverses? Rojas et al. show that shrinkage bias can be amplified by the subsequent inversion. We test this directly with the shrinkage baseline. If a well-tuned ridge beats unregularized selection, that refutes the transplanted theory, and we would report it as such.
Proposed Experiments
The synthetic study uses linear SEMs with $q \in \{1, 2, 3\}$ hidden confounders and 4 to 12 candidate proxies, built so that several disjoint proxy sets are valid. We control the signal strength of each set and the alignment of each set's weakest singular direction with the effect coordinate. We sweep $n$ from 50 to 10,000 and report effect MSE, selection regret relative to an oracle that knows the true per-set MSE, and tail quantiles of the absolute error. One decisive check is the misaligned regime: there the directional rule should beat condition-number selection, while in the aligned regime the two should tie. A second decisive check is the predicted sign of the shrinkage effect.
For real data we use the mouse obesity gene expression set from Xie et al., where the true effects are unknown. We therefore evaluate stability: agreement of the estimated effects across disjoint valid proxy sets, and subsampling variance at fractional sample sizes. We also check that the genes Gstm2 and Sirpa, reported as recovered by Proxy-Rank, remain recovered under our selection. Lower subsampling variance without losing those genes counts as support. A shift in which genes are flagged would be a substantive finding that we would analyze rather than suppress.
References
- Automating the Selection of Proxy Variables of Unmeasured Confounders
- Training sequence design for MIMO channels: an application-oriented approach
- Condition number bounds for causal inference
- MSE-optimal adjustment sets in linear Gaussian causal models with finite sample size
- Estimator selection: End-performance metric aspects
- Reward-oriented Causal Representation Learning
- Fast Proxy Experiment Design for Causal Effect Identification