Aligned Confounding: Balancing Only the Shared Subspace of Hospital Treatment Policies

When retrospective data are pooled across many hospitals, each hospital chooses treatments by its own rules. Each set of rules is a separate source of bias, because sicker or different patients end up in different treatment arms. Standard corrections treat these sources as unrelated, or they balance every covariate direction at once. Both fail as the number of hospitals grows. The first spends statistical power on every source separately. The second discards information that predicts outcomes. Wireless engineers met the same structure when many transmitters share one channel. Their answer, interference alignment, was to arrange the unwanted signals so they all fall into one small shared subspace that the receiver can discard, while the useful signal keeps its full dimension. This note proposes the clinical analogue. We estimate the low-dimensional subspace in which all hospitals' treatment-selection rules jointly live. We then remove imbalance only inside that subspace, and we add an explicit constraint that the directions which predict outcomes survive intact.

In two sentences: if many hospitals' treatment policies differ from one another only along a few shared covariate directions, then correcting selection bias needs to touch only those few directions, not every covariate and not every hospital separately. We find those directions with a convex low-rank estimate, borrowed from the rank-minimization view of interference alignment, and we forbid the estimate from absorbing outcome-predictive directions. The method only addresses measured confounding. It does nothing about unmeasured confounding, and we say so throughout.

Core Claim

The claim has two parts. The first is empirical: in multi-center registries, the differences between centers' treatment policies have low rank relative to the number of centers. We call this "aligned confounding." The second is methodological: when aligned confounding holds, balancing only the estimated $r$-dimensional selection subspace gives lower individual treatment-effect error than three alternatives. These are balancing all directions, balancing each center's propensity separately, and the adversarial single-subspace method of Spiked-CFR. The gain should grow with the number of centers $K$ for fixed rank $r$.

The novel ingredient is not low-rank propensity estimation by itself. Reduced-rank regression is standard statistics, and our ablation is designed to check whether it is enough. The novel ingredient is the signal-preservation constraint taken from interference alignment, applied to pooled multi-site selection.

The closest prior work is Spiked-CFR. It posits that selection bias in high-dimensional text embeddings concentrates in a $k$-dimensional "spike." It finds that spike with a minimax game on the space of orthonormal projections, in which an adversary picks the projection with the largest treated-versus-control Wasserstein gap. That paper supplies our motivation. It shows that balancing all directions converges at the slow empirical rate $n^{-1/D}$, where $D$ is the representation dimension, and that this over-balancing destroys prognostic signal.

Our proposal differs in three ways. First, it treats a pooled registry as $K$ distinct selection sources rather than one, and it asks whether those sources align. Second, it finds the shared subspace by convex nuclear-norm estimation rather than adversarial gradient steps. Third, it imposes a hard constraint on outcome-predictive directions rather than hoping the minimax objective leaves them alone.

The method source is Interference Alignment as a Rank Constrained Rank Minimization. That paper casts alignment as minimizing the rank of interference at each receiver while keeping the desired signal full rank. It relaxes rank to the nuclear norm (the sum of singular values) and the full-rank requirement to a minimum-eigenvalue constraint. It also reports that the nuclear norm drives interference to a genuinely low rank, whereas squared-error "leakage" minimization spreads residual energy across all dimensions. That last finding is the reason we expect plain shrinkage of propensity models to underperform.

Research Program

Let $x \in \mathbb{R}^p$ be a patient's covariates. Let $b_k \in \mathbb{R}^p$ be the coefficients by which center $k$'s treatment policy deviates from the pooled policy, collected into $B = [b_1, \dots, b_K]$. Let $G \in \mathbb{R}^{p \times m}$ hold coefficients of an outcome model for $m$ arms or outcome components, fitted on held-out folds. The selection subspace is estimated by

$$ \min_{B} \; \sum_{k=1}^{K} \ell_k(b_k) + \lambda \|B\|_* \quad \text{subject to} \quad \sigma_{\min}\big((I - \Pi_B)\, G\big) \ge \delta, $$

where $\ell_k$ is center $k$'s logistic propensity loss and $\|B\|_*$ is the nuclear norm. $\Pi_B$ is the projection onto the top-$r$ left singular vectors of $B$, and $\sigma_{\min}$ is the smallest singular value. The threshold $\delta$ requires that outcome-predictive directions remain recoverable outside the selection subspace.

The constraint is non-convex because $\Pi_B$ depends on $B$. Following Papailiopoulos et al., we solve the problem by alternating between two steps. One is a convex nuclear-norm step with the projection held fixed. The other is a projection update.

The estimated subspace $S$ then plugs into a counterfactual-regression learner, meaning a neural network that predicts outcomes under each treatment from a shared representation. Its Wasserstein balance penalty is applied only to the projection of the representation onto $S$. Following the rate argument in Spiked-CFR, the balance term should then converge like $n^{-1/r}$ rather than $n^{-1/D}$. For fixed $r$, adding centers adds selection sources without adding balancing dimensions.

We take one constraint seriously from the feasibility analysis of Bresler et al.. That paper shows the asymptotic promise of alignment collapses into a hard dimension-counting limit when resources are finite. By analogy, we do not assume that alignment always exists. We will map empirically where aligned balancing helps and where it stops helping as $K$, $r$, $p$, and per-center sample size vary. We will report the region where the selection rank approaches $\min(p, K)$ as the regime in which the method should reduce to ordinary balancing. This is an empirical phase diagram, not a theorem.

We also state the method's limits directly. The approach cannot rescue designs like Zhou et al.'s pancreaticoduodenectomy comparison, where each treatment came from a different hospital so that center and treatment are inseparable. It requires every center to give both treatments to some patients. It does nothing for unmeasured frailty or surgeon judgment, which that paper rightly flags. The adjusted analysis of Dumas, Chevret et al. shows a case where propensity weighting on measured covariates removed an apparent harm of non-invasive ventilation (weighted OR 1.10, p=0.562). That result is not a contradiction. It is the measured-confounding regime our method targets, and the claim here concerns statistical efficiency when there are many centers, not identification.

Baselines

The naive reference is an outcome regression pooled across all centers that ignores both treatment selection and center. We also run the same regression with center indicator variables, which is the classical fixed-effects adjustment.

The strongest classical competitor is augmented inverse-probability weighting. It combines a propensity model with an outcome model so that the estimate stays consistent if either one is correctly specified. Here the propensity model is a center-stratified logistic regression with center-by-covariate interactions, and weights are trimmed at a fixed quantile. This is the natural epidemiological answer to multi-center selection. Its known weakness is high-variance weights under imbalance, which UTVAE documents for deep latent models.

Among representation learners, we run plain counterfactual regression with global Wasserstein balancing. We also run Spiked-CFR with its adversarial subspace search and the same $r$. This is the strongest direct competitor, because it already exploits low-dimensional selection.

We add HyperTLearner from HyperITE. It generates each arm's network weights from a shared hypernetwork, and it is the strongest reported method in the file for small cohorts. On IHDP, the file reports a PEHE-out (error in individual effects on held-out units) of 1.12 for HyperTLearner versus 1.32 for a T-Learner. For average-effect error on IHDP, the file reports UTVAE at 0.62 against CEVAE at 1.03. The file reports no number for Spiked-CFR on IHDP or ACIC-2016, so no reported result was found on Lacuna for that comparison, and we will run it ourselves.

Two ablations isolate the key ingredient. In the first, the nuclear-norm subspace is estimated without the signal-preservation constraint, which is just reduced-rank propensity regression. In the second, the subspace comes from ordinary principal components of the covariates.

All methods share identical folds, the same backbone network where applicable, and the same hyperparameter search budget of 50 trials, selected by a validation criterion that does not see counterfactuals. All methods also get the same compute. Beating the field means a lower PEHE than both Spiked-CFR and the unconstrained ablation across seeds, with paired confidence intervals excluding zero. On standard IHDP, where there is only one "center," we expect parity, and we treat parity there as a sanity check rather than a target.

Why not just…?

Why not just use center-stratified propensity weighting? It is consistent, simple, and trusted. It estimates $K$ separate selection models, however, so its variance grows with $K$ while per-center samples shrink. This is exactly the fragmentation HyperITE documents across treatment arms. The experiment that settles this sweeps $K$ at fixed total sample size. If stratified augmented weighting keeps up with our method as $K$ grows, aligned balancing adds nothing, and we will report that.

Why not just fit a reduced-rank propensity model and stop? This is the most serious threat, because low-rank multi-task logistic regression is already known. The signal-preservation constraint matters only when selection and prognosis share directions, as in confounding by indication, where the sickest patients get the aggressive treatment. In that case an unconstrained low-rank subspace will absorb prognostic directions, and balancing it will throw away outcome signal. We therefore vary the overlap angle between the selection and prognostic subspaces in simulation. If the unconstrained ablation matches the full method even at high overlap, the contribution reduces to known statistics, and we will say so.

Why not just use Spiked-CFR on the pooled data? It already balances a low-dimensional spike. It does not, however, model multiple selection sources, and its adversarial search is unstable when samples are small. We compare against it at matched $r$, and we report seed-to-seed variance of the recovered subspace, measured by principal angles to the true subspace, alongside PEHE.

Why not just residualize covariates on center indicators? That is the literal zero-forcing projection suggested by Downlink Interference Alignment. It removes differences in center means, not differences in how centers select treatment as a function of covariates. It is included as a baseline, and we expect it to fail whenever policies differ in slope rather than intercept.

Proposed Experiments and Decisive Checks

The first set of experiments is semi-synthetic, built on real covariates and real center membership. We take covariates and hospital identifiers from the multi-center eICU database and from MIMIC-IV. We simulate center-specific treatment policies whose deviation matrix has controlled rank $r$ in $\{1, 2, 4\}$ across $K$ from 5 to 100 hospitals, and we simulate outcomes with controlled overlap between the prognostic and selection subspaces. We repeat the exercise on IHDP and ACIC-2016 covariates by assigning synthetic center labels. The metrics are PEHE, average-effect error, and the principal-angle error of the recovered subspace.

The decisive check is a pair of curves. If aligned confounding is the right lens, the PEHE gap between our method and both global balancing and stratified weighting should widen with $K$ at fixed $r$. That gap should vanish as $r$ approaches $K$. The gap over the unconstrained ablation should grow with subspace overlap. If any of these patterns fails to appear, the corresponding part of the claim is falsified.

The second experiment tests the alignment hypothesis directly on real data. For a recorded intervention in eICU chosen by protocol before analysis, we estimate the singular-value spectrum of the hospital deviation matrix with cross-validated nuclear-norm regularization and compare it against a permutation null. A flat spectrum would mean real hospital policies do not align, which would undercut the premise outside simulation.

The third experiment uses negative-control outcomes, meaning outcomes the treatment cannot plausibly affect. Effects estimated on them should be near zero, and residual effects are reported as bias that remains after adjustment. A residual effect would indicate either confounding along directions outside the estimated subspace or unmeasured confounding, which this method cannot fix.