Letting the Proxies Decide How Many Senses a Confounder Has

Doctors' notes often mention the very things that bias observational estimates: a patient's frailty, an unrecorded diagnosis, a clinician's hunch. Recent work turns such notes into causal adjustments by asking a language model a yes/no question about a named condition in two separate documents. One answer serves as a proxy for treatment and the other as a proxy for outcome. A proxy here is an observed variable that carries information about the hidden confounder, the unrecorded factor that drives both treatment and outcome. The weakness is that someone must already know what the confounder is and name it in the prompt. This proposal removes that requirement. Each document type is clustered into an open-ended, data-determined number of meaning groups, the way non-parametric word sense induction lets the number of senses of a word grow with the data. The number of groups is chosen by a signal that only exists in the causal setting: how much one document's groups predict the other document's groups.

The core point is that both fields share one problem and that each can supply what the other lacks. Sense induction fails because it cannot tell when not to split. On natural text, every unsupervised method loses to the trivial rule "one cluster per word." Proximal causal inference with discrete proxies fails in two ways. If the proxies have fewer categories than the confounder has states, identification breaks. If they have many redundant categories, the estimator becomes numerically unstable. Two conditionally independent views of the same hidden state settle the question. A split is real only if it shows up in the other view. We propose to build this cross-view stopping rule, use it for open-vocabulary proximal adjustment on clinical text, and test as a secondary question whether the same rule loosens the one-cluster ceiling in word sense induction.

Core Claim

Suppose two notes A and B are written about the same patient before treatment, and their contents depend on each other only through a hidden discrete confounder $U$ with $m$ states. Cluster each note's content into $k_A$ and $k_B$ groups, and call the resulting categorical variables $Z$ (treatment proxy, from A) and $W$ (outcome proxy, from B). Then, within each treatment arm $t$, the cross-view table $P(W, Z \mid T=t)$ factors through $U$ and has rank at most $m$. The proximal estimator needs that rank to be exactly $m$, a "completeness" condition. Under-splitting drops the rank below $m$ and biases the estimate. Over-splitting keeps the rank at $m$ but adds near-zero singular values, which blow up the variance when we solve for the outcome bridge function $h$. In the discrete case $h$ is defined by

$$P(Y=1 \mid Z=z, T=t) = \sum_{w} h(w,t)\, P(W=w \mid Z=z, T=t),$$

and the effect is then $E[h(W,1) - h(W,0)]$. Solving this system means inverting the matrix $P(W \mid Z, T=t)$, so its smallest retained singular value controls the error.

We therefore choose the clustering resolution so that the table's statistically supported rank is fully captured and no clusters are spent on noise. A split is kept only if it raises the number of singular values of $P(W,Z \mid T=t)$ that sit above a bootstrap noise floor. This makes the number of senses an output of cross-view dependence rather than of embedding geometry. That is exactly the "when to stop" signal that In the LLM era, Word Sense Induction remains unsolved shows is missing from purely geometric induction.

Relation to Prior Work

The closest work is Proximal Causal Inference With Text Data. It prompts zero-shot language models (FLAN-T5, OLMo-7B-Instruct) for binary indicators of pre-named conditions in two notes and feeds them into a two-stage linear proximal estimator. It also proves that both proxies must come from different documents, and it offers an odds-ratio test to abort when proxies look dependent. That paper supplies our identification setup, our two-note design, which we keep rather than try to circumvent, and our falsification check. We differ in three ways. The concept is never named. The proxies are multi-category rather than binary. And the number of categories is learned, which the original leaves to future work.

Text and Causal Inference: A Review motivates the project. It documents that text adjustment relies on fixed-$k$ topic models or dense embeddings, and that small representational choices swing the estimates. Our rule replaces an arbitrary $k$ with one that is tied to identification. Deep Multi-Modal Structural Equations is the strongest generative alternative. It infers a continuous Gaussian confounder from several unstructured proxies with a product-of-experts variational model. We instead commit to a discrete latent, where identification reduces to linear algebra that can be checked rather than to non-convex training.

Bayesian Nonparametric Adjustment of Confounding plays a constraining-and-inspiring role. It shows that tying confounder selection to joint relevance for treatment and outcome beats fitting each in isolation, though only on tabular covariates. Our cross-view criterion is a text-side analogue of "relevance to both sides." Using Embeddings to Correct for Unobserved Confounding similarly found that representations learned with the causal task beat pre-trained, two-stage ones. Our method sits between the two: the clustering is unsupervised within each view, but its resolution is set by a causal criterion. The clustering engine itself is DP-means, as used in Infinite Mixture Chaining. There a single distance threshold $\lambda$ interpolates between one prototype and one cluster per exemplar. We reuse that machinery and replace the way $\lambda$ is chosen.

Research Program

The pipeline has three parts. First, each sentence of each pre-treatment note is embedded with a frozen clinical-capable encoder. DP-means is run separately on note type A and note type B over a path of thresholds $\lambda$, and each patient's note is summarized by its dominant cluster. We will also test a soft version that represents each note as a cluster-frequency vector. Second, for each pair of thresholds we form the treatment-stratified cross-view tables. We compute the bootstrap noise floor for their singular values and retain the coarsest pair whose supported rank equals the maximal supported rank along the path. Clusters whose rows lie in the span of others are merged. Third, the discrete bridge equation is solved by regularized least squares. The odds-ratio falsification check from the text-proximal paper is applied, and we report bootstrap intervals that rerun the whole pipeline, including the threshold selection, to avoid the forking-paths problem.

The analysis will state the conditions under which the selected cardinality is consistent for $m$: conditional independence of the views given $U$, and a minimum singular-value gap. It will also characterize what happens when note content depends on things other than $U$, for example template or author effects shared across both notes. That case is the main threat, because shared style inflates the cross-view rank with non-confounding structure. We will test it directly by injecting shared-author structure.

As a secondary question, the same rule is applied to word sense induction. The two views are the left and the right context of each token, encoded separately, and a lemma is split only when the split is supported in both views. We expect this to be harder than the clinical setting because the two context windows are not truly independent given the sense, so we treat it as an exploratory test of whether the stopping signal transfers, not as a headline claim.

Baselines

The naive reference point is the unadjusted difference in means, together with its sense-induction counterpart, one cluster per lemma. On SemCor-WSI, one cluster per lemma reaches 73.6 F-B³ according to In the LLM era, Word Sense Induction remains unsolved. An oracle reference adjusts on the true simulated confounder and bounds what any proxy method can reach.

The strongest existing method on the causal question is the text-proximal estimator of Proximal Causal Inference With Text Data. We run it twice. In the first run the prompt names the true condition, which gives it the advantage of knowing what the confounder is. In the second run the prompt names a plausible but wrong or generic condition, which reflects realistic ignorance. Both runs use the same note pairs, the same simulated treatments and outcomes, and the same bootstrap. That paper reports qualitatively low bias versus a biased naive backdoor adjustment, but no reported result found on Lacuna for its exact bias values on MIMIC-III, so we will rerun it ourselves.

The generative competitor is Deep Multi-Modal Structural Equations. It gets the same sentence embeddings as input and a tuning budget for latent dimension and learning rate equal in compute to our threshold sweep. Its reported synthetic-image result is an average treatment effect error of 0.096 against 4.067 for inverse-probability weighting. That comes from a different benchmark, so it shows the method is strong but does not provide a number to beat here.

Two further text baselines are backdoor adjustment with a doubly robust estimator on pooled embeddings of both notes, and proximal estimation with fixed-$k$ clusters, with $k$ chosen by within-view BIC or silhouette or set to a topic-model-style 50.

The ablation that isolates the key ingredient keeps everything identical but chooses $\lambda$ by within-view criteria instead of the cross-view rank. Beating the realistic-ignorance text-proximal run and the fixed-$k$ run on bias, while staying within the oracle-named run's confidence interval, counts as success. Matching the within-view ablation counts as failure of the core idea. On SemCor-WSI, any reliable improvement over 73.6 without dictionary constraints would be meaningful. BERT-Wikt's 76.0, which uses Wiktionary must-link constraints, is the supervised ceiling.

Why not just…?

Why not just ask a language model to list the patient's relevant conditions, then cluster those answers? This is the real threat, and we include it as a baseline. A frontier model's free-text condition summaries are normalized, clustered, and passed into the same proximal estimator. The evidence against it is indirect. On sense induction over natural distributions, GPT-4o scores 66.9 F-B³, below the one-cluster rule, and it degrades on lemmas with many instances (In the LLM era, Word Sense Induction remains unsolved). That suggests open-ended model output over-splits and drifts. Its free-text labels also still need a cardinality decision, so our rule can be applied on top of them. If the model's summaries plus our stopping rule beat embeddings plus our rule, we will report that the stopping rule, not the clustering engine, is the contribution.

Why not just adjust for the full embeddings with a doubly robust estimator? That approach is backdoor adjustment on a noisy measurement of the confounder. Measurement error does not average away in backdoor adjustment, and the text-proximal paper found naive backdoor on model predictions stayed significantly biased. With high-dimensional inputs, weighting methods also collapse: plain inverse weighting reached only 0.190 correlation with true effects on text treatments in Causal Risk Minimization for High-Dimensional Treatments. Our experiments include this baseline, and we expect it to stay biased as label noise grows.

Why not just pick $k$ with BIC or silhouette? Within-view criteria measure how well clusters fit one note's geometry, not whether the clusters track the shared hidden state. This is precisely the over-splitting failure behind the one-cluster result. The within-view ablation answers this alternative directly. If it matches our method, the cross-view idea adds nothing, and we will say so.

Why not just use a continuous latent model? It may work, which is why we run it as a baseline. Its latent space carries no guarantee of matching the confounder, and the gap between a continuous latent and a discrete clinical state is an empirical question that the composite-confounder settings below are designed to expose.

Proposed Experiments

The main testbed follows the semi-synthetic MIMIC-III design of Proximal Causal Inference With Text Data: real nursing-note and echocardiogram-report pairs, with treatments and outcomes simulated from a chosen confounder. We extend it so the confounder is hidden from the method and harder to name. The confounder cardinality $m$ is set to 2, 3, 4, and 6 by crossing two or three of the recorded conditions into composite states. Confounding strength is varied, and the rare-state frequency is varied to stress the singular-value gap. A stress condition injects shared author or template effects across both notes.

Metrics are absolute bias and root mean squared error of the average treatment effect over 200 simulation replicates, coverage of 95% bootstrap intervals, the selected cardinality versus the true $m$, and the rate at which the odds-ratio check correctly aborts.

The decisive checks are three. First, the method's bias should track the oracle-named text-proximal run and clearly beat the wrong-name run. Second, the selected cardinality should rise with the true $m$ while fixed-$k$ and within-view choices should not. Third, under injected shared-style structure, the method should either over-estimate rank in a detectable way or be stopped by the falsification check. A silent failure there would be a negative result we will report. On SemCor-WSI we report F-B³ for the left/right-context version of the stopping rule against one cluster per lemma, the within-view DP-means ablation, and BERT-Wikt.