Some causal questions stay unanswerable even in a perfectly randomized experiment. A trial tells us the average effect of a drug, but not what fraction of patients it actually helps. That would require seeing each patient both treated and untreated. The best we can do is compute bounds, meaning the lowest and highest values consistent with the data. Measuring baseline covariates narrows those bounds. If age and a lab value predict almost perfectly whether a patient recovers in each arm, the set of possible "who benefits" answers shrinks sharply. Covariates are not free, though. Each one costs money to collect in a follow-up trial, and each one added to a finite sample adds estimation noise. Tightening Causal Bounds via Covariate-Aware Optimal Transport states this second point in its limitations section.
This note proposes treating covariate choice as an optimization problem whose objective is the width of the causal bound, not predictive accuracy. Given a pilot trial with many candidate baseline covariates and a budget of $k$, the method searches for the $k$ covariates that make the bound narrowest. It then reports a bound whose validity survives the search itself. That second requirement is the easy part to get wrong. Plug-in estimates of bounds are already biased toward being too narrow, because they are built from min and max operations. Choosing among thousands of subsets for the narrowest estimated interval makes that bias worse. A method without a correction will therefore report bounds that are sharp and wrong.
Core Claim
We work in randomized trials with pre-treatment covariates. Here any covariate subset yields a valid bound in the population, so the choice of subset is purely a question of statistics and computation. There is no causal-structure risk: pre-treatment covariates cannot be colliders or post-treatment variables in a randomized design.
In this regime we claim three things. First, the bound-width objective differs from predictive objectives in a way that can be characterized. Second, a screen-then-swap discrete search guided by a smooth surrogate of that objective finds subsets that give materially narrower bounds than predictive selection at the same budget. Third, selecting on one data split, scoring with an upper-confidence width, and reporting on a held-out split restores nominal coverage, which in-sample selection loses.
To make the objective concrete, take a binary outcome and the query "fraction who benefit," $P(Y(1)=1, Y(0)=0)$. Let $p_d(x) = P(Y=1 \mid X_S = x, T=d)$ be the recovery probability in arm $d$ given the chosen covariates $X_S$. The sharp bounds are the classical Fréchet–Hoeffding limits within each covariate stratum, averaged over strata. Their width for subset $S$ is
$$ \Delta(S) = \mathbb{E}\big[\min\{p_1(X_S),\, 1-p_0(X_S)\} - \max\{0,\, p_1(X_S)-p_0(X_S)\}\big]. $$
This quantity shrinks when the covariates push either arm's recovery probability toward 0 or 1. It is flat or kinked elsewhere. Predictive loss such as log-loss instead rewards improvement everywhere. The two objectives agree roughly but not exactly, and the gap between them is what the proposal tests.
For continuous outcomes and correlation-type queries, the same logic applies with the mirror-relaxed transport bound of Tightening Causal Bounds via Covariate-Aware Optimal Transport standing in for $\Delta(S)$.
Research Program
The work builds on four lines of prior work, each playing a different role.
Motivation from optimal transport. Tightening Causal Bounds via Covariate-Aware Optimal Transport turns covariate-conditional transport into an unconditional problem by adding a penalty $\eta \|Z(0)-Z(1)\|^2$ to the transport cost. It proves that bounds tighten as $\eta$ grows. It also concedes that irrelevant covariates hurt finite-sample performance, and it floats, without implementing, a LASSO-style penalty. Its covariate vector is fixed. We choose it.
Motivation from causal dimensionality reduction. Bounds on Causal Effects and Application to High Dimensional Data splits a high-dimensional confounder vector into an observed part and a latent part with a known prior. It shows this trades a little width for large sample savings: 1,920 samples versus more than 30,000 for exact point computation. The split, however, is chosen by hand. We automate the analogous choice, without needing a known prior, because randomization removes confounding.
Search method from feature selection. Population-Robust Feature Selection via Generalized Welfare Optimization does budget-$k$ selection by screening candidates, then running a validated one-swap local search. It ranks candidate swaps with a quadratic surrogate before refitting, and it reports that discrete search beats continuous relaxation. We reuse that search skeleton but replace its predictive teacher-compression objective with $\Delta(S)$.
Smoothing method from data fusion. Data Fusion for Partial Identification of Causal Effects smooths min/max bound envelopes with a Boltzmann operator so they can be differentiated. We use the same smoothing so that swap gains can be scored by a first-order surrogate, avoiding a refit for every candidate.
The pipeline has four stages.
Stage 1: splitting and nuisance fitting. The pilot data are split into a selection fold and a reporting fold. On the selection fold, arm-specific outcome models are fit on a screened pool. Screening uses marginal association with outcome extremeness in either arm, together with group-lasso screening across the two arms.
Stage 2: discrete search. Local search starts from a greedy solution and uses one-swap moves. Each candidate swap is scored by the smoothed width's change, estimated from the fitted models. Only the top candidates are refit.
Stage 3: upper-confidence scoring. Selection minimizes an upper-confidence width, $\hat\Delta(S) + c\,\widehat{\mathrm{se}}(S)$, with the standard error estimated by cross-fitting inside the selection fold. This discourages subsets whose narrowness is a small-sample artifact.
Stage 4: honest reporting. On the reporting fold, bounds are estimated for the chosen subset only. We use debiased estimators of the smoothed bound functionals and widen by the smoothing error, so the reported interval is conservative.
We will also prove a simple result: under a margin condition, the kink-weighted objective and arm-wise log-loss choose different subsets. Constructing that separating case also serves as a synthetic stress test.
Baselines
Naive reference: no covariates. The first reference point uses no covariates at all: unconditional Fréchet bounds for binary outcomes, or unconditional transport bounds for continuous ones. They are computed on the full pilot data. Any selection method must beat this on width while matching its coverage. Tightening Causal Bounds via Covariate-Aware Optimal Transport reports that its covariate-aware bounds reduce interval width for the correlation of potential outcomes by 26% relative to this unconditional baseline on the STAR data. That number anchors how much covariates can buy in a real trial.
All covariates. The strongest existing method on this question conditions on every covariate using the mirror relaxation of Tightening Causal Bounds via Covariate-Aware Optimal Transport for continuous outcomes, and flexible arm-specific models plugged into the Fréchet formula for binary ones. It ignores the budget, so it is an upper benchmark on usable information rather than a fair budget-$k$ competitor. We run it on the same folds with the same nuisance learners and tuning budget. The proposal "beats" it if, at small $k$, the selected subset attains width within a stated tolerance of the all-covariate width, or narrower in small samples where the all-covariate estimate is noisy, while keeping nominal coverage.
Predictive selection at the same budget. The main budget-matched competitor is PopFS from Population-Robust Feature Selection via Generalized Welfare Optimization. It is run with $\alpha=1$ and the two treatment arms treated as populations, with the same screening pool and swap budget. We also include arm-wise lasso top-$k$. Selected sets are passed to the identical honest reporting stage. Beating this baseline means narrower held-out width at equal $k$, across a majority of semi-synthetic draws, with no loss of coverage.
Continuous relaxation. An L1-penalized relaxation of the mirror transport cost is rounded to $k$ covariates. This is exactly the approach Tightening Causal Bounds via Covariate-Aware Optimal Transport conjectures but does not evaluate.
Ablations. Two ablations isolate the key ingredients. The first removes honesty: it selects and reports on the same data with a point-width score, which should undercover. The second removes the surrogate: it runs the same search with a full refit for every swap, to measure what the surrogate costs in quality and saves in time.
Why not just…?
Why not just condition on predicted outcome probabilities from all covariates? For binary outcomes, $\Delta$ depends on the covariates only through the pair $(p_1, p_0)$. Conditioning on that learned two-dimensional score is therefore lossless in the population. This is a real threat when all covariates are available at reporting time, and we include it as the all-covariate benchmark. It does not answer the budgeted question, because computing the score requires measuring every covariate in the next trial. It also does not remove the finite-sample optimism of plug-in bounds. The experiments report the score-based width alongside the budget-$k$ width, so readers can see exactly how much the budget costs.
Why not just pick the most predictive covariates and bound on those? Predictive selection rewards outcome accuracy everywhere. Bound width rewards pushing arm-specific probabilities toward 0 or 1 near the kinks of the Fréchet formula. When the two objectives coincide, predictive selection should do about as well, and we expect that on many real covariate pools. The collider argument often raised against predictive selection does not apply here, because pre-treatment covariates in a randomized trial are safe. The contest is therefore settled empirically. If sharpness-targeted search fails to beat PopFS on held-out width in the semi-synthetic benchmarks, the honest conclusion is that predictive selection plus honest reporting suffices. That would itself be a useful, publishable negative result.
Why not just use a continuous L1 relaxation? Population-Robust Feature Selection via Generalized Welfare Optimization found discrete search beat the continuous relaxation baseline in all 16 comparisons. It also found the relaxation took hours or failed to finish within 24 hours on 2,110 features, while discrete search took 3–12 minutes. That evidence is from predictive selection, not bounds, so we run the relaxation directly rather than assume it fails.
Why not skip selection and use DRCFS-style causal feature selection? DRCFS: Doubly Robust Causal Feature Selection tests for direct causal parents of an outcome using debiased estimates of controlled direct effects, reaching 0.973 accuracy on the ANDES network. Its target is a parent set. A covariate that is not a parent can still tighten bounds by predicting the outcome, and a weak parent may not tighten them at all. The question it answers is a different one.
Proposed Experiments
Semi-synthetic benchmarks with known answers. We take real covariate pools from IHDP and ACIC 2019, as used by Partial Identification with Noisy Covariates: A Robust Optimization Approach, and the Kang–Schafer design. On these we simulate randomized treatment and both potential outcomes from a known joint model. This lets us compute the true sharp width for every subset and the true fraction who benefit.
Metrics. For each budget $k$ and sample size, we measure held-out width, coverage of the true identified set, coverage of the true query, and wall-clock time.
Real data. On the STAR demonstration data we cannot know coverage. We instead report width relative to unconditional transport, against the 26% reduction reported by Tightening Causal Bounds via Covariate-Aware Optimal Transport, as a function of $k$.
Decisive checks. Three results would settle the proposal.
The first concerns the ablation without honesty. It should show coverage falling measurably below nominal as the candidate pool grows, while the honest pipeline stays at nominal. If in-sample selection does not undercover, the validity half of the claim is unneeded.
The second concerns the constructed separating case, where covariates predict well only away from the kinks. There, sharpness-targeted search must beat PopFS on width. On the real covariate pools, we report the fraction of draws in which it does. The main claim stands only if the advantage is not confined to the constructed case.
The third concerns budget cost. At a small budget, the selected subset should recover most of the width reduction of the all-covariate score. If it recovers little, budgeted bounding is not worth recommending in practice.
References
- Tightening Causal Bounds via Covariate-Aware Optimal Transport
- Bounds on Causal Effects and Application to High Dimensional Data
- Population-Robust Feature Selection via Generalized Welfare Optimization
- Data Fusion for Partial Identification of Causal Effects
- DRCFS: Doubly Robust Causal Feature Selection
- Partial Identification with Noisy Covariates: A Robust Optimization Approach