Outcome-Specific Treatment Balancing for Multi-Outcome Causal Estimation

When a hospital or an ad platform wants to know what a treatment would do, it rarely cares about a single outcome. A drug changes survival and side-effect risk. A promotion changes visits and purchases. The standard neural recipe for estimating such effects from observational logs is adversarial balancing: an encoder learns features from which a second network, the discriminator, cannot tell who was treated. The point is to stop the model from memorizing the historical rule that decided who got treated. The standard recipe for several outcomes is multi-task learning, where one shared encoder feeds a separate prediction head per outcome. Combining the two seems natural, and two papers already do it. Both apply a single, global balancing pressure to the shared features.

This note argues that the global choice is wrong whenever a covariate plays different causal roles for different outcomes. Suppose a patient's insurance status drives treatment choice and also affects readmission, but has no bearing on tumor response. For readmission it is a confounder that must be kept. For tumor response it is an instrument, a variable that only affects who gets treated, and it should be removed because it only adds variance. One global discriminator must either erase it, which biases the readmission estimate, or keep it, which makes the tumor-response estimate noisier. We propose giving each outcome head its own learned view of the shared features, with its own discriminator. Each discriminator is conditioned on that head's outcome, so balancing removes only what is irrelevant to that outcome. The claim is specific and testable. The advantage over global balancing should grow with the fraction of treatment-predictive covariates whose role differs across outcomes. The advantage over training one de-confounded model per outcome should grow as some outcomes become sparsely labeled.

What exists and where this sits

Three papers supply the method. Adversarial De-confounding in Individualised Treatment Effects Estimation splits the representation into instrument, confounder and adjustment parts and applies the gradient reversal layer only to the confounder part. A gradient reversal layer passes values unchanged going forward and flips the sign of the discriminator's gradient going back into the encoder. That paper shows that balancing indiscriminately hurts. Its partition is defined for a single outcome, and it reports strong sensitivity to the balancing weight. On the Estimation of Treatment Effect with Text Covariates (CTAM) conditions the discriminator on the outcome, so features that predict treatment but not outcome are the ones removed. We reuse that conditioning, one copy per outcome. Gradient Surgery for Multi-Task Learning and RotoGrad supply the competing view that the trouble is conflicting gradients rather than conflicting invariance targets. We test that view directly as a baseline rather than adopting it.

Two papers motivate the gap by doing the global version. Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials (MTAL) shares layers across tumor types and uses one true-or-false discriminator on factual versus generated outcomes. Its tasks are subpopulations sharing one outcome, not distinct outcomes with distinct causal structure. Adaptive Adversarial Multi-task Representation Learning makes shared features indistinguishable across tasks, not across treatment arms. It also reports that this invariance hurts when tasks are unrelated, a negative result that constrains our design: we deliberately do not force cross-task invariance.

Two multi-outcome uplift systems show demand without balancing. Multi-Treatment Multi-Task Uplift Modeling for Enhancing User Growth (MTMT) uses mixture-of-experts sharing with no de-biasing. Large-Scale Notification Dispatch with Bundle Treatments and Multi-Outcome Uplift Optimization sidesteps confounding by training on randomized data. Bayesian Inference of Individualized Treatment Effects using Multi-task Gaussian Processes (CMGP) treats the treated and control outcome surfaces as related tasks, which is a different use of "multi-task" from ours.

Core Claim

A shared encoder produces features $z = \Phi(x)$. Each outcome $k$ owns a soft gate $m_k(x) \in [0,1]^d$, produced by a small gating network, and sees only $z_k = m_k(x) \odot z$. A per-outcome discriminator $D_k$ tries to predict treatment $t$ from the pair $(z_k, y_k)$, where $y_k$ is the observed value of outcome $k$. Its gradient is reversed into both the gate and the encoder. The objective is

$$\min_{\Phi, m, h} \max_{D} \sum_{k=1}^{K} \Big[ \mathcal{L}_{Y}\big(h_k(z_k, t), y_k\big) - \lambda\, \mathcal{L}_{T}\big(D_k(z_k, y_k), t\big) \Big] + \mu \sum_{j \ne k} \overline{m_j \odot m_k}\,\big|_{\text{treatment-predictive dims}}.$$

Here $h_k$ is outcome $k$'s head, $\mathcal{L}_Y$ is the factual prediction loss (masked when $y_k$ is missing), and $\mathcal{L}_T$ is the discriminator's cross-entropy. The last term is a light penalty with weight $\mu$. It discourages heads from sharing dimensions that a probe flags as treatment-predictive, while leaving shared adjustment features free to be reused.

The encoder stays shared. Adjustment variables, which predict outcomes regardless of treatment, therefore pool statistical strength across outcomes, and that is the reason to prefer this over K separate models. Balancing happens only inside each head's view, so no outcome's confounders get stripped on another outcome's behalf, and that is the reason to prefer it over global balancing. Conditioning on $y_k$ follows CTAM. A feature that predicts treatment but carries no information about $y_k$ remains detectable by $D_k$ even after the discriminator sees $y_k$, so it gets pushed out of $z_k$. Features informative about $y_k$ face weaker pressure. This attenuates erasure of confounders; it does not guarantee clean separation, and the experiments measure how well the gates recover the true roles.

Research Program

The first piece of work is a benchmark, because no standard multi-outcome observational benchmark with known per-outcome causal roles appears in the literature we surveyed. We build it in three tiers.

The first tier keeps IHDP's real covariates and generates K = 3 outcome surfaces. Each covariate is assigned a role (confounder, instrument, adjustment, or irrelevant) separately per outcome. The treatment rule depends on the union of confounders and instruments. A single knob $\rho \in [0,1]$ sets the fraction of treatment-predictive covariates whose role differs across outcomes. At $\rho = 0$ every outcome has the same confounders and the global approach should be optimal. Following Really Doing Great at Estimating CATE?, we generate each configuration under both a nonlinear-interaction effect surface and a purely additive one, because that paper shows rankings can reverse between the two. ACIC 2016 covariates serve as a second covariate source.

The second tier applies the same construction to the News text covariates with 4 and 16 treatments. This tests the method where nearly-instrumental word features are abundant.

The third tier uses CRITEO Uplift, a randomized ad dataset with two binary outcomes, visit and conversion. We induce confounding by subsampling treated and control users with a covariate-dependent rule that uses different features for each outcome's natural drivers. Because the full data is randomized, effect estimates on held-out randomized data remain an honest ground truth. Throughout, we also vary label sparsity by hiding a growing fraction of one outcome's labels.

Analysis has two parts. On synthetic data we read the learned gates and report how well each $m_k$ separates outcome k's true instruments from its confounders, measured by AUROC of gate value against true role. On all data we track the angle between the reversed adversarial gradient and the outcome gradients on the shared encoder. Gap 1 in the literature hypothesizes that conflicts there cause failures. If our per-outcome design removes most of that conflict without any gradient surgery, that is evidence the conflict was a symptom of mismatched invariance targets rather than an independent optimization problem.

Baselines

The naive reference is a shared encoder with one head per outcome and no balancing at all: hard parameter sharing in the style of a multi-headed TARNet. It shows how much confounding hurts in each configuration. If it is close to everything else, the benchmark is not confounded enough and must be made harder before any other comparison means anything.

The strongest existing method on the balancing side is the selective de-confounding model of Adversarial De-confounding in Individualised Treatment Effects Estimation (SNet+), trained separately per outcome. It is the most direct competitor because it already solves the role problem for one outcome; it just does not share. On its own synthetic benchmark it reports a PEHE-out of 0.4008, against 0.4116 for SNet and 2.5233 for a T-learner. PEHE measures how accurately the model estimates each individual's treatment effect, and PEHE-out is that error on held-out individuals. Beating it means lower PEHE on every outcome at matched compute, plus a gap that widens as one outcome's labels become sparse. For IHDP-style data the closest reported figure is CMGP's √PEHE of 0.9 in Bayesian Inference of Individualized Treatment Effects using Multi-task Gaussian Processes. For News, CTAM reports PEHE 2.06 and MTAL reports an ATE error of 5.12 on News-16, where ATE is the average effect over the whole population. Those numbers are single-outcome; on our multi-outcome variants there is no reported result found on Lacuna, so all methods are rerun by us.

The strongest multi-task competitor is global balancing: one shared encoder, one discriminator, one gradient reversal layer. We run it in two forms. The first is the single-outcome Chauhan partition applied with one confounder block shared by all heads. The second is an MTAL-style true-or-false discriminator following Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials. A third variant adds PCGrad projection between the adversarial and outcome gradients. On CRITEO we also run MTMT; its reported QINI is 0.164 against 0.0779 for FlexTENet on the original randomized data, from Multi-Treatment Multi-Task Uplift Modeling for Enhancing User Growth. QINI scores how well a model ranks users by uplift. Those figures are not comparable to our confounded subsample, so we retrain MTMT there. Beating these baselines means a gain that is negligible at $\rho = 0$ and grows monotonically with $\rho$. A uniform gain would suggest extra capacity rather than mechanism.

Three ablations isolate the key ingredient: per-outcome discriminators without gates, gates without conditioning on $y_k$, and per-outcome gates with a single shared discriminator.

Every method gets the same encoder width and depth (the gated model's parameter count is matched by widening baselines), the same training budget, and the same tuning grid over the balancing weight. That grid is searched with the same selection rule for all methods: factual validation loss plus a held-out treatment-classifier check on the balanced features, since true counterfactuals are unavailable. We also report each method's best-in-hindsight λ, so that sensitivity is visible rather than hidden. As an additional tuning arm, we apply the augmented-Lagrangian schedule of Adaptive Adversarial Multi-task Representation Learning to both our model and global balancing. All results average 10 seeds and 100 data draws for the IHDP tier.

Why not just…?

Why not just train one de-confounded model per outcome? This is the real threat. With abundant labels for every outcome it may match us, because the per-outcome role problem disappears when nothing is shared. The case for sharing is statistical: adjustment variables common to all outcomes are learned from every outcome's labels. The label-sparsity sweep settles this. If separate SNet+ models match us even when one outcome keeps only 10% of its labels, the shared encoder is not earning its keep, and we would report that sharing adds nothing here.

Why not just keep global balancing and fix the optimization with gradient surgery? PCGrad removes the component of one gradient that conflicts with another. It cannot change what a single discriminator is asking for. If a covariate must be removed for outcome A and kept for outcome B, a projected gradient still drives one shared invariance target. The prediction that separates the two explanations is that global balancing plus PCGrad should improve over plain global balancing at $\rho = 0$, where the trouble is pure optimization, but should keep losing to our method as $\rho$ grows. If PCGrad closes the gap at high $\rho$, our diagnosis is wrong.

Why not just skip balancing and rely on multi-task regularization? Sharing a backbone reduces overfitting to noise, but confounding is a systematic property of the treatment rule, not noise. Global adversarial models already beat non-adversarial ones in Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials and Adversarial De-confounding in Individualised Treatment Effects Estimation. The naive baseline checks this holds in our benchmarks, especially under the additive effect surface, where the benchmarking critique suggests flexible networks can lose their edge.

Decisive checks

The proposal stands or falls on three results. First, the gain over global balancing must be near zero at $\rho = 0$ and rise with $\rho$ under both the nonlinear and additive effect surfaces. Second, the learned gates must separate per-outcome instruments from confounders on synthetic data with AUROC clearly above chance; if they do not, any PEHE gain is not coming from the stated mechanism. Third, on the confounded CRITEO subsample, effect estimates must rank users better than global balancing and MTMT when scored against the randomized hold-out. If the first check fails, the role-mismatch story is wrong. If the second fails but the others pass, we have a better estimator without understanding why, and we would say so.