At the core of classical causal inference lies a convenient simplification: the binary treatment assumption. By framing decisions as a choice between two states—treatment versus control, drug versus placebo, or action versus inaction—the mathematics of counterfactual estimation becomes remarkably clean. In this binary paradigm, every individual has exactly two potential outcomes, and estimating a causal effect reduces to measuring the contrast between these two parallel worlds.
However, real-world decision-making rarely fits into a neat binary box. Clinicians must determine the optimal dosage of a drug along a continuous spectrum, policymakers must choose between multiple distinct intervention strategies, and environmental scientists must evaluate how complex mixtures of continuous exposures jointly impact public health. When we force these multi-valued and continuous action spaces into binary frameworks, the underlying statistical, geometric, and optimization mechanisms begin to break down.
The Comfort of the Binary Paradigm
To understand why scaling beyond binary treatments is difficult, we must first look at why the binary assumption is so powerful. In observational studies, the primary challenge is confounding: individuals who receive a treatment often differ fundamentally from those who do not.
Modern deep learning approaches solve this by learning balanced representations that strip away confounding features. For example, in the framework developed by Liuyi Yao et al. in Representation Learning for Treatment Effect Estimation from Observational Data., neural networks force the feature distributions of the treated and control groups to overlap in a shared latent space. Similarly, when treatment assignments are partially unobserved, Milan Kuzmanovic et al. utilize adversarial learning in Estimating Conditional Average Treatment Effects with Missing Treatment Information to correct for a "dual covariate shift" between observed and missing groups. Even when filtering out noise, the disentanglement framework by Ahmad Saeed Khan et al. in On the Effects of Irrelevant Variables in Treatment Effect Estimation with Deep Disentanglement relies on this clean, two-sided division to isolate causal drivers.
In all these cases, the geometry is simple: you are minimizing the distance between two distributions. But if the treatment space expands to multiple discrete choices or a continuous dosage, "balancing" distributions becomes an intractable multi-domain or continuous conditional density matching problem. There is no longer a single counterfactual distribution to target.
The Optimization Bottleneck in Policy Learning
The binary limitation becomes even more acute when the goal shifts from estimating effects to learning an Optimal Treatment Regime (OTR)—a rule that maps patient characteristics to the best possible action.
In a binary setting, maximizing an individual's expected outcome is fundamentally a sign-estimation or binary classification problem. If the Conditional Average Treatment Effect (CATE) is positive, the optimal rule assigns the treatment; if negative, it assigns the control. Because of this, powerful statistical frameworks have been designed to pinpoint the exact decision boundary where the treatment effect crosses zero:
- Variable Selection: The Sequential Advantage Selection (SAS) method by Ailin Fan et al. in Sequential advantage selection for optimal treatment regime isolates the specific patient covariates that alter the treatment's advantage over the control.
- Effect Curves: The semi-parametric approach by Wenchuan Guo et al. in Estimation of Optimal Individualized Treatment Rules Using a Covariate-Specific Treatment Effect Curve With High-Dimensional Covariates maps a curve over high-dimensional covariates to identify exactly which subgroups benefit from an intervention.
- Concordance Learning: The Sparse Concordance-assisted Learning (SCAL) framework by Shuhan Liang et al. in Sparse Concordance-assisted Learning for Optimal Treatment Decision uses convex surrogate losses to rank individual treatment benefits.
When these methods confront multi-action spaces or continuous dosages, the optimization landscape changes completely. The decision boundary is no longer a simple threshold where an effect changes sign; it becomes a complex, high-dimensional surface. As Zhengling Qi et al. note in Estimation of Individualized Decision Rules Based on an Optimized Covariate-Dependent Equivalent of Random Outcomes, moving away from binary choices transforms a manageable optimization problem into a highly non-convex, discontinuous multi-action space challenge that standard surrogate losses cannot easily solve.
Why Structural Assumptions Collapse
Beyond optimization, the structural assumptions that anchor causal identification strategies frequently fall apart outside the binary domain.
Consider Instrumental Variables (IV), which are used to estimate causal effects in the presence of unmeasured confounding by leveraging an external source of randomness (like a genetic variant or a lottery). In A Semiparametric Instrumental Variable Approach to Optimal Treatment Regimes Under Endogeneity, Eric Tchetgen Tchetgen establishes robust guarantees for learning optimal regimes under endogeneity, but the mathematical proofs rely strictly on binary treatment assignments. Similarly, the neural framework introduced by Jonas Schweisthal et al. in Learning Representations of Instruments for Partial Identification of Treatment Effects maps complex instruments into discrete latent spaces to bound treatment effects.
These IV methods work in binary settings because they can partition the population into discrete compliance archetypes (such as "compliers," "always-takers," and "never-takers"). When a treatment becomes continuous, an individual can exhibit an infinite variety of compliance behaviors, destroying the discrete structural partitions required to identify or bound the causal effect.
A similar scaling bottleneck plagues adaptive experimentation and incentive design. The active learning algorithm developed by Ojash Neopane et al. in Logarithmic Neyman Regret for Adaptive Estimation of the Average Treatment Effect achieves an exponential variance reduction (logarithmic regret) when dynamically allocating individuals to treatments, but it is explicitly optimized for a single treatment-versus-control choice. When Jiachun Li et al. design mechanisms in Aligning Incentives to Balance Covariates in Experiments with Selection Bias to nudge self-selecting participants into balanced experimental groups, the incentive alignment relies on a binary choice architecture. Expanding these active allocation and incentive mechanisms to high-dimensional or continuous action spaces causes the underlying variance reduction math to degrade rapidly.
The Mirror Problem: Outcome Restrictions
The mathematical rigidity of the binary assumption is not unique to treatments; it represents a symmetrical bottleneck affecting outcome variables. Causal frameworks are frequently locked into a specific data type on either the treatment side or the outcome side due to their underlying statistical link functions:
| Framework | Treatment Type | Outcome Type | Core Bottleneck Beyond This Setup |
|---|---|---|---|
| Shin et al. | Multivariate Continuous | Continuous | Fails to handle binary endpoints (e.g., hospitalization status) or count data. |
| Song et al. | Binary | Dichotomous (Binary) | Requires entirely different statistical link functions to maintain accuracy for continuous endpoints. |
| Li et al. | Binary | Binary | Bounding mechanics for handling truncation-by-death and missingness rely on discrete outcome structures. |
As shown above, even when a method successfully breaks the binary barrier on one side—such as the Bayesian framework developed by Heejun Shin et al. to handle multivariate continuous environmental exposures—it often remains rigidly bound to a continuous outcome model. Conversely, methods designed by Yang Song et al. and Wei Li et al. to handle complex trial dynamics or missing data are structurally constrained by dichotomous outcomes. Crossing these boundaries requires completely rewriting the underlying statistical assumptions.
Bridging the Gap via Shared Embeddings
To move past these binary constraints, modern causal inference is shifting toward representation methods that treat multi-valued or continuous spaces not as isolated discrete choices, but as continuous, interconnected topologies.
A clear example of this paradigm shift is found in longitudinal settings where multiple complex treatment strategies must be evaluated simultaneously over time. Instead of evaluating each policy in isolation, Wenxin Chen et al. developed PEQ-Net in Smooth Multi-Policy Causal Effect Estimation in Longitudinal Settings.
[Traditional Multi-Arm] -> Policy A (Isolated) \
-> Policy B (Isolated) --> High Variance, No Shared Info
-> Policy C (Isolated) /
[Continuous Embedding] -> Policy A \
-> Policy B --> [ Shared Latent Space ] --> Smooth Information Sharing
-> Policy C /
Instead of treating distinct policies as independent arms, PEQ-Net embeds different policies into a shared mathematical latent space. Because similar policies sit close to one another in this embedding space, the model can smoothly share information across related treatment strategies. This geometric smoothing drastically reduces statistical variance and provides a conceptual blueprint for the field: by mapping discrete actions or continuous dosages into structured latent spaces, causal inference can finally transition from binary simplifications to the continuous complexity of real-world environments. Meanwhile, meta-learning frameworks like the SA-Learner developed by Yan-Yu Chen et al. in SA-Learner: Surrogate-Assisted Meta-Learner with Missing Outcomes underscore that extending these heterogeneous treatment effect estimators to complex treatment structures remains the vital next step for handling real-world data missingness.
Key Foundational Work
On Estimation of Optimal Treatment Regimes for Maximizingt-Year Survival Probability (2016) — contributed 1 concept elements to this direction.
Sequential advantage selection for optimal treatment regime (2016) — contributed 1 concept elements to this direction.
A Bayesian prediction model between a biomarker and the clinical endpoint for dichotomous variables (2014) — contributed 1 concept elements to this direction.