Causal inference operates on a hierarchy of increasingly difficult questions. At the baseline lies observation (What does a symptom tell us about a disease?). Above that sits intervention (What will happen if we administer a drug?). At the summit lies counterfactual reasoning (Would the patient have survived if we had not administered the drug?).
According to Pearl’s Causal Hierarchy, answering counterfactual questions generally requires a fully specified Structural Causal Model (SCM). Observational and interventional data are usually insufficient because multiple distinct SCMs can generate identical statistical distributions while predicting completely different counterfactual outcomes. This is the fundamental impasse of counterfactual identification.
A rapidly growing body of research resolves this impasse by leveraging monotonicity constraints. By assuming that causal relationships move in a consistent direction, researchers can collapse the latent ambiguity of alternative worlds, mapping population-level observations directly to individual-level counterfactuals.
The Intuition: Collapsing Latent Ambiguity
To understand why monotonicity unlocks counterfactuals, consider a binary treatment (a drug) and a binary outcome (survival). In a standard population, an individual could belong to one of four latent strata:
- Compliers/Respondents: They survive only if they take the drug.
- Defiers: They survive only if they do not take the drug.
- Always-Survivors: They survive regardless.
- Never-Survivors: They die regardless.
If we only observe population statistics, we cannot distinguish a population composed entirely of Always-Survivors and Never-Survivors from a population where Compliers and Defiers perfectly cancel each other out. Because we cannot isolate these latent groups, individual counterfactual probabilities (such as the Probability of Necessity and Sufficiency) are unidentifiable and can only be bounded.
Monotonicity changes this by asserting a directional constraint: the treatment can never harm an individual. This simple assumption completely eliminates the "Defier" stratum.
By ruling out contradictory individual responses, the mathematical mapping between observational data and latent counterfactual states becomes injective. As Kawakami et al. (2025) demonstrate, this directional consistency allows researchers to precisely calculate individual-level causal variance and probabilities of causation directly from population-level cumulative distribution functions (CDFs).
Non-Crossing Trajectories and Rank Invariance
When moving from binary variables to continuous or high-dimensional spaces, monotonicity manifests as a non-crossing potential outcome trajectories assumption.
Imagine plotting how different individuals respond to varying dosages of a treatment. Without constraints, these individual response curves can crisscross arbitrarily. An individual who responds exceptionally well to a low dose might respond poorly to a high dose. This chaotic crossing is precisely what makes continuous counterfactuals unidentifiable.
Unconstrained Trajectories (Unidentifiable) Monotonic / Non-Crossing (Identifiable)
Outcome Outcome
^ ^
| /---\ | /--- Individual A
| / \ Individual A | /
| /-------\----------------> | /---- Individual B
|/ \ Individual B | /
+------------------------> +------------------------>
Dosage Dosage
Kawakami et al. (2024) formalize a generalized monotonicity framework where these individual trajectories are constrained never to intersect. If Individual A responds more favorably than Individual B at one treatment level, they must respond more favorably across all treatment levels.
This non-crossing property guarantees rank invariance: an individual’s relative position (quantile) within the population distribution remains constant across all hypothetical treatment states.
This insight bridges the gap between deep learning and causal theory. If rank is invariant across counterfactual worlds, we can determine an individual's counterfactual outcome simply by mapping their observed outcome to its corresponding quantile in the target treatment's conditional distribution. Xie et al. (2023) leverage this mechanism to prove that counterfactual outcomes can be uniquely identified via quantile regression, bypassing the need to ever recover the exact mathematical equations of the true SCM.
Triangular Monotonic Flows
In machine learning, this theoretical principle is operationalized by embedding monotonicity directly into generative architectures, such as Normalizing Flows and SCMs.
To uniquely reconstruct a counterfactual, a model must take an observed state, invert the causal mechanism to isolate the exact historical exogenous noise ($\epsilon$) for that specific individual, and then run the mechanism forward under a new intervention. For arbitrary functions, this inversion is impossible.
To solve this, Le et al. (2025) and Chen et al. (2025) introduce architectures constrained by a triangular monotonically increasing structure.
Exogenous Noise (ε) --------[ Bijective / Monotonic ]--------> Observed Data (X)
^ |
|-------------------[ Unique Inversion ]-----------------------|
These models enforce two structural priors:
- Causal Ordering (Triangular Jacobian): Variables are arranged in a strict topological sequence where each variable depends only on its causal predecessors and its own exogenous noise.
- Monotonicity: The function mapping the exogenous noise to the variable is strictly increasing.
When these two conditions are met, the mapping from noise to data becomes a bijection. Given a known causal ordering, the model can uniquely disentangle the underlying exogenous noise from observational data without requiring experimental interventions. This architectural guarantee enables deep twin networks and causal flows to perform exact counterfactual reasoning at scale, even on categorical variables (Vlontzos et al. (2022)) and high-dimensional vector fields.
Applications and Bounding
Monotonicity constraints are not merely computational conveniences; they are deeply rooted in domain knowledge. As Imbens (2019) notes, economic and behavioral theories frequently dictate shape restrictions—such as diminishing marginal utility or non-negative treatment responses—that are naturally expressed as monotonicity constraints.
When applied to complex causal graphs, these constraints act as mathematical levers:
- Instrumental Variables: In instrumental variable settings, Imbens (2014) highlights how the monotonicity principle ensures that an instrument uniformly affects treatment selection, enabling the identification of Local Average Treatment Effects (LATE).
- Unobserved Confounding: When unobserved confounders threaten internal validity, Jun et al. (2020) show that combining monotone treatment response with monotone treatment selection yields sharp, informative upper bounds on causal risk, eliminating the need for the often-unrealistic assumption of complete unconfoundedness.
- Mediation and Truncation: In mediation analysis, assuming that a treatment moves intermediate mediators in a single direction allows researchers to decompose path-specific probabilities of causation (Kawakami et al. (2025)) and resolve truncation-by-death challenges in clinical trials (Wang et al. (2015)).
By formalizing the intuitive notion that causes act in predictable directions, monotonicity transforms counterfactual inference from an intractable philosophical problem into an algorithmically solvable optimization task.
Key Foundational Work
On the Identification of a Class of Linear Models (2007) — contributed 1 concept elements to this direction.
Sensitivity Analyses Comparing Outcomes Only Existing in a Subset Selected Post‐Randomization, Conditional on Covariates, with Application to HIV Vaccine Trials (2005) — contributed 1 concept elements to this direction.