Modified Vector-Valued Treatment Policy

• Modified Vector-Valued Treatment Policy (MVTP) is a causal framework where each unit’s observed multivariate treatment vector is modified through a deterministic map rather than applying a fixed treatment for all.
• It relaxes positivity requirements by comparing natural treatments with their policy-induced modifications, thus enabling robust causal effect estimation in multivariate and time-varying settings.
• Applications in mechanical ventilation illustrate how targeted modifications of ventilator settings can lead to improved patient outcomes by aligning interventions with realistic clinical practices.

Searching arXiv for the cited papers and related MVTP/LMTP work. Modified Vector-Valued Treatment Policy (MVTP) denotes an intervention framework in which treatment is represented as a vector $A \in \mathbb{R}^k$, and the intervention acts by modifying each unit’s observed, or “natural,” treatment vector through a deterministic map $q : (\mathcal X,\mathcal A)\to\mathcal A$, rather than assigning a single fixed treatment vector to the entire population. In explicit usage, the term appears in work on mechanical ventilation, where multiple continuous ventilator settings are shifted jointly and the target causal estimand is $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$ (Jiang et al., 13 Jul 2025). Closely related work on modified treatment policies (MTPs), longitudinal modified treatment policies (LMTPs), and structural nested mean models (SNMMs) shows that this vector-valued formulation fits naturally within a broader causal framework for realistic interventions on multivariate and time-varying treatments (Díaz et al., 22 May 2026).

1. Formal definition and conceptual scope

In the one-stage formulation, the observed data are $\{(X_i,A_i,Y_i)\}_{i=1}^n$, where $X_i \in \mathcal X \subseteq \mathbb R^p$, $A_i \in \mathcal A \subseteq \mathbb R^k$, and $Y_i \in \mathbb R$ (Jiang et al., 13 Jul 2025). A Modified Vector-Valued Treatment Policy is a deterministic map

$$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$

so that the modified treatment for unit $i$ is $q(X_i,A_i)$, and the corresponding mean potential outcome is

$q : (\mathcal X,\mathcal A)\to\mathcal A$0

The principal causal contrast is often $q : (\mathcal X,\mathcal A)\to\mathcal A$1, which is a risk difference when $q : (\mathcal X,\mathcal A)\to\mathcal A$2 is binary (Jiang et al., 13 Jul 2025).

This formulation differs from the average dose response function in a specific way. Instead of asking what would happen if everyone were assigned the same fixed vector $q : (\mathcal X,\mathcal A)\to\mathcal A$3, MVTPs ask what would happen if each unit’s observed treatment vector were modified according to a rule $q : (\mathcal X,\mathcal A)\to\mathcal A$4. In the scalar MTP literature, this modification-based perspective was introduced precisely because interventions on observed natural treatment values are often more realistic than population-wide deterministic assignments, and because they relax the most stringent positivity requirements (Jiang et al., 2023). The vector-valued generalization preserves that logic: treatment is modified jointly, but still relative to what was actually observed.

The term “vector-valued” has two distinct but related senses in the literature. In causal inference, it refers to a multi-component treatment vector such as a ventilator setting vector or a multivariate dose (Jiang et al., 13 Jul 2025). In sequential decision-making, related work uses vector-valued or matrix-valued objects to represent multiple admissible actions or multiple value functions in dynamic treatment regimes (Yazzourh et al., 19 Mar 2026). This suggests that MVTP is best viewed as a causal-intervention framework first, with downstream links to richer policy representations.

2. Identification, support, and the target distribution

The causal interpretation of $q : (\mathcal X,\mathcal A)\to\mathcal A$5 relies on consistency, a support condition tailored to the policy, and conditional exchangeability of related populations. In the MVTP formulation for vector-valued ventilator settings, the assumptions are stated as follows. Consistency requires that if $q : (\mathcal X,\mathcal A)\to\mathcal A$6, then $q : (\mathcal X,\mathcal A)\to\mathcal A$7. Positivity requires

$q : (\mathcal X,\mathcal A)\to\mathcal A$8

Conditional exchangeability of related populations requires that for each $q : (\mathcal X,\mathcal A)\to\mathcal A$9, letting $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$0, the conditional laws of $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$1 and $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$2 coincide (Jiang et al., 13 Jul 2025).

The support condition is weaker than the positivity condition required for a multivariate average dose response function. It only requires support along the policy-induced mapping $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$3, not over all treatment vectors. This is especially important for multivariate treatments with infeasible or rarely observed combinations, such as ventilator settings constrained by clinical guidelines (Jiang et al., 13 Jul 2025). The same support logic appears in the scalar MTP literature, where the modified treatment must remain in an enforceable set and where the target population under the policy is characterized through the hypothetical distribution of $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$4 (Jiang et al., 2023).

Under block-wise smooth invertibility of $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$5, the mean policy value can be written as an expectation under a policy-induced joint distribution $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$6. In the MVTP formulation,

$\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$7

where $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$8 (Jiang et al., 13 Jul 2025). The scalar MTP work makes the same point in a complementary way: the fundamental estimation task is to approximate expectations under the target distribution of $\mu^q \equiv \mathbb{E}[Y^{q(X,A)}]$9 using the observed sample from $\{(X_i,A_i,Y_i)\}_{i=1}^n$0 (Jiang et al., 2023). This target-distribution view is central because it turns MVTP estimation into a distributional transport problem rather than a generalized propensity score problem.

3. Estimation with balancing weights and augmented estimators

A prominent estimation strategy for MVTPs is based on energy balancing weights. In the ventilator paper, the MVTP problem is reduced to a binary-treatment Average Treatment effect on the Treated setup by constructing an augmented population in which $\{(X_i,A_i,Y_i)\}_{i=1}^n$1 corresponds to the observed $\{(X_i,A_i,Y_i)\}_{i=1}^n$2 distribution and $\{(X_i,A_i,Y_i)\}_{i=1}^n$3 corresponds to the policy-induced $\{(X_i,A_i,Y_i)\}_{i=1}^n$4 distribution (Jiang et al., 13 Jul 2025). The resulting ATT equals $\{(X_i,A_i,Y_i)\}_{i=1}^n$5, so weighting methods for binary treatments can be repurposed for MVTP estimation.

The scalar MTP paper formalizes the same idea through the target distribution $\{(X_i,A_i,Y_i)\}_{i=1}^n$6 and an error decomposition: $\{(X_i,A_i,Y_i)\}_{i=1}^n$7 This identifies the first term as the component controlled by balancing weights (Jiang et al., 2023).

The imbalance metric used there is the weighted energy distance between the weighted empirical distribution of $\{(X_i,A_i,Y_i)\}_{i=1}^n$8 and the empirical distribution of $\{(X_i,A_i,Y_i)\}_{i=1}^n$9. It has both a characteristic-function representation and a bias bound: when the weighted energy distance is small, the confounding bias term is small for sufficiently smooth $X_i \in \mathcal X \subseteq \mathbb R^p$0 (Jiang et al., 2023). Penalized energy balancing weights are then obtained by minimizing the empirical energy distance plus a quadratic penalty on the weights. In the MVTP ventilator study, a kernelized version with a Gaussian kernel performed best in simulations, and augmented estimators combined these weights with outcome regression fitted by SuperLearner (Jiang et al., 13 Jul 2025).

The doubly robust structure follows the usual augmented form. The scalar MTP work gives

$X_i \in \mathcal X \subseteq \mathbb R^p$1

and establishes asymptotic guarantees under regularity conditions (Jiang et al., 2023). In the MVTP setting, the same form applies with vector-valued $X_i \in \mathcal X \subseteq \mathbb R^p$2, and the practical nuisance-estimation problem becomes the modeling of $X_i \in \mathcal X \subseteq \mathbb R^p$3 and the construction of weights that align the observed and policy-induced joint distributions.

Sensitivity analysis is also built around the augmented binary-treatment representation. In the ventilator application, a marginal sensitivity model indexed by $X_i \in \mathcal X \subseteq \mathbb R^p$4 bounds the odds-ratio discrepancy between the observed pseudo-treatment propensity and the unobserved propensity that could depend on potential outcomes, and this is linked to bounds on ideal versus observed weights (Jiang et al., 13 Jul 2025). This provides a direct way to ask how strong unmeasured confounding would have to be to overturn an estimated MVTP effect.

4. Longitudinal and history-dependent extensions

The one-stage MVTP sits inside a broader LMTP framework for time-varying interventions. In the longitudinal formulation, the observed data are

$X_i \in \mathcal X \subseteq \mathbb R^p$5

with $X_i \in \mathcal X \subseteq \mathbb R^p$6, and an LMTP is defined recursively by

$X_i \in \mathcal X \subseteq \mathbb R^p$7

so that each treatment at time $X_i \in \mathcal X \subseteq \mathbb R^p$8 modifies the natural value that would have arisen under the policy up to time $X_i \in \mathcal X \subseteq \mathbb R^p$9 (Díaz et al., 2020). That paper already states that $A_i \in \mathcal A \subseteq \mathbb R^k$0 denotes “a vector of intervention variables such as treatment and/or censoring status,” so the longitudinal machinery is already compatible with vector-valued treatments (Díaz et al., 2020).

More recent work extends LMTPs to depend on the natural history of treatment rather than only the current natural value. The generalized formulation is

$A_i \in \mathcal A \subseteq \mathbb R^k$1

where $A_i \in \mathcal A \subseteq \mathbb R^k$2 is the history of natural treatments under the intervention (Díaz et al., 22 May 2026). This extension is needed for delay interventions and grace-period policies, because once a treatment has been overwritten by an intervention, a current-time-only rule no longer has access to the original natural value. The paper develops augmented-data sequential regression, efficient influence functions, TMLE, and SDR estimators for this setting, with $A_i \in \mathcal A \subseteq \mathbb R^k$3 inference under doubly robust rate conditions (Díaz et al., 22 May 2026).

Structural Nested Mean Models provide an additional layer for characterizing heterogeneity. In that framework, treatment at time $A_i \in \mathcal A \subseteq \mathbb R^k$4 is explicitly allowed to be “possibly multidimensional with discrete and/or continuous components,” and the regime-specific blip function is

$A_i \in \mathcal A \subseteq \mathbb R^k$5

This allows time- and history-specific heterogeneous MTP effects to be parameterized under exchangeability or under parallel trends assumptions (Shahn, 26 Sep 2025). For MVTPs, this means that vector-valued treatment modifications can be embedded in semiparametric blip models, rather than only summarized through a marginal $A_i \in \mathcal A \subseteq \mathbb R^k$6.

5. Mechanical ventilation as the canonical MVTP application

The term MVTP is used explicitly in the study of mechanical ventilation with observational ICU data from MIMIC-III (Jiang et al., 13 Jul 2025). There, $A_i \in \mathcal A \subseteq \mathbb R^k$7 comprises 97 pre-treatment variables from the first 24 hours of ICU stay, $A_i \in \mathcal A \subseteq \mathbb R^k$8 is a vector of ventilator settings from the second 24 hours, and $A_i \in \mathcal A \subseteq \mathbb R^k$9 is in-hospital mortality. The framework is designed to address infeasible treatment combinations, stringent positivity assumptions, and interpretability concerns in settings with multiple continuous ventilator parameters (Jiang et al., 13 Jul 2025).

One treatment vector is

$Y_i \in \mathbb R$0

and the derived scalar mechanical power is

$Y_i \in \mathbb R$1

Two policies were contrasted. The first scales tidal volume only,

$Y_i \in \mathbb R$2

and the second scales airway pressures only,

$Y_i \in \mathbb R$3

Both induce the same proportional change in total mechanical power, but through different coordinates of the treatment vector (Jiang et al., 13 Jul 2025).

The empirical finding was that lowering airway pressures may yield greater reductions in patient mortality compared to proportional adjustments of tidal volume alone (Jiang et al., 13 Jul 2025). In a second analysis,

$Y_i \in \mathbb R$4

and the policy

$Y_i \in \mathbb R$5

preserved respiratory-system compliance and minute ventilation while modifying driving pressure. In that setting, controlling for respiratory-system compliance and minute ventilation, there was a significant benefit of reducing driving pressure in patients with acute respiratory distress syndrome (Jiang et al., 13 Jul 2025).

These applications make the interpretability of MVTP especially clear. The policies are stated as clinically recognizable modifications of actual practice, not as unsupported assignments to a universal treatment vector. They also illustrate a recurring feature of MVTP analyses: the policy family is typically indexed by a scalar $Y_i \in \mathbb R$6, and estimation is combined with support diagnostics to identify ranges of $Y_i \in \mathbb R$7 for which balancing remains credible (Jiang et al., 13 Jul 2025).

6. Relation to adjacent policy-learning frameworks and current limitations

MVTP should be distinguished from several neighboring ideas. In set-valued policy learning, a policy outputs a set $Y_i \in \mathbb R$8 of plausible treatments rather than a modified natural treatment vector, and the key coverage requirement is

$Y_i \in \mathbb R$9

That framework supplies uncertainty-aware treatment sets and can encode the set $$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$0 as a binary vector in $$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$1, but its primary target is coverage of optimal actions, not the causal effect of a modification map $$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$2 (Fuentes-Vicente et al., 19 May 2026). In near-equivalent Q-learning for dynamic treatment regimes, the output is a family of $$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$3-optimal policies obtained from action sets satisfying

$$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$4

which yields matrix-valued backward recursion and regions of treatment indifference rather than modified natural-treatment interventions (Yazzourh et al., 19 Mar 2026). These are conceptually adjacent, but not the same estimand.

There is also a broader decision-theoretic backdrop in finite-horizon Markov decision process formulations for medical treatment. One example models multi-modality cancer therapy with a vector-valued clinical state $$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$5 for toxicity and tumor progression, scalarized reward

$$q : (\mathcal X,\mathcal A)\to\mathcal A,\qquad q(x,a)\in\mathcal A,$$6

and constrained modality histories (Maass et al., 2017). That line of work formalizes sequential treatment optimization over vector-valued states, whereas MVTP formalizes causal effects of modifying natural treatments. The connection is structural rather than definitional.

Current limitations are consistent across the literature. Exchangeability assumptions remain strong. In the ventilator application, sensitivity analysis addresses unmeasured confounding, but not other biases such as measurement error or outcome-model misspecification (Jiang et al., 13 Jul 2025). Large policy shifts can violate support, and energy-balancing diagnostics are then intended to flag unreliable regions rather than rescue identifiability (Jiang et al., 13 Jul 2025). In longitudinal settings with natural-history dependence, explicit augmentation over treatment histories is computationally demanding, particularly for continuous or multivariate treatments, and further work is needed on density-ratio or Riesz-representer methods in those cases (Díaz et al., 22 May 2026). Functional-treatment extensions reinforce the same point from another direction: once treatment leaves finite-dimensional Euclidean space, careful definitions of population averages and stronger regularity conditions become necessary (Jiang et al., 9 Feb 2026).

Taken together, these developments place MVTP at the intersection of realistic intervention design, multivariate causal identification, and semiparametric estimation. Its defining feature is not merely that treatment is vector-valued, but that the intervention modifies the observed treatment vector in a support-aware, policy-interpretable way.