Individualized Treatment Regimes

• Individualized treatment regimes (ITRs) are decision rules that map patient-specific covariates to treatment assignments to maximize expected clinical outcomes.
• Researchers employ various methodologies—such as regression-based plug‐in methods, outcome weighted learning, and Bayesian approaches—to derive and validate these rules.
• Recent work expands ITRs through transfer learning, multi-outcome optimization, and emphasis on interpretability to ensure clinical usability and external validity.

Individualized treatment regimes (ITRs), also called individualized treatment rules, are decision rules that map patient-specific covariates to treatment assignments so as to optimize an expected outcome or utility. In the single-stage case, an ITR is the simplest special case of a dynamic treatment regime, but the contemporary literature treats the object much more broadly: value-based policy learning, semiparametric blip modeling, Bayesian decision rules, transfer learning under covariate shift, and extensions to censored survival, competing risks, combination therapies, and strategically reported preferences all fall within the ITR framework (Wang et al., 3 Jan 2025, Dong et al., 2021, Zhao et al., 2023).

1. Formal definition and causal structure

An ITR is typically written as a mapping

$d:\mathcal{X}\to\mathcal{A},$

where $X\in\mathcal{X}\subset\mathbb{R}^p$ denotes baseline covariates and $\mathcal{A}$ is the treatment space. The literature represented here uses several common treatment codings: $\{0,1\}$, $\{-1,+1\}$, and finite multi-arm sets. In all cases, the decision problem is defined through potential outcomes. With binary treatment, one writes $Y^0$ and $Y^1$, or $Y^*(-1)$ and $Y^*(1)$, and evaluates a regime by the value

$V(d)=\mathbb{E}\big[Y^{d(X)}\big].$

Under the usual causal conditions—consistency, SUTVA, conditional exchangeability or no unmeasured confounding, and positivity—the optimal rule is the treatment assignment that maximizes the relevant conditional mean outcome for each covariate profile (Logan et al., 2017, Bian et al., 2022).

A standard representation separates the outcome surface into a treatment-free component and a treatment contrast. For discrete outcomes, one semiparametric specification is

$X\in\mathcal{X}\subset\mathbb{R}^p$0

where $X\in\mathcal{X}\subset\mathbb{R}^p$1 is a nuisance baseline term and $X\in\mathcal{X}\subset\mathbb{R}^p$2 is the blip function. Because $X\in\mathcal{X}\subset\mathbb{R}^p$3 does not depend on treatment, treatment choice is driven by $X\in\mathcal{X}\subset\mathbb{R}^p$4. With binary treatment and an increasing link, the optimal ITR reduces to

$X\in\mathcal{X}\subset\mathbb{R}^p$5

or, in linear blip form, $X\in\mathcal{X}\subset\mathbb{R}^p$6 (Bian et al., 2022).

A major design axis is the restriction placed on the class of admissible rules. Several papers emphasize interpretable linear regimes of the form

$X\in\mathcal{X}\subset\mathbb{R}^p$7

or, for identifiability in a survival-transfer setting,

$X\in\mathcal{X}\subset\mathbb{R}^p$8

These classes deliberately trade unrestricted optimality for transparency, clinical usability, and stable optimization (Wang et al., 3 Jan 2025, Zhao et al., 2023).

2. Principal estimation paradigms

A first major family is regression-based estimation. In these methods, one models either the full conditional mean $X\in\mathcal{X}\subset\mathbb{R}^p$9 or the blip $\mathcal{A}$0, and then derives the rule by plug-in maximization. This includes Q-learning, A-learning, and dynamic weighted least squares variants. In the generalized dynamic weighted ordinary least squares framework, the outcome model is decomposed as

$\mathcal{A}$1

and the blip parameters defining the rule are estimated with balancing weights satisfying a generalized propensity-score balancing property. The appeal of this class is that the estimated $\mathcal{A}$2 directly induces a treatment rule and, under the stated conditions, enjoys double robustness with respect to the treatment model and the treatment-free outcome model (Dong et al., 2021).

A second family is direct value optimization by weighted classification. Outcome Weighted Learning formulates ITR estimation as minimizing a weighted classification loss, using the clinical outcome as the weight. Residual Weighted Learning modifies this by weighting misclassification errors with residuals from an outcome regression on clinical covariates excluding treatment assignment, rather than with the raw outcomes. The method uses a smoothed ramp loss and a difference-of-convex algorithm, and it extends to continuous, binary, and count outcomes. The residualization step is designed to reduce finite-sample instability and to focus the optimization on treatment-related heterogeneity rather than on large baseline components unrelated to treatment choice (Zhou et al., 2015).

A third family combines doubly robust estimation with penalization for tailoring-variable discovery. For discrete outcomes, a penalized doubly robust framework re-expresses the estimating equations as iteratively reweighted generalized linear models. The construction targets the blip parameters, enforces strong heredity through modified adaptive weights, and yields variable selection with oracle-type properties under the stated conditions. This literature distinguishes tailoring variables—covariates whose effects depend on treatment—from variables that only enter the baseline mean or are irrelevant to treatment assignment (Bian et al., 2022).

A fourth family is Bayesian prediction-based policy learning. In the BART framework, one models the conditional outcome distribution $\mathcal{A}$3, computes posterior predictive utilities for each treatment, and assigns the treatment maximizing posterior expected utility. The same posterior draws support uncertainty quantification for patient-specific treatment differences, for the value of the implemented ITR, and even for the value of the Optimal ITR. The literature also describes a “Fit-the-fit” strategy, in which a simple tree is fitted to the BART-estimated treatment differences to obtain a clinically interpretable approximation of a black-box rule (Logan et al., 2017).

3. Transfer learning and target-population optimality

A central development in recent ITR work is the distinction between a source population, where treatment and outcome are observed, and a target population, where the covariate distribution differs and one still wants a target-optimal rule. This problem is explicit in transfer learning from randomized experiments to real-world data and in source-target ICU data fusion. The core issue is covariate shift: $\mathcal{A}$4 so a rule that is optimal under the source covariate distribution may be suboptimal for the target population even if treatment effect mechanisms are transportable (Wu et al., 2021, Wang et al., 3 Jan 2025).

The standard remedy is weighting or calibration. One line of work uses sampling-score weights or balancing weights to calibrate the experimental covariate distribution to the real-world distribution, and then learns the ITR under the reweighted trial sample. Another line uses entropy balancing to construct calibration weights $\mathcal{A}$5 so that weighted moments in the source align with moments in the target. In the sepsis transfer-learning framework, these weights are combined with augmented inverse probability weighting to define a calibrated AIPW value estimator, and the optimal linear ITR is the maximizer of that estimated target-population value (Wu et al., 2021, Wang et al., 3 Jan 2025).

In right-censored survival transfer learning, the value function is generalized to survival functionals such as survival probability at time $\mathcal{A}$6 and restricted mean survival time. The relevant nuisance structure includes the source propensity score, sampling score, censoring survival, outcome regression, and an additional conditional survival-based nuisance $\mathcal{A}$7. The efficient influence function yields a doubly robust estimator of $\mathcal{A}$8, and the optimal rule is learned by maximizing the resulting estimated value over a parametric ITR class. This framework establishes an $\mathcal{A}$9 rate of convergence for the parameter indexing the optimal regime, while the estimated optimal value itself is consistent and asymptotically normal, including under cross-fitting with flexible machine learning nuisance estimators (Zhao et al., 2023).

These transfer formulations formalize a point often obscured in homogeneous-population ITR work: external validity is not a post hoc evaluation issue but part of the decision target itself. A plausible implication is that the “optimal ITR” is always relative to a specified population distribution, not merely to a conditional mean model.

4. Nonstandard outcomes and observation schemes

Much of the recent literature is organized around departures from the canonical setting of a continuous uncensored outcome measured once per subject. For discrete outcomes, the main complication is that log and logit links produce nonlinear estimating equations, and direct penalization of classical A-learning equations becomes cumbersome. The proposed solution is to build doubly robust estimating functions that can be implemented as weighted Poisson or weighted logistic GLMs via IRGLM, preserving both computational tractability and variable-selection structure (Bian et al., 2022).

For survival data, censoring changes both the estimand and the inferential machinery. In the transfer-learning survival framework, the value function accommodates $\{0,1\}$0 and $\{0,1\}$1, corresponding respectively to survival probability and RMST. Identification requires conditionally independent censoring, and the efficient-influence-function estimator augments IPW/IPCW terms with outcome regression and transport corrections. This brings survival ITRs into a semiparametric efficiency and orthogonality framework more often associated with modern causal survival analysis than with classical regime estimation (Zhao et al., 2023).

Crossover designs create a different opportunity. In a $\{0,1\}$2 crossover, each subject receives both treatments, so the reward for policy learning can be the within-subject treatment response difference

$\{0,1\}$3

after correcting for carryover or period effects. The resulting crossover GOWL objective is a weighted classification problem in which $\{0,1\}$4 weights misclassification and the sign of $\{0,1\}$5 determines which treatment label is correct. The paper establishes Fisher consistency and global consistency for this construction, showing that crossover designs can be especially informative for precision medicine because they supply within-subject counterfactual information (Nguyen et al., 2019).

Competing risks and correlated data require yet another reformulation. In that setting, the target rule may average over cause-specific treatment effects using $\{0,1\}$6, yielding a weighted regime, or may optimize only with respect to the most likely cause, yielding a greedy regime. The corresponding blip-based benefit function is estimated by weighted AFT-GEE, which extends dynamic weighted survival modeling to clustered data. The method is doubly robust in the stated sense—consistent if either the treatment-free outcome model or the weighting models are correct—and explicitly incorporates uncertainty about the unobserved cause of failure at decision time (Dolmatov et al., 26 Sep 2025).

5. Extensions beyond the standard binary single-outcome setting

Recent work pushes ITRs beyond the single binary treatment and single primary outcome template. For combination therapies, treatment is represented as a binary vector $\{0,1\}$7, and the central modeling problem is the exponential growth of feasible treatment combinations. The Double Encoder Model addresses this with a nonparametric representation

$\{0,1\}$8

where $\{0,1\}$9 is a covariate encoder and $\{-1,+1\}$0 is a treatment encoder with additive and interaction components. Under budget constraints, the individualized assignment problem becomes a multi-choice knapsack problem over estimated subject-treatment values and treatment costs (Xu et al., 2023).

When multiple outcomes matter, one line of work keeps the primary outcome as the optimization target but incorporates secondary outcomes through agreement with their optimal rules. The proposed fusion penalty encourages the ITR for the primary outcome to agree with estimated secondary-outcome ITRs, using disagreement indicators or their surrogate losses. The theory shows that the agreement rate between the estimated primary-outcome ITR and the optimal secondary-outcome rules converges faster than if the secondary outcomes are ignored, while non-asymptotic properties are derived for both the value function and the misclassification rate (Gao et al., 2024).

Preference-sensitive treatment assignment introduces an entirely different difficulty: the key covariate may be strategically manipulated. In the binary-preference setting, the treatment rule is a map

$\{-1,+1\}$1

where $\{-1,+1\}$2 is the probability of treatment 1 for true preference type $\{-1,+1\}$3. If individuals know the rule in advance, they may misreport their preferences. The relevant contribution here is that the welfare-maximizing ITR under strategic reporting is strategy-proof, in the sense that individuals do not have a strong incentive to lie even if they know the optimal ITR a priori. Identification then depends on specialized designs such as strictly strategy-proof randomized controlled trials and doubly randomized preference trials, and the resulting statistical treatment rules have maximum regret converging to zero at a square-root rate (Kido, 2023).

These extensions show that the formal core of an ITR—mapping covariates to treatment—remains intact even when the treatment space becomes combinatorial, the outcome vector becomes multivariate, or the covariates become endogenous. What changes is the definition of the relevant value function and the structure of the admissible rule class.

6. Interpretability, dimension reduction, and local explanation

Interpretability is not a peripheral desideratum in this literature; it often determines the rule class itself. Linear threshold rules remain a dominant template because they are clinically legible and easy to operationalize. Variable-selection methods sharpen this by distinguishing prognostic variables from tailoring variables and by enforcing strong heredity, so that treatment interactions are not retained without their associated main effects. In the discrete-outcome setting, modified adaptive lasso weights are designed specifically to retain that structure (Bian et al., 2022).

A complementary strategy is dimension reduction targeted at treatment effect heterogeneity rather than at outcome prediction per se. In high-dimensional observational studies, one recent approach assumes a central mean subspace $\{-1,+1\}$4 such that

$\{-1,+1\}$5

This reduced representation is estimated by gradient kernel dimension reduction on a pseudo-outcome tied to the treatment contrast, while kernel-based covariate balancing is performed in the full covariate space. The final ITR is then learned by augmented outcome-weighted learning in the reduced space, and the method is shown to achieve universal consistency under the stated conditions (Son et al., 11 Jan 2026).

Another response to the interpretability-flexibility tension is to keep the global outcome learner fully flexible but explain its treatment decision locally. The Locally Interpretable Individualized Treatment Rule method trains a black-box model for the outcome surface, uses a $\{-1,+1\}$6-VAE to generate realistic local synthetic covariates, and fits a mixture of interpretable experts to approximate the black-box response locally around each patient. The resulting rule is not a single global scoring function but a subject-specific local linear rule, with local coefficients that directly quantify how prognostic and prescriptive covariates drive the treatment recommendation (Charvadeh et al., 12 Feb 2026).

Bayesian “fit-the-fit” procedures occupy a middle ground: the primary estimator can be a flexible BART model, but the final deployed rule may be a shallow tree fitted to the BART-estimated treatment differences. This suggests that interpretability in ITR research is best viewed not as a single model class but as a design principle that can be enforced globally, sparsely, dimensionally, or locally (Logan et al., 2017).

7. Empirical domains, assumptions, and recurring points of dispute

The empirical range of ITR methodology is broad. Recent applications include mechanical ventilation decisions for sepsis patients using transfer learning between MIMIC-III and eICU-CRD, sodium bicarbonate therapy for severe metabolic acidaemia using RCT and observational ICU data, IL-7 administration strategies in HIV, crossover-based dietary interventions in adolescents, conditioning regimen choice in hematopoietic cell transplantation, optimal drug combinations in colorectal cancer PDX data, preference-based assignment to vocabulary versus mathematics training, transthoracic echocardiography in ICU sepsis, depression treatment rules fused across multiple outcomes, HCV-positive kidney allocation, and precision side-effect management in breast cancer (Wang et al., 3 Jan 2025, Zhao et al., 2023, Dong et al., 2021, Nguyen et al., 2019, Logan et al., 2017, Xu et al., 2023, Kido, 2023, Son et al., 11 Jan 2026, Gao et al., 2024, Dolmatov et al., 26 Sep 2025, Charvadeh et al., 12 Feb 2026).

These applications also clarify the assumptions under which ITRs are identified and transportable. Across papers, the recurrent conditions are SUTVA or consistency, no unmeasured confounding or conditional exchangeability, positivity, and, when relevant, transportability and common support for source-target learning, conditionally independent censoring for survival outcomes, exclusion restrictions for preference trials, and adequate carryover control in crossover designs. Many methodological differences are downstream of which of these assumptions are plausible in a given design rather than of superficial differences in optimization (Zhao et al., 2023, Wang et al., 3 Jan 2025, Kido, 2023, Nguyen et al., 2019).

Several recurrent misconceptions are addressed implicitly by this literature. One is that an ITR is simply a prediction model for $\{-1,+1\}$7; in fact, most frameworks target either the value function or the treatment contrast, and nuisance components that improve outcome prediction may still be irrelevant for treatment choice. Another is that treatment rules estimated in randomized trials automatically generalize to practice; transfer-learning work shows that external validity can fail under covariate shift even when internal validity is strong. A third is that interpretability and flexibility are mutually exclusive; the literature instead contains linear blip models, sparse penalized rules, tree approximations to Bayesian policies, and subject-specific local explanations for black-box models. This suggests that the field’s main fault line is not interpretability versus accuracy, but rather how aggressively one restricts the rule class while preserving the decision target.

In its current form, the ITR literature is therefore best understood as a family of decision-theoretic and causal-learning frameworks centered on the same primitive object—a covariate-indexed treatment rule—but adapted to increasingly complex outcome types, sampling schemes, and deployment constraints. The unifying objective is stable across these variants: optimize clinically relevant potential-outcome functionals for the population of interest while making the rule estimable, transportable, and, when necessary, interpretable.