Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Uplift modeling with continuous treatments: A predict-then-optimize approach

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper extends uplift modeling to continuous treatment doses: estimate conditional average dose responses, then solve a constrained dose-allocation problem as an integer linear program.

desk verdict Useful formalization of continuous-treatment uplift with a flexible ILP allocation layer, but the experimental validation has an unflagged overlap violation and thin support. read the letter →

arxiv 2412.09232 v2 pith:L5DD2CO5 submitted 2024-12-12 cs.LG

Simon De Vos , Christopher Bockel-Rickermann , Stefan Lessmann , Wouter Verbeke This is my paper · ORCID
classification cs.LG
keywords upliftmodelingcontinuoustreatmentsconditionalaveragedoseresponseintegerlinearprogrammingcausalmachinelearningallocationfairnessconstraintspredict-then-optimize
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Uplift modeling normally assumes a binary treatment decision: treat or do not treat. This paper extends the predict-then-optimize recipe to continuous treatment doses, such as a medication dose or training hours, by splitting the task into an inference step that estimates conditional average dose response (CADR) curves with off-the-shelf causal machine learning, and an optimization step that solves the dose-allocation problem as an integer linear program. The ILP formulation makes it straightforward to add constraints such as budget and group fairness, or to change the objective to incorporate instance-dependent costs and benefits, all without retraining the prediction model. Experiments on semi-synthetic IHDP data show that the framework outperforms a greedy ranking heuristic at partial budgets, and that the most accurate dose-response estimator is not always the one that yields the best treatment allocations. The paper's broader claim is that uplift modeling should be defined as constrained treatment allocation, not merely treatment-effect estimation.

What carries the argument

The central object is the conditional average dose response (CADR) function $\mu(s, x)$ and its derived dose effect CADE $\tau_s(x) = \mu(s, x) - \mu(0, x)$. The paper discretizes the dose interval $[0,1]$ into $\delta$ bins, forming a vector of CADE estimates per entity, and feeds these into an integer linear program that chooses one dose per entity to maximize total policy value under a budget constraint, with optional fairness constraints (disparate treatment and disparate outcome, each with a slack parameter) and a customizable linear objective. The ILP is what carries the optimization; a greedy ranking heuristic, which treats each dose as a separate treatment and takes the argmax CADE per entity, is the baseline that the ILP must beat.

What would settle it

Run the same predict-then-optimize pipeline on a dataset in which all true dose-response curves are monotonic with diminishing returns: if the ILP's prescribed policy value never exceeds the greedy heuristic's and the fairness constraints never change any assigned dose, then the paper's case that an ILP formulation is necessary for continuous-dose allocation would be falsified.

Watch

Extended reading notes

Core claim

The paper claims that continuous-valued treatments in uplift modeling should be handled as a dose-allocation problem rather than a CATE ranking problem. The central discovery, demonstrated experimentally, is that an ILP built on discretized CADR and CADE estimates can allocate doses under a budget and fairness constraints, and that the choice of allocation objective changes who receives what dose in ways a pure uplift ranking would miss. The paper also establishes that prediction accuracy (MISE) and allocation quality (AUUC, regret) can diverge: the estimator with the best dose-response fit can produce worse treatment portfolios than a less globally accurate one, because only errors in dose regions that are actually selected matter. This motivates evaluating estimators on downstream policy value rather than curve-fit error alone.

Load-bearing premise

The experimental validation assumes that the semi-synthetic IHDP dose-response curves (Equations 30-33) are representative of real-world continuous treatments, including the non-monotonic shapes that drive the observed fairness-value trade-offs; the paper itself flags that experiments use a single dataset and data-generating process (Section 6).

Editorial extensions

If this is right

  • Dose allocation with continuous treatments can be solved by composing off-the-shelf CADR estimators with an ILP, so adding fairness terms or cost-sensitivity requires only re-solving the optimization, not retraining the prediction model.
  • Policy value and curve-fit error can diverge, so practitioners should select dose-response estimators on downstream allocation metrics such as AUUC or regret rather than MISE alone.
  • Fairness constraints reduce total uplift, and the reduction grows when the protected groups differ more in their true average treatment effects.
  • A cost-sensitive objective changes the treatment portfolio, prioritizing high benefit-to-cost entities over high-uplift ones, and yields higher value under tight budgets.
  • At budgets where the true optimal policy stops increasing dose due to non-monotonic dose responses, both the ILP and the heuristic flatten the uplift curve, indicating that overspending is wasteful.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If real-world dose responses are mostly monotonic with diminishing returns, the demonstrated advantage of the ILP over greedy ranking could shrink substantially; testing the framework on such data would clarify the scope of the claimed benefit.
  • The paper finds that estimated outcome fairness can diverge from true outcome fairness; a natural extension is to enforce fairness constraints using uncertainty sets or conformal intervals around the CADR estimates.
  • The same ILP formulation could be embedded in a decision-focused learning pipeline as a differentiable layer, allowing the dose-response estimator to be trained to maximize allocation value rather than curve-fit accuracy.
  • The framework is single-phase; extending it to sequential or multi-treatment settings would require re-solving the ILP per decision epoch or replacing it with a dynamic policy.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-stage ('predict-then-optimize') framework for uplift modeling with continuous treatments. In the first stage, conditional average dose responses (CADRs) and conditional average dose effects (CADEs) are estimated with off-the-shelf causal machine-learning estimators. In the second stage, the treatment allocation problem is cast as a dose-allocation integer linear program (ILP) with a budget constraint and optional fairness constraints, and the objective can be adapted to instance-dependent benefits. The paper also offers a formal definition of uplift modeling, compares four CADR estimators on a semi-synthetic IHDP benchmark, and reports experiments on fairness trade-offs and cost-sensitive allocation.

Significance. The conceptual separation of CADE estimation from constrained dose allocation is useful and timely, and the ILP formulation in Eqs. (11)-(18) is a clean, flexible way to impose budget, fairness, and cost-related constraints without retraining the prediction model. The identification argument in Appendix B follows standard assumptions, and the authors provide code. If the empirical claims were better supported, the framework would be a solid modular contribution to the prescriptive-analytics literature. However, the current experimental evidence rests on a single semi-synthetic setting, and two formal gaps—the overlap-violating baseline dose and the unformalized cost-sensitive objective—make the main empirical conclusions uncertain.

major comments (4)
  1. [Appendix C, Eq. (31); Assumption 3; Appendix B, Eq. (29)] The semi-synthetic data-generating process violates the Overlap assumption at exactly the baseline dose used to define the CADE. Eq. (31) sets t = 1/(1+exp(-2*tilde(t))), whose support is (0,1), so p(S=0|X)=0. Yet the CADE is defined as tau_s(x) = E[Y(s)-Y(0)|X=x], and the identification result in Eq. (29) requires E[Y|X=x,S=0]. Under Assumption 3 this quantity is not identified, so every CADR estimator must extrapolate to S=0, and the ILP in Eqs. (11)-(18) explicitly includes dose 0 as an allocation option. This is not a purely technical point: an additive error at the baseline propagates into every estimated CADE, so the optimized allocations, AUUC values, and fairness trade-offs in Section 5 could be artifacts of the extrapolation rather than evidence for the framework. Please either define the reference dose at an interior value supported by the DGP, add a randomized baseline at S=0, or explicitly acknowledge the extrapolation and test its sensitivity.
  2. [Section 3.3, Eqs. (6)-(8) and (19)-(21); Section 5.3.3] The claimed cost-sensitive objective is never formally defined. In Eqs. (6)-(8), the value V_i is the CADE dot product scaled by the benefit b_i, and treatment costs C appear only in the budget constraint Eq. (12). Experiment 3 states that the ILP objective is adapted to account for instance-dependent costs C and benefits b, but no net-value objective such as V_i - Psi_i is introduced, and Eqs. (19)-(21) contain no cost term. As written, the 'cost-sensitive' policy in Figure 7 is a benefit-weighted uplift policy under a budget constraint, not a policy maximizing the cost-sensitive value that the text claims. Please define the objective explicitly and state how C enters it, or reframe the experiment accordingly.
  3. [Section 5.3, Table 3 and Figure 3] The experimental support consists of point estimates from a single semi-synthetic realization. The DGP in Eqs. (30)-(32) contains noise terms, yet no error bars, standard errors, or repeated-seed results are reported for MISE, AUUC, or the value curves. Estimator rankings in Table 3 (e.g., S-Learner (mlp) versus S-Learner (rf)) could plausibly change under different draws. Please add repeated trials or bootstrap confidence intervals, or otherwise quantify the variability of the reported comparisons.
  4. [Section 5.3.1, Table 3] The AUUC evaluation budget of 140 appears to be selected post hoc after observing in Figure 3 that all curves flatten around that budget. Using a data-dependent evaluation point can favor estimators whose errors happen to fall outside the selected region and weakens the cross-estimator comparison. Please report AUUC over the full budget range, pre-specify the evaluation budget, or justify the value 140 independently of the observed curves.
minor comments (4)
  1. [Sections 3.2-3.3] The dimensions are inconsistent: D = {(d-1)/delta : d=1,...,delta+1} has delta+1 dose levels, but the text says the vector tau_hat(x) contains delta CADE estimates and defines pi as a map into {0,1}^delta. Since Eq. (17) sums over all s in D, the notation should be corrected, for example by defining delta+1 estimates or by excluding the baseline dose from the policy vector.
  2. [Appendix C, Eqs. (30)-(32)] There are several small typos in the notation: the sets are introduced as S_bin,1 and S_bin,2, but Eqs. (30)-(32) use S_dis,1 and S_bin,2, and the subscripts are not typeset consistently. These should be cleaned up for readability.
  3. [Abstract and github link] The repository URL in the abstract is written as 'github.com/SimonDeV os/UMCT' with an extra space, and there is a typo 'topk' in the abstract. Both should be corrected.
  4. [Table 3] The caption says the best results are bold and second-best are italic, but it is not clear how ties and the multiple metrics are handled; the selection rule should be stated explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity; derivation is a genuine predict-then-optimize pipeline with independent benchmarks.

full rationale

The derivation chain is self-contained. CADR/CADE identification (Eq. 29) follows the standard Consistency/Ignorability/Overlap assumptions and is credited to an external source (Neal, 2020), not to the authors' prior work. The predictive methods (S-learner, DRNet, VCNet) are existing off-the-shelf estimators fitted to factual outcomes; none is fitted or calibrated to the ILP objective, AUUC, regret, or fairness slack. The ILP (Eqs. 11-18) maximizes a linear objective whose coefficients are the CADR/CADE estimates, so the allocation is a genuine function of those estimates rather than an inverse construction. The evaluation explicitly separates expected, prescribed, and optimal policy values (Eqs. 19-21), using ground-truth CADRs only as a benchmark; this is evaluation protocol, not a fitted input. The paper's self-citations (e.g., Bockel-Rickermann et al., 2023/2024; De Vos et al., 2023) are contextual and do not carry the central claim. The skeptical concern about the semi-synthetic DGP—Eq. 31 has support (0,1), while Assumption 3 requires overlap on [0,1], so the baseline dose 0 requires extrapolation—is a possible violation of an identifying assumption and a correctness/validity risk, not a circularity step: it does not make any prediction equal to an input by construction. The paper also acknowledges its single-dataset limitation, which further supports treating this as a generalizability concern rather than circularity. No load-bearing step reduces to self-citation, definitional equivalence, or renamed fitting.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework itself introduces no new causal entities or free-floating parameters beyond standard user choices (number of bins, fairness slacks) and an unspecified cost/benefit setup in Experiment 3. The main load-bearing assumptions are the standard causal identification assumptions plus the representativeness of the single semi-synthetic data generator.

free parameters (3)
  • Number of dose bins δ = 10
    Chosen by the authors (Appendix D); discretization granularity affects the ILP solution space and the results.
  • Fairness slack parameters εDT, εDO = varied over {1.0, 0.1, 0.01}
    User-chosen tolerances; the demonstrated fairness-value trade-offs depend directly on these values.
  • Benefit vector b and cost matrix C = unspecified
    In Experiment 3, instance-dependent benefits and costs are used but their data-generating process is not described in the paper.
assumptions (4)
  • domain assumption Consistency, Ignorability, Overlap (Assumptions 1-3)
    Standard causal inference assumptions needed to identify the CADE via Eq 29 (Appendix B).
  • domain assumption Treatment costs are directly proportional to dose levels
    Stated in Section 3.3; used in the ILP budget constraint and experiments, though the paper notes it is modular.
  • ad hoc to paper The semi-synthetic DGP (Eqs 30-33) is representative of real-world dose-response relationships
    The paper's experimental conclusions about fairness trade-offs and estimator comparison rely on this single data generator; the authors acknowledge limited generalizability in Section 6.
  • domain assumption Gurobi ILP solver returns global optima for the given instances
    The paper states the solver converged, but this is a practical solver guarantee, not a mathematical proof for all problem sizes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uplift modeling with continuous treatments: A predict-then-optimize approach." pith.science (2026). https://pith.science/paper/L5DD2CO5

@misc{pith2026241209232,
  author       = {Pith},
  title        = {Pith review of: Uplift modeling with continuous treatments: A predict-then-optimize approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5DD2CO5}},
  note         = {Machine review of arXiv:2412.09232}
}
read the original abstract

The goal of uplift modeling is to recommend actions that optimize specific outcomes by determining which entities should receive treatment. One common approach involves two steps: first, an inference step that estimates conditional average treatment effects (CATEs), and second, an optimization step that ranks entities based on their CATE values and assigns treatment to the top k within a given budget. While uplift modeling typically focuses on binary treatments, many real-world applications are characterized by continuous-valued treatments, i.e., a treatment dose. This paper presents a predict-then-optimize framework to allow for continuous treatments in uplift modeling. First, in the inference step, conditional average dose responses (CADRs) are estimated from data using causal machine learning techniques. Second, in the optimization step, we frame the assignment task of continuous treatments as a dose-allocation problem and solve it using integer linear programming (ILP). This approach allows decision-makers to efficiently and effectively allocate treatment doses while balancing resource availability, with the possibility of adding extra constraints like fairness considerations or adapting the objective function to take into account instance-dependent costs and benefits to maximize utility. The experiments compare several CADR estimators and illustrate the trade-offs between policy value and fairness, as well as the impact of an adapted objective function. This showcases the framework's advantages and flexibility across diverse applications in healthcare, lending, and human resource management. All code is available on github.com/SimonDeVos/UMCT.

Figures

Figures reproduced from arXiv: 2412.09232 by the authors.

Figure 1
Figure 1. This DAG represents the assumed causal relationships between variables in the training data. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Overview of predict-then-optimize approach for UM with continuous treatment e [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. This figure shows U presc for four dose-response estimators as the available budget levels increase. Doses are allocated with an ILP and a heuristic approach. avoid overtreating certain entities— due to the non-monotonic nature of CADRs. From a budget of 140 onwards, per estimator, the heuristic and ILP approach find the same solution (i.e., selection of the dose with maximum estimated CADE) and they converge in ter… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The effect of slack parameters ϵDT and ϵDO on, respectively, disparate treatment (allocation fairness) and disparate outcome (outcome fairness). Both panels use a logarithmic scale. 10−2 10−1 100 DT 10−2 10−1 100 DO 0.2 0.4 0.6 0.8 Normalized Uprescr [PITH_FULL_IMAG…
Figure 5
Figure 5. Figure 5: The x-axis represents the fairness constraint parameter for disparate treatment ( [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Fairness constraints and varying ground-truth ATEs. This figure examines the link between the e [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Comparison of cost-sensitive and cost-insensitive optimization, illustrating di [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Effect of available dose bins on U presc and calculation time 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: This figure presents the ground-truth dose responses in blue, with the blue dots representing the factual [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: Runtime of the ILP and heuristic as a function of problem size. The original dataset size corresponds with [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Principles and Guidelines for Randomized Controlled Trials in AI Evaluation

    cs.CY 2026-05 unverdicted novelty 6.0 of 10

    The authors adapt established RCT validity principles from other fields into a standardized framework with 33 guidelines tailored to AI evaluation contexts.

  2. Hidden Representation Clustering with Multi-Task Representation Learning towards Robust Online Budget Allocation

    cs.LG 2025-06 conditional novelty 5.0 of 10

    Budget allocation by clustering users in a learned hidden representation space and optimizing per cluster improves order volume and gross merchandise volume by up to 0.65% relative to individual-level baselines in Mei...

Reference graph

Works this paper leans on

15 extracted references · 8 canonical work pages · cited by 2 Pith papers

  1. [5]

    Fair Off-Policy Learning from Observational Data

    Fair off-policy learning from observational data. arXiv preprint arXiv:2303.08516 . Gubela, R.M., Lessmann, S.,

    work page Pith review arXiv

  2. [7]

    Kleinberg, J., Mullainathan, S., Raghavan, M.,

    On fairness and stability: Is estimator variance a friend or a foe? arXiv preprint: 2302.04525 . Kleinberg, J., Mullainathan, S., Raghavan, M.,

    arXiv

  3. [9]

    arXiv preprint arXiv:2307.13565

    Decision-focused learning: Foundations, state of the art, benchmark and future opportunities. arXiv preprint arXiv:2307.13565 . Nabi, R., Malinsky, D., Shpitser, I.,

    arXiv

  4. [10]

    arXiv preprint arXiv:2103.07861

    Vcnet and functional targeted regularization for learning causal effects of continuous treatments. arXiv preprint arXiv:2103.07861 . 24 Uplift modeling with continuous treatments: A predict-then-optimize approach A Preprint Olaya, D., Coussement, K., Verbeke, W.,

    arXiv

  5. [11]

    arXiv preprint arXiv:2407.03094

    Conformal prediction for causal effects of continuous treatments. arXiv preprint arXiv:2407.03094 . Schwab, P., Linhardt, L., Bauer, S., Buhmann, J.M., Karlen, W.,

    arXiv

  6. [12]

    (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc

    Observational-interventional priors for dose-response learning, in: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc. URL: https://proceedings.neurips.cc/paper_files/paper/2016/file/ aff1621254f7c1be92f64550478c56e6-Paper.pdf. Vanderschueren, T., Berrevoets, J., Ve...

    work page 2016

  7. [13]

    arXiv preprint arXiv:2405.02183

    Metalearners for ranking treatment effects. arXiv preprint arXiv:2405.02183 . Vasquez, C.O., De Weerdt, J., vanden Broucke, S.,

    arXiv

  8. [14]

    Thorax 78, 983–989

    Estimating individual treatment effects on copd exacerbations by causal machine learning on randomised controlled trials. Thorax 78, 983–989. URL: https://thorax.bmj.com/content/78/10/983, doi:doi:10.1136 /thorax-2022-219382, arXiv:https://thorax.bmj.com/content/78/10/983.full.pdf. Zhan, B., Liu, C., Li, Y ., Wu, C.,

    work page 2022

Show all 15 references
  1. [2014]

    Example-dependent cost-sensitive logistic regression for credit scoring, in: 2014 13th International conference on machine learning and applications, IEEE. pp. 263–269. Barocas, S., Hardt, M., Narayanan, A.,

    2014

  2. [2015]

    Archives of toxicology 89, 2059–2068

    Optimal experimental designs for dose–response studies with continuous endpoints. Archives of toxicology 89, 2059–2068. Höppner, S., Baesens, B., Verbeke, W., Verdonck, T.,

  3. [2016]

    arXiv preprint arXiv:1609.05807

    Inherent trade-offs in the fair determination of risk scores. arXiv preprint arXiv:1609.05807 . Kozodoi, N., Jacob, J., Lessmann, S.,

    arXiv

  4. [2021]

    COM(2021) 206 final

    Proposal for a Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts. COM(2021) 206 final. Fernández-Loría, C., Provost, F.,

    2021

  5. [2022]

    arXiv preprint arXiv:2202.01336

    Exploring transformer backbones for heterogeneous treatment effect estimation. arXiv preprint arXiv:2202.01336 . 25 Uplift modeling with continuous treatments: A predict-then-optimize approach A Preprint Zhao, Z., Bai, Y ., Xiong, R., Cao, Q., Ma, C., Jiang, N., Wu, F., Kuang, K.,

    arXiv

  6. [2023]

    arXiv preprint arXiv:2309.03731

    Learning continuous- valued treatment effects through representation balancing. arXiv preprint arXiv:2309.03731 . Bockel-Rickermann, C., Vanderschueren, T., Verdonck, T., Verbeke, W.,

    arXiv

  7. [2024]

    arXiv preprint arXiv:2406.08206

    Sources of gain: Decomposing performance in conditional average dose response estimation. arXiv preprint arXiv:2406.08206 . Brooks-Gunn, J., Liaw, F.r., Klebanov, P.K.,

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.