betaFeedback

WorksMathematics

Sharp Bounds on Continuous Treatment Effects under Unobserved Confounding

Abstract

We study partial identification of conditional treatment effects with a continuous treatment under unobserved confounding. We target the conditional average potential outcome (CAPO) E[Y_x|z] and the conditional average treatment effect (CATE) between two treatment levels. We assume a bounded outcome and Lipschitz continuity of the response functions and of the propensity scores given the latent confounder, and show that these conditions imply the smooth treatment response and smooth treatment selection conditions of Kim et al. (2018). The combined bound of Kim et al. uses only two reference points on the identified diagonal and is generally loose. Bounding through every diagonal reference point instead gives explicit bounds whose average is the sharp identified set of the CAPO at every treatment, inside or outside the observational support. Capping the difference between the upper and lower bounds gives sharp causal bounds on the CATE. These bounds are informative without monotonicity or overlap assumptions.

Introduction

Evaluating treatment effects from observational data is a central task in causal inference, with applications to policy evaluation, healthcare, and reinforcement learning. The challenge is that observed treatment–outcome correlations generally do not imply causal effects: unobserved confounding can induce spurious relations, leaving the target effect non-identifiable, i.e., underdetermined by the observational data together with causal knowledge about the underlying system [32]. Several conditions and algorithms have been developed to decide whether a causal effect is identifiable [32, 6].

For example, under the back-door condition [32 Ch. 3.3.1], or equivalently ignorability [33], the effect is identified by covariate adjustment and can be estimated at standard rates [33, 4]. In practice, however, the available causal assumptions are often insufficient to determine the target effect from observational data, in which case the effect is said to be non-identifiable. The following example illustrates this challenge.

Example 1. Consider a system M\mathcal{M} with treatment X←−U/2X \leftarrow-U/2 and outcome Y←−X2+U2Y \leftarrow-X^2+U^2, where U∼N(0,1)U \sim\mathcal{N}(0,1) is unobserved. Figure 1a shows samples (blue) from the resulting observational distribution P(X,Y)P(X,Y) alongside samples (orange) from the interventional distribution Px(Y)P_x(Y) obtained by randomly assigning XX over [−2,2][-2,2]. The two deviate significantly: the unobserved confounder UU induces a spurious correlation between XX and YY, making some treatments appear more effective than they actually are. Treatment effects Ex[Y]E_x[Y] are not identifiable here [16, 35]. To witness, consider the alternative system M′\mathcal{M}' with the same treatment X←−U/2X \leftarrow-U/2 but outcome Y←X2+12U2Y \leftarrow X^2+\frac{1}{2}U^2. Both models share the causal relationships (X→YX \to Y and X←U→YX \leftarrow U \to Y) and induce the identical observational law P(X,Y)P(X,Y) (substituting U=−2XU=-2X gives Y=3X2Y=3X^2 in either model), yet they disagree on the effect: intervening do⁡(X←x)\operatorname{do}(X \leftarrow x) yields Ex[Y]=1−x2\mathbb{E}_x[Y]=1-x^2 under M\mathcal{M} but Ex[Y]=x2+12\mathbb{E}_x[Y]=x^2+\frac{1}{2} under M′\mathcal{M}' (Figure 1b). Both models thus agree that XX affects YY and are confounded in the same way; they disagree on which treatment is preferable, with x=0x=0 optimal under M\mathcal{M} but the worst choice under M′\mathcal{M}'. The observational data alone cannot tell them apart. ■

Two confounded systems M and M′ showing observational and interventional laws

Figure 1. Two confounded systems that induce the same observational law P(X,Y)P(X,Y) (blue; regression E[Y∣x]E[Y \mid x] in dark blue) yet different interventional laws Px(Y)P_x(Y) (orange), illustrating non-identifiability: (a) M\mathcal{M} with Y←−X2+U2Y \leftarrow-X^2+U^2 has a causal effect Ex[Y]=1−x2E_x[Y]=1-x^2 (dark orange curve); (b) M′\mathcal{M}' with Y←X2+12U2Y \leftarrow X^2+\frac{1}{2}U^2 has a causal effect Ex[Y]=x2+12E_x[Y]=x^2+\frac{1}{2} (dark orange curve).

When identifiability fails, a common alternative approach is partial identification: bounding the effect over all confounding patterns compatible with the data ([24, 3]). For discrete treatments and outcomes, sharp bounds follow from programs over a discretized latent domain ([3, 44]) or from generative-model approximations ([41, 42]). However, this enumeration does not carry over to continuous treatments, where three lines of work exist.

Sensitivity models bound how far hidden confounding may distort the treatment density: the continuous marginal sensitivity model (CMSM) of Jesson et al. [17], its sharp and generalized versions ([13, 1]), the δ\delta-MSM ([26]), and exposure-dependent and marginal-structural variants ([7, 10]). Several bound the conditional average potential outcome (CAPO) pointwise, but all are stated relative to the observed treatment density and require it to be positive at the query. Optimization-based bounds search a response-function class consistent with the data, usually given an instrument ([20, 14, 31, 2]). Smoothness bounds restrict how fast counterfactual means vary ([36]). Following Hausman and Newey [15], Kim et al. [21] bound the effect of treatment on the treated assuming it is Lipschitz in the intervened treatment (smooth treatment response, STR) and in the received treatment (smooth treatment selection, STS), and Khan et al. [19] use Lipschitz smoothness for off-policy evaluation without overlap. App. A provides more details on related work.

We work in the smoothness line, in the setting of Kim et al. [21] extended to covariates. Our targets are the CAPO τ(x,z)=E[Yx∣z]\tau(x,z)=\mathbb{E}[Y_x \mid z] and the conditional average treatment effect (CATE) τ(x1,z)−τ(x0,z)\tau(x_1,z)-\tau(x_0,z) between two treatment levels. The bounds of Kim et al. [21] are sharp under STR together with a monotonicity assumption. Under STR and STS, however, sharp bounds remain an open challenge: we show that their combined bound is not sharp in general.

This paper aims to address this challenge. Our contributions are threefold. (1) Structural conditions for STR and STS. We derive both conditions from Lipschitz continuity of the response functions and of the propensity scores given the latent confounder (Prop. 2), with a selection cost that depends only on the observed treatment density (Lem. 1). (2) The bounds of Kim et al. [21] are loose. Their combined STR+STS bound uses only two reference points on the identified diagonal (Prop. 8) and is not sharp in general, even within their own model class (Prop. 9). (3) Sharp bounds on the CAPO and the CATE. Combining the STR and STS bounds through every diagonal reference point gives a valid bound (Thm. 6) that is the sharp identified set of the CAPO at every treatment, inside or outside the observational support (Thm. 10). Capping the difference of the upper and lower bounds gives the sharp identified set of the CATE, without any monotonicity assumption (Thm. 11). All proofs and experimental details are in the appendix.

Notations. We denote variables by capital letters (XX) and their values by lowercase (xx); ∣X∣|X| is the cardinality of a set XX. P(X)P(X) is the distribution of XX, P(x)P(x) abbreviates probability P(X=x)P(X=x), and P(X)P(X) stands for the probability P(X∈X)P(X\in\mathcal{X}) that variables XX take a value in a set X\mathcal{X}. For reals a,ba,b we write a∨b≜max⁡{a,b}a\vee b\triangleq\max\{a,b\} and a∧b≜min⁡{a,b}a\wedge b\triangleq\min\{a,b\}.

Lipschitz continuity for partial identification

Our analysis focuses on the structural causal models (SCMs) ([32, 5]) of a contextual bandit, with a treatment XX, an outcome YY, covariates ZZ, and an unobserved confounder UU affecting all three. In a model M\mathcal{M}, structural functions set Z←fZ(U)Z\leftarrow f_Z(U), X←fX(Z,U,εX)X\leftarrow f_X(Z,U,\varepsilon_X), and Y←fY(X,Z,U,εY)Y\leftarrow f_Y(X,Z,U,\varepsilon_Y), where UU, εX\varepsilon_X, and εY\varepsilon_Y are mutually independent exogenous variables. The assignment noise εX\varepsilon_X makes the treatment random given (Z,U)(Z,U), so that it can have a conditional density p(x∣z,u)p(x\mid z,u); this is needed only for the structural condition (A3) below. The learner does not know M\mathcal{M}; it only observes samples of (X,Z,Y)(X,Z,Y), summarized by the observational distribution P(X,Y,Z)≜P(X,Y,Z∣M)P(X,Y,Z)\triangleq P(X,Y,Z\mid\mathcal{M}). An intervention do⁡(X←x)\operatorname{do}(X\leftarrow x) (for short, do⁡(x)\operatorname{do}(x)) replaces fXf_X by the constant xx, inducing the submodel Mx\mathcal{M}_x. The potential outcome Yx(u)Y_x(u) is the solution for YY in Mx\mathcal{M}_x, and the interventional distribution is the joint distribution Px(X,Y,Z)≜P(Yx,Zx∣M)P_x(X,Y,Z)\triangleq P(Y_x,Z_x\mid\mathcal{M}) over potential outcomes.

Our goal is to infer the conditional average potential outcome (CAPO) τ(x,z)≜Ex[Y∣z]\tau(x,z)\triangleq\mathbb{E}_x[Y\mid z] at every treatment xx and covariate value zz (output) from the observational distribution P(X,Z,Y)P(X,Z,Y) (input). We also bound the conditional average treatment effect (CATE) τ(x1,z)−τ(x0,z)=E[Yx1−Yx0∣z]\tau(x_1,z)-\tau(x_0,z)=\mathbb{E}[Y_{x_1}-Y_{x_0}\mid z] between two treatment levels Thm. 11. Under unobserved confounding, τ\tau is not identifiable: many distinct SCMs reproduce the same P(X,Z,Y)P(\mathbf{X},\mathbf{Z},Y) yet disagree on the target query. Consequently, we aim at partial identification: deriving informative bounds over the CAPO and CATO from the observational distribution.

Without restrictions on the unobserved mechanisms, the data leave YxY_x free at treatments that are rarely or never observed at zz, and no estimator can improve on the trivial range bound there. An informative bound requires restricting how fast the unobserved mechanisms may vary ([21, 36]). We posit one for each unobserved mechanism: the outcome and the treatment assignment.

Each latent stratum induces a response function hu(x,z)≜E[fY(x,z,u,ϵY)]h_u(x,z)\triangleq\mathbb{E}[f_Y(x,z,u,\epsilon_Y)], the expected outcome of do⁡(X←x)\operatorname{do}(X\leftarrow x) at covariates zz within stratum uu. Treatments and covariates take values in Borel sets X⊆Rn\mathcal{X}\subseteq\mathbb{R}^n and Z⊆Rn\mathcal{Z}\subseteq\mathbb{R}^n, which need not be bounded. With constants LY,LX>0L_Y,L_X>0, we assume the following.

(A1) (Compact outcome domain.) The outcome takes values in a compact interval Y=[y‾,y‾]⊂R\mathbb{Y}=[\underline{y},\overline{y}]\subset\mathbb{R}.

(A2) (Lipschitz response functions.) For all u∈Uu\in\mathbb{U}, z∈Zz\in\mathcal{Z}, and x,x′∈Xx,x'\in\mathcal{X},

∣hu(x,z)−hu(x′,z)∣≤LY∥x−x′∥.(1)\left|h_u(x,z)-h_u(x',z)\right|\leq L_Y\left\|x-x'\right\|. \tag*{(1)}

(A3) (Lipschitz propensity density.) For all u∈Uu\in\mathbb{U} and z∈Zz\in\mathcal{Z}, the treatment has a probability density p(x∣z,u)p(x\mid z,u) satisfying, for all x,x′∈Xx,x'\in\mathcal{X},

∣p(x∣z,u)−p(x′∣z,u)∣≤LX∥x−x′∥.(2)\left|p(x\mid z,u)-p(x'\mid z,u)\right|\leq L_X\left\|x-x'\right\|. \tag*{(2)}

We write diam⁡(Y)≜y‾−y‾\operatorname{diam}(\mathbb{Y})\triangleq\overline{y}-\underline{y} for the outcome range. Every response function output fY(x,z,u,ϵY)f_Y(x,z,u,\epsilon_Y) lies in [y‾,y‾][\underline{y},\overline{y}]. The constants LYL_Y and LXL_X are sensitivity parameters the analyst must prespecify.

Smoothness of Treatment on the Treated

Assumptions (A2) and (A3) restrict mechanisms that the data never reveal. To connect them to the data, we rewrite the target through a function whose values we partially identify. Specifically, let gz(x,x′)≜E[Yx∣x′,z]g_z(x,x')\triangleq\mathbb{E}[Y_x\mid x',z] denote the effect of treatment on the treated (ETT) at treatment x′x': the mean outcome under do⁡(x)\operatorname{do}(x) of the units that observationally received x′x' ([32]). It is the function studied by Kim et al. [21], here conditional on zz. We call xx the intervened treatment and x′x' the received treatment, and regard gzg_z as a function on X×Sz\mathcal{X}\times S_z, where Sz≜supp⁡P(X∣z)S_z\triangleq\operatorname{supp}P(X\mid z) is the observational support at zz. Conditioning on the received treatment X=x′X=x' and applying the law of total expectation expands the target over the observational treatment distribution,

τ(x,z)=∫gz(x,x′) dP(x′∣z).(3)\tau(x,z)=\int g_z(x,x')\,\mathrm{d}P(x'\mid z). \tag*{(3)}

References

  1. [1]Baitairian, J.-B., Sebastien, B., Jreich, R., Kat-sahian, S., and Guilloux, A. (2025). Sharp bounds for continuous-valued treatment effects with unobserved confounders. Biometrical Journal, 67(5):e70084.DOI
  2. [2]Balazadeh Meresht, V., Syrgkanis, V., and Krishnan, R. G. (2022). Partial identification of treatment effects with implicit generative models. In Advances in Neural Information Processing Systems, volume 35, pages 22816–22829.
  3. [3]Balke, A. and Pearl, J. (1997). Bounds on treatment effects from studies with imperfect compliance. Journal of the American Statistical Association, 92(439):1172–1176.DOI
  4. [4]Bang, H. and Robins, J. M. (2005). Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4):962–973.pubmed.ncbi.nlm.nih.gov/16401269
  5. [5]Bareinboim, E., Correa, J. D., Ib eling, D., and Icard, T. (2022). On pearl’s hierarchy and the foundations of causal inference. In Probabilistic and causal inference: the works of judea pearl, pages 507–556.DOI
  6. [6]Bareinboim, E. and Pearl, J. (2016). Causal inference and the data-fusion problem. Proceedings of the National Academy of Sciences, 113:7345–7352.DOI
  7. [7]Bonvini, M., Kennedy, E. H., Ventura, V., and Wasserman, L. (2022). Sensitivity analysis for marginal structural models. arXiv preprint arXiv:2210.04681.arxiv.org/abs/2210.04681
  8. [8]Carolei, A. et al. (1997). The international stroke trial (ist): a randomized trial of aspirin, subcutaneous heparin, both, or neither among 19435 patients with acute ischaemic stroke. The Lancet, 349:1569–1581.
  9. [9]Chen, B., Kumor, D., and Bareinboim, E. (2017). Identification and model testing in linear structural equation models using auxiliary variables. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 757–766. PMLR.arxiv.org/abs/1612.03451
  10. [10]Dalal, A. and Tchetgen Tchetgen, E. J. (2025). Partial identification of causal effects for endogenous continuous treatments. In arXiv preprint arXiv:2508.13946.arxiv.org/abs/2508.13946
  11. [11]Dudík, M., Langford, J., and Li, L. (2011). Doubly robust policy evaluation and learning. In Proceedings of the 28th International Conference on International Conference on Machine Learning, pages 1097–1104. Omnipress.arxiv.org/abs/1103.4601
  12. [12]Frangakis, C. and Rubin, D. (2002). Principal stratification in causal inference. Biometrics, 1(58):21–29.DOI
  13. [13]Frauen, D., Melnychuk, V., and Feuerriegel, S. (2023). Sharp bounds for generalized causal sensitivity analysis. In Advances in Neural Information Processing Systems, volume 36, pages 40556–40586.arxiv.org/abs/2305.16988
  14. [14]Gunsilius, F. (2020). A path-sampling method to partially identify causal effects in instrumental variable models. arXiv preprint arXiv:1910.09502.arxiv.org/abs/1910.09502
  15. [15]Hausman, J. A. and Newey, W. K. (2016). Individual heterogeneity and average welfare. Econometrica, 84(3):1225–1248.
  16. [16]Huang, Y. and Valtorta, M. (2006). Pearl’s calculus of intervention is complete. In Dechter, R. and Richardson, T., editors, Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence, pages 217–224. AUAI Press, Corvallis, OR.
  17. [17]Jesson, A., Douglas, A., Manshausen, P., Solal, M., Meinshausen, N., Stier, P., Gal, Y., and Shalit, U. (2022). Scalable sensitivity and uncertainty analyses for causal-effect estimates of continuous-valued interventions. In Advances in Neural Information Processing Systems, volume 35, pages 13892–13907.DOI
  18. [18]Kallus, N. and Zhou, A. (2018). Confounding-robust policy improvement. In Advances in neural information processing systems, pages 9269–9279.arxiv.org/abs/1805.08593
  19. [19]Khan, S., Saveski, M., and Ugander, J. (2024). Off-policy evaluation beyond overlap: Sharp partial identification under smoothness. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research. PMLR.arxiv.org/abs/2305.11812
  20. [20]Kilbertus, N., Kusner, M. J., and Silva, R. (2020). A class of algorithms for general instrumental variable models. In Advances in Neural Information Processing Systems, volume 33.arxiv.org/abs/2006.06366
  21. [21]Kim, W., Kwon, K., Kwon, S., and Lee, S. (2018). The identification power of smoothness assumptions in models with counterfactual outcomes. Quantitative Economics, 9(2):617–642.
  22. [22]Kocaoglu, M., Snyder, C., Dimakis, A. G., and Vish-wanath, S. (2018). Causalgan: Learning causal implicit generative models with adversarial training. In International Conference on Learning Representations.arxiv.org/abs/1709.02023
  23. [24]Manski, C. F. (1990). Nonparametric bounds on treatment effects. The American Economic Review, 80(2):319–323.
  24. [26]Marmarelis, M. G., Haddad, E., Jesson, A., Jahanshad, N., Galstyan, A., and Ver Steeg, G. (2023). Partial identification of dose responses with hidden confounders. In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, volume 216 of Proceedings of Machine Learning Research, pages 1368–1379. PMLR.arxiv.org/abs/2204.11206
  25. [27]McShane, E. J. (1934). Extension of range of functions. Bulletin of the American Mathematical Society, 40(12):837–842.DOI
  26. [28]Munos, R., Stepleton, T., Harutyunyan, A., and Bellemare, M. (2016). Safe and efficient off-policy reinforcement learning. In Advances in Neural Information Processing Systems.arxiv.org/abs/1606.02647
  27. [29]Namkoong, H., Keramati, R., Yadlowsky, S., and Brunskill, E. (2020). Off-policy policy evaluation for sequential decisions under unobserved confounding. Advances in Neural Information Processing Systems, 33:18819–18831.arxiv.org/abs/2003.05623
  28. [30]Nasr-Esfahany, A., Alizadeh, M., and Shah, D. (2023). Counterfactual identifiability of bijective causal models. In International Conference on Machine Learning, pages 25733–25754. PMLR.
  29. [31]Padh, K., Zeitler, J., Watson, D., Kusner, M., Silva, R., and Kilbertus, N. (2023). Stochastic causal programming for bounding treatment effects. In Conference on Causal Learning and Reasoning, pages 142–176. PMLR.arxiv.org/abs/2202.10806
  30. [32]Pearl, J. (2000). Causality: Models, Reasoning, and Inference. Cambridge University Press, NY. 2nd edition, 2009.DOI
  31. [33]Rosenbaum, P. and Rubin, D. (1983). The central role of propensity score in observational studies for causal effects. Biometrika, 70:41–55.
  32. [34]Sandercock, P., Niewada, M., and Członkowska, A. (2011). International stroke trial database (version 2). Edinburgh DataShare, University of Edinburgh, Department of Clinical Neurosciences. https://doi.org/10.7488/ds/104.
  33. [35]Shpitser, I. and Pearl, J. (2006). Identification of joint interventional distributions in recursive semi-Markovian causal models. In Proceedings of the Twenty-First National Conference on Artificial Intelligence, pages 1219–1226.
  34. [36]Tan, J., Blanchet, J., and Syrgkanis, V. (2024). Consistency of neural causal partial identification. In Advances in Neural Information Processing Systems (NeurIPS), volume 37.arxiv.org/abs/2405.15673
  35. [37]Thomas, P. and Brunskill, E. (2016). Data-efficient off-policy policy evaluation for reinforcement learning. In International Conference on Machine Learning, pages 2139–2148.arxiv.org/abs/1604.00923
  36. [38]Wang, Y. and Blei, D. M. (2019). The blessings of multiple causes. Journal of the American Statistical Association, 114(528):1574–1596.
  37. [39]Whitney, H. (1934). Analytic extensions of differentiable functions defined in closed sets. Transactions of the American Mathematical Society, 36(1):63–89.
  38. [40]Wright, S. (1934). The method of path coefficients. The annals of mathematical statistics, 5(3):161–215.DOI
  39. [41]Xia, K., Lee, K.-Z., Bengio, Y., and Bareinboim, E. (2021). The causal-neural connection: Expressiveness, learnability, and inference. Advances in Neural Information Processing Systems, 34:10823–10836.arxiv.org/abs/2107.00793
  40. [42]Xia, K., Pan, Y., and Bareinboim, E. (2022). Neural causal models for counterfactual identification and estimation. In Proceedings of the 11th Eleventh International Conference on Learning Representations.arxiv.org/abs/2210.00035
  41. [44]Zhang, J., Tian, J., and Bareinboim, E. (2022). Partial counterfactual identification from observational and experimental data. In International Conference on Machine Learning, pages 26548–26558. PMLR.arxiv.org/abs/2110.05690

Paper details

vlnc:W_u2khy4xu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents