betaFeedback

WorksMathematics

Uncertainty Quantification for Topological Treatment Effects

Abstract

We present a unified framework for uncertainty quantification over filtration scale, extending the topological average treatment effect (TATE) analysis of \citet{kim-lee} to conditional mean confidence and individual effect prediction. For the conditional topological average treatment effect (CTATE), we derive multiplier confidence bands at a fixed covariate value, simultaneous over filtration scale, and prove validity in a correctly specified finite-dimensional regression regime, with explicit control of smoothing bias and nuisance-estimation error. For the individual topological treatment effect (ITTE), we integrate existing weighted and functional split-conformal methods to obtain finite-sample marginal prediction coverage under known target-to-calibration likelihood ratios, with a total-variation bound for estimated ratios. We also extend the current TATE silhouette stability analysis to the diagram-independent bound ∥Φr(D)−Φr(D′)∥∞≤(1+2r)W1(D,D′)\|Φ_r(D)-Φ_r(D')\|_\infty \le (1+2r)W_1(D,D') for r≥1r\geq1, recovering the known r=1r=1 constant, and propagate diagram approximation error to prediction bands. The installable Python library tcda\_uq integrates mean confidence, individual prediction, and topological preprocessing, with tutorials and reproduction materials. Simulations and a PACCE lesion-measurement application illustrate the distinct coverage targets and the importance of declaring the target population.

Introduction

Treatment effects are often summarized through changes in scalar outcomes, such as survival time, blood pressure, or total tumor burden. In many applications, however, the outcome contains additional structure that may itself be scientifically relevant. For example, longitudinal cancer studies may record multiple lesion measurements across visits. An intervention can then affect not only the overall level of these measurements, but also the geometric pattern they form, potentially revealing early structural changes in treatment response before changes in conventional summary measures become apparent. Similar questions arise when outcomes are naturally represented as point clouds, images, networks, or other structured objects. Reducing such data to a single scalar may therefore discard features of the treatment response that are of substantive interest.

Persistent homology provides one way to summarize this structure by describing features that persist across multiple filtration scales. This makes it possible to ask causal questions about structural aspects of an outcome rather than only about a conventional scalar endpoint. Building on the mean-effect framework of Kim and Lee [ [10], we represent each potential outcome by the power-weighted silhouette of its persistence diagram and study treatment-induced differences between these functional summaries. The resulting objects are continuous functions of filtration scale, allowing both the magnitude and the scale at which structural differences appear to be retained. Formally, we work in the separable Banach space C(T)C(T) with the supremum norm, where TT is a compact filtration interval.

Defining a topological treatment effect, however, is only part of the inferential problem. For such effects to be used empirically, one must also quantify how much uncertainty surrounds them. This distinction is particularly important because different scientific questions require fundamentally different uncertainty statements. At the population level, the topological average treatment effect (TATE) [10] describes the average treatment-induced difference between potential topological summaries. Conditioning this contrast on baseline characteristics gives the conditional topological average treatment effect (CTATE). These are deterministic mean curves, so the relevant inferential objects are simultaneous confidence bands that quantify estimation uncertainty over the entire filtration interval. Kim and Lee [ [10] already identify TATE, derive its efficient augmented inverse-probability-weighted (AIPW) estimator, establish functional weak convergence, and discuss simultaneous confidence bands. Their Remark 2 describes cross-fitting, and their Corollary 5.4 uses multiplier calibration for a supremum test. These results provide the mean-effect foundations of our framework. Fixed-covariate CTATE inference additionally uses established pseudo-outcome regression ideas [16].

A different uncertainty problem arises when the question concerns a new individual. We take the individual topological treatment effect (ITTE), the contrast between that unit’s two potential summary curves, as an explicit prediction target in this topological framework. Even if the population-level quantities were known without estimation error, the effect for a new unit would remain random because treatment responses can vary across individuals. A confidence band for a mean effect therefore cannot generally be interpreted as a prediction band for an individual effect. Individual prediction must instead account for the distribution of unit-to-unit variation in addition to uncertainty introduced by estimation. The central aim of this paper is to develop uncertainty-quantification procedures appropriate to these distinct topological causal targets: simultaneous confidence bands for TATE and CTATE, and marginal prediction bands for the ITTE of a new unit.

For mean effects, our construction uses the functional influence process of the AIPW estimator of Kim and Lee [ [10], implemented with cross-fitting. Cross-fitting and orthogonality address nuisance-estimation error under suitable rate conditions [7]. Simultaneous inference over the filtration interval, however, requires more than pointwise asymptotic normality or double robustness: it requires process-level convergence and valid calibration of the supremum of the estimation error. Following their functional weak-convergence analysis, we state these requirements explicitly and verify a bounded, uniformly Lipschitz sufficient regime that also checks multiplier validity and variance consistency. For CTATE, we distinguish a general linear-smoother formulation from a fully specified fixed-dimensional regression case, making explicit the role of second-stage approximation bias. We also clarify the relationship with existing functional-data procedures: the simultaneous bands of Liebl and Reimherr [ [13] require their own process regularity conditions, while the interval-wise procedure of Pini and Vantini [ [14] is a testing procedure rather than a confidence band unless an appropriate family of tests is inverted.

Individual prediction requires a different construction. Functional conformal prediction supplies prediction sets for curve-valued responses [8], while weighted conformal prediction accommodates covariate shift when the relevant likelihood ratio is known [19]. In causal applications this shift arises naturally: observed outcomes within one treatment arm and potential outcomes for the target population generally have different covariate distributions [11]. We therefore construct a weighted functional split-conformal band separately for each potential topological outcome and combine the two arm-specific bands by a union bound. The resulting band gives finite-sample marginal coverage for the ITTE when the target-to-calibration likelihood ratios are known. The construction does not require independence between the two potential outcomes and does not identify their joint law. Related arm-differencing logic appears in causal conformal inference, including Schröder et al. [ [15]; our measurable-transformation argument establishes the applicability of this existing machinery to the topological summaries considered here rather than a new general conformal principle.

The weighting step also exposes an important practical limitation. Under weak overlap, a target unit can receive sufficiently large weight that the full-population conformal radius becomes infinite. Overlap weighting instead targets a population concentrated where both treatment arms are represented [12], yielding bounded weights but changing the population to which the prediction statement applies. Bounded weights alone do not guarantee a finite radius at every sample size, since sufficient calibration mass is still required. Moreover, estimated likelihood ratios do not inherit the exact finite-sample guarantee merely because they are estimated on independent data. We quantify this loss explicitly through a total-variation discrepancy between the intended and induced target distributions.

The main contribution is a unified framework for uncertainty quantification that develops both CTATE confidence and ITTE prediction alongside population TATE confidence. The TATE estimand, identification, AIPW representation, and functional weak-convergence foundation follow Kim and Lee [ [10]. We restate these foundations to make the framework self-contained and extend the TATE inference with explicit studentized calibration and sufficient conditions that verify variance consistency and multiplier validity under cross-fitting (Theorems 3–4).

The conditional-mean contribution extends their marginal TATE analysis to simultaneous confidence for CTATE at a prespecified covariate value. Proposition 5 establishes coverage under explicit conditions on the smoothing bias, first-stage remainder, convergence rate, and multiplier process. Corollary 6 then verifies a concrete sufficient regime for correctly specified finite-dimensional regression of cross-fitted AIPW pseudo-outcomes. Thus the conditional contribution includes both a general coverage statement under process assumptions and a proved regression case. It uses established pseudo-outcome regression tools [16]; the arm-outcome regressions in Kim and Lee [ [10]Example 5.1] serve instead as nuisance functions for marginal TATE estimation.

The individual-effect contribution takes ITTE as an explicit prediction target and combines weighted functional conformal calibration and arm differencing to obtain finite-sample marginal coverage under known likelihood ratios, with an explicit total-variation correction for estimated ratios. The conformal principles are inherited from the literature cited above; their integration and coverage analysis here concern the topological individual contrast. Uniformity throughout is over the filtration interval, with CTATE confidence at a fixed covariate value and ITTE prediction marginal over the declared target population.

We also provide the installable, reusable Python library tcda_uq, which brings topological preprocessing, mean confidence, individual prediction, and target-specific evaluation into a common interface with executable tutorials and reproduction materials (§ 5.6). Supporting this framework, we also extend the silhouette stability analysis of Kim and Lee [ [10]Theorem 5.3] by proving, for every fixed r≥1r \ge1, the diagram-independent bound

∥Φr(D)−Φr(D′)∥∞≤(1+2r)W1(D,D′).\lVert\Phi_r(D) - \Phi_r(D') \rVert_\infty\le(1+2r)W_1(D,D').

Appendix C.5 of Kim and Lee [ [10] already gives the same constant 3 at r=1r=1; for r>1r>1, the bound above removes the dependence on diagram lifetime bounds in their constant. We use it to propagate known persistence-diagram approximation errors into prediction bands. We further provide simultaneous prediction over prespecified collections of homology degrees and silhouette exponents and a deterministic extension from gridwise bands to the full filtration interval.

We examine these guarantees and their assumptions through synthetic benchmarks and an application to longitudinal lesion measurements from the PACCE trial [9]. It illustrates how topological summaries and their uncertainty bands can be brought to an applied setting, while also demonstrating why statistical validity cannot be separated from the definition of the target population.

Several limitations are important for interpreting the resulting uncertainty statements. The guarantees concern treatment effects on a prespecified topological summary rather than on the complete persistence diagram or the underlying outcome: silhouettes are not injective, so a zero silhouette effect does not imply equality of diagrams. The ITTE prediction guarantee is marginal over a new target unit and does not provide conditional coverage at every covariate value or identify the joint law of the two potential outcomes. Finally, confidence and prediction should not be interchanged. In particular, we show that a shrinking confidence band for the mean effect has asymptotically zero individual prediction coverage when the individual-effect distribution has no atom at the mean curve. Thus increasingly precise estimation of an average topological effect does not, by itself, resolve uncertainty about the treatment effect experienced by a new individual.

The remainder of the paper is organized as follows. § 2 defines the topological representation, causal assumptions, and inferential targets. § 3 develops simultaneous confidence bands for mean effects, and § 4 develops individual prediction bands. § 5 considers overlap populations, propagation of persistence-diagram approximation error, joint prediction, and extension from discrete grids to continuous filtration intervals. The numerical and application sections evaluate the procedures while distinguishing the sampling population, inferential target, and coverage event. § 8 concludes with a synthesis of the framework and outlines open research directions, while the appendices provide notation, proofs, additional experiments, and a register of the mathematical claims.

Problem setup and estimands

We begin by connecting the three ingredients of the framework: the potential-outcomes formulation of the causal problem, the topological representation of structured outcomes, and the causal quantities defined from that representation. The basic idea is straightforward. Each unit has two potentially structured outcomes, one under each treatment condition. We map each of these outcomes to a continuous curve that summarizes its topological structure and then define treatment effects as contrasts between the resulting curves. This section makes each step precise and establishes the notation used throughout the paper.

Throughout, 0<α<10 < \alpha< 1; armwise errors also lie in (0,1)(0,1).

Observed data and causal assumptions

Let (Xi,Ai,Yi0,Yi1)(X_i, A_i, Y_i^0, Y_i^1), i=1,…,ni = 1,\ldots,n, be independent and identically distributed full-data units. Here, XiX_i denotes baseline covariates, Ai∈{0,1}A_i \in\{0,1\} the treatment assignment, and Yi0,Yi1Y_i^0, Y_i^1 the potential outcomes under control and treatment, respectively. Unlike in conventional causal analyses, YiaY_i^a need not be scalar or even finite-dimensional: it may be a point cloud, image, network, or another structured object. Only one of the two potential outcomes is observed, so the observed data for unit ii are Oi=(Xi,Ai,Yi)O_i = (X_i, A_i, Y_i), Yi=YiAiY_i = Y_i^{A_i}.

We work under the standard assumptions that connect these observed data to the potential-outcome distributions of interest. Specifically, we assume consistency, Yi=YiAiY_i = Y_i^{A_i}, conditional exchangeability, (Y0,Y1)⊥A∣X(Y^0, Y^1) \perp A \mid X, and strict positivity, 0<π(X)<10 < \pi(X) < 1 almost surely, where π(x)=P(A=1∣X=x)\pi(x) = P(A = 1 \mid X = x) is the propensity score. Regular conditional distributions are assumed to exist; for example, the underlying spaces may be taken to be standard Borel spaces.

Together, these assumptions identify the conditional marginal distribution of each potential outcome from the corresponding observed treatment arm. They do not, however, identify the dependence between Y0Y^{0} and Y1Y^{1}. This distinction will matter later: mean treatment effects can be identified from the two arm-specific marginal laws, whereas the joint distribution of the two potential outcomes for the same individual generally cannot.

Some of our inferential results require a stronger form of positivity. Where explicitly stated, we assume that for some ϵ>0\epsilon> 0,

ϵ≤π(X)≤1−ϵalmost surely.(1)\epsilon\le\pi(X) \le1-\epsilon\quad\text{almost surely}. \tag*{(1)}

We refer to (1) as strong overlap. Strict positivity, together with consistency and conditional exchangeability, identifies the arm-specific potential-outcome distributions. The stronger condition (1) is imposed where required for the asymptotic inference results of § 3. For the overlap-weighted target considered in § 5.1, positivity is required only on the support of the target population.

The causal assumptions above tell us when the observed treatment arms can be used to learn about the potential outcomes. They do not yet specify what aspect of a structured potential outcome will serve as the response in our causal analysis. We turn to that representation next.

The topological representation

When an outcome is structured, a treatment may alter more than its overall level. It may change the geometry, connectivity, or organization of the object itself. A point cloud, for example, may develop or lose clusters and loops; analogous structural changes can occur in images, networks, and other complex data. Our goal is therefore to convert each structured potential outcome into a common mathematical object that preserves information about such structure while remaining amenable to statistical analysis.

Persistent homology provides the first step in this conversion. Starting from an outcome, we construct a filtration, a nested family of objects indexed by a scale parameter tt. As tt varies, topological features such as connected components, loops, and higher-dimensional cavities may appear and disappear. A persistence diagram records each such feature through its birth and death scales. Features that survive across a wide range of scales have large persistence and represent comparatively persistent topological structure, whereas short-lived features appear closer to the diagonal of the diagram.

A persistence diagram is already a useful summary of the original structured outcome, but it remains a collection of points whose cardinality may vary from one unit to another. For the functional inferential framework developed below, it is convenient to take one further step and represent each diagram by a continuous curve on a common domain. We use the power-weighted persistence silhouette for this purpose. The complete representation can therefore be viewed as the prespecified pipeline

structured outcome ⟶\longrightarrow filtration ⟶\longrightarrow persistence diagram ⟶\longrightarrow silhouette curve in C(T)C(T).

This final step is important for what follows: after the transformation, every unit is represented by a function on the same filtration interval, so causal contrasts can be defined and studied pointwise in filtration scale while retaining their functional nature.

Readers seeking further mathematical background on topological representations and their role in causal analysis may consult Kim and Lee [2026, Section 2] and {citet}`Souto and Diamantis {citep}`[2026, Sections 2.2–3]. We give here the definitions needed for the present framework.

Fix a nondegenerate compact filtration interval T=[t−,t+]T = [t_-, t_+], a homology degree dd, and an exponent r≥1r \ge1. The degree dd specifies the type of topological feature under consideration (for example, connected components when d=0d = 0 and loops when d=1d = 1), while rr determines how strongly the representation emphasizes features with greater persistence.

Let D={(bj,ej)}j=1NDD = \{(b_j,e_j)\}_{j=1}^{N_D}, bj<ej<∞b_j < e_j < \infty, be a finite multiset of off-diagonal birth–death pairs from the degree-dd persistence diagram. Associate with each pair the triangular tent function

λj(t)=max⁡{0,min⁡(t−bj,ej−t)}.\lambda_j(t) = \max\{0,\min(t-b_j,e_j-t)\}.

The tent is zero outside the lifetime of the feature, increases after its birth bjb_j, and decreases toward zero as tt approaches its death eje_j. Its contribution therefore records both where in the filtration the feature exists and how long it persists.

The power-weighted silhouette combines these individual contributions into the single curve

Φr(D)(t)=∑j(ej−bj)rλj(t)∑j(ej−bj)r.(2)\Phi_r(D)(t) = \frac{\sum_j(e_j-b_j)^r\lambda_j(t)}{\sum_j(e_j-b_j)^r}. \tag*{(2)}

Thus, Φr(D)\Phi_r(D) is a persistence-weighted average of the individual tents. The exponent rr controls their relative importance: increasing rr gives greater weight to long-lived features relative to short-lived ones. We set Φr(∅)=0\Phi_r(\varnothing) = 0, discard zero-persistence points, and treat the removal or finite truncation of essential classes as part of the prespecified representation pipeline.

The silhouette construction follows Chazal et al. [ [6] and Kim and Lee [ [10]. Each silhouette is nonnegative and 1-Lipschitz, as recorded in their Lemma 2.1. Moreover, if every persistence is bounded above by cc, then ∥Φr(D)∥∞≤c/2\|\Phi_r(D)\|_\infty\le c/2. These properties are useful because they place the topological summaries in a well-behaved common function space and later provide regularity for simultaneous inference over filtration scale.

Formally, let Gd,r:Y⟶C(T)G_{d,r} : \mathcal{Y} \longrightarrow C(T) denote the resulting prespecified Borel map from the original outcome space to the space of continuous functions on TT. Applying the same transformation to the two potential outcomes gives

ϕi,da=Gd,r(Yia),a∈{0,1},\phi_{i,d}^{a} = G_{d,r}(Y_i^a), \qquad a \in\{0,1\},

while consistency implies that the observed topological response is

ϕi,d=Gd,r(Yi)=ϕi,dAi.\phi_{i,d} = G_{d,r}(Y_i) = \phi_{i,d}^{A_i}.

The original causal problem has therefore been transformed into a functional one: unit ii has two potential topological curves, ϕi,d0\phi_{i,d}^{0} and ϕi,d1\phi_{i,d}^{1}, of which only ϕi,dAi\phi_{i,d}^{A_i} is observed.

This transformation also makes clear the scope of the causal quantities studied below. They describe treatment-induced changes in the prespecified topological summary, not treatment effects on the complete original outcome or on the persistence diagram itself; see Souto and Diamantis [ [18] for further discussion. We assume

E∥ϕda∥∞<∞,a=0,1,E\|\phi_{d}^{a}\|_\infty< \infty, \qquad a = 0,1,

so that population mean curves are well defined as Bochner expectations in C(T)C(T). Stronger moment and regularity conditions are introduced only when required by particular inferential results. Likewise, alternative vectorizations of persistence diagrams would require their own state spaces and regularity arguments and are not obtained by a purely formal substitution in (2).

With the potential outcomes now represented by the pair of curves

(ϕi,d0,ϕi,d1)∈C(T)×C(T),(\phi_{i,d}^{0},\phi_{i,d}^{1}) \in C(T) \times C(T),

the remaining question is which contrast between these curves is scientifically relevant. Just as ordinary causal inference distinguishes population-average, conditional-average, and individual treatment effects, the topological setting leads naturally to three corresponding functional targets. The next subsection defines these quantities and clarifies the different uncertainty statements associated with each.

Three targets

Once each potential outcome has been represented by a curve in C(T)C(T), treatment effects can be defined by comparing the two potential topological summaries. The appropriate comparison depends on the scientific question. We distinguish three targets: the average structural effect in the target population, the average structural effect among units with specified baseline characteristics, and the treatment-induced structural change for an individual unit. These correspond, respectively, to TATE, CTATE, and ITTE. The first two are population-level summaries: they average the contrast between the potential topological curves either marginally or conditionally on covariates. The third is fundamentally different. It is the unit-specific contrast between the two potential curves and therefore remains a random object when considering a new individual.

Definition 1 (Mean and individual topological effects). For the fixed pair (d,r)(d,r), define

ψd(t)=E{ϕd1(t)−ϕd0(t)},(3)\psi_d(t) = E\{\phi_d^1(t) - \phi_d^0(t)\}, \tag*{(3)}
τd(t,x)=E{ϕd1(t)−ϕd0(t)∣X=x},(4)\tau_d(t,x) = E\{\phi_d^1(t) - \phi_d^0(t) \mid X = x\}, \tag*{(4)}
δi,d(t)=ϕi,d1(t)−ϕi,d0(t).(5)\delta_{i,d}(t) = \phi_{i,d}^1(t) - \phi_{i,d}^0(t). \tag*{(5)}

Equation (3) restates the TATE introduced by Kim and Lee [ [10]; (4) gives its conditional-mean analogue, and (5) defines the individual contrast studied here. In each case, the effect is itself a curve indexed by the filtration scale (t)(t): its value at a given (t)(t) describes the treatment-induced difference in the topological summary at that scale. Consequently, these targets are naturally paired with different statistical procedures: population-mean curves require simultaneous confidence bands, whereas individual contrasts require predictive intervals.

The identification of conditional means follows the standard argument of Kim and Lee [ [10]. Choosing measurable conditional-mean versions μa,d(t,x)=E{ϕda(t)∣X=x}\mu_{a,d}(t,x) = E\{\phi_d^a(t) \mid X = x\} with continuous paths in tt, standard causal assumptions identify these quantities from the observed data: for PXP_X-almost every xx,

μa,d(t,x)=E{ϕd(t)∣X=x,A=a}.(6)\mu_{a,d}(t,x) = E\{\phi_d(t) \mid X = x, A = a\}. \tag*{(6)}

The identification in (6) holds PXP_X-almost everywhere. Inference at a prespecified covariate value xx requires an identified version of the conditional mean, justified by suitable continuity or regression-model assumptions. In particular, inference at points outside the covariate support is necessarily model-dependent. Because individual silhouettes are 1-Lipschitz, their conditional means μa,d(⋅,x)\mu_{a,d}(\cdot,x) can be chosen 1-Lipschitz in tt, implying that ψd\psi_d, τd(⋅,x)\tau_d(\cdot,x), and δi,d\delta_{i,d} are 2-Lipschitz.

Crucially, the joint distribution of the individual contrast δi,d\delta_{i,d} is not point-identified from the arm-specific marginals alone. However, the prediction bands developed in § 4 remain valid across every valid coupling of these marginal distributions that satisfies the stated assumptions.

AIPW estimation and independent conformal calibration

We use the AIPW score and estimator of Kim and Lee [ [10], with the standard cross-fitting extension described in their Remark 2 and in Chernozhukov et al. [ [7]. Write η=(π,μ0,d,μ1,d)\eta= (\pi,\mu_{0,d},\mu_{1,d}) and define

References

  1. [1]Konrad Abramowicz, Alessia Pini, Lina Schelin, Sara Sjostedt de Luna, Aymeric Stamm, and Simone Vantini. Domain selection and familywise error rate for functional data: A unified framework. Biometrices, 79(2):1119–1132, 2023. doi: 10.1111/biom.13669. URL https://doi.org/10.1111/biom.13669.
  2. [2]Henry Adams, Tegan Emerson, Michael Kirby, Rachel Neville, Chris Peterson, Patrick Shipman, Sofya Chepushtanova, Eric Hanson, Francis Motta, and Lori Ziegelmeier. Persistence images: A stable vector representation of persistent homology. Journal of Machine Learning Research, 18 (8):1–35, 2017. URL http://jmlr.org/papers/v18/16-337.html. Henry Adams, Tegan Emerson, Michael Kirby, Rachel Neville, Chris Peterson, Patrick Shipman, Sofya Chepushtanova, Eric Hanson, Francis Motta, and Lori Ziegelmeier. Erratum to Persistence Images: A Stable Vector Representation of Persistent Homology, n.d. URL https://jmlr.org/papers/volume18/16-337/pi_erratum.pdf. Erratum supplied with the 2017 JMLR article.
  3. [3]Ahmed M. Alaa, Zaid Ahmad, and Mark van der Laan. Conformal meta-learners for predictive inference of individual treatment effects. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 47682–47703. Curran Associates, Inc., 2023. doi: 10.52202/075280-2065. URL https://proceedings.neurips.cc/paper_files/paper/2023/file/94ab02a30b0e4a692a42ccd0b4c55399-Paper-Conference.pdf.DOI
  4. [4]Heejung Bang and James M. Robins. Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4):962–973, 2005. doi: 10.1111/j.1541-0420.2005.00377.x. URL https://doi.org/10.1111/j.1541-0420.2005.00377.x.
  5. [5]Rina Foygel Barber, Emmanuel J. Candès, Aaditya Ramdas, and Ryan J. Tibshirani. The limits of distribution-free conditional predictive inference. Information and Inference: A Journal of the IMA, 10(2):455–482, 2021. doi: 10.1093/imaiai/iaaa017. URL https://doi.org/10.1093/imaiai/iaaa017.
  6. [6]Frédéric Chazal, Brittany Terese Fasy, Fabrizio Lecci, Alessandro Rinaldo, and Larry Wasserman. Stochastic convergence of persistence landscapes and silhouettes. Journal of Computational Geometry, 6(2):140–161, 2015. doi: 10.20382/jocg.v6i2a8. URL https://doi.org/10.20382/jocg.v6i2a8.
  7. [7]Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1):C1–C68, 2018. doi: 10.1111/ectj.12097. URL https://doi.org/10.1111/ectj.12097.
  8. [8]Jacopo Diquigiovanni, Matteo Fontana, and Simone Vantini. Conformal prediction bands for multivariate functional data. Journal of Multivariate Analysis, 189:104879, May 2022. ISSN 0047-259X. doi: 10.1016/j.jmva.2021.104879. URL http://dx.doi.org/10.1016/j.jmva.2021.104879.
  9. [9]J. Randolph Hecht, Edith Mitchell, Tarek Chidiac, Carroll Scroggin, Christopher Hagenstad, David Spigel, John Marshall, Allen Cohn, David McCollum, Philip Stella, Robert Deeter, Seta Shahin, and Rafael G. Amado. A randomized phase IIIB trial of chemotherapy, bevacizumab, and panitumumab compared with chemotherapy and bevacizumab alone for metastatic colorectal cancer. Journal of Clinical Oncology, 27(5):672–680, 2009. doi: 10.1200/JCO.2008.19.8135. URL https://doi.org/10.1200/JCO.2008.19.8135.
  10. [10]Kwangho Kim and Hajin Lee. Topological causal effects. In The Fourteenth International Conference on Learning Representations, 2026. doi: 10.48550/arXiv.2603.02289. URL https://proceedings.iclr.cc/paper_files/paper/2026/hash/4be03be633b57362bb034547490ee196-Abstract-Conference.html. Also available as arXiv:2603.02289 [stat.ME].DOI
  11. [11]Lihua Lei and Emmanuel J. Candès. Conformal inference of counterfactuals and individual treatment effects. Journal of the Royal Statistical Society Series B: Statistical Methodology, 83(5):911–938, November 2021. ISSN 1467-9868. doi: 10.1111/rssb.12445. URL http://dx.doi.org/10.1111/rssb.12445.
  12. [12]Fan Li, Kari Lock Morgan, and Alan M. Zaslavsky. Balancing covariates via propensity score weighting. Journal of the American Statistical Association, 113(521):390–400, 2018. doi: 10.1080/01621459.2016.1260466. URL https://doi.org/10.1080/01621459.2016.1260466.
  13. [13]Dominik Liebl and Matthew Reimherr. Fast and fair simultaneous confidence bands for functional parameters. Journal of the Royal Statistical Society Series B: Statistical Methodology, 85(3):842–868, July 2023. ISSN 1467-9868. doi: 10.1093/jrsssb/qkad026. URL http://dx.doi.org/10.1093/jrsssb/qkad026.
  14. [14]A. Pini and S. Vantini. Interval-wise testing for functional data. Journal of Nonparametric Statistics, 29(2):407–424, April 2017. ISSN 1029-0311. doi: 10.1080/10485252.2017.1306627. URL http://dx.doi.org/10.1080/10485252.2017.1306627.
  15. [15]Maresa Schröder, Dennis Frauen, Jonas Schweithal, Konstantin Hess, Valenty n Melnychuk, and Stefan Feuerriegel. Conformal prediction for causal effects of continuous treatments. In Advances in Neural Information Processing Systems, volume 38, pages 1500–1539, 2025. doi: 10.52202/085713-0057. URL https://proceedings.neurips.cc/paper_files/paper/2025/hash/02435e97d53c423a68322cb3c42403fd-Abstract-Conference.html. Also available as arXiv:2407.03094v4 [stat.ME].DOI
  16. [16]Vira Semenova and Victor Chernozhukov. Debiased machine learning of conditional average treatment effects and other causal functions. The Econometrics Journal, 24(2):264–289, 2021. doi: 10.1093/ectj/utaa027. URL https://doi.org/10.1093/ectj/utaa027.
  17. [17]Eduardo Soares, Plamen Angelov, Sarah Bia so, Michele Higa Froes, and Daniel Kanda Abe. SARS-CoV-2 CT-scan dataset: A large dataset of real patients CT scans for SARS-CoV-2 identification. medRxiv, 2020. doi: 10.1101/2020.04.24.20078584. URL https://www.medrxiv.org/content/10.1101/2020.04.24.20078584v2. Version 2 (dataset paper).DOI
  18. [18]Hugo Gobato Souto and Ioannis Diamantis. A mathematical framework for topological causal data analysis, 2026. URL https://arxiv.org/abs/2607.28161. arXiv:2607.28161 [stat.ME].
  19. [19]Ryan J. Tibshirani, Rina Foygel Barber, Emmanuel Candès, and Aaditya Ramdas. Conformal prediction under covariate shift. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper_files/paper/2019/hash/8fb21ee7a2207526da55a679f0332de2-Abstract.html.arxiv.org/abs/1904.06019
  20. [20]Aad W. van der Vaart and Jon A. Wellner. Weak Convergence and Empirical Processes: With Applications to Statistics. Springer, New York, 1996. doi: 10.1007/978-1-4757-2545-2. URL https://doi.org/10.1007/978-1-4757-2545-2.

Paper details

vlnc:W_sg4kmykk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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