pith. sign in

arxiv: 2509.00472 · v3 · submitted 2025-08-30 · 📊 stat.ML · cs.LG· math.ST· stat.TH

Partially Functional Dynamic Backdoor Diffusion-based Causal Model

Xinwen Liu , Lei Qian , Song Xi Chen , Niansheng Tang This is my paper

Pith reviewed 2026-05-18 19:39 UTC · model grok-4.3

classification 📊 stat.ML cs.LGmath.STstat.TH
keywords causal inferencediffusion modelsspatio-temporal datafunctional databackdoor adjustmentstructural causal modelscounterfactual estimationdynamic confounding
0
0 comments X

The pith

A diffusion-based causal model preserves effects under basis expansion for data with dynamic spatio-temporal confounders.

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

This paper introduces a generative model for estimating causal effects in complex settings where unmeasured confounders vary over space and time and data comes at multiple resolutions. It formalizes a structural causal model using conditional autoregressive processes for the confounders and represents functional observations through basis expansions whose coefficients function as ordinary nodes in the causal graph. The model embeds backdoor adjustment directly into a diffusion process and supplies error bounds on the resulting counterfactuals. Readers interested in environmental or health data would care because existing causal tools break down when confounders are dynamic and variables are functions rather than fixed scalars, as shown in the air pollution example.

Core claim

The Partially Functional Dynamic Backdoor Diffusion-based Causal Model formalizes a novel structural causal model that captures spatio-temporal dependencies in latent confounders through conditional autoregressive processes, represents functional variables via basis expansion coefficients treated as standard graph nodes, and integrates valid backdoor adjustment into a diffusion-based generative process, while providing theoretical guarantees on the preservation of causal effects under basis expansion and error bounds for counterfactual estimates.

What carries the argument

Integration of valid backdoor adjustment into a diffusion-based generative process on a structural causal model whose functional variables are represented by basis-expansion coefficients acting as ordinary graph nodes.

If this is right

  • The model supplies theoretical guarantees that causal effects remain unchanged when functional variables are replaced by their basis-expansion coefficients.
  • Error bounds are available for the counterfactual estimates produced by the diffusion process.
  • Performance exceeds prior methods on both synthetic benchmarks and a real air-pollution dataset for observational, interventional, and counterfactual tasks.
  • Non-stationary and multi-resolution spatio-temporal systems become tractable for causal queries.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the basis-expansion representation holds, the same machinery could apply to other functional data domains such as time-series sensor readings or image-based covariates.
  • The conditional autoregressive structure on confounders suggests a route to testing sensitivity to the specific autoregressive order chosen.
  • Improved counterfactual accuracy in environmental applications could support better-targeted interventions for pollution control.

Load-bearing premise

The chosen conditional autoregressive processes fully capture the spatio-temporal dynamics of the latent confounders and the basis-expansion coefficients preserve all relevant causal relations when treated as ordinary nodes.

What would settle it

A simulation study in which the ground-truth causal effect is known exactly, yet the model's estimated counterfactual distribution lies outside the derived error bounds, would falsify the preservation guarantee.

Figures

Figures reproduced from arXiv: 2509.00472 by Lei Qian, Niansheng Tang, Song Xi Chen, Xinwen Liu.

Figure 1
Figure 1. Figure 1: DAG with three nodes (left) and SCM with three exogenous and endogenous nodes (right) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: PFST-DSCM with 33 exogenous and endogenous nodes (where nodes [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
read the original abstract

Causal inference in spatio-temporal settings is critically hindered by unmeasured confounders with complex spatio-temporal dynamics and the prevalence of multi-resolution data. While diffusion models present a promising avenue for estimating structural causal models, existing approaches are limited by assumptions of causal sufficiency or static confounding, failing to capture the region-specific, temporally dependent nature of real-world latent variables or to directly handle functional variables. We bridge this gap by introducing the Partially Functional Dynamic Backdoor Diffusion-based Causal Model (PFD-BDCM), a unified generative framework designed to simultaneously tackle causal inference with dynamic confounding and functional data. Our approach formalizes a novel structural causal model that captures spatio-temporal dependencies in latent confounders through conditional autoregressive processes, represents functional variables via basis expansion coefficients treated as standard graph nodes, and integrates valid backdoor adjustment into a diffusion-based generative process. We provide theoretical guarantees on the preservation of causal effects under basis expansion and derive error bounds for counterfactual estimates. Experiments on synthetic data and a real-world air pollution case study demonstrate that PFD-BDCM outperforms existing methods across observational, interventional, and counterfactual queries. This work provides a rigorous and practical tool for robust causal inference in complex spatio-temporal systems characterized by non-stationarity and multi-resolution data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces the Partially Functional Dynamic Backdoor Diffusion-based Causal Model (PFD-BDCM), a generative framework for causal inference in spatio-temporal settings with unmeasured dynamic confounders and functional/multi-resolution data. It formalizes a novel SCM that models spatio-temporal latent confounding via conditional autoregressive processes, represents functional variables through basis expansion coefficients treated as ordinary graph nodes, and embeds valid backdoor adjustment inside a diffusion generative process. The authors claim theoretical guarantees on preservation of causal effects under basis expansion together with error bounds for counterfactual estimates, and report that PFD-BDCM outperforms existing methods on synthetic data and a real-world air-pollution case study across observational, interventional, and counterfactual queries.

Significance. If the representation of basis coefficients as graph nodes preserves identifiability and the claimed error bounds hold under the autoregressive confounding dynamics, the work would supply a practical tool for causal queries in non-stationary spatio-temporal systems that current diffusion-based or static-confounder methods do not address. The combination of functional data handling, dynamic backdoor adjustment, and diffusion generation is novel and could influence both causal inference and generative modeling literature.

major comments (2)
  1. [Abstract / SCM formalization] Abstract and the paragraph on formalization of the novel SCM: the assertion that basis-expansion coefficients can be treated as standard graph nodes while preserving causal effects under conditional autoregressive latent confounding lacks the function-space conditions (completeness of the basis, orthogonality with respect to the measure induced by the autoregressive kernel) needed to ensure the backdoor criterion survives truncation error and the dynamic component. Without these conditions the claimed theoretical guarantees on preservation of causal effects are not yet established.
  2. [Theoretical guarantees section] The derivation of error bounds for counterfactual estimates (mentioned in the abstract) must be checked against the interaction between the finite basis truncation and the autoregressive process parameters; if the bounds are derived under an assumption that the truncation error is independent of the latent dynamics, that assumption needs explicit justification because it is load-bearing for the central identifiability claim.
minor comments (1)
  1. [Notation / Preliminaries] Notation for the conditional autoregressive process and the diffusion schedule should be introduced with a single consistent table or diagram early in the paper to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the theoretical foundations of our work. We address each major comment below and will revise the manuscript accordingly to strengthen the formalization and guarantees.

read point-by-point responses

  1. Referee: [Abstract / SCM formalization] Abstract and the paragraph on formalization of the novel SCM: the assertion that basis-expansion coefficients can be treated as standard graph nodes while preserving causal effects under conditional autoregressive latent confounding lacks the function-space conditions (completeness of the basis, orthogonality with respect to the measure induced by the autoregressive kernel) needed to ensure the backdoor criterion survives truncation error and the dynamic component. Without these conditions the claimed theoretical guarantees on preservation of causal effects are not yet established.

    Authors: We appreciate this observation. Our derivation of causal effect preservation under basis expansion implicitly relies on a complete orthogonal basis with respect to the inner product induced by the autoregressive measure, which ensures the backdoor criterion is preserved after truncation. However, we agree that these function-space conditions were not stated with sufficient explicitness in the SCM formalization. In the revised manuscript, we will add a dedicated paragraph in the theoretical section specifying completeness, orthogonality with respect to the autoregressive kernel, and how these guarantee survival of the backdoor criterion under truncation error and dynamic confounding. revision: yes

  2. Referee: [Theoretical guarantees section] The derivation of error bounds for counterfactual estimates (mentioned in the abstract) must be checked against the interaction between the finite basis truncation and the autoregressive process parameters; if the bounds are derived under an assumption that the truncation error is independent of the latent dynamics, that assumption needs explicit justification because it is load-bearing for the central identifiability claim.

    Authors: Thank you for this precise comment. The error bounds in Section 4 are derived under the orthogonality of the chosen basis, which renders the truncation error uncorrelated with the latent autoregressive dynamics. We acknowledge that the interaction with autoregressive parameters was not analyzed in full detail. In the revision, we will expand the proof to explicitly justify the independence assumption via the basis properties or, alternatively, derive refined bounds that incorporate any residual dependence on the autoregressive coefficients, thereby making the identifiability claim more robust. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation chain remains self-contained

full rationale

The abstract and available description formalize a novel SCM using conditional autoregressive processes for spatio-temporal latent confounders, treat basis expansion coefficients as graph nodes, and embed backdoor adjustment within a diffusion generative process, followed by claimed theoretical guarantees on causal effect preservation and error bounds for counterfactuals. No equations or self-citations are exhibited that reduce the guarantees, bounds, or predictions directly to fitted parameters or prior inputs by construction. The central claims rest on the formalization and derivation steps rather than tautological renaming or load-bearing self-reference, making the chain independent of its own outputs.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claim rests on the representation of latent confounders as conditional autoregressive processes and the treatment of basis-expansion coefficients as ordinary nodes in the causal graph; these modeling choices are introduced without independent external validation beyond the claimed theoretical guarantees.

free parameters (2)
  • autoregressive process parameters
    Parameters governing the conditional autoregressive dynamics of latent confounders are estimated from data and enter the generative process.
  • diffusion model parameters
    Parameters of the diffusion-based generative process are fitted and used to produce counterfactual estimates.
axioms (2)
  • domain assumption Causal effects are preserved under basis expansion of functional variables
    Invoked when treating basis coefficients as standard graph nodes; location: abstract description of the novel SCM.
  • domain assumption Valid backdoor adjustment can be integrated into the diffusion generative process
    Central modeling choice that enables the causal claims.
invented entities (1)
  • PFD-BDCM framework no independent evidence
    purpose: Unified generative model for dynamic confounding and functional causal inference
    New postulated model whose validity is supported only by internal theoretical guarantees and experiments described in the abstract.

pith-pipeline@v0.9.0 · 5754 in / 1565 out tokens · 37286 ms · 2026-05-18T19:39:15.104336+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    D., Imbens, G

    Angrist, J. D., Imbens, G. W., and Rubin, D. B. (1996). Identification of causal effects using instrumental variables. Journal of the American Statistical Association , 91:444--455

    work page 1996

  2. [2]

    Besag, J. (1974). Spatial interaction and the statistical analysis of lattice systems. Journal of the royal statistical society series b-methodological , 36:192--225

    work page 1974

  3. [3]

    Chao, P., Blöbaum, P., and Kasiviswanathan, S. P. (2023). Interventional and counterfactual inference with diffusion models

    work page 2023

  4. [4]

    Efficacy of china's clean air actions to tackle pm2.5 pollution between 2013 and 2020

    Geng, G., Liu, Y., Liu, Y., and et al (2024). Efficacy of china's clean air actions to tackle pm2.5 pollution between 2013 and 2020. Nature Geoscience , 17(10)

    work page 2024

  5. [5]

    Assessment to china's recent emission pattern shifts

    Guan, Y., Shan, Y., Huang, Q., and et al (2021). Assessment to china's recent emission pattern shifts. Earth's Future , 9(11)

    work page 2021

  6. [6]

    Hill, J. L. (2011). Bayesian nonparametric modeling for causal inference. Journal of Computational and Graphical Statistics , 20(1):217--240

    work page 2011

  7. [7]

    Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems , 33:6840--6851

    work page 2020

  8. [8]

    Imbens, G. W. and Rubin, D. B. (2015). Causal Inference for Statistics, Social, and Biomedical Sciences . Cambridge University Press

    work page 2015

  9. [9]

    LaLonde, R. J. (1986). Evaluating the econometric evaluations of training programs with experimental data. The American Economic Review , pages 604--620

    work page 1986

  10. [10]

    Persistent growth of anthropogenic nmvoc emissions in china during 1990-2017: Dynamics, speciation, and ozone formation potentials

    Li, M., Zhang, Q., Zheng, B., and et al (2019). Persistent growth of anthropogenic nmvoc emissions in china during 1990-2017: Dynamics, speciation, and ozone formation potentials. Atmospheric Chemistry and Physics , 19(13):8897--8913

    work page 2019

  11. [11]

    Sample-efficient reinforcement learning via counterfactual-based data augmentation

    Lu, C., Huang, B., Wang, K., and et al (2020). Sample-efficient reinforcement learning via counterfactual-based data augmentation

    work page 2020

  12. [12]

    Nasr-Esfahany, A., Alizadeh, M., and Shah, D. (2023). Counterfactual identifiability of bijective causal models

    work page 2023

  13. [13]

    and Kiciman, E

    Nasr-Esfahany, A. and Kiciman, E. (2023). Counterfactual (non-) identifiability of learned structural causal models

    work page 2023

  14. [14]

    Ozcan, B. (2013). The nexus between carbon emissions, energy consumption and economic growth in middle east countries: A panel data analysis. Energy Policy , 62:1138--1147

    work page 2013

  15. [15]

    Pearl, J. (2009). Causal inference in statistics: An overview. Statistics Surveys , 3:96--146

    work page 2009

  16. [16]

    Pearl, J., Glymour, M., and Jewell, N. (2016). Causal inference in statistics: A primer. Wiley

    work page 2016

  17. [17]

    Causal discovery with continuous additive noise models

    Peters, J., Mooij, J., Janzing, D., and et al (2013). Causal discovery with continuous additive noise models. Journal of Machine Learning Research , 15

    work page 2013

  18. [18]

    Rosenbaum, P. R. and Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika , 70:41--55

    work page 1983

  19. [19]

    D., and Sontag, D

    Shalit, U., Johansson, F. D., and Sontag, D. (2017). Estimating individual treatment effect: Generalization bounds and algorithms. International Conference on Machine Learning , pages 3076--3085

    work page 2017

  20. [20]

    Shimizu, T. (2023). Diffusion model in causal inference with unmeasured confounders. In 2023 IEEE Symposium Series on Computational Intelligence (SSCI) , pages 683--688

    work page 2023

  21. [21]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and et al (2015). Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of Machine Learning Research , 37:2256--2265

    work page 2015

  22. [22]

    Song, J., Meng, C., and Ermon, S. (2021). Denoising diffusion implicit models. International Conference on Learning Representations

    work page 2021

  23. [23]

    Song, X., Tang, N., and Chow, S. (2012). A bayesian approach for generalized random coefficient structural equation models for longitudinal data with adjacent time effects. Computational Statistics & Data Analysis , 56(12):4190--4203

    work page 2012

  24. [24]

    Strobl, E. V. and Lasko, T. A. (2023). Identifying patient-specific root causes with the heteroscedastic noise model. Journal of Computational Science , 72:102099

    work page 2023

  25. [25]

    G., and Zhu, H

    Tang, N., Chow, S.-M., Ibrahim, J. G., and Zhu, H. (2017). Bayesian sensitivity analysis of a nonlinear dynamic factor analysis model with nonparametric prior and possible nonignorable missingness. Psychometrika , 82(4):875–903

    work page 2017

  26. [26]

    Xu, J., Guan, Y., Oldfield, J., Guan, D., and Shan, Y. (2024). China carbon emission accounts 2020-2021. Applied Energy , 360

    work page 2024

  27. [27]

    Zhu, Y., Liang, Y., and Chen, S. X. (2021). Assessing local emission for air pollution via data experiments. Atmospheric Environment , 252:118323

    work page 2021