arxiv: 2605.00398 · v1 · submitted 2026-05-01 · 💻 cs.LG · physics.ao-ph· stat.ML

Recognition: unknown

M-CaStLe: Uncovering Local Causal Structures in Multivariate Space-Time Gridded Data

J. Jake Nichol , Michael Weylandt , G. Matthew Fricke , Jhayron Perez-Carrasquilla , Melanie E. Moses

Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:12 UTC · model grok-4.3

classification 💻 cs.LG physics.ao-phstat.ML

keywords causal discoveryspace-time datagridded datamultivariate analysisstencil learningcausal graphsatmospheric modeling

0 comments

The pith

M-CaStLe extends local causal discovery to jointly recover within-variable and cross-variable structures in multivariate space-time grids.

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

The paper presents M-CaStLe as a generalization of the earlier CaStLe meta-algorithm to multivariate settings. It tackles the problem of causal graph discovery when the number of grid cells greatly exceeds the number of time observations per cell. By restricting candidate parents to a fixed-size space-time neighborhood around each location and pooling replicates across space, the method increases the effective sample size enough to make discovery tractable. The resulting stencil graph is further split into reaction and spatial components to support interpretation in applications such as atmospheric chemistry and climate dynamics.

Core claim

M-CaStLe generalizes the local embedding and parent-identification phases of CaStLe to jointly model local within-variable and cross-variable space-time causal structures in gridded data. By constraining candidate parents to a constant-size space-time neighborhood and pooling spatial replicates, M-CaStLe increases effective sample size to make discovery tractable in high-dimensional settings. The resulting multivariate stencil graph is decomposed into reaction and spatial graphs to aid interpretation. In controlled benchmarks it recovers ground-truth structures more accurately than alternatives; in real-world case studies it identifies physically meaningful dynamics such as phase-dependent耦合

What carries the argument

The multivariate space-time stencil graph that encodes local causal parents across multiple variables and neighboring grid cells.

If this is right

In a multivariate vector-autoregression benchmark with known ground truth, M-CaStLe recovers the causal structure more accurately than prior approaches.
On an advective-diffusive-reaction PDE verification problem, the method reconstructs the derived physical reference structure.
In a low-temporal-sample atmospheric chemistry case study, M-CaStLe identifies important physical dynamics.
Applied to El Niño Southern Oscillation reanalysis data, the algorithm detects phase-dependent ocean-atmosphere coupling.

Where Pith is reading between the lines

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

The reaction-versus-spatial decomposition could be used to initialize hybrid physics-informed causal models that respect both data-driven links and known governing equations.
If the locality assumption is only approximately true, the method might still serve as a fast screening step before applying more expensive global causal searches on a reduced variable set.
The pooling of spatial replicates suggests a natural extension to other replicated spatiotemporal domains such as sensor arrays or image sequences where stationarity holds locally.

Load-bearing premise

Space-time locality and stationarity hold so that candidate parents can be restricted to a fixed neighborhood and spatial replicates can be pooled without bias or missed long-range effects.

What would settle it

A controlled multivariate space-time dataset containing known long-range causal links or strong non-stationarity in which M-CaStLe misses true edges that a global method recovers.

Figures

Figures reproduced from arXiv: 2605.00398 by G. Matthew Fricke, Jhayron Perez-Carrasquilla, J. Jake Nichol, Melanie E. Moses, Michael Weylandt.

**Figure 2.** Figure 2: A schematic diagram of the input, computational phases, and output of M-CaStLe. Similar to [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Results of the VAR benchmark analysis comparing M-CaStLe, Cartesian-CaStLe, and direct non [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

**Figure 4.** Figure 4: Angle estimation performance of M-CaStLe on the advection–diffusion–reaction PDE verification [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

**Figure 5.** Figure 5: Causal graphs recovered from six-hourly E3SMv2-SPA output over the Pinatubo plume region ( [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

**Figure 6.** Figure 6: M-CaStLe results obtained using DYNOTEARS illustrating causal relationships between sea sur [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

**Figure 7.** Figure 7: Extent of number of links with 30 Replicates. Horizontal bars show, for each Number of Variables [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗

**Figure 8.** Figure 8: The relationship between link coefficients and the number of links present. As the number of [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗

**Figure 9.** Figure 9: In simple chains of multivariate stencils, even with an extremely large number of variables, recall [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗

**Figure 10.** Figure 10: Time series causal graph for the ADR model. Nodes represent species [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗

**Figure 11.** Figure 11: Histogram showing the frequency distribution of [PITH_FULL_IMAGE:figures/full_fig_p032_11.png] view at source ↗

read the original abstract

Causal graph discovery for space-time systems is challenging in high-dimensional gridded data, which often has many more grid cells than temporal observations per cell. The Causal Space-Time Stencil Learning (CaStLe) meta-algorithm was developed to address that niche under space-time locality and stationarity assumptions, but it is currently limited to univariate analyses. In this work, we present M-CaStLe. M-CaStLe generalizes the local embedding and parent-identification phases of CaStLe to jointly model local within-variable and cross-variable space-time causal structures in gridded data. Like CaStLe, by constraining candidate parents to a constant-size space-time neighborhood and pooling spatial replicates, M-CaStLe increases effective sample size to make discovery tractable in high-dimensional settings. We further decompose the resulting multivariate stencil graph into reaction and spatial graphs to aid interpretation in complex settings. We study M-CaStLe in four settings: a multivariate space-time vector autoregression benchmark with known ground truth, an advective-diffusive-reaction partial differential equation verification problem with derived physical reference structure, an atmospheric chemistry case study in a low-temporal-sample regime, and an El Ni\~{n}o Southern Oscillation study on reanalysis data, identifying phase-dependent ocean--atmosphere coupling. Across these settings, M-CaStLe more accurately recovers multivariate causal structure in controlled settings and identifies important physical dynamics in real-world case studies. Overall, M-CaStLe advances causal discovery for multivariate space-time systems while retaining interpretability at the grid level.

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.

Desk Editor's Note private letter to a colleague

M-CaStLe gives a clean multivariate extension of CaStLe plus a reaction-spatial split, but the ENSO results rest on a stationarity assumption that the paper does not test.

read the letter

M-CaStLe generalizes the local embedding and parent-identification steps from the earlier CaStLe work so they handle multiple variables at once, and it splits the recovered stencil into separate reaction and spatial subgraphs. That split is the part that actually helps interpretation in gridded physical data. The controlled tests back this up: on the multivariate VAR benchmark and the advective-diffusive-reaction PDE, the method recovers the known structure more accurately than the univariate baseline, which is the clearest evidence the paper supplies. Those experiments match the method's assumptions, so the gains look real there. The atmospheric chemistry and ENSO case studies are presented as finding meaningful dynamics, including phase-dependent coupling in the ocean-atmosphere system. The trouble is the pooling of spatial replicates. The algorithm needs that step to get enough samples, but it only works if the local causal structure is the same everywhere. ENSO is a classic case where coupling strength varies by region, so the pooled graph could be an average that does not match any actual location. The paper does not show a sensitivity check or a comparison against non-pooled local fits, so the real-data claims stay weaker than the benchmark ones. This is aimed at researchers who already work with space-time gridded data in climate or atmospheric modeling and want grid-level causal graphs they can read. It is worth sending to referees because the multivariate extension and the benchmark results are substantive, but any review should press on whether the stationarity assumption holds in the heterogeneous cases.

Referee Report

2 major / 2 minor

Summary. The paper introduces M-CaStLe as a multivariate generalization of the CaStLe meta-algorithm for causal discovery in high-dimensional space-time gridded data. It extends local embedding and parent identification to jointly recover within-variable and cross-variable structures by restricting candidate parents to a fixed-size space-time neighborhood and pooling spatial replicates under locality and stationarity assumptions. The resulting multivariate stencil graph is decomposed into reaction and spatial components for interpretability. Evaluations cover a multivariate VAR benchmark with ground truth, an advective-diffusive-reaction PDE with derived reference structure, an atmospheric chemistry case study, and an ENSO reanalysis application claiming phase-dependent ocean-atmosphere coupling. The central claim is improved multivariate causal recovery in controlled settings and identification of important physical dynamics in real data.

Significance. If the empirical results and assumptions hold, the work meaningfully extends causal discovery tools to multivariate spatio-temporal grids, a setting where sample size is typically limited by the number of time steps. The use of controlled benchmarks with known ground truth and the decomposition into interpretable reaction/spatial graphs are clear strengths that support potential utility in physical sciences applications. The approach retains the locality-based tractability of the original CaStLe while adding cross-variable modeling.

major comments (2)

[Methods (locality and pooling construction) and ENSO case study] The pooling of spatial replicates to boost effective sample size (described in the generalization of the parent-identification phase) is valid only under the assumption that the underlying process is spatially stationary, so that all grid cells share identical local causal structure. This assumption is load-bearing for the real-world claims: the ENSO reanalysis case study is known to exhibit spatially heterogeneous and phase-dependent coupling strengths, so pooling risks producing averaged stencil graphs that do not correspond to any actual local mechanism. The headline assertion that M-CaStLe “identifies important physical dynamics” therefore rests on unverified stationarity in the real-data sections.
[Experiments (ENSO reanalysis subsection)] In the ENSO application, the recovered phase-dependent ocean-atmosphere couplings are presented as evidence of physical insight, yet no sensitivity analysis, stationarity diagnostic, or comparison against non-pooled local fits is reported. Without such checks, it is impossible to determine whether the multivariate stencil graphs reflect true local dynamics or artifacts of the pooling step, directly weakening the cross-setting claim of accurate recovery and physical relevance.

minor comments (2)

The decomposition of the stencil graph into separate reaction and spatial graphs is introduced without an explicit algorithmic description or pseudocode; a short formal definition or worked example would improve reproducibility.
Figure captions and axis labels in the benchmark results could more explicitly indicate which metrics correspond to within-variable versus cross-variable recovery to aid interpretation of the multivariate improvement.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important considerations regarding the stationarity assumptions underlying our pooling approach and the need for additional validation in the ENSO case study. We address each major comment point-by-point below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Methods (locality and pooling construction) and ENSO case study] The pooling of spatial replicates to boost effective sample size (described in the generalization of the parent-identification phase) is valid only under the assumption that the underlying process is spatially stationary, so that all grid cells share identical local causal structure. This assumption is load-bearing for the real-world claims: the ENSO reanalysis case study is known to exhibit spatially heterogeneous and phase-dependent coupling strengths, so pooling risks producing averaged stencil graphs that do not correspond to any actual local mechanism. The headline assertion that M-CaStLe “identifies important physical dynamics” therefore rests on unverified stationarity in the real-data sections.

Authors: We agree that the spatial stationarity assumption is fundamental to the validity of pooling spatial replicates, as described in the methods. The manuscript states this assumption explicitly when generalizing CaStLe. For the ENSO application, separate analyses per phase address temporal non-stationarity to some degree, but we acknowledge that spatial heterogeneity could lead to averaged structures. In the revision, we will expand the discussion of this assumption's implications for the ENSO results and add a sensitivity analysis comparing pooled stencil graphs against those obtained from spatial subregions to assess robustness. revision: yes
Referee: [Experiments (ENSO reanalysis subsection)] In the ENSO application, the recovered phase-dependent ocean-atmosphere couplings are presented as evidence of physical insight, yet no sensitivity analysis, stationarity diagnostic, or comparison against non-pooled local fits is reported. Without such checks, it is impossible to determine whether the multivariate stencil graphs reflect true local dynamics or artifacts of the pooling step, directly weakening the cross-setting claim of accurate recovery and physical relevance.

Authors: The referee is correct that the current manuscript lacks these diagnostics for the ENSO results. To address this directly, the revised version will include a stationarity diagnostic (e.g., spatial homogeneity tests on recovered causal strengths) and, where sample sizes permit, comparisons of pooled results against non-pooled local fits on representative grid cells. These additions will clarify whether the identified phase-dependent couplings are robust or potentially influenced by pooling, thereby supporting a more cautious interpretation of physical relevance. revision: yes

Circularity Check

0 steps flagged

M-CaStLe generalizes prior CaStLe under stated locality assumptions; no derivation reduces to self-fit or load-bearing self-citation.

full rationale

The manuscript presents M-CaStLe as a direct generalization of the univariate CaStLe meta-algorithm by extending local embedding and parent-identification phases to the multivariate case, while retaining the same space-time neighborhood constraint and spatial pooling. These steps are algorithmic extensions justified by the same explicit assumptions (locality and stationarity) rather than by any equation that re-derives its own inputs. Claims of superior recovery rest on separate controlled benchmarks (VAR with known ground truth, PDE with derived reference structure) and case studies whose evaluation metrics are computed independently of the fitting procedure itself. No self-citation chain is invoked to establish uniqueness or to substitute for empirical verification, and no prediction is statistically forced by a parameter fit to the identical data. The work therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumptions of space-time locality and stationarity inherited from the prior CaStLe paper; no new free parameters or invented entities are described in the abstract.

axioms (1)

domain assumption Space-time locality and stationarity assumptions allow restriction of candidate parents to a fixed-size neighborhood and pooling of spatial replicates.
Invoked to make high-dimensional discovery tractable; stated in the abstract as the basis for the meta-algorithm.

pith-pipeline@v0.9.0 · 5614 in / 1253 out tokens · 25473 ms · 2026-05-09T20:12:38.017174+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

290 extracted references · 233 canonical work pages · 3 internal anchors

[1]

Journal of climate , volume=

The effective number of spatial degrees of freedom of a time-varying field , author=. Journal of climate , volume=
[2]

Bellman , title =

Richard E. Bellman , title =. 1957 , address =

1957
[3]

Estimating High-Dimensional Directed Acyclic Graphs with the

Markus Kalisch and Peter B. Estimating High-Dimensional Directed Acyclic Graphs with the. 2007 , JOURNAL =

2007
[4]

Inferring causation from time series in earth system sciences.Nature Com- munications, 10(1):2553, 2019

Runge, Jakob and Bathiany, Sebastian and Bollt, Erik and Camps-Valls, Gustau and Coumou, Dim and Deyle, Ethan and Glymour, Clark and Kretschmer, Marlene and Mahecha, Miguel D. and Munoz-Mari, Jordi and Nes, Egbert H. van and Peters, Jonas and Quax, Rick and Reichstein, Markus and Scheffer, Marten and Scholkopf, Bernhard and Spirtes, Peter and Sugihara, Ge...

work page doi:10.1038/s41467-019-10105-3 2019
[5]

Detecting and quantifying causal associations in large nonlinear time series datasets

Runge, Jakob and Nowack, Peer and Kretschmer, Marlene and Flaxman, Seth and Sejdinovic, Dino , journal =. 2019 , title =. doi:10.1126/sciadv.aau4996 , url =

work page doi:10.1126/sciadv.aau4996 2019
[6]

, journal =

Nowack, Peer and Runge, Jakob and Eyring, Veronika and Haigh, Joanna D. , journal =. 2020 , title =. doi:10.1038/s41467-020-15195-y , url =

work page doi:10.1038/s41467-020-15195-y 2020
[7]

2010 , title =

Spirtes, Peter , journal =. 2010 , title =

2010
[8]

2009 , title =

Pearl, Judea , isbn =. 2009 , title =

2009
[9]

2009 , keywords =

Pearl, Judea , journal =. 2009 , keywords =. doi:10.1214/09-ss057 , pages =

work page doi:10.1214/09-ss057 2009
[10]

and Min, Jae S

Prosperi, Mattia and Guo, Yi and Sperrin, Matt and Koopman, James S. and Min, Jae S. and He, Xing and Rich, Shannan and Wang, Mo and Buchan, Iain E. and Bian, Jiang , journal =. 2020 , title =. doi:10.1038/s42256-020-0197-y , url =

work page doi:10.1038/s42256-020-0197-y 2020
[11]

Page 19 of 26 Stephen R Cole and Constantine E Frangakis

Rubin, Donald B. , journal =. 1974 , title =. doi:10.1037/h0037350 , pages =

work page doi:10.1037/h0037350 1974
[12]

Donges, Jaroslav Hlinka, Nikola Jajcay, Martin Vejmelka, David Hartman, Norbert Marwan, Milan Palu s , and J \"u rgen Kurths

Runge, Jakob and Petoukhov, Vladimir and Donges, Jonathan F. and Hlinka, Jaroslav and Jajcay, Nikola and Vejmelka, Martin and Hartman, David and Marwan, Norbert and Paluš, Milan and Kurths, Jürgen , journal =. 2015 , title =. doi:10.1038/ncomms9502 , pmid =

work page doi:10.1038/ncomms9502 2015
[13]

1993 , title =

Spirtes, Peter and Glymour, Clark and Scheines, Richard , series =. 1993 , title =. doi:10.1007/978-1-4612-2748-9 , volume =

work page doi:10.1007/978-1-4612-2748-9 1993
[14]

2018 , keywords =

Pearl, Judea and Mackenzie, Dana , isbn =. 2018 , keywords =

2018
[15]

Spirtes and K

Spirtes, Peter and Zhang, Kun , journal =. 2016 , title =. doi:10.1186/s40535-016-0018-x , pmid =

work page doi:10.1186/s40535-016-0018-x 2016
[16]

2015 , title =

Runge, Jakob , journal =. 2015 , title =. doi:10.1103/physreve.92.062829 , pmid =. 1508.03808 , pages =

work page doi:10.1103/physreve.92.062829 2015
[17]

2018 , title =

Hannart, Alexis and Naveau, Philippe , journal =. 2018 , title =. doi:10.1175/jcli-d-17-0304.1 , eprint =

work page doi:10.1175/jcli-d-17-0304.1 2018
[18]

and Flato, Gregory M

Eyring, Veronika and Cox, Peter M. and Flato, Gregory M. and Gleckler, Peter J. and Abramowitz, Gab and Caldwell, Peter and Collins, William D. and Gier, Bettina K. and Hall, Alex D. and Hoffman, Forrest M. and Hurtt, George C. and Jahn, Alexandra and Jones, Chris D. and Klein, Stephen A. and Krasting, John P. and Kwiatkowski, Lester and Lorenz, Ruth and ...

work page doi:10.1038/s41558-018-0355-y 2019
[19]

and Kurths, Jürgen , journal =

Runge, Jakob and Donner, Reik V. and Kurths, Jürgen , journal =. 2015 , title =. doi:10.1103/physreve.91.052909 , pmid =. 1506.05822 , pages =

work page doi:10.1103/physreve.91.052909 2015
[20]

and Migliavacca, Mirco and Perez-Priego, Oscar and El-Madany, Tarek and Carrara, Arnaud and Mahecha, Miguel D

Krich, Christopher and Runge, Jakob and Miralles, Diego G. and Migliavacca, Mirco and Perez-Priego, Oscar and El-Madany, Tarek and Carrara, Arnaud and Mahecha, Miguel D. , journal =. 2020 , title =. doi:10.5194/bg-17-1033-2020 , pages =

work page doi:10.5194/bg-17-1033-2020 2020
[21]

An algorithm for fast recovery of sparse causal graphs.Social Science Computer Review, 9(1):62–72, 1991

Spirtes, Peter and Glymour, Clark , journal =. 1991 , title =. doi:10.1177/089443939100900106 , pages =

work page doi:10.1177/089443939100900106 1991
[22]

and Hyvarinen, Aapo and Kerminen, Antti , journal =

Shimizu, Shohei and Hoyer, Patrik O. and Hyvarinen, Aapo and Kerminen, Antti , journal =. 2006 , title =

2006
[23]

Review of Causal Discovery Methods Based on Graphical Models

Glymour, Clark and Zhang, Kun and Spirtes, Peter , journal =. 2019 , title =. doi:10.3389/fgene.2019.00524 , pmid =

work page doi:10.3389/fgene.2019.00524 2019
[24]

A Survey of Learning Causality with Data: Problems and Methods

Guo, Ruocheng and Cheng, Lu and Li, Jundong and Hahn, P. Richard and Liu, Huan , journal =. 2020 , title =. doi:10.1145/3397269 , eprint =

work page doi:10.1145/3397269 2020
[25]

& Hochberg, Y

Benjamini, Yoav and Hochberg, Yosef , journal =. 1995 , title =. doi:10.1111/j.2517-6161.1995.tb02031.x , pages =

work page doi:10.1111/j.2517-6161.1995.tb02031.x 1995
[26]

2014 , title =

Deng, Yi and Ebert‐Uphoff, Imme , journal =. 2014 , title =. doi:10.1002/2013gl058646 , pages =

work page doi:10.1002/2013gl058646 2014
[27]

2012 , title =

Ebert‐Uphoff, Imme and Deng, Yi , journal =. 2012 , title =. doi:10.1029/2012gl053269 , number =

work page doi:10.1029/2012gl053269 2012
[28]

2016 , title =

Hannart, A and Pearl, J and Otto, F E L and Naveau, P and Ghil, M , journal =. 2016 , title =. doi:10.1175/bams-d-14-00034.1 , pages =

work page doi:10.1175/bams-d-14-00034.1 2016
[29]

1982 , title =

Florens, J P and Mouchart, M , journal =. 1982 , title =. doi:10.2307/1912602 , pages =

work page doi:10.2307/1912602 1982
[30]

2018 , title =

Runge, Jakob , booktitle =. 2018 , title =

2018
[31]

and Seth, Anil K

Barnett, Lionel and Barrett, Adam B. and Seth, Anil K. , journal =. 2009 , title =. doi:10.1103/physrevlett.103.238701 , pmid =. 0910.4514 , pages =

work page doi:10.1103/physrevlett.103.238701 2009
[32]

2008 , author =

Advances in Neural Information Processing Systems , url =. 2008 , author =

2008
[33]

Schreiber, Measuring information transfer, Physical review letters 85 (2) (2000) 461

Schreiber, Thomas , journal =. 2000 , title =. doi:10.1103/physrevlett.85.461 , pmid =. nlin/0001042 , pages =

work page doi:10.1103/physrevlett.85.461 2000
[34]

Scheines, Richard , keywords =
[35]

2017 , title =

Gong, Mingming and Zhang, Kun and Schölkopf, Bernhard and Glymour, Clark and Tao, Dacheng , journal =. 2017 , title =

2017
[36]

van and Scheffer, Marten and Brovkin, Victor and Lenton, Timothy M

Nes, Egbert H. van and Scheffer, Marten and Brovkin, Victor and Lenton, Timothy M. and Ye, Hao and Deyle, Ethan and Sugihara, George , journal =. 2015 , title =. doi:10.1038/nclimate2568 , pages =

work page doi:10.1038/nclimate2568 2015
[37]

1986 , title =

Glymour, Clark and Scheines, Richard , journal =. 1986 , title =. doi:10.1007/bf00413966 , pages =

work page doi:10.1007/bf00413966 1986
[38]

and Thompson, Matthew and Barron, Mace G

Carriger, John F. and Thompson, Matthew and Barron, Mace G. , journal =. 2021 , title =. doi:10.1002/ieam.4443 , pmid =

work page doi:10.1002/ieam.4443 2021
[39]

, isbn =

Hernán, Miguel Angel and Robins, James M. , isbn =. 2020 , title =

2020
[40]

34th Conference on Neural Information Processing Systems (NeurIPS 2020) , keywords =

Heskes, Tom and Sijben, Evi , title =. 34th Conference on Neural Information Processing Systems (NeurIPS 2020) , keywords =

2020
[41]

34th Conference on Neural Information Processing Systems (NeurIPS 2020), , keywords =

Frye, Christopher and Rowat, Colin and Feige, Ilya , title =. 34th Conference on Neural Information Processing Systems (NeurIPS 2020), , keywords =

2020
[42]

Huang, Yu and Fu, Zuntao and Franzke, Christian L. E. , journal =. 2020 , title =. doi:10.1063/5.0007670 , pmid =

work page doi:10.1063/5.0007670 2020
[43]

and Runge, Jakob , journal =

Kretschmer, Marlene and Coumou, Dim and Donges, Jonathan F. and Runge, Jakob , journal =. 2016 , title =. doi:10.1175/jcli-d-15-0654.1 , pages =

work page doi:10.1175/jcli-d-15-0654.1 2016
[44]

and Dickinson, Robert E

Zhuang, Yizhou and Fu, Rong and Santer, Benjamin D. and Dickinson, Robert E. and Hall, Alex , journal =. 2021 , title =. doi:10.1073/pnas.2111875118 , pmid =

work page doi:10.1073/pnas.2111875118 2021
[45]

2019 , title =

Lodi, Sara and Phillips, Andrew and Lundgren, Jens and Logan, Roger and Sharma, Shweta and Cole, Stephen R and Babiker, Abdel and Law, Matthew and Chu, Haitao and Byrne, Dana and Horban, Andrzej and Sterne, Jonathan A C and Porter, Kholoud and Sabin, Caroline and Costagliola, Dominique and Abgrall, Sophie and Gill, John and Touloumi, Giota and Pacheco, An...

work page doi:10.1093/aje/kwz100 2019
[46]

, journal =

Rubin, Donald B. , journal =. 2019 , title =. doi:10.1080/24709360.2019.1670513 , pages =

work page doi:10.1080/24709360.2019.1670513 2019
[47]

Econometrica , author =

Granger, C W J , journal =. 1969 , title =. doi:10.2307/1912791 , pages =

work page doi:10.2307/1912791 1969
[48]

The large-sample distribution of the likelihood ratio for testing composite hypotheses

Wright, Sewall , journal =. 1934 , title =. doi:10.1214/aoms/1177732676 , pages =

work page doi:10.1214/aoms/1177732676 1934
[49]

1921 , title =

Wright, Sewall , journal =. 1921 , title =

1921
[50]

Journal of the American Statistical Association , volume =

Angrist, Joshua D. and Imbens, Guido W. and Rubin, Donald B. , journal =. 2012 , title =. doi:10.1080/01621459.1996.10476902 , pages =

work page doi:10.1080/01621459.1996.10476902 2012
[51]

2017 , title =

Peters, Jonas and Janzing, Dominik and Schlkopf, Bernhard , isbn =. 2017 , title =

2017
[52]

1995 , title =

Pearl, Judea , doi =. 1995 , title =

1995
[53]

and Feldl, Nicole and Weijer, Wilbert and Veneziani, Milena , journal =

Kaufman, Zachary S. and Feldl, Nicole and Weijer, Wilbert and Veneziani, Milena , journal =. 2020 , title =. doi:10.1175/jcli-d-19-0525.1 , pages =

work page doi:10.1175/jcli-d-19-0525.1 2020
[54]

and Brookhart, M

Rassen, Jeremy A. and Brookhart, M. Alan and Glynn, Robert J. and Mittleman, Murray A. and Schneeweiss, Sebastian , journal =. 2009 , title =. doi:10.1016/j.jclinepi.2008.12.005 , pmid =

work page doi:10.1016/j.jclinepi.2008.12.005 2009
[55]

and Grassberger, P

Arnhold, J. and Grassberger, P. and Lehnertz, K. and Elger, C.E. , journal =. 1999 , title =. doi:10.1016/s0167-2789(99)00140-2 , eprint =

work page doi:10.1016/s0167-2789(99)00140-2 1999
[56]

doi:10.1126/science.1227079

Sugihara, George and May, Robert and Ye, Hao and Hsieh, Chih-hao and Deyle, Ethan and Fogarty, Michael and Munch, Stephan , journal =. 2012 , title =. doi:10.1126/science.1227079 , pmid =

work page doi:10.1126/science.1227079 2012
[57]

1995 , title =

Pearl, Judea , journal =. 1995 , title =. doi:10.2307/2337329 , pages =

work page doi:10.2307/2337329 1995
[58]

Jolliffe, and David B

Hannachi, A. and Jolliffe, I. T. and Stephenson, D. B. , journal =. 2007 , title =. doi:10.1002/joc.1499 , pages =

work page doi:10.1002/joc.1499 2007
[59]

, journal =

Hall, Nancy S. , journal =. 2007 , title =. doi:10.1007/s10739-006-9119-z , pmid =

work page doi:10.1007/s10739-006-9119-z 2007
[60]

1986 , title =

Holland, Paul W , journal =. 1986 , title =. doi:10.2307/2289064 , pages =

work page doi:10.2307/2289064 1986
[61]

2005 , title =

Rubin, Donald B , journal =. 2005 , title =. doi:10.1198/016214504000001880 , pages =

work page doi:10.1198/016214504000001880 2005
[62]

2020 , author =

Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence , url =. 2020 , author =

2020
[63]

2020 , author =

Advances in Neural Information Processing Systems , doi =. 2020 , author =

2020
[64]

Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins

Robins, James , journal =. 1986 , title =. doi:10.1016/0270-0255(86)90088-6 , url =

work page doi:10.1016/0270-0255(86)90088-6 1986
[65]

1998 , title =

Pearl, Judea , journal =. 1998 , title =. doi:10.1177/0049124198027002004 , url =

work page doi:10.1177/0049124198027002004 1998
[66]

Fisher, R.A. , url =. 1925 , title =

1925
[67]

arXiv preprint arXiv:1210.4852 , year=

Pearl, Judea , journal =. 2012 , title =. 1210.4852 , keywords =

work page arXiv 2012
[68]

2018 , journal =

Tarka, Piotr , journal =. 2018 , title =. doi:10.1007/s11135-017-0469-8 , pmid =

work page doi:10.1007/s11135-017-0469-8 2018
[69]

, isbn =

Pearl, J. , isbn =. 2000 , title =

2000
[70]

2001 , title =

Pearl, Judea , journal =. 2001 , title =

2001
[71]

1956 , title =

Wiener, Norbert , journal =. 1956 , title =

1956
[72]

2014 , title =

Vejmelka, Martin and Pokorná, Lucie and Hlinka, Jaroslav and Hartman, David and Jajcay, Nikola and Paluš, Milan , journal =. 2014 , title =. doi:10.1007/s00382-014-2244-z , pages =

work page doi:10.1007/s00382-014-2244-z 2014
[73]

2014 , title =

Runge, Jakob and Petoukhov, Vladimir and Kurths, Jürgen , journal =. 2014 , title =. doi:10.1175/jcli-d-13-00159.1 , pages =

work page doi:10.1175/jcli-d-13-00159.1 2014
[74]

2017 , author =

Proceedings of the 31st International Conference on Neural Information Processing Systems , isbn =. 2017 , author =

2017
[75]

2018 , title =

Runge, Jakob , url =. 2018 , title =

2018
[76]

2011 , author =

Proceedings of the Twenty-Seventh Conference on Uncertainty in Artificial Intelligence , isbn =. 2011 , author =

2011
[77]

1956 , title =

Reichenbach, Hans , editor =. 1956 , title =

1956
[78]

2015 , title =

Gong, Mingming and Zhang, Kun and Schoelkopf, Bernhard and Tao, Dacheng and Geiger, Philipp , booktitle =. 2015 , title =

2015
[79]

2017 , title =

Zhang, Kun and Schölkopf, Bernhard and Spirtes, Peter and Glymour, Clark , journal =. 2017 , title =. doi:10.1093/nsr/nwx137 , pmid =

work page doi:10.1093/nsr/nwx137 2017
[80]

1953 , title =

Dorn, Harold F , journal =. 1953 , title =. doi:10.2105/ajph.43.6\_pt\_1.677 , pmid =

work page doi:10.2105/ajph.43.6 1953

Showing first 80 references.