betaFeedback

WorksMathematics

PP-stationary structural equation modeling: Causal effect identification under temporally incomplete observations

Abstract

Causal modeling of discrete-time processes often involves the assumption that the modeled processes are stationary. A general class of linear causal models for discrete-time stationary processes are structural equation processes (SEP), which generalize the well known structural equation model for i.i.d. observations. In this paper, we extend the SEP model from stationary to PP-stationary processes, which are non-stationary processes whose PP-periodically sampled subprocesses are jointly stationary. We will use this model to approach the problem of causal effect identification under temporally incomplete observations from a graphical modeling perspective. To have a notion of generic identifiability, we restrict our analysis to PP-stationary structural vector autoregressive (SVAR) processes, which are SVAR processes whose underlying coefficients can vary PP-periodically over time. Our findings allow to detect, in any given temporally incomplete observational setting, if a set of PP-periodic instrumental subprocesses qualifies to identify a causal effect between PP-periodic subprocesses, and if this is the case, that effect admits an explicit expression as a rational function evaluated at the observational spectral density.

Introduction

Causal inference [33, 42] is about modeling cause and effect relations in the presence of randomness. The fundamental concept in causal inference is that of a causal model. A causal model has an associated directed graph whose vertices represent the entities between which causal relations are to be modeled, and whose directed links specify the immediate cause and effect relations. In addition to the graph, a causal model contains a set of functions compatible with the directed links, and a random noise object for each vertex. If the model is well defined, it gives rise to the observational distribution, which is a probability distribution over a collection of random variables indexed by the vertices of the graph. The question of what information about the causal model can be recovered from the observational distribution drives a major stream of research in causal inference, including the present paper.

Time is intrinsic to causal relations in that an effect cannot precede its cause, i.e., cause and effect are ordered by time. When defining a causal model which obeys the ordering of time, a discrete set of time points is often assumed. Structural equation processes [35, 36] provide a broad class of linear causal models for discrete-time processes. The qualitative causal structure of an SEP is encoded by a finite directed graph. Every vertex on that graph has an independent stationary noise process, and every directed link has an associated infinite sequence, which quantifies how the past of the causing process linearly influences the present state of the caused process. The structural equation process defines a linear equation system, which is uniquely solved by the (stationary) observational process of the SEP.

In an SEP, the influence of one process on another is time invariant, i.e., the sequence describing how the latter process at a given time point is linearly influenced by the past of the former does not depend on the time point. In particular, stationary SEPs are not capable of modeling time-varying interactions. In the current paper, we extend the class of structural equation processes to a subclass of non-stationary discrete-time processes, which we call PP-stationary, and which allow to model periodically varying interactions between processes. For the purpose of defining PP-stationarity, we declare a subsampling to be a subset S⊂ZS \subset\mathbb{Z}, and we say that it has periodicity PP if t∈St \in S implies that also t±P∈St \pm P \in S. Note that there are exactly PP-many subsamplings with periodicity PP. In this work, a process XX is called PP-stationary if the subprocesses XS=(X(t):t∈S)X_S = (X(t) : t \in S), where SS indexes the PP-periodic subsamplings, are jointly stationary. To give an example, imagine we measure the temperature at some fixed location twice a year, once during summer and once during winter. The resulting temperature time series will turn out non-stationary if at that location winter and summer are significantly different from each other. Nonetheless, the time series consisting only of the winter and summer measurements, respectively, might still be stationary, as we would expect summer and winter, respectively, to be similar from year to year. If, in addition, the dependence between winter and summer does not change over time, the entire time series, which includes both winter and summer values, will be 2-stationary.

In this paper, we define a class of linear causal models for PP-stationary processes whose components are indexed by some finite index set VV. To stay in line with previous terminology [35, 34, 36], we term them PP-stationary SEPs. A PP-stationary SEP consists of a graph G=(V×{0,…,P−1},D)G = (V \times\{0, \ldots, P - 1\}, D), a link filter for every link on GG, and a noise process for every vertex on GG. A vertex on GG represents a PP-periodic subsampling of some component process vv. Consequently, a link between two vertices expresses a linear influence of one PP-periodic subprocess onto another PP-periodic subprocess.

Our main motivation to introduce PP-stationary SEPs is to provide a graphical model for handling temporally incomplete observations. For the biyearly temperature measurements, the subprocess consisting only of the summer or winter values is a simple example of temporally incomplete observation. However, our notion of temporal incompleteness allows for much more general observational scenarios.

The underlying observation to our approach is that any temporally incomplete observation of an PP-stationary SEP can be represented, for an appropriate integer RR, as a finite collection of RR-periodic subsamplings. Then we will see that every PP-stationary SEP has an equivalent RPRP-stationary SEP, which is equipped with a latent structure rendering every missing subsample as latent. In this setting, we analyze the problem of identifying causal effects, i.e., link filters, in PP-stationary SEPs from temporally incomplete observations. Following previous work [34, 36], we will study the problem in the frequency domain, where the algebraic relations are more tractable. The identification of a given effect in the PP-stationary SEP using a set of instrumental processes requires certain entries in the spectral density. If the observations are temporally incomplete, one faces the problem that the spectral density can generally not be computed, so identification via instrumental processes seems infeasible. However, by providing a linear recovery formula (Proposition 12), we can characterize (Definition 13) the temporally incomplete observational settings for which it is possible to recover a given entry in the spectral density (using that linear formula). Accordingly, in a setting of temporally incomplete observations, we detect a given effect to be identifiable if there is a suitable set of instrumental processes while the temporal incompleteness of the observations is such that all the entries in the spectral density that are required for identification via that set of instrumental processes can be recovered (Theorem 18). For technical reasons, we restrict our analysis to PP-stationary structural vector autoregressive processes, i.e., SVAR processes whose coefficients are allowed to vary periodically over time with periodicity PP.

To exemplify the problem and our way of approaching it, consider the simple setup of three stationary SVAR processes, whose causal relations are given by u→v→wu \to v \to w. Suppose we are interested in the effect associated with the link v→wv \to w. If we had temporally complete observations for vv and ww, we could identify this effect from the spectral density via linear regression in the frequency domain. However, this approach cannot be pursued if the observations for vv and ww are temporally incomplete, since in this case the spectral density cannot be computed. Assuming temporally complete observations for the instrumental process uu, it will follow from the theory we develop in this paper that the spectral dependence between uu and vv, and between uu and ww can be recovered, so that the effect v→wv \to w can be identified by the usual instrumental process formula for the frequency domain. A more interesting case, discussed in Example 9 and pictured in Figure 7, demonstrates that our findings allow to identify effects even if none of the involved processes is fully observable.

Figure 7. a.) The summary graph for a 2-stationary SVAR process. A vertex aka_k represents the process defl⁡2∘sh⁡k(Xa)\operatorname{defl}_2 \circ\operatorname{sh}^k(X_a), and a bidirectional link represents two directed links, i.e., one in each direction. b.) The observational structure, where a vertex of the form ak,ja_{k,j} represents the subprocess defl⁡4∘sh⁡2k+j(Xa)\operatorname{defl}_4 \circ\operatorname{sh}^{2k+j}(X_a). The fully and palely colored nodes represent the observed and unobserved subprocesses, respectively. The process ll is completely unobserved.

The rest of the paper is organized as follows. In § 2, we recall some basic terminology for linear causal models of discrete-time stationary processes. In § 3, we define PP-stationary SEPs and SVAR processes and show how an PP-stationary SVAR process can equivalently by described as an PP-stationary SEP. Furthermore, we define the embedding operator that identifies for every PP-stationary SEP an equivalent RPRP-stationary SEP. In Section 4, we present the algebraic structure of PP-stationary SEPs in the frequency domain. The main result of this section will be used to determine if a given temporally incomplete observational pattern allows to linearly recover a specified entry in the spectral density. In § 4, we formalize the problem of identifying causal effects in PP-stationary SEPs from temporally incomplete data in the frequency domain, i.e., we will make precise what we mean by a temporally incomplete observations and we will provide a notion of rational identifiability for PP-stationary SVAR processes (under a given temporally incomplete observational pattern). Using those terms, we will formulate the identifiability result.

Related work

We approach causal inference in the framework of Structural Causal Models presented and developed in, e.g., [33, 28, 6, 42]. Since we focus on linear models, much of the work and theory developed here is inspired and influenced by the theory and terminology used in the literature on linear Structural Equation Modeling, see, e.g., [4, 41, 46, 47], and especially those contributions that take an algebraic perspective on the matter, see, e.g., [12, 13, 19, 11, 44].

Our paper is particularly related to those contributions which incorporate time as a discrete index into the modeling of causal relations. Early contributions include those related to Granger Causality, see [22, 23, 25, 26, 27], which underwent a reformulation in terms of concepts and terminology from graphical modeling, see [8, 9, 14, 18, 15, 17, 16]. This line of work has then been further developed, for instance, in [40, 39, 38, 21, 30].

The theory presented in the present paper builds on [35, 37, 34], which develop the notion of (stationary) structural equation processes to quantify process-level causal relations (in the frequency domain). An algebraic approach to structure and effect identification from observational data, in the spirit of the contributions mentioned above, was developed in [36].

A specific from of temporally incomplete observations are undersampled time series, i.e., when observations are gathered less frequently than causal effects happen. An introduction to the problems related to inferring causal relations from undersampled time series can be found in [10]. Specific approaches to inferring the causal structure from an undersampled time series model have been developed in, e.g., [31, 29]. The special case where the underlying time series model is linear is investigated, for example, in [24, 45]. Specifically, they consider SVAR processes of order one (causal effects happen either instantaneously or with a time lag of one time unit) generated by a non-Gaussian white noise process. In [45], the observations are allowed to be gathered at two different sampling rates, which is also known as the mixed-frequency case. Another approach to the identification of VAR models (without instantaneous effects) is presented in [2, 1], where the authors consider stable VAR processes of some arbitrary finite order pp, and which are assumed to be sampled at two different frequencies. The fast frequency is assumed to be the actual time scale of the process, i.e., the frequency with which causal effects happen. Under the assumption that at least one process is observed at the actual time scale, they show that for almost every possible VAR parameter configuration, that parameter can be uniquely recovered from the second moments of the observed process (which is generally temporally incomplete).

In this work, we consider linear time series models whose coefficients are allowed to vary periodically over time. Such models generally yield non-stationary processes. What we call an PP-stationary SVAR process in this work is a special instance of a time varying VAR process, see, e.g., [43, 7]. A recent contribution to non-stationary undirected graphical modeling can be found in [3]. Our approach to weaken the stationarity assumption in a discrete-time setting is similar to the one taken in [20], which develops a causal discovery algorithm for so-called semi-stationary Markov chains with an underlying periodically varying causal structure.

Basic notation

Let VV be some finite set and K\mathbb{K} a field, then KV\mathbb{K}^{V} denotes the K\mathbb{K}-vector space of dimension ∣V∣|V| and whose coordinates are indexed by VV. Let UU be another finite set, then KV×U\mathbb{K}^{V \times U} denotes the space of matrices whose rows are indexed by VV and whose columns are indexed by UU. In case V=UV = U, we write IVI_V for the identity matrix. For a matrix A=[Av,u]v∈V,u∈U∈KV×UA = [A_{v,u}]_{v\in V,u\in U} \in\mathbb{K}^{V\times U}, and subsets V′⊂VV' \subset V and U′⊂UU' \subset U, we define the submatrix of AA corresponding to the entries V′×U′V' \times U' as [A]V′,U′:=[Av,u]v∈V′,u∈U′∈KV′×U′[A]_{V',U'} := [A_{v,u}]_{v\in V',u\in U'} \in\mathbb{K}^{V'\times U'}. The transpose of AA is denoted A⊤:=[Av,u]u∈U,v∈V∈KU×VA^\top:= [A_{v,u}]_{u\in U,v\in V} \in\mathbb{K}^{U\times V}. Since we will often encounter block-matrices, it will be useful to have a dedicated notation for them. Let U,V,I,JU, V, I, J be finite sets, then define BMat⁡(K,(U,V),(I,J))\operatorname{BMat}(\mathbb{K},(U,V),(I,J)) to be the set of matrices AA which have the following two level block structure:

A=[Au,v]u∈U,v∈V,Au,v=[Au,v(i,j)]i∈I,j∈J∈KI×J.A = [A_{u,v}]_{u\in U,v\in V}, \qquad A_{u,v} = [A_{u,v}^{(i,j)}]_{i\in I,j\in J} \in\mathbb{K}^{I\times J}.

If U=VU = V or I=JI = J, we will abbreviate the corresponding set of block-matrices by BMat⁡(K,V,(I,J))\operatorname{BMat}(\mathbb{K},V,(I,J)) and BMat⁡(K,(U,V),I)\operatorname{BMat}(\mathbb{K},(U,V),I) respectively, and if both U=VU = V and I=JI = J, then we abbreviate the corresponding set of block-matrices by BMat⁡(K,U,I)\operatorname{BMat}(\mathbb{K},U,I).

For two integers p,q∈Zp,q \in\mathbb{Z}, define [p,q]:={i∣p≤i≤q}[p,q] := \{i \mid p \le i \le q\}, and for n≥1n \ge1, we write [n][n] to abbreviate [1,n][1,n]. For a positive integer P≥1P \ge1, we define the set of PP-multiples to be PZ:={kP∣k∈Z}P\mathbb{Z} := \{kP \mid k \in\mathbb{Z}\}, and for m∈[0,P−1]m \in[0,P-1], we define PZ−m:={kP−m∣k∈Z}P\mathbb{Z}-m := \{kP-m \mid k \in\mathbb{Z}\}.

For K\mathbb{K} either R\mathbb{R} or C\mathbb{C}, the space ℓp(Z,K)\ell^p(\mathbb{Z},\mathbb{K}) consists of all sequences λ:Z→K\lambda:\mathbb{Z}\to\mathbb{K} such that ∥λ∥pp:=∑k∈Z∣λ(k)∣p<∞\|\lambda\|_p^p := \sum_{k\in\mathbb{Z}}|\lambda(k)|^p < \infty. We use ℓp(Z,KV×U)\ell^p(\mathbb{Z},\mathbb{K}^{V\times U}) to denote the space of sequences Λ:Z→KV×U\Lambda:\mathbb{Z}\to\mathbb{K}^{V\times U} such that [Λ]v,u:=(k∈Z↦[Λ(k)]v,u)∈ℓp(Z,K)[\Lambda]_{v,u}:=(k\in\mathbb{Z}\mapsto[\Lambda(k)]_{v,u})\in\ell^p(\mathbb{Z},\mathbb{K}) for all (v,u)∈V×U(v,u)\in V\times U. In this work, we will call any element in Λ∈ℓ1(Z,KV×U)\Lambda\in\ell^1(\mathbb{Z},\mathbb{K}^{V\times U}) such that Λ(k)=0\Lambda(k)=0 if k<0k<0, a filter, a term often used in the time series literature. For two sequences λ,μ∈ℓp(Z,K)\lambda,\mu\in\ell^p(\mathbb{Z},\mathbb{K}), the convolution resp. tilted convolution are defined as follows:

λ∗μ:=(k↦∑l∈Zλ(l)μ(k−l)),λ∗′μ:=(k↦∑l∈Zλ(k+l)μ(l)).\lambda* \mu:= \left(k\mapsto\sum_{l\in\mathbb{Z}}\lambda(l)\mu(k-l)\right), \qquad\lambda*' \mu:= \left(k\mapsto\sum_{l\in\mathbb{Z}}\lambda(k+l)\mu(l)\right).

The transpose of a sequence Λ\Lambda is the element-wise transpose, Λ⊤:=(k↦Λ(k)⊤)\Lambda^\top:= (k\mapsto\Lambda(k)^\top). If V=UV=U, then IVI_V denotes the identity filter, i.e., IV(0)=IVI_V(0)=I_V, and IV(k)=0I_V(k)=0 for all k≠0k\ne0. Furthermore, for a fixed set SS and a subset S′⊂SS'\subset S, we write 1S′:S→{0,1}\mathbf{1}_{S'}:S\to\{0,1\} for the indicator function, i.e., 1S′(x)=1\mathbf{1}_{S'}(x)=1 if and only if x∈S′x\in S'.

Preliminaries

In this section, we review the basic terminology for directed graphs, discrete-time stationary processes, and linear causal models of processes, i.e., structural equation processes [35, 34, 36].

Directed graphs

References

  1. [1]Brian D. O. Anderson, Manfred Deistler, Elisabeth Felsenstein, Bernd Funovits, Lukas Koelbl, and Mohsen Zamani. Multivariate AR systems and mixed frequency data: G-identifiability and estimation. Econometric Theory, 32(4):793–826, 2016.DOI
  2. [2]Brian D. O. Anderson, Manfred Deistler, Elisabeth Felsenstein, Bernd Funovits, Peter Zadrozny, Michael Eichler, Weitian Chen, and Mohsen Zamani. Identifiability of regular and singular multivariate autoregressive models from mixed frequency data. In 2012 IEEE 51st IEEE Conference on Decision and Control (CDC), pages 184–189. IEEE, 2012.DOI
  3. [3]Sumanta Basu and Suhasini Subba Rao. Graphical models for nonstationary time series. Ann. Statist., 51(4):1453–1483, 2023. MR4634345.arxiv.org/abs/2109.08709
  4. [4]Kenneth A. Bollen. Structural Equations with Latent Variables. Wiley-Interscience, 1989.
  5. [5]Peter J. Brockwell and Richard A. Davis. Time Series: Theory and Methods. Springer New York, NY, 2009.
  6. [6]Peter Bühlmann, Jonas Peters, and Jan Ernest. CAM: Causal additive models, high-dimensional order search and penalized regression. Ann. Statist., 42(6):2526–2556, 2014. MR3269988.DOI
  7. [7]Rainer Dahlhaus. Fitting time series models to nonstationary processes. Ann. Statist., 25(1):1–37, 1997. MR1429916.DOI
  8. [8]Rainer Dahlhaus. Graphical interaction models for multivariate time series. Metrika, 51(2):157–172, 2000. MR1790908.
  9. [9]Rainer Dahlhaus and Michael Eichler. Causality and graphical models for time series. In Peter J. Green, Nils L. Hjort, and Sylvia Richardson, editors, Highly Structured Stochastic Systems, volume 27 of Oxford Statistical Science Series, pages 115–137. Oxford University Press, Oxford, 2003. MR2082408.
  10. [10]David Danks and Sergey Plis. Learning causal structure from undersampled time series. In NeurIPS Workshop on Causality, 2013.
  11. [11]Mathias Drton, Rina Foygel, and Seth Sullivant. Global identifiability of linear structural equation models. Ann. Statist., 39(2):865–886, 2011.DOI
  12. [12]Mathias Drton, Bernd Sturmfels, and Seth Sullivant. Algebraic factor analysis: Tetrads, pentads and beyond. Probab. Theory Related Fields, 138(3-4):463–493, 2007.
  13. [13]Mathias Drton, Bernd Sturmfels, and Seth Sullivant. Lectures on Algebraic Statistics, volume 39 of Oberwolfach Seminars. Birkhäuser Verlag, Basel, 2009.
  14. [14]Michael Eichler. A graphical approach for evaluating effective connectivity in neural systems. Philosophical Transactions of the Royal Society B: Biological Sciences, 360(1457):953–967, 2005.pubmed.ncbi.nlm.nih.gov/16087440
  15. [15]Michael Eichler. Granger causality and path diagrams for multivariate time series. Journal of Econometrics, 137(2):334–353, 2007.DOI
  16. [16]Michael Eichler. Causal inference in time series analysis. In Carlo Berzuini, Philip Dawid, and Luisa Bernardinelli, editors, Causality: Statistical Perspectives and Applications, pages 327–354. John Wiley & Sons, Chichester, 2012.DOI
  17. [17]Michael Eichler. Graphical modelling of multivariate time series. Probab. Theory Related Fields, 153(1-2):233–268, 2012. MR2925716.
  18. [18]Michael Eichler and Vanessa Didelez. Causal reasoning in graphical time series models. In Proceedings of the 23rd Conference on Uncertainty in Artificial Intelligence (UAI), pages 109–116, 2007.
  19. [19]Rina Foygel, Jan Draisma, and Mathias Drton. Half-trek criterion for generic identifiability of linear structural equation models. Ann. Statist., 40(3):1682–1713, 2012.DOI
  20. [20]Shanyun Gao, Raghavendra Addanki, Tong Yu, Ryan A. Rossi, and Murat Kocaoglu. Causal discovery in semi-stationary time series. In Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.arxiv.org/abs/2407.07291
  21. [21]Andreas Gerhardus and Jakob Runge. High-recall causal discovery for autocorrelated time series with latent confounders. In Advances in Neural Information Processing Systems 33 (NeurIPS), pages 12615–12625, 2020.arxiv.org/abs/2007.01884
  22. [22]John Geweke. Inference and causality in economic time series models. In Zvi Griliches and Michael D. Intriligator, editors, Handbook of Econometrics, volume 2, pages 1101–1144. North-Holland, Amsterdam, 1984.
  23. [23]John F. Geweke. Measures of conditional linear dependence and feedback between time series. J. Amer. Statist. Assoc., 79(388):907–915, 1984. MR775894.DOI
  24. [24]Mingming Gong, Kun Zhang, Bernhard Schölkopf, Dacheng Tao, and Philipp Geiger. Discovering temporal causal relations from subsampled data. In Proceedings of the 32nd International Conference on Machine Learning (ICML), volume 37 of Proceedings of Machine Learning Research, pages 1898–1906. PMLR, 2015.
  25. [25]Clive W. J. Granger. Economic processes involving feedback. Information and Control, 6(1):28–48, 1963.DOI
  26. [26]Clive W. J. Granger. Investigating causal relations by econometric models and cross-spectral methods. Econometrica, 37(3):424–438, 1969.
  27. [27]Clive W. J. Granger. Testing for causality: A personal viewpoint. Journal of Economic Dynamics and Control, 2:329–352, 1980.
  28. [28]Patrik O. Hoyer, Dominik Janzing, Joris M. Mooij, Jonas Peters, and Bernhard Schölkopf. Nonlinear causal discovery with additive noise models. In Advances in Neural Information Processing Systems 21 (NeurIPS), pages 689–696, 2008.
  29. [29]Antti Hyttinen, Sergey Plis, Matti Järvisalo, Frederick Eberhardt, and David Danks. Causal discovery from subsampled time series data by constraint optimization. In Proceedings of the 8th International Conference on Probabilistic Graphical Models (PGM), volume 52 of Proceedings of Machine Learning Research, pages 216–227, 2016.
  30. [30]Aapo Hyvärinen, Kun Zhang, Shohei Shimizu, and Patrik O. Hoyer. Estimation of a structural vector autoregression model using non-Gaussianity. J. Mach. Learn. Res., 11(5):1709–1731, 2010.
  31. [31]Mingzhou Liu, Xinwei Sun, Lingjing Hu, and Yizhou Wang. Causal discovery from subsampled time series with proxy variables. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.
  32. [32]Helmut Lütkepohl. New Introduction to Multiple Time Series Analysis. Springer Berlin, Heidelberg, 2005.
  33. [33]Judea Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, Cambridge, 2nd edition, 2009. MR2548166.
  34. [34]Nicolas-Domenic Reiter. Structural equation processes - A linear model for causal inference in discrete-time. PhD thesis, TU Berlin, 2025.
  35. [35]Nicolas-Domenic Reiter, Andreas Gerhardus, Jonas Wahl, and Jakob Runge. Causal inference on process graphs, part i: The structural equation process representation. arXiv preprint arXiv:2305.11561, 2023.
  36. [36]Nicolas-Domenic Reiter, Jonas Wahl, Andreas Gerhardus, and Jakob Runge. Causal inference on process graphs: Causal structure and effect identification. Bernoulli, 32(2), 2026.
  37. [37]Nicolas-Domenic Reiter, Jonas Wahl, Gabriele C. Hegerl, and Jakob Runge. Asymptotic uncertainty in the estimation of frequency domain causal effects for linear processes. arXiv preprint arXiv:2406.18191, 2024.
  38. [38]Jakob Runge. Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets. In Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), pages 1388–1397. PMLR, 2020.
  39. [39]Jakob Runge, Sebastian Bathiany, Erik Bollt, Gustau Camps-Valls, Dim Coumou, Ethan Deyle, Clark Glymour, Marlene Kretschmer, Miguel D. Mahecha, Jordi Muñoz Marí, Egbert H. van Nes, Jonas Peters, Rick Quax, Markus Reichstein, Marten Scheffer, Bernhard Schölkopf, Peter Spirtes, George Sugihara, Jie Sun, Kun Zhang, and Jakob Zscheischler. Inferring causation from time series in earth system sciences. Nature Communications, 10:2553, 2019.DOI
  40. [40]Jakob Runge, Peer Nowack, Marlene Kretschmer, Seth Flaxman, and Dino Sejdinovic. Detecting and quantifying causal associations in large nonlinear time series datasets. Science Advances, 5(11):eaau4996, 2019.DOI
  41. [41]Shohei Shimizu, Patrik O. Hoyer, Aapo Hyvärinen, and Antti Kerminen. A linear non-Gaussian acyclic model for causal discovery. J. Mach. Learn. Res., 7:2003–2030, 2006.
  42. [42]Peter Spirtes, Clark Glymour, and Richard Scheines. Causation, Prediction, and Search. MIT Press, Cambridge, MA, 2nd edition, 2000.
  43. [43]Tata Subba Rao. The fitting of non-stationary time-series models with time-dependent parameters. J. R. Stat. Soc. Ser. B Stat. Methodol., 32(2):312–322, 1970.DOI
  44. [44]Seth Sullivant, Kelli Talaska, and Jan Draisma. Trek separation for Gaussian graphical models. Ann. Statist., 38(3):1665–1685, 2010.DOI
  45. [45]A Tank, E B Fox, and A Shojaie. Identifiability and estimation of structural vector autoregressive models for subsampled and mixed-frequency time series. Biometrika, 106(2):433–452, 06 2019.DOI
  46. [46]Sewall Wright. Correlation and causation. Journal of Agricultural Research, 20(7):557–585, 1921.
  47. [47]Sewall Wright. The method of path coefficients. Ann. Math. Statist., 5(3):161–215, 1934.DOI

Paper details

vlnc:W_nnaz2qs5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

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