A Linear Non-Gaussian Acyclic Model for Causal Discovery

Shohei ShimizuPatrik O. HoyerAapo HyvärinenAntti Kerminen

article2006JMLR2,076 citations

Introduces the LiNGAM framework, which uses independent component analysis and non-Gaussian error distributions to identify the complete, directed causal structure of continuous observational data without requiring prior variable ordering.

Listen

Understanding cause-and-effect relationships is essential for predicting the consequences of organizational interventions and strategic policy decisions. However, conducting direct, controlled experiments is often prohibitively expensive, unethical, or logistically impossible, forcing leaders to rely on observational data. Traditional statistical discovery methods for continuous data assume standard bell-curve (Gaussian) distributions, which generally fail to identify the true directional flow of cause and effect and leave decision-makers with multiple indistinguishable, competing models.

The article demonstrates that full causal discovery—identifying both the exact directional order and the numerical strength of relationships without prior knowledge of time sequencing—is mathematically possible for continuous observational data. It establishes this capability under the framework of a Linear Non-Gaussian Acyclic Model, which relies on the condition that external disturbance or noise variables do not follow a bell-curve distribution.

To accomplish this, the article introduces an analytical algorithm based on independent component analysis, a statistical technique that separates mixed signals into independent, non-bell-curve components. The method resolves core mathematical ambiguities by pairing variables to disturbances and ordering them into a directional chain where causes strictly precede effects. The authors also formulated hypothesis tests to prune statistically insignificant connections and assess overall model fit. The credibility of the framework was tested across 1,000 synthetic trials across various system sizes (3 to 100 variables) and sample scopes (200 to 10,000 data points), as well as on 22 real-world economic and environmental time-series datasets.

Key findings show that non-Gaussian distributions provide sufficient information to uniquely determine the full causal structure and parameter values without any pre-specified variable ordering. In extensive simulations, the algorithm estimated underlying causal connection strengths near-flawlessly as sample sizes grew. When pruning unnecessary connections at sample sizes of 5,000 and above, the method achieved statistical power between 90% and 97% for correctly identified connections, keeping total classification errors below 10%. In real-world time-series evaluations, the method successfully recovered the true forward time order in well-behaved data, identified reverse causality in financial random-walk series, and flagged zero-effect relationships where no true dependencies existed.

These findings mean organizations can extract directional causal insights directly from historical and observational data, significantly reducing the cost, time, and operational risk associated with running large-scale physical experiments. Unlike traditional covariance-based techniques that yield ambiguous results, this approach provides a single, actionable model of causal mechanisms. This allows decision-makers to simulate interventions and evaluate compliance or performance impacts with much higher precision.

Practitioners should consider using this analytical discovery method as an exploratory first step to formulate data-driven causal hypotheses prior to committing capital to physical trials. When applying the model, analysts must implement the recommended statistical pruning and model-fit tests to filter out weak or spurious links. Further analysis and validation against domain knowledge are necessary before executing critical operational decisions.

Confidence in the results is high when the underlying assumptions hold: the relationships must be linear, free of unobserved common causes (confounders), and driven by non-bell-curve noise. Readers should exercise caution, as computational optimization can occasionally settle into local errors, and violations of key assumptions—such as non-linear dynamics or missing confounding factors—will degrade model reliability.

Department of Computer Science Helsinki Institute for Information Technology The Institute of Statistical Mathematics University of Helsinki non-gaussian linear non-Gaussian acyclic model causal discovery causal inference

cdt15/lingam Shimizu et al (2006).pdf

Paper: A New Learning Algorithm for Blind Signal Separation, Shun-ichi Amari et al. (1995). Introduces natural gradient learning for independent component analysis using non-Gaussian cumulants, providing the core statistical unmixing foundation directly exploited by the LiNGAM causal discovery framework.
Paper: Learning Bayesian networks: The combination of knowledge and statistical data, David Heckerman et al. (1994). Establishes foundational Bayesian network structure learning principles and the challenge of Markov equivalence classes, which LiNGAM overcomes by incorporating non-Gaussian disturbance assumptions.
Paper: A Tutorial on Learning with Bayesian Networks, David Heckerman (1999). Presents a comprehensive tutorial on learning graphical models and causal relationships from observational data under standard causal Markov assumptions.
Paper: A Bayesian method for the induction of probabilistic networks from data, Gregory F. Cooper et al. (1992). Develops the landmark Bayesian network structure induction framework that defined score-based causal DAG discovery prior to continuous non-Gaussian identification techniques.
Paper: Direct and Indirect Effects, Judea Pearl (2001). Provides the mathematical and structural equation framework for defining and identifying causal path effects in acyclic systems without unmeasured confounding.

Paper: The max-min hill-climbing Bayesian network structure learning algorithm, Ioannis Tsamardinos et al. (2006). Introduces the Max-Min Hill-Climbing algorithm, offering a premier hybrid baseline for continuous and discrete DAG structure discovery to contrast against ICA-based identification.
Paper: Learning Bayesian Networks with the bnlearn R Package, Marco Scutari (2009). Implements constraint-based and score-based network structure learning algorithms in R, providing the standard computational suite used alongside LiNGAM for causal graph benchmarking.
Paper: Invariant Risk Minimization, Martin Arjovsky et al. (2019). Extends linear structural equation modeling to multi-environment settings by learning invariant causal representations under intervention and distribution shift.
Paper: Identifying Weight-Variant Latent Causal Models, Yuhang Liu et al. (2026). Generalizes linear causal model identifiability to latent variable structures where causal parameters vary across contexts, building upon linear structural discovery foundations.
Paper: Counterfactual Fairness, Matt J. Kusner et al. (2017). Applies structural equation modeling and latent causal factor identification to enforce fairness constraints in downstream predictive models.

Cover for A Linear Non-Gaussian Acyclic Model for Causal Discovery

Abstract

In recent years, several methods have been proposed for the discovery of causal structure from non-experimental data. Such methods make various assumptions on the data generating process to facilitate its identification from purely observational data. Continuing this line of research, we show how to discover the complete causal structure of continuous-valued data, under the assumptions that (a) the data generating process is linear, (b) there are no unobserved confounders, and (c) disturbance variables have non-Gaussian distributions of non-zero variances. The solution relies on the use of the statistical method known as independent component analysis, and does not require any pre-specified time-ordering of the variables. We provide a complete Matlab package for performing this LiNGAM analysis (short for Linear Non-Gaussian Acyclic Model), and demonstrate the effectiveness of the method using artificially generated data and real-world data.

Topics

non-gaussian

Keywords

linear non-Gaussian acyclic model causal discovery causal inference causal ordering independent component analysis covariance structure analysis unobserved confounders non-Gaussian distributions acyclic graphs

1. Introduction
2. Linear Causal Networks
3. Model Identification Using Independent Component Analysis
4. LiNGAM Discovery Algorithm
5. Permutation Algorithms for Large Dimensions
5.1 Permuting the Rows of W
5.2 Permuting B to Get a Causal Order
6. Statistical Tests for Pruning Edges
6.1 Wald Test for Examining Significance of Edges
6.2 A Chi-Square Test for Evaluating the Overall Fit of the Estimated Model
6.2.1 MOMENT STRUCTURES OF MODELS
6.2.2 SOME TEST STATISTICS TO EVALUATE A MODEL FIT
6.2.3 A DIFFERENCE CHI-SQUARE TEST FOR MODEL COMPARISON OF NESTED MODELS
6.3 A Method for Pruning Edges
7. Simulations
7.1 Estimation of B
7.2 Pruning Edges
8. Examples With Real-World Data
9. Conclusions
Acknowledgments
Appendix A. Proof of Uniqueness of Row Permutation
Appendix B. An Example of the Permutation Problem
Appendix C. ML Derivation of Objective Function for Finding the Correct Row Permutation
Appendix D. Asymptotic Variance of ICA
Appendix E. Exact Form of J = ∂σ 2 ( τ ) / ∂τ T
References

Knowls

Knowl 1 — Linear Non-Gaussian Acyclic Model
model/method
The Linear Non-Gaussian Acyclic Model (LiNGAM) describes a data generating process over $m$ continuous observed random variables $x = [x_1, \dots, x_m]^T$ defined by structural linear equations:

$x_i = \sum_{k(j) < k(i)} b_{ij} x_j + e_i + c_i,$

where:
- $k(i)$ is a causal ordering function of the variables such that no later variable causes an earlier variable, ensuring the directed graph of relations is acyclic (DAG).
- $b_{ij}$ represents the linear connection strength from $x_j$ to $x_i$ .
- $c_i$ is an optional constant intercept term.
- $e_i$ are continuous disturbance (noise) terms with non-zero variances and non-Gaussian distributions.
- The disturbance variables $e_1, \dots, e_m$ are mutually independent, satisfying $p(e_1, \dots, e_m) = \prod_{i=1}^m p_i(e_i)$ , which implies causal sufficiency (absence of unobserved confounders).
In matrix notation, subtracting the means yields $x = Bx + e$ , where $B$ is an $m \times m$ coefficient matrix with zeros on the diagonal that can be permuted to strict lower triangular form by simultaneous row and column permutations corresponding to $k(i)$ .
Knowl 2 — Identifiability of LiNGAM Causal Structure via ICA
theoretical result
Under the LiNGAM assumptions, the full causal network structure and all connection strengths $B$ are uniquely identifiable from observational data without prior knowledge of the causal ordering.

From the mean-centered model $x = Bx + e$ , solving for $x$ gives:

$x = (I - B)^{-1} e = A e,$

where $A = (I - B)^{-1}$ is the mixing matrix and $W = A^{-1} = I - B$ is the demixing matrix. Because the disturbances $e_i$ are non-Gaussian and independent, standard linear Independent Component Analysis (ICA) identifies $A$ (and $W$ ) up to column/row permutations and component scalings: $W_{\text{ICA}} = P_{\text{ICA}} W$ , where $P_{\text{ICA}}$ is a random permutation matrix.

Because $B$ is strictly lower triangular under the true causal ordering $P_d$ , the demixing matrix can be written as $W = P_d M P_d^T$ , where $M$ is lower triangular with non-zero diagonal entries. For any lower triangular matrix $M$ with non-zero diagonal entries, a permutation $P_1 M P_2^T$ has only non-zero entries on the main diagonal if and only if $P_1 = P_2$ . Consequently, there exists exactly one row permutation of $W_{\text{ICA}}$ that produces non-zero entries along the main diagonal, uniquely resolving ICA's permutation indeterminacy and identifying the causal ordering and parameter matrix $B$ .
Knowl 3 — LiNGAM Discovery Algorithm
algorithm
The LiNGAM discovery algorithm estimates the causal ordering and the connection strength matrix $B$ from an observational data matrix $X$ of size $m \times n$ (with $m \ll n$ ).
Input: Continuous data matrix $X$ of size $m \times n$
Output: Estimated connection matrix $\hat{B}$ , causal permutation matrix $P$
Subtract sample mean from each row of $X$
Apply ICA decomposition to obtain $X = A S$ , where $S$ contains independent components
Compute demixing matrix $W = A^{-1}$
Find row permutation of $W$ yielding $\tilde{W}$ that minimizes $\sum_{i=1}^m 1/|\tilde{W}_{ii}|$
Divide each row $i$ of $\tilde{W}$ by its diagonal element $\tilde{W}_{ii}$ to obtain $\tilde{W}'$
Compute initial connection matrix estimate $\hat{B} = I - \tilde{W}'$
Find permutation matrix $P$ to minimize upper triangular sum of squares $\sum_{i \le j} (P \hat{B} P^T)_{ij}^2$
Set upper triangular and diagonal elements of $P \hat{B} P^T$ to 0
return $\hat{B}$ , $P$
Knowl 4 — Row Permutation and Causal Ordering for High Dimensions
algorithm
For high-dimensional networks where exhaustive search is infeasible, row permutation of $W$ and causal ordering of $\hat{B}$ are solved via linear assignment and iterative pruning.
1. Row Permutation of $W$ : The maximization of diagonal magnitudes is formulated as a linear assignment problem minimizing $\sum_{i=1}^m C_{\phi(i), i}$ , where $C_{ij} = 1 / |W_{ij}|$ and $\phi$ is a permutation. This is solved with worst-case computational complexity $O(m^3)$ .
2. Causal Ordering of $\hat{B}$ via Iterative Pruning:
Input: Estimated connection matrix $\hat{B}$ of size $m \times m$
Output: Permuted strictly lower triangular matrix $\tilde{B}$
Set the $m(m+1)/2$ smallest absolute value elements of $\hat{B}$ to 0
loop
    Test DAGness: search for a row with all zeros, append index to ordering list $p$ , remove that row and column from matrix, repeat until empty or no zero row found
    if DAGness test succeeds then
        Apply permutation $p$ to rows and columns of $\hat{B}$ to produce $\tilde{B}$
        return $\tilde{B}$
    else
        Set the next smallest absolute value element of $\hat{B}$ to 0
    end if
end loop
Knowl 5 — Wald Test for Edge Significance
model/method
To prune weak or spurious edges in the estimated network, a Wald test evaluates the null hypothesis $H_0: b_{ij} = 0$ (equivalent to $H_0: w_{ij} = 0$ ) against $H_1: b_{ij} \ne 0$ .

The Wald test statistic is:

$\text{Wald}_{ij} = \frac{\hat{w}_{ij}^2}{\text{avar}(\hat{w}_{ij})},$

where $\hat{w}_{ij}$ is the estimated element of $W$ and $\text{avar}(\hat{w}_{ij})$ is its asymptotic variance derived from the estimating equations of the ICA estimator (e.g., FastICA).

Under $H_0$ , the statistic $\text{Wald}_{ij}$ asymptotically follows a chi-square distribution with 1 degree of freedom ( $\chi^2(1)$ ). An edge is judged non-significant if the probability of observing a value greater than or equal to $\text{Wald}_{ij}$ exceeds a significance threshold $\alpha$ (e.g., 0.05).
Knowl 6 — Chi-Square Goodness-of-Fit and Difference Tests
equation
Overall model fit to observational second-order moments is evaluated by comparing empirical moments $m_2 = \frac{1}{n} \sum_{j=1}^n \text{vec}^+(x_j x_j^T)$ against model-implied moments $\sigma_2(\tau) = \text{vec}^+\{E(xx^T)\}$ , where $\tau$ contains the free parameters of $B$ and disturbance variances $E(e_i^2)$ , and $\text{vec}^+(\cdot)$ vectorizes the non-duplicate entries of a symmetric matrix.

The weighted least-squares distance is:

$F(\hat{\tau}) = \{m_2 - \sigma_2(\hat{\tau})\}^T \hat{M} \{m_2 - \sigma_2(\hat{\tau})\},$

where $\hat{M} = \hat{V}^{-1} - \hat{V}^{-1} \hat{J} (\hat{J}^T \hat{V}^{-1} \hat{J})^{-1} \hat{J}^T \hat{V}^{-1}$ , $\hat{V}$ is the sample covariance matrix of $m_2$ , and $\hat{J} = \partial \sigma_2(\tau)/\partial \tau^T |_{\tau=\hat{\tau}}$ .

The Yuan-Bentler corrected test statistic for overall fit is:

$T_2 = \frac{n F(\hat{\tau})}{1 + F(\hat{\tau})},$

which asymptotically follows $\chi^2(u - v)$ under $H_0: E(m_2) = \sigma_2(\tau)$ , where $u$ is the number of distinct second moments and $v$ is the number of free parameters in $\tau$ .

For two nested models with $q$ and $q-1$ edges, the difference statistic $T_2(q) - T_2(q-1)$ asymptotically follows $\chi^2(1)$ to test whether pruning an edge significantly degrades overall model fit.
Knowl 7 — Combined Statistical Pruning Procedure
algorithm
To prune the fully connected initial DAG estimate while maintaining overall model validity, individual Wald tests are combined with nested model difference tests and chi-square goodness-of-fit tests.
Input: Estimated lower triangular matrix $\tilde{B}$ , significance level $\alpha$ (e.g., 0.05)
Output: Pruned DAG connection matrix $\tilde{B}_{\text{pruned}}$
Identify all non-significant edges in strictly lower triangular part of $\tilde{B}$ using individual Wald tests at level $\alpha$
Order candidate non-significant edges by ascending significance (smallest Wald statistic first)
for each candidate non-significant edge do
    Create a candidate model with this edge set to 0
    Compute nested difference statistic $T_2(q) - T_2(q-1)$ against preceding model
    Compute overall fit statistic $T_2(q-1)$ of candidate model
    if difference test accepts null (fit change not significant) and overall fit test accepts null (model fits data) then
        Prune the edge (keep entry set to 0)
    else
        Retain the edge in the model
    end if
end for
return $\tilde{B}_{\text{pruned}}$

Knowl 8 — Empirical Edge Pruning Performance on Synthetic Data

data/table

Across 1000 simulation trials on synthetic linear acyclic models with non-Gaussian disturbances generated via power non-linearities, the combined statistical pruning method (evaluated at $\alpha = 0.05$ ) demonstrates increasing edge identification accuracy and power as sample size $n$ increases.

Condition	True Pos.	False Neg.	True Neg.	False Pos.	Sum of False Pos./Neg.
Dim. = 5, $n = 1000$	8101 (90.5%)	849 (9.5%)	921 (87.7%)	129 (12.3%)	978 (9.8%)
Dim. = 5, $n = 5000$	8556 (95.6%)	394 (4.4%)	943 (89.8%)	107 (10.2%)	501 (5.0%)
Dim. = 5, $n = 10000$	8691 (97.1%)	259 (2.9%)	972 (92.6%)	78 (7.4%)	337 (3.4%)
Dim. = 10, $n = 1000$	27825 (80.6%)	6698 (19.4%)	8171 (78.0%)	2306 (22.0%)	9004 (20.0%)
Dim. = 10, $n = 5000$	31623 (91.6%)	2900 (8.4%)	9350 (89.2%)	1127 (10.8%)	4027 (8.9%)
Dim. = 10, $n = 1000$	32477 (94.1%)	2046 (5.9%)	9466 (90.4%)	1011 (9.6%)	3057 (6.8%)

True positives and negatives were counted strictly within the lower triangular support. For $n \ge 5000$ , statistical power exceeds $90\%$ and total error rates (false positives plus false negatives) remain under $10\%$ across both 5- and 10-variable networks.

Knowl 9 — Causal Discovery in Autoregressive Time Series
empirical result
A stationary autoregressive process of order $p$ , $X_t = \sum_{k=1}^p \phi_k X_{t-k} + a_t$ with independent non-Gaussian white noise $a_t$ , conforms to LiNGAM when represented over sliding time windows $[X_t, X_{t+1}, \dots, X_{t+m-1}]$ as multivariate samples (though order $p \ge 2$ introduces latent confounding for the first $p$ variables in a window).

Empirical evaluation over 22 real-world economic and environmental time series revealed four distinct outcome behaviors:
1. Correct Causal Ordering (5 cases): Observed in stationary series with non-Gaussian innovations (e.g., monthly precipitation PEAS), correctly recovering AR parameters.
2. Reverse Causal Ordering (9 cases): Observed in non-stationary random walks $X_t = X_{t-1} + a_t$ (e.g., daily IBM stock prices), where time direction is mathematically symmetric and unidentifiable, yielding estimated first AR coefficients near 1 in either direction.
3. Zero Connection Matrix (1 case): Observed in independent white noise series (e.g., daily precipitation PRECIP), where all coefficients are statistically non-significant.
4. Inconsistent Ordering (7 cases): Caused by strong model violations, such as unmodelled non-linearities, strong seasonal/trend non-stationarity, or external confounding (e.g., atmospheric $\text{CO}_2$ MLCO2).
Knowl 10 — Limitations and Failure Modes of LiNGAM
limitation
The LiNGAM framework is subject to several core limitations and failure modes:
- Gaussianity: If disturbances $e_i$ are Gaussian, ICA mixing matrices are identifiable only up to orthogonal rotations, leaving the causal direction indistinguishable beyond the Markov equivalence class.
- Unobserved Confounding: Latent common causes $f$ alter the disturbance structure to $e = G f + e'$ , introducing statistical dependencies among disturbance terms that violate mutual independence $p(e_1, \dots, e_m) = \prod p_i(e_i)$ .
- Non-stationarity and Random Walks: Processes with unit roots ( $X_t = X_{t-1} + a_t$ ) create symmetrical forward-backward statistical representations, rendering temporal direction unidentifiable.
- Non-convex ICA Optimization: Algorithms such as FastICA can become trapped in local optima, producing inconsistent estimates across different random initializations when sample size is low or model assumptions fail.

Coverage note — Omitted the explicit tensor and matrix derivative derivations for the asymptotic variance formulas (Appendices D and E) and the maximum likelihood derivation of the linear assignment objective (Appendix C), as these serve as mathematical derivations supporting the stated test statistics and algorithms rather than standalone conceptual contributions.

References

1.Y. Benjamini and Y. Hochberg. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B, 57:289–300, 1995.
2.K. A. Bollen. Structural Equations with Latent Variables. John Wiley & Sons, 1989.
3.G. E. P. Box and G. M. Jenkins. Time Series Analysis: forecasting and control. Holden-Day, Oakland, California, USA, revised edition, 1976.
4.P. J. Brockwell and R. A. Davis. Time Series: Theory and Methods. Springer-Verlag, New York, USA, 1987.
5.M. W. Browne. Asymptotically distribution-free methods for the analysis of covariance structures. British Journal of Mathematical and Statistical Psychology, 9:665–672, 1984.
6.R. E. Burkard and E. Cela. Linear assignment problems and extensions. In P. M. Pardalos and D. Z. Du, editors, Handbook of Combinatorial Optimization - Supplement Volume A, pages 75–149. Kluwer, 1999.
7.J.-F. Cardoso and B. H. Laheld. Equivariant adaptive source separation. IEEE Trans. on Signal Processing, 44:3017–3030, 1996.
8.J.-F. Cardoso and A. Souloumiac. Blind beamforming for non Gaussian signals. IEE Proceedings-F, 140(6):362–370, 1993.
9.P. Comon. Independent component analysis – a new concept? Signal Processing, 36:287–314, 1994.
10.Y. Dodge and V. Rousson. On asymptotic properties of the correlation coefficient in the regression setting. The American Statistician, 55(1):51–54, 2001.
11.B. Efron and R. Tibshirani. An Introduction to the Bootstrap. Chapman & Hall, New York, 1993.
12.D. Geiger and D. Heckerman. Learning gaussian networks. In Proceedings of the 10th Annual Conference on Uncertainty in Artificial Intelligence (UAI-94), pages 235–243, 1994.
13.V. P. Godambe. Estimating functions. Oxford University Press, New York, 1991.
14.J. Himberg, A. Hyvärinen, and F. Esposito. Validating the independent components of neuroimaging time-series via clustering and visualization. Neuroimage, 22:1214–1222, 2004.
15.Y. Hochberg. A sharper Bonferroni procedure for multiple tests of significance. Biometrika, 4: 800–802, 1988.
16.Y. Hochberg and A. C. Tamhane. Multiple comparison procedures. John Wiley & Sons, New York, 1987.
17.S. Holm. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6:65–70, 1979.
18.P. O. Hoyer, S. Shimizu, A. Hyvärinen, Y. Kano, and A. J. Kerminen. New permutation algorithms for causal discovery using ICA. In Proceedings of International Conference on Independent Component Analysis and Blind Signal Separation, Charleston, SC, USA, pages 115–122, 2006a.
19.P. O. Hoyer, S. Shimizu, and A. J. Kerminen. Estimation of linear, non-gaussian causal models in the presence of confounding latent variables. In Proc. the third European Workshop on Probabilistic Graphical Models (PGM2006), 2006b. In press.
20.L. Hu, P. M. Bentler, and Y. Kano. Can test statistics in covariance structure analysis be trusted? Psychological Bulletin, 112:351–362, 1992.
21.R. J. Hyndman. Time series data library, 2005. URL http://www-personal.buseco.monash.edu.au/~hyndman/TSDL/. [June 2005].
22.A. Hyvärinen. One-unit contrast functions for independent component analysis: A statistical analysis. In Neural Networks for Signal Processing VII (Proceedings of IEEE Workshop on Neural Networks for Signal Processing), pages 388–397, 1997.
23.A. Hyvärinen. Fast and robust fixed-point algorithms for independent component analysis. IEEE Trans. on Neural Networks, 10(3):626–634, 1999.
24.A. Hyvärinen, J. Karhunen, and E. Oja. Independent Component Analysis. Wiley Interscience, 2001.
25.M. Kawanabe and K. R. Müller. Estimating functions for blind separation when sources have variance dependencies. Journal of Machine Learning Research, 6:453–482, 2005.
26.National Statistics, 2005. URL http://www.statistics.gov.uk/. [June 2005].
27.J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, 2000.
28.D. T. Pham and P. Garrat. Blind separation of mixture of independent sources through a quasi-maximum likelihood approach. Signal Processing, 45:1457–1482, 1997.
29.S. Shimizu, A. Hyvärinen, P. O. Hoyer, and Y. Kano. Finding a causal ordering via independent component analysis. Computational Statistics & Data Analysis, 50(11):3278–3293, 2006a.
30.S. Shimizu, A. Hyvärinen, Y. Kano, and P. O. Hoyer. Discovery of non-gaussian linear causal models using ICA. In Proc. the 21st Conference on Uncertainty in Artificial Intelligence (UAI-2005), pages 526–533, 2005.
31.S. Shimizu, A. Hyvärinen, Y. Kano, P. O. Hoyer, and A. J. Kerminen. Testing significance of mixing and demixing coefficients in ICA. In Proceedings of International Conference on Independent Component Analysis and Blind Signal Separation, Charleston, SC, USA, pages 901–908, 2006b.
32.S. Shimizu and Y. Kano. Use of non-normality in structural equation modeling: Application to direction of causation. Journal of Statistical Planning and Inference, 2006. In press.
33.R. J. Simes. An improved Bonferroni procedure for multiple tests of significance. Biometrika, 73: 751–754, 1986.
34.P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction, and Search, 2nd ed. MIT Press, 2000.
35.Statistical Software Information, 2005. URL http://www-unix.oit.umass.edu/~statdata/. [June 2005].
36.P. Tichavský, Z. Koldovský, and E. Oja. Performance analysis of the FastICA algorithm and Cramèr-Rao bounds for linear independent component analysis. IEEE Trans. on Signal Processing, 54(4):1189–1203, 2006.
37.K-H. Yuan and P. M. Bentler. Mean and covariance structure analysis: Theoretical and practical improvements. Journal of the American Statistical Association, 92(438):767–774, 1997.

Citation

MLA

Shimizu, S., et al. “A Linear Non-Gaussian Acyclic Model for Causal Discovery”. 2006, http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.163.2163.

APA

Shimizu, S., Hoyer, P. O., Hyvärinen, A., & Kerminen, A. (2006). A Linear Non-Gaussian Acyclic Model for Causal Discovery. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.163.2163

Chicago

Shimizu, S., P. O. Hoyer, A. Hyvärinen, and A. Kerminen. 2006. “A Linear Non-Gaussian Acyclic Model for Causal Discovery”. Preprint. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.163.2163.

Harvard

Shimizu, S. et al. (2006) “A Linear Non-Gaussian Acyclic Model for Causal Discovery”. Available at: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.163.2163.

Vancouver

1. Shimizu S, Hoyer PO, Hyvärinen A, Kerminen A (2006) A Linear Non-Gaussian Acyclic Model for Causal Discovery.

BibTeX

@article{shimizu2006linear,
  title = {A Linear Non-Gaussian Acyclic Model for Causal Discovery},
  author = {Shimizu, Shohei and Hoyer, Patrik O. and Hyvärinen, Aapo and Kerminen, Antti},
  year = {2006},
  url = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.163.2163}
}

Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

A Linear Non-Gaussian Acyclic Model for Causal Discovery

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Linear Non-Gaussian Acyclic Model

Knowl 2 — Identifiability of LiNGAM Causal Structure via ICA

Knowl 3 — LiNGAM Discovery Algorithm

Knowl 4 — Row Permutation and Causal Ordering for High Dimensions

Knowl 5 — Wald Test for Edge Significance

Knowl 6 — Chi-Square Goodness-of-Fit and Difference Tests

Knowl 7 — Combined Statistical Pruning Procedure