Discovering Unconfounded Causal Relationships Using Linear Non-Gaussian Models
Causal relationships among a set of observed variables are often modeled using directed acyclic graph (DAG) structures, and learning such structures from data is known as the causal discovery problem. We here consider the learning of linear non-Gaussian acyclic models  with hidden variables . Estimation of such models is computationally challenging and hence only possible when the number of variables is small. We present an algorithm for obtaining partial but in the large sample limit correct information about pairwise total causal effects in such a model. In particular, we obtain consistent estimates of the total effects for all variable pairs for which there exist an unconfounded superset of observed variables. Simulations show that the estimated pairwise total effects are good approximations of the true total effects.
KeywordsCausal Discovery Latent Variables Non-Gaussianity
Unable to display preview. Download preview PDF.
- 2.Darmois, G.: Analyse générale des liaisons stochastiques. RIIS 21 (1953)Google Scholar
- 3.Eriksson, J., Koivunen, V.: Identifiability, Separability, and Uniqueness of Linear ICA Models. IEEE Signal Processing Letters 11(7) (2004)Google Scholar
- 4.Gretton, A., Fukumizu, K., Teo, C.H., Song, L., Schölkopf, B., Smola, A.J.: A Kernel Statistical Test of Independence. In: Adv. NIPS (2008)Google Scholar
- 5.Hoyer, P.O., Shimizu, S., Kerminen, A.J., Palviainen, M.: Estimation of causal effects using linear non-Gaussian models with hidden variables. IJAR 49 (2008)Google Scholar
- 7.Kawahara, Y., Bollen, K., Shimizu, S., Washio, T.: GroupLiNGAM: Linear non-Gaussian acyclic models for sets of variables. arXiv:1006.5041v1 (2010)Google Scholar
- 9.Shimizu, S., Hoyer, P.O., Hyvärinen, A., Kerminen, A.: A linear non-gaussian acyclic model for causal discovery. JMLR 7 (2006)Google Scholar
- 10.Skitovitch, W.P.: On a property of the normal distribution. DAN SSSR 89 (1953)Google Scholar