Embedded Files
Post date: May 25, 2012 5:05:59 PM
Jonas Peters
Abstract:
We consider structural equation models (SEMs) in which variables can be written as a function of their parents and noise terms (the latter are assumed to be jointly independent). Corresponding to each SEM, there is a directed acyclic graph (DAG) G_0 describing the relationships between the variables. In Gaussian SEMs with linear functions, the graph can be identified from the joint distribution only up to Markov equivalence classes (assuming faithfulness). It has been shown, however, that this constitutes an exceptional case. In the case of linear functions and non-Gaussian noise, the DAG becomes identifiable. In this work we present two alternative directions of deviating from the general linear Gaussian case: (i) apart from few exceptions identifiability also holds for non-linear functions and arbitrarily distributed additive noise. And (ii), if we require all noise variables to have the same variances, again, the DAG can be recovered from the joint distribution.
Our results can be applied to the problem of causal inference. If the data follow one of the two model assumptions and given that all variables are observed, the causal structure can be inferred from observational data only. We present practical methods for solving this task.
Papers:
Page updated
Report abuse