Causal Dantzig: fast inference in linear structural equation models with hidden variables under additive interventions (Q2414102)

This paper develops a new approach for causal inference in a linear Structural Equation Model setting when data from distinct observational environments are available. The environments can have an arbitrary and unknown intervention effect on all predictor variables and the method exploits invariance of the conditional distribution of the target variable, given its causal parents, across environments. Arbitrary interventions on all variables are allowed, except direct interventions on the target variable. A notion of inner-product invariance is introduced. A causal Dantzig estimator is created, with unregularized and regularized versions. Regularization is based on a regularization problem similar to that of the Dantzig selector and the regularized causal Dantzig inherits its variable selection, shrinkage and regularization properties. Bounds for the regularized causal Dantzig involve a causal cone invertibility factor. Identifiability conditions for the causal parameter are discussed and asymptotic confidence intervals are derived in the low-dimensional setting. In the high-dimensional setting finite-sample bounds are derived. Their performance is investigated on simulated datasets. For two environments the causal Dantzig estimator is compared with instrumental variable regression and it is shown that it is consistent in certain settings in which instrumental variable regression fails. A two-stage alternative to running the causal Dantzig directly on a high-dimensional dataset is proposed, based on doing preselection. The first stage would screen out as many variables as possible, except for the parental set of the target variable. In the second stage, the causal Dantzig with or without regularization will run on the active set of the first stage.