We consider estimation problem in structural vector autoregressive model which disturbance has non-Gaussian distribution. We call this model as non-Gaussian vector autoregressive (NG-SVAR) model. Since the estimation problem of this model is closely related to the independent component analysis (ICA) developed in machine learning and signal processing we apply the theory of ICA to our estimation problem. However, since we do not know the true non-Gaussian distribution in practice, we cannot construct the exact loglikelihood function. In this paper we propose a pseudo maximum loglikelihood estimator instead. It is shown that our estimator is statistical efficient from view point of semiparametric statistics. Furthermore, we show that our estimator has satisfactory performance by Monte Carlo experiment and empirical example in small sample.