Information theory has been introduced to test nonlinear interrelation structures in multivariate time series. In this paper, we propose an information theoretic statistic for testing direct nonlinear dependence. The statistic is defined as the difference between general conditional mutual information and linear conditional mutual information measures. The general conditional mutual information is estimated by the correlation integral. The significance of the test statistic is determined by means of bootstrap method based on surrogate data. The size and power properties of the test are examined by simulation examples.