Statistical analysis of analog circuits usually relies on the standard Monte Carlo method to estimate the yield of a circuit. However, this method is limited by a slow convergence rate which leads to a prohibitive number of simulations to reach a given accuracy. In this paper, we propose to combine the kernelbased distribution estimator with the control variates method in order to obtain an accurate yield estimation with only a few hundred simulations. With respect to the auxiliary variable needed for the control variates method, we propose a quick modeling technique based on local sensitivities.