In this letter, we study the ergodic secrecy sum-rate of a multiuser downlink system achievable by using the regularized zero-forcing precoding based on imperfect channel estimation. The channels between the base station and users are modeled as fading channels with different spatial correlations and line-of-sight components. By using the limiting distribution of a new class of random matrix, we derive a large-system approximation of the ergodic secrecy sum-rate in the multiuser downlink system. The large-system approximation provides a reliable performance prediction, which enables us to identify the optimal regularization parameter efficiently.