We consider the problem of downlink traffic control in Multi-code Code Division Multiple Access (MC-CDMA) systems, which support multiple classes of services with diverse QoS requirements. Prior solutions proposed for this problem have largely focused on call admission control at the connection level while neglecting the stochastic behavior of mobile subscribers and channel conditions. We quantitatively demonstrate that these statistical factors, in particular log-normal shadowing in propagation and voice activity factors, have a significant impact on the connection-level performance. Furthermore, we show that conventional data services can be best handled at the packet level as background transmissions by taking advantage of these statistical variations, which leads to significantly better utilization of the scarce wireless spectrum.