We propose the quality-of-service (QoS) driven power allocation schemes for parallel fading channels when considering imperfect channel estimations. In particular, the parallel communication model plays a fundamental role in physical-layer evolutions of wireless networks. By integrating information theory with the concept of effective capacity, our proposed schemes aim at maximizing the system throughput subject to a given delay constraint. Solving the original non-convex problem by a 2-dimensional convex optimization approach, we develop the optimal allocation algorithms under different QoS and power constraints. Consistent with our previous work assuming perfect channel state information (CSI), our analyses considering imperfect CSI demonstrate that when the QoS constraint becomes more and more stringent, the optimal effective capacity decreases from the ergodic capacity to the zero-outage capacity. Moreover, our results indicate that the channel estimation error has a significant impact on QoS provisioning, especially when the delay constraint is stringent. Specifically, as long as the channel estimation is not perfect, a positive zero-outage capacity is unattainable. On the other hand, our simulations also suggest that a larger number of parallel channels can provide higher throughput and more stringent QoS, while offering better robustness against the imperfectness of CSI.