Non-orthogonal multiple access (NOMA) has recently drawn attentions on its ability to fairly serve multiple users on the same radio resource with a desirable performance. However, achievable NOMA gain is primarily limited by channel gain disparity and successive interference cancellation (SIC) receiver characteristics. Accordingly, we introduce an imperfect SIC receiver model considering the power disparity and sensitivity constraints, delay tolerance, and residual interference due to detection and estimation errors. Then, a generic cluster formation (CF) and Power-Bandwidth Allocation (PBA) is formulated as a mixed-integer non-linear programming (MINLP) problem for downlink (DL) heterogeneous networks (HetNets). After dividing the MINLP problem into mixed-integer and non-linear subprobems, we first transform CF into a multi-partite matching, which is solved sequentially using bi-partite matching techniques. For sumrate maximization, max-min fairness, and energy & spectrum efficiency objectives, we secondly put highly non-convex joint PBA into a convex form using geometric programming (GP). Extensive simulations unleash the potential of NOMA to handle large number of users, traffic offloading, and user fairness.