There have been no studies evaluating the dominant factor responsible for generating self-similar network traffic, although many researchers have discussed the causes of self-similarity. In this paper, we have isolated and compared the effects of the application, transport and network layers on self-similarity. We first focus on the interaction between the transport and network layers only. With a simple computer simulation scenario we show that in realistic bottleneck scenarios with only long greedy TCP flows present, the TCP flows fully utilize the bottleneck link, which make the aggregated TCP flow short-range dependent while each flow that share the bottleneck link itself is long range dependent. We also show that in special situations, when the TCP flows become synchronized, the bottleneck link may become underutilized giving space to long-range dependence to surface. We show the connection between the network layer parameters necessary for synchronization to occur, and verify our results with simulation. To examine the effect of the application layer, we introduce web traffic besides the long greedy flows into the bottleneck link. We show that link utilization is again a key factor determining the self-similar properties of the aggregated flow, but has much less impact on the self-similar properties of the aggregated web sessions and long greedy flows. Finally, we point out the most influential effects and the conditions when they dominate the self-similar properties of aggregated TCP flows.