MPTCP is a promising transport technique to boost throughput of wireless multi-homed device by supporting multiple concurrent transmissions through heterogeneous wireless interfaces. As the number of concurrent subflows increase, MPTCP can achieve a linearly increasing throughput performance, but it is unclear how much improvement in the end-to-end delay performance can be attained from additional subflows. In this paper, we develop an analytical framework to understand the end-to-end delay performance of MPTCP that accounts for TCP dynamics and subflow interactions. Interestingly, it turns out that the delay performance can be even degraded with additional subflows. Considering MPTCP in heterogeneous wireless networks of Wi-Fi and LTE, we formulate a cost minimization problem subject to the end-to-end delay constraint. Based on the insight obtained from our model, we approximate the problem and develop a greedy scheme that splits traffic to minimize the cost while satisfying the delay constraints. Through simulations, we demonstrate that our proposed scheme outperforms the conventional MPTCP, and significantly improves the delay performance while lowering the user cost.