In recent years, network coding has been investigated as a way to improve throughput and power efficiency of wireless networks by mixing of various traffic flows. The scenario under consideration is a half-duplex decode-and-forward relay channel. The receiving stations decode the error-correcting-coded bits using Log-Likelihood Ratio (LLR). Here, it is important to set initial LLR to decode the coded data accurately. In this paper, we present how to set the initial LLR for belief propagation decoding with network coding to achieve better decoding error rate performance. We calculate initial LLR based not only on channels from relay station to receiving stations, but also on all channels from sending stations to receiving stations when stations decode error-correcting coded data. When we use turbo code or low-density parity-check (LDPC) code as error-correcting code, we decode the coded data accurately by the use of accurate initial LLR. We provide computer simulation results both under perfect and imperfect channel state information. We show that the proposed method improves bit error rate (BER) performance and that the proposed method decreases the number of decoding iterations.