Retrieving data from mobile source vehicles is a crucial routine operation for a wide spectrum of vehicular network applications, including road surface monitoring and sharing. Network coding has been widely exploited and is an effective technique for diffusing information over a network. The use of network coding to improve data availability in vehicular networks is explored in this paper. With random linear network codes, simple replication is avoided, and instead, a node forwards a coded block that is a random combination of all data received by the node. We use a network-coding-based approach to improve data availability in vehicular networks. To determine the feasibility of this approach, we conducted an empirical study with extensive simulations based on two real vehicular GPS traces, both of which contain records from thousands of vehicles over more than a year. We observed that, despite significant improvement in data availability, there is a serious issue with linear correlation between the received codes. This reduces the data-retrieval success rate. By analyzing the real vehicular traces, we discovered that there is a strong community structure within a real vehicular network. We verify that such a structure contributes to the issue of linear dependence. Then, we point out opportunities to improve the network-coding-based approach by developing community-aware code-distribution techniques.