Location details of social users are important in diverse applications ranging from news recommendation systems to disaster management. However, user location is not easy to obtain from social networks because many users do not bother to provide this information or decline to do so due to privacy concerns. Thus, it is useful to estimate user locations from implicit information in the network. For this purpose, many location prediction models have been proposed that exploit different network features. Unfortunately, these models have not been benchmarked on common datasets using standard metrics. We fill this gap and provide an in-depth empirical comparison of eight representative prediction models using five metrics on four real-world large-scale datasets, namely Twitter, Gowalla, Brightkite, and Foursquare. We formulate a generalized procedure-oriented location prediction framework which allows us to evaluate and compare the prediction models systematically and thoroughly under extensive experimental settings. Based on our results, we perform a detailed analysis of the merits and limitations of the models providing significant insights into the location prediction problem.