File-type Identification (FTI) is an important problem in digital forensics, intrusion detection, and other related fields. Using state-of-the-art classification techniques to solve FTI problems has begun to receive research attention, however, general conclusions have not been reached due to the lack of thorough evaluations for method comparison. This paper presents a systematic investigation of the problem, algorithmic solutions and an evaluation methodology. Our focus is on performance comparison of statistical classifiers (e.g. SVM and kNN) and knowledge-based approaches, especially COTS (Commercial Off-The-Shelf) solutions which currently dominate FTI applications. We analyze the robustness of different methods in handling damaged files and file segments. We propose two alternative criteria in measuring performance: 1) treating file-name extensions as the true labels, and 2) treating the predictions by knowledge based approaches on intact files, these rely on signature bytes as the true labels (and removing these signature bytes before testing each method). In our experiments with simulated damages in files, SVM and kNN substantially outperform all the COTS solutions we tested, improving classification accuracy very substantially -- some COTS methods cannot identify damaged files at all.