BACKGROUND: Untargeted mass spectrometry (MS)-based metabolomics data often contain missing values that reduce statistical power and can introduce bias in biomedical studies. However, a systematic assessment of the various sources of missing values and strategies to handle these data has received little attention. Missing data can occur systematically, e.g. from run day-dependent effects due to limits of detection (LOD); or it can be random as, for instance, a consequence of sample preparation. METHODS: We investigated patterns of missing data in an MS-based metabolomics experiment of serum samples from the German KORA F4 cohort (n = 1750). We then evaluated 31 imputation methods in a simulation framework and biologically validated the results by applying all imputation approaches to real metabolomics data. We examined the ability of each method to reconstruct biochemical pathways from data-driven correlation networks, and the ability of the method to increase statistical power while preserving the strength of established metabolic quantitative trait loci. RESULTS: Run day-dependent LOD-based missing data accounts for most missing values in the metabolomics dataset. Although multiple imputation by chained equations performed well in many scenarios, it is computationally and statistically challenging. K-nearest neighbors (KNN) imputation on observations with variable pre-selection showed robust performance across all evaluation schemes and is computationally more tractable. CONCLUSION: Missing data in untargeted MS-based metabolomics data occur for various reasons. Based on our results, we recommend that KNN-based imputation is performed on observations with variable pre-selection since it showed robust results in all evaluation schemes.