Emerging technologies now allow for mass spectrometry-based profiling of thousands of small molecule metabolites (‘metabolomics’) in an increasing number of biosamples. While offering great promise for insight into the pathogenesis of human disease, standard approaches have not yet been established for statistically analyzing increasingly complex, high-dimensional human metabolomics data in relation to clinical phenotypes, including disease outcomes. To determine optimal approaches for analysis, we formally compare traditional and newer statistical learning methods across a range of metabolomics dataset types. In simulated and experimental metabolomics data derived from large population-based human cohorts, we observe that with an increasing number of study subjects, univariate compared to multivariate methods result in an apparently higher false discovery rate as represented by substantial correlation between metabolites directly associated with the outcome and metabolites not associated with the outcome. Although the higher frequency of such associations would not be considered false in the strict statistical sense, it may be considered biologically less informative. In scenarios wherein the number of assayed metabolites increases, as in measures of nontargeted versus targeted metabolomics, multivariate methods performed especially favorably across a range of statistical operating characteristics. In nontargeted metabolomics datasets that included thousands of metabolite measures, sparse multivariate models demonstrated greater selectivity and lower potential for spurious relationships. When the number of metabolites was similar to or exceeded the number of study subjects, as is common with nontargeted metabolomics analysis of relatively small cohorts, sparse multivariate models exhibited the most-robust statistical power with more consistent results. These findings have important implications for metabolomics analysis in human disease.