The goal of many LC-MS proteomic investigations is to quantify and compare the abundance of proteins in complex biological mixtures. However, the output of an LC-MS experiment is not a list of proteins, but a list of quantified spectral features. To make protein-level conclusions, researchers typically apply ad hoc rules, or take an average of feature abundance to obtain a single protein-level quantity for each sample. We argue that these two approaches are inadequate. We discuss two statistical models, namely, fixed and mixed effects Analysis of Variance (ANOVA), which views individual features as replicate measurements of a protein's abundance, and explicitly account for this redundancy. We demonstrate, using a spike-in and a clinical data set, that the proposed models improve the sensitivity and specificity of testing, improve the accuracy of patient-specific protein quantifications, and are more robust in the presence of missing data.