OBJECTIVE: To determine the optimal methods for measuring mismatch negativity (MMN), an auditory event-related potential (ERP), and quantify sources of MMN variance in a multisite setting. METHODS: Reliability of frequency, duration, and double (frequency + duration) MMN was determined from eight traveling subjects, tested on two occasions at eight laboratory sites. Deviant-specific variance components were estimated for MMN peak amplitude and latency measures using different ERP processing methods. Generalizability (G) coefficients were calculated using two-facet (site and occasion), fully-crossed models and single-facet (occasion) models within each laboratory to assess MMN reliability. RESULTS: G-coefficients calculated from two-facet models indicated fair (0.4 < G<=0.6) duration MMN reliability at electrode Fz, but poor (G < 0.4) double and frequency MMN reliability. Single-facet G-coefficients averaged across laboratory resulted in improved reliability (G > 0.5). MMN amplitude reliability was greater than latency reliability, and reliability with mastoid referencing significantly outperformed nose-referencing. CONCLUSIONS: EEG preprocessing methods have an impact on the reliability of MMN amplitude. Within site MMN reliability can be excellent, consistent with prior single site studies. SIGNIFICANCE: With standardized data collection and ERP processing, MMN can be reliably obtained in multisite studies, providing larger samples sizeswithin rare patient groups.