Comparisons of seismic moment M0 in Harvard's centroid moment tensor (CMT) catalog with M0 estimated from regional surface waves reveal a systematic discrepancy of about a factor of 2 for earthquakes located in western China. Causes of the discrepancy are investigated with analyses of observations and with calculations of Green's function responses for relevant Earth models. The PREM model used in CMT inversions is characterized by a thin, high-velocity crust, unlike tectonic structures in western China, while ASIA, used for modeling surface waves, was developed for this region. Analyses of teleseismic body wave synthetics for models similar to PREM and ASIA show that medium properties, source radiation patterns, and the presence of interfering phases introduce systematic amplitude variations. Inversions of synthetic data were carried out to simulate features of observed M0 residuals, log<M0(reg)/M0(CMT)>, which are (1) a bias of -0.24 ¿ 0.04 log units for dip-slip events with mb 6 to 6.8, (2) a tendency for bias to increase with decreasing mb, and (3) a mechanism dependence, where residuals are large for high-angle, dip-slip events and small for vertical strike-slip events. Residuals for small events scatter more, probably due to larger errors in source depth and mechanism affecting both M0 estimates. Results of inverting synthetics mimic magnitude and mechanism dependences but underestimate the bias in feature 1. Rather than impugn the integrity of CMT catalogs, this study sheds light on uncertainties introduced by limitations of the methodology.