Coda waves are often considered to be generated by backscattering of primary waves from randomly distributed heterogeneities in the crust; however, modeling and experimental work indicates that much of the coda may be due to scattering near the receiver. Knowing where the coda is generated is important for several reasons. Coda waves have been used to characterize site amplification, scattering and attenuation throughout the crust, and velocity changes associated with large earthquakes; however, a lack of knowledge of where coda waves are generated can lead to ambiguous interpretations. In this study we analyze 26 s of coda waves recorded at up to 78 stations using slowness stacking on two source arrays and find that at nearly all stations, regardless of distance from the source, nearly all the arrivals are within one hemisphere, most are strongly clustered near the direct arrival, and the clusters are, in general, symmetrically disposed about the array axis (fault plane). This finding suggests that the coda consists primarily of waves scattered near the stations rather than waves scattered throughout the crust. This result holds even at stations very close to the sources where the amount of coda examined is about 6 times the direct S travel time, and it suggests that much of the coseismic velocity decrease associated with the 1989 Loma Prieta earthquake <Ellsworth et al., 1992> occurred in the shallow crust near the stations. One possible mechanism for inducing such a change is a decrease in the mean stress from the mainshock slip; however, on the basis of static stress modeling, we find that this mechanism would produce a velocity increase in the region where velocity was observed to decrease. We conclude that the velocity decrease was more likely caused by crack opening or crack coalescence due to strong shaking from the mainshock.¿ 1997 American Geophysical Union