Ground-motion relations for earthquakes that occur in subduction zones are an important input to seismic-hazard analyses in many parts of the world. In the Cascadia region (Washington, Oregon, northern California, and British Columbia), for example, there is a significant hazard from megathrust earthquakes along the subduction interface and from large events within the subducting slab. These hazards are in addition to the hazard from shallow earthquakes in the overlying crust. We have compiled a response spectra database from thousands of strong-motion recordings from events of moment magnitude (M) 5-8.3 occurring in subduction zones around the world, including both interface and in-slab events. The 2001 M 6.8 Nisqually and 1999 M 5.9 Satsop earthquakes are included in the database, as are many records from subduction zones in Japan (Kyoshin-Net data), Mexico (Guerrero data), and Central America. The size of the database is four times larger than that available for previous empirical regressions to determine ground-motion relations for subduction-zone earthquakes. The large dataset enables improved determination of attenuation parameters and magnitude scaling, for both interface and in-slab events. Soil response parameters are also better determined by the data. We use the database to develop global ground-motion relations for interface and in-slab earthquakes, using a maximum likelihood regression method. We analyze regional variability of ground-motion amplitudes across the global database and find that there are significant regional differences. In particular, amplitudes in Cascadia differ by more than a factor of 2 from those in Japan for the same magnitude, distance, event type, and National Earthquake Hazards Reduction Program (NEHRP) soil class. This is believed to be due to regional differences in the depth of the soil profile, which are not captured by the NEHRP site classification scheme. Regional correction factors to account for these differences are proposed for Cascadia and Japan. The results of this study differ significantly from previous analyses based on more limited data and have important implications for seismic-hazard analysis. The ground-motion relations predict that a great megathrust earthquake (M ???8) at a fault distance of about 100 km would produce pseudoacceleration (PSA), 5% damped, horizontal component on soil sites of about 110 cm/sec2 at 0.5 Hz, 660 cm/sec2 at 2.5 Hz, and 410 cm/sec2 at 5 Hz, with a peak ground acceleration of about 180 cm/ sec2 . These damaging levels of motion would be experienced over a very large area, corresponding to a rectangular area about 300 km wide by 500 km long. Large in-slab events (M 7.5) would produce even higher PSA values within 100 km of the fault, but the in-slab motions attenuate much more rapidly with distance. Thus the hazard posed by moderate to large in-slab events such as the 2001 Nisqually earthquake is modest compared to that of a Cascadia megathrust earthquake of M ???8, in terms of the area that would experience damaging levels of ground motion.