Twenty-seven earthquakes in the northwest Pacific, with mb ≧ 5.4, are used to study upper-mantle attenuation. The data were recorded at seismometer arrays in Australia, Canada, India, and Scotland. The short-period instruments used in the arrays are sufficiently broadband, and the signal-to-noise enhancement resulting from beamforming the data is good enough, that seismic noise is above system noise out to between 3 and 5 Hz, and in some cases to 8 Hz. The instrument response has been removed and a power law (ω−2) farfield source correction has been assumed and applied. The remaining fall-off in the spectra above 1 Hz (which is above the corner frequency fc) is assumed to be due to attenuation, including losses due both to anelasticity and to scattering. An ω−3 spectral fall-off, preferred by some, would result in lower estimates of attenuation. Conservative estimates of the average attenuation along the whole path have been made from the spectral fall-off. Despite the imperfect knowledge of the source spectra, the high sensitivity of spectral amplitudes to attenuation at high frequencies yields quite reliable and stable attenuation estimates. The scatter in the data is small, resulting in estimates of mean t* (tAV*) with an average standard error of only 21%, where tAV* is defined as the ratio of the travel time T to the effective mean quality factor QAV. Only 14 of the 64 measurements show a tAV* above 0.5 sec within the given error bounds. All estimates of tAV* and QAV refer to P-wave attenuation only. The results for tAV* from the northwest Pacific subduction zone cannot be explained simply by the traditional high-Q subducting slab in a low-Q upper mantle. A low-Q region in the uppermost 100 km (approximately) of the subduction zone has been identified. Source-corrected tAV* yield upper-mantle QAV estimates of about 1200 for predominantly oceanic paths (Australia and India) and 1500 for mixed continental and oceanic paths (Canada and Scotland).