We present the results of a seismological investigation of the frequency-dependent amplitude variations across Iceland using data from the HOTSPOT array currently deployed there. The array is composed of 30 broad-band PASSCAL instruments. We use the parameter t(*), defined in the usual manner from spectral ratios (Halderman and Davis 1991), to compare observed S-wave amplitude variations with those predicted due to both anelastic attenuation and diffraction effects. Four teleseismic events at a range of azimuths are used to measure t(*). A 2-D vertical cylindrical plume model with a Gaussian-shaped velocity anomaly is used to model the variations. That part of t(*) caused by attenuation was estimated by tracing a ray through IASP91, then superimposing our plume model velocity anomaly and calculating the path integral of 1/vQ. That part of t(*) caused by diffraction was estimated using a 2-D finite difference code to generate synthetic seismograms. The same spectral ratio technique used for the data was then used to extract a predicted t(*). The t(*) variations caused by anelastic attenuation are unable to account for the variations we observe, but those caused by diffraction do. We calculate the t(*) variations caused by diffraction for different plume models and obtain our best-fit plume, which exhibits good agreement between the observed and measured t(*). The best-fit plume model has a maximum S-velocity anomaly of - 12 per cent and falls to 1/e of its maximum at 100 km from the plume centre. This is narrower than previous estimates from seismic tomography, which are broadened and damped by the methods of tomography. This velocity model would suggest greater ray theoretical traveltime delays than observed. However, we find that for such a plume, wave-front healing effects at frequencies of 0.03-0.175 Hz (the frequency range used to pick S-wave arrivals) causes a 40 per cent reduction in traveltime delay, reducing the ray theoretical delay to that observed.