Large modifications of seismic waves are produced by variations of material properties near the Earth's surface and by both surface and buried topography. These modifications, usually referred to as "site response", in general lead to larger motions on soil sites than on rock-like sites. Because the soil amplifications can be as large as a factor of ten, they are important in engineering applications that require the quantitative specification of ground motions. This has been recognised for years by both seismologists and engineers, and it is hard to open an earthquake journal these days without finding an article on site response. What is often missing in these studies, however, are discussions of the uncertainty of the predicted response. A number of purely observational studies demonstrate that ground motions have large site-to-site variability for a single earthquake and large earthquake-location-dependent variability for a single site. This variability makes site-specific, earthquake-specific predictions of site response quite uncertain, even if detailed geotechnical and geological information is available near the site. Predictions of site response for average classes of sites exposed to the motions from many earthquakes can be made with much greater certainty if sufficient empirical observations are available.