In regions of low to moderate seismicity, accelerometric data may be scarce or nonexistant. This makes it difficult to use empirical or semi-empirical methods for estimating the strong ground motion of a potentially damaging earthquake. When a few small events are nevertheless well recorded locally, we propose to use a simple empirical Green's function method to obtain rough strong-motion estimates for a hypothetical magnitude 5 to 7 earthquake. The method is equivalent to applying a source-dependent linear filter to the small-event records. At low frequencies, tests with synthetics computed for a circular fault model show that the method may be valid at distances of a few fault lengths from the source, but it fails at predicting the near-field terms at shorter distances. Spatial aliasing problems, due to the sampling of the large-event fault surface with a finite series of subfaults, put another limit at high frequencies. Applying the method to an earthquake sequence in central California, we also show that uncertainties in the small-event source parameters may cause important errors in the strong-motion prediction. Despite these quite severe limitations, we consider that rough estimations of strong ground motion are possible by making a series of hypotheses on the source process. We show, for example, using a small earthquake recorded at the surface and in a 500-m-deep borehole, how the SH strong-motion acceleration spectra could be estimated at depth, next to a mine, and at the surface, for a hypothetical M 6.5 earthquake in the sediment-filled valley of the upper Rhine Grabben.