This study uses a database of <3,500 seismograms recorded on rock sites in the northeastern United States/southeastern Canada for events in the magnitude range from 3.5 to 6.5 at distances >800 km to evaluate published attenuation models for the region. The models include functional forms that are linear (Atkinson 1989; Atkinson and Mereu 1992), bilinear (Street and Turcotte 1975; Atkinson and Mereu 1992; Boatwright and Seekins 2011) and trilinear (Atkinson and Boore 1995; Atkinson 2004) in shape. Each model is evaluated by examining residual distance trends in ground-motion data after correction of the data for the attenuation effects according to that model. None of the models provide a satisfactory description of attenuation over all distance ranges. A simple 1/R geometric spreading model at all distances and frequencies (with an associated Q ∼ 2,000) does as well as any model in providing (on average) low bias at near-source distances while being pegged by the amplitude levels observed at regional distances—even though more complex models provide a statistically better fit to the attenuation shape overall. We need to rethink the concepts used in deriving attenuation models for the development of ground-motion prediction equations such that the model shapes are not driven so strongly by regional observations. Any attenuation model that is not significantly constrained by data at hypocentral distances from 10 to 50 km is unlikely to provide reliable estimates of near-source ground motions or source parameters.