Limitations of the existing earthquake ground motion database lead to scaling of records to obtain seismograms consistent with a ground motion target for structural design and evaluation. In the engineering seismology community, acceptable limits for 'legitimate' scaling vary from one (no scaling allowed) to 10 or more. The concerns expressed by detractors of scaling are mostly based on the knowledge of, for example, differences in ground motion characteristics for different earthquake magnitude-distance (Mw-Rclose) scenarios, and much less on their effects on structures. At the other end of the spectrum, proponents have demonstrated that scaling is not only legitimate but also useful for assessing structural response statistics for Mw-Rclose scenarios. Their studies, however, have not investigated more recent purposes of scaling and have not always drawn conclusions for a wide spectrum of structural vibration periods and strengths. This article investigates whether scaling of records randomly selected from an Mw-Rclose bin (or range) to a target fundamental-mode spectral acceleration (Sa) level introduces bias in the expected nonlinear structural drift response of both single-degree-of-freedom oscillators and one multi-degree-of-freedom building. The bias is quantified relative to unscaled records from the target Mw-Rclose bin that are 'naturally' at the target Sa level. We consider scaling of records from the target Mw-Rclose bin and from other Mw-Rclose bins. The results demonstrate that scaling can indeed introduce a bias that, for the most part, ca be explained by differences between the elastic response spectra of the scaled versus unscaled records. Copyright ?? 2007 John Wiley & Sons, Ltd.