We combine earthquake spectra from multiple studies to investigate whether the increase in stress drop with depth often observed in the crust is real, or an artifact of decreasing attenuation (increasing Q) with depth. In many studies, empirical path and attenuation corrections are assumed to be independent of the earthquake source depth. We test this assumption by investigating whether a realistic increase in Q with depth (as is widely observed) could remove some of the observed apparent increase in stress drop with depth. We combine event spectra, previously obtained using spectral decomposition methods, for over 50,000 earthquakes (M0 to M5) from 12 studies in California, Nevada, Kansas and Oklahoma. We find that the relative high-frequency content of the spectra systematically increases with increasing earthquake depth, at all magnitudes. By analyzing spectral ratios between large and small events as a function of source depth, we explore the relative importance of source and attenuation contributions to this observed depth dependence. Without any correction for depth-dependent attenuation, we find a systematic increase in stress drop, rupture velocity, or both, with depth, as previously observed. When we add an empirical, depth-dependent attenuation correction, the depth dependence of stress drop systematically decreases, often becoming negligible. The largest corrections are observed in regions with the largest seismic velocity increase with depth. We conclude that source parameter analyses, whether in the frequency or time domains, should not assume path terms are independent of source depth, and should more explicitly consider the effects of depth-dependent attenuation.