The secular variation of the Earth's magnetic field results from the effects of magnetic induction in the fluid outer core and from the effects of magnetic diffusion in the core and the mantle. Adequate observations to map the magnetic field at the core-mantle boundary extend back over three centuries, providing a model of the secular variation at the core-mantle boundary. Here we consider how best to analyze this time-dependent part of the field. We propose that the first step should be to seek the steady core flow that best fits the field maps, isolating what we call the steady secular variation, the part of the secular variation explained by steady core flow. To calculate steady core flow over long time periods, we introduce an adaptation of our earlier method of calculating the flow in order to achieve greater numerical stability. We perform this procedure for the periods 1840--1990 and 1690--1840 and find that well over 90% of the variance of the time-dependent field can be explained by simple steady core flow. The core flows obtained for the two intervals are broadly similar to each other and to flows determined over much shorter recent intervals. Although we can explain a large part of the signal with these flows, none of them provides an adequate explanation of the secular variation. In particular, the fit to the original observations from which the field models were derived is inadequate. We examine the residual secular variation, the part of the signal that remains. We argue that it is dominated by the effects of nonsteady flow and magnetic diffusion. We do not find any compelling evidence for torsional oscillations as the dominant ingredient of this nonsteady part of the secular variation. ¿ American Geophysical Union 1992