S U M M A R YSeismic intensity observations contain sufficient information about the earthquake source to quantitatively constrain its scalar seismic moment, M,, and hence moment magnitude, M, within useful limits. This is valuable, especially in the pre-instrumental and early instrumental seismic eras, but also in the modern era. This study is limited to stable continental regions where intensity data are especially important for seismic hazard assessment, but the methodology is generic and can be applied to other tectonically active iegions. It builds on regression techniques developed in the Part I analysis (Johnston 1996) and applies them to isoseismal area (Ai) data. Derived regression relations for modified Mercalli isoseismal areas for levels felt to VIII yieid a predicted log(M,) or M value within specified la uncertainty bounds. Standard linear and polynomial regressions are tested against a new functional regression form proposed by Frankel ( 1994), which contains both geometrical spreading and anelastic attenuation terms. Goodness-of-fit statistics are similar, but the Frankel regression form is preferred because it is derived from physical principles of wave propagation.The Afel,-AIV regressions, with most data at epicentral distance r > 100 km, are controlled by surface-wave (L,) geometrical spreading and attenuation characteristics. For the A,,, and A,,,, regressions, r is mainly less than 100 km and body-wave propagation dominates, although near-source site and path effects are significant. The A, and A,, regressions are transitional between the L, and body-wave domains. With either the Frankel or quadratic regression, individual isoseismal areas can constrain the source event's moment magnitude within an estimated &0.30-0.45 M units except at the regression extremes. If a suite of isoseismal areas is available for the same earthquake, these uncertainties can be approximately halved by weighted averaging. This means that when good-quality isoseismal data exist for an event, they are capable of constraining its seismic moment estimate within the same order of uncertainty as individual instrumental magnitude readings. Hence, historical seismicity data may usefully be combined with instrumental data in seismic hazard analyses. A hierarchy of methods to recover an earthquake's M , or M combines the instrumental results of Part I with the isoseismal area results of this study. Finally, regressions on I, , , and number of recording stations provide the means to estimate M to within --+0.45-0.55 M units when no other data are available.