The effect of non-proportional loading on ductile failure by void growth is investigated using micromechanical unit cell model simulations. A cubic unit cell containing a concentric spherical void is deformed along a piecewise radial loading path in stress space, with a step change in loading direction at an intermediate strain, before failure occurs by the onset of strain localization at the macro-scale. The strain to failure is determined using loss of ellipticity of the macroscopic equilibrium equations, which depends on the tangent stiffness tensor estimated from the cell model simulations. The failure loci, showing the variation of the equivalent strain to failure as a function of the intermediate strain for given initial and subsequent loading paths, are compared with the predictions of continuum models of ductile failure. It is shown that widely used continuum damage models predict a linear variation of the failure strain with the intermediate strain, although the numerically obtained failure loci and strongly non-linear. The reason for this discrepancy can be attributed to the phenomenological damage accumulation rule and failure criterion used in the continuum damage models. It is shown that a failure criterion based on the onset of plastic instability in a porous material, together with a micromechanics-based void growth law, predicts the correct shapes of the failure loci under non-proportional loading. In particular, it is shown that the experimentally observed effect of the Lode parameter on ductility is primarily due to the Lode dependence of the failure criterion, and not the damage growth law as assumed in the continuum damage models. The instability-based failure model is also shown to predict the correct shapes of the failure loci under proportional loading, including both triaxial and plane stress loading paths.