All population projections contain some quite critical assumptions, especially when continued for more than one projection period as is usual. The most important implicit assumption is usually one of population homogeneity. That such an assumption is critical is obviously the case with an attempt to project total numbers without any disaggregation. To take an extreme example, consider a population equally divided between two strains at the start of the projection, one strain of which is doubling every ten years and one of which is halving every ten years. The current growth rate would be zero, the long-term growth rate would be doubling every ten years. Such considerations have been implicitly recognised by most projectors since the earliest component projections. Much of the heterogeneity with respect to mortality is removed by taking age and sex (and perhaps racial characteristics) into account. This is routinely done in most projections. However, it is well known that mortality is strongly associated with other characteristics, such as regular smoking. Why do projections not control for this characteristic? One obvious reason is a paucity of data. A further reason is that smoking habits are not immutably fixed, whereas date of birth is, as is sex to all intents and purposes. The introduction of smoking status would lead to the need to specify the ways in which such status changed. This would probably make the projection process less efficient. Thus the desire for homogeneous sub-groups is tempered by the fixity of the characteristics involved. If occupation and education were roughly equivalent in terms of reducing heterogeneity, education would be preferred as being a relatively more fixed attribute, at least for adults.