Selection processes could create or exaggerate differences, or they could offset or disguise them. Identifying and analytically accounting for selection effects (and reciprocal causation) is a substantial, continuing, task. Researchers should either adopt strategies that attenuate some of the associated distortions or, at least, qualify the inferences likely to be drawn. But first they need to recognize the problem and define its reach. Research on health differences in old age should consistently consider the possible operation of selection processes. In most cases, a model representing the genesis of health and mortality differences should include a proper representation of selection processes. To understand and better model selection, research directed at this issue is also advisable.
Research Need 4: Identify and quantify the various selection processes that affect health differences among racial and ethnic groups.
The key reason for the neglect or superficial treatment of selection is the lack of information about the processes underlying it. What data and how much of it should be collected depend on the pertinent selection processes and on the availability of alternative procedures to account for their influence. A good example of the additional information that can be collected involves immigrant surveys that follow individuals from their places of origin to their destinations and through the assimilation process, tracking changes in health, socioeconomic and political status, and other factors (Jasso et al., 2004). (Other suggestions about alternative data collection strategies to account for immigration selection are given by Palloni and Morenoff .)
In particular cases, the data needed could be obtained through modifications of conventional study designs to promote a broader range of comparisons. To understand immigrant selection, for instance, one needs to be able to compare movers and stayers in the origin population, possibly even comparing migrants with siblings, relatives, or in-laws. A transnational study design, though it involves logistical and financial obstacles, is therefore far more useful than a study design based only on the receiving population.
With appropriate data, models could be formulated to yield ranges of estimates for differences as well as confidence intervals. A number of estimation procedures do exist for dealing with a reduced class of selection processes. These should be deployed judiciously, if not to adjust estimates of differences to account for selection processes, at least to provide an idea of the sensitivity of these estimates to assumptions about selection. It is always better to supply information about uncertainty, even if this blurs the scientific conclusions, than to convey the impression of robustness while ignoring potentially relevant effects.