established without invoking some minimum assumptions or restrictions. This is because of the impossibility of observing the counterfactual (i.e., observing that individual if he or she had not been subject to that particular series of stressful events). Yet if we want to understand the impact of various policies or say anything meaningful about the relative importance of various determinants, we need to be able to make statements about plausible causality.
Even the best social science has important limitations. Good natural experiments are few and far between and it is difficult if not impossible to control for all relevant aspects of context simultaneously. Furthermore, much of the empirical literature apparently relevant to understanding the determinants of racial and ethnic differences in health is not based on solid experimental data but rather it is based on associations among observable qualitative and quantitative data.
Conversations on causality both within and across disciplines are not always easy (Bachrach and McNicoll, 2003; Moffitt, 2003). There is no settled and accepted set of principles for addressing causal questions within the social sciences and different disciplines have different levels of tolerance for various kinds of assumptions. Plausible causality is best established by combining clear models of behavior with high-quality data. Findings that have been replicated in a number of studies using a multiplicity of different approaches are usually the most convincing.
The evidence we consider on these factors comes from studies of different types—each with their own sets of strengths and weaknesses—reflecting not only the variety of disciplines interested in the nexus between health, aging, and race and ethnicity, but also the wide assortment of research questions that must be considered. To verify differences between groups, either reliable vital registration (for mortality) or large representative surveys are needed. To explain these differences, one approach that researchers take, understandably, is to investigate variables in the same large surveys. This has the presumed advantage that accurately measured differences are being explained, but it also has a number of disadvantages. The variables available may not include all those of interest, or their measures may not be the most appropriate. Racial and ethnic identification may be problematic. The racial and ethnic groups of interest, if they are minorities and have not been properly oversampled, may be represented by too few cases.
These disadvantages might be overcome with a survey specifically designed to investigate differences, but other disadvantages are inherent in studies that rely on analysis of data from even well-planned surveys. Such studies are nonexperimental, and interpreting the empirical associations that may be uncovered is not straightforward. Causal inferences drawn from such associations are hazardous. Relationships may be spurious: a given determinant and the health outcome under analysis may appear linked