Mapping "Race": Critical Approaches to Health Disparities Research

Chapter 4: Ethical Dilemmas in Statistical Practice: The Problem of Race in Biomedicine

53 Jay S. Kaufman Most discussions of ethics in statistical practice revolve around declaring conflicts of interest due to funding sources and engaging in honest descriptions of data and procedures (American Statistical Association 1983, 1999; International Statistical Institute 1985; Jowell 1981). This is all well and good, but such a narrow focus on greed and fabrication ignores some of the more interesting and intractable ethical problems necessarily involved in inference and prediction, especially when social groups are transformed into statistical categories. Racial comparisons therefore provide an interesting case study of some of these problems, as they are frequently the targets of statistical analysis in clinical medicine and public health, and they are imbued with all kinds of baggage due to our understandings about racial groups in the social world (Marshall 1993). The history of modern statistics is intimately intertwined with the quantitative demonstration of racial differences, with obvious implications for the justification and maintenance of existing social hierarchies. For example, the increasingly intricate quantifications of late nineteenth-century craniology propelled the development of regression techniques by Francis Galton, while the nascent field of intelligence testing gave rise to factor analysis (Gould 1981). The development of psychometrics as a quantitative discipline was motivated by the task of reifying intelligence as a measurable trait and arraying racial or ethnic groups in a ranked fashion across this singular dimension of innate capacity. The ethical implications of this work can be seen in their application to a wide variety of social policies, including immigration restrictions and forced sterilizations (Lewontin, Rose, and Kamin 1993). Ethical Dilemmas in Statistical Practice The Problem of Race in Biomedicine Chapter 4 54 Jay S. Kaufman The problems we face in such comparisons are more fundamental than those related to the assumptions and technical limitations of specific statistical models, however. Even the simplest descriptions of disparity invoke conundrums in the choice of contrasts, which can paint dramatically different pictures of the state of the world (Harper et al. 2010; Scanlon 2006). For example, the simple technique of epidemiologic standardization is a ubiquitous feature of surveillance because it facilitates “fair” comparisons between populations that are imbalanced by another factor (for example, age). The procedure involves taking a weighted average of the stratum-specific measures, where the weights are taken from some population that is defined as the common standard. The seemingly innocuous choice of the standard population, however, can lead to drastically different impressions. An illustration of this phenomenon occurred when the U.S. government switched from the 1970 census to the 2000 census population standard for its official statistics, with the consequence that the magnitude of racial disparity in mortality decreased overnight. This occurred because the year 2000 population was older, and because ratio measures of disparity are more modest in older age, so a heavier weighting at the older end of the population reduced the magnitude of disparity overall (Krieger and Williams 2001). This is not to argue that the 1970 standard population is any more or less valid, only that any standard is a necessary fiction; the real world remains stubbornly unadjusted. Categorization Just as standard populations are fictional representations of our world, so are virtually all categorized variables in statistical analyses. For example, while there really is something called height, which can be measured objectively along a continuum, there are no objective standards for what constitutes a “tall” or a “short” person. Human variability is similarly continuous or near continuous in most respects, and so categorizations of study subjects into “Black” and “White,” or “poor” and “non-poor,” or “hypertensive” and “normortensive,” are all to some extent arbitrary, and often inflected by myths and traditions that end up being imposed on individuals as essential characteristics that come to define them (Little 1998). Even the categories “male” and “female,” which obviously have some biological significance in the natural world, are nonetheless carefully tidied into a dichotomy when Mother Nature is not so abhorrent of ambiguity (Dreger 1998). Epidemiologists and biostatisticians have a long tradition of favoring categorized variables, much more so than is common practice in psychology or economics . For the left-hand-side (response) variable in a regression model, this is Ethical Dilemmas in Statistical Practice 55 probably due to the fact that we traditionally study “disease,” which is some kind of clinical categorization of health status. Some disease outcomes are reasonably concrete, such as a fractured bone or infection by a parasite. Others, however...