Uncertainty in the Disease Burden Attributable to Risk Factors
The assessments of the disease burden attributable to selected risk factors reported in chapter 4 are affected by additional sources of uncertainty, beyond the uncertainty in DALY estimates for specific disease and injury outcomes discussed earlier. A full uncertainty analysis of such burden estimates has not yet been carried out, but would involve assessment of the following additional types of uncertainty:
-
uncertainty in the estimated distributions of population risk exposure;
-
uncertainty in estimates of relative risks for cause-specific mortality and incidence associated with specific exposures, for which a significant source of uncertainty is the extrapolation of relative risks measured at other ages to older age groups;
-
uncertainty associated with estimating joint effects of risk factors.
Uncertainty in exposure and in both the existence and magnitude of hazardous effect always affects quantitative risk assessment. In one taxonomy, risk assessment uncertainty can be divided into parameter uncertainty and model uncertainty (National Research Council 1994). Parameter uncertainty is often quantifiable using random variable methods, for example, uncertainty due to sample size or measurement error. Model uncertainty is due to gaps in scientific theory, measurement technology, and data. It includes uncertainty in causal relationships or the form of the exposure-response relationship (for instance, threshold versus continuous or linear versus nonlinear), the level of bias in measurement, and so on. Defined broadly, model uncertainty also includes extrapolation of exposure or hazard from one population to another. Model uncertainty dominates uncertainty in risk assessment, a result of difficulty in carrying out direct studies on exposure, hazard, and background disease burden. This has motivated innovative assumptions and extrapolations even in the case of the most widely studied risk factors like smoking (Peto and others 1992).
Uncertainty around disease causation (Evans 1978; Hill 1965) was, in practice, secondary to uncertainty around hazard size, for example, relative risk, because when causality was uncertain, the estimates of relative risk needed for risk assessment were also unknown or uncertain. For example, whether the relationships between physical inactivity and lower back pain or between alcohol and violence are causal has equivalent questions on the magnitude of hazard of each risk for the disease outcome. Collective scientific knowledge from disciplines such as social and behavioral sciences, physiology and neuroscience, and epidemiology would confirm the possibility of a causal relationship in the foregoing cases, but would shift the uncertainty to hazard size. As a result, for some risk factors, we could only quantify the contribution to a subset of disease outcomes because epidemiological studies did not provide enough information for all risk factor and disease pairs, even when the causal relationship was believed or suspected.
Estimates of hazard in individual epidemiological studies were adjusted for confounding as much as possible. Extrapolation of hazard from a limited number of studies to other populations has received less attention. While the robustness of proportional measures of risk has been confirmed for more proximal factors in studies across populations (Eastern Stroke and Coronary Heart Disease Collaborative Research Group 1998; Horton 2000; Law, Wald, and Thompson 1994), their extrapolation is an important source of uncertainty for more distal risks (such as childhood sexual abuse) or those whose effects are heterogeneous (for example, alcohol and injuries versus alcohol and cancer).
Direct exposure data for many risk factors are limited both because of measurement difficulties and because of underinvestment in risk factor surveillance. To allow maximum use of available data, such risk factors were represented using indirect or aggregate indicators, for instance, smoking impact ratio for accumulated hazards of smoking, weight-for-age for childhood undernutrition, and use of solid fuels for indoor air pollution. Furthermore, for some risks multiple data sources allowed limiting the range of exposure estimates. For example, in the absence of alcohol surveys, information on total alcohol production, trade, and unrecorded consumption provided upper bounds on the fraction of the population that would be in the highest consumption category. Finally, some of the risk factors examined in chapter 4 were represented using continuous exposure variables such as high blood pressure. Others used categorical variables, for example, indoor smoke from household use of solid fuels, childhood underweight, and physical inactivity, even though the health effects occur along a continuum. This choice reflected the availability of exposure data and hazard estimates in categories. In such cases, the contribution to disease within the baseline category would not have been captured.
In addition to uncertainty in exposure and hazard, the uncertainty of estimated population attributable fractions (PAFs) is determined by the analytical properties of the PAF relationship. In particular, the PAF relationship is an increasing concave function of relative risk and exposure level, approaching 100 percent asymptotically, that is, the rate of increase declines with increasing relative risk or prevalence (figure 5.18). Therefore, if a risk factor or group of risk factors individually or jointly account for large fractions of specific diseases, the PAFs are more robust to uncertainty in inputs. Finally, there is uncertainty in mortality and disease burden estimates to which the estimated PAF are applied (see the previous section).
[Figure
5.18]
The findings in chapter 4 should therefore be considered within the context of limited available data and viewed as subject to uncertainty, which varies across risk factors and geographical regions. For further discussion of sources and quantification of uncertainty for specific risk factors see Ezzati and others (2004).
