Estimation of relative risks from individual and ecological correlation studies
B. Raja Rao,Richard D. Day,Gary M. Marsh
DOI: https://doi.org/10.1080/03610929208830776
1992-01-01
Abstract:The problem discussed in the present paper is that of estimating relative risks from ecological correlation studies, where the variables concerned are averages measured on groups of people. Some biases encountered in such estimation procedures are discussed when analyzing aggregated and non-aggregated data with a particular reference to the shape of the dose-response relationship. In order to place the problem in its proper perspective, some relevant literature is summarized and some empirical observations are made by means of four examples concerning the relative risks between cigarette consumption and lung, bladder, oesophagus cancer and coronary heart disease mortality. In the present paper, the accent is placed on the distribution of exposure to a carcinogen. We consider a population or group of subjects, who are free of a disease and who are exposed to a risk factor to a level X, which is suspected to cause the disease in the population. The level X of the risk factor is a random variable having some probability distribution with a probability function or a probability density function (p.d.f.) f(x), over some range a ≥ x ≥ b. In order to illustrate the point, let us suppose that an investigator is interested in examining the influence of the distribution of exposure to a specified carcinogen, say arsenic, on the incidence of lung cancer, in a population of subjects. These subjects may be exposed to varying amounts of arsenic, which gives a probability distribution of exposure to the risk factor. As another example, we may imagine a population of smokers, smoking different quantities of cigarettes or a population of workers in a coke oven, inhaling varying amounts of the fumes on their jobs. The linear and the exponential risk models are considered for the disease incidence and it is demonstrated that in the case of a linear risk, the relative risks can be estimated either from individual studies or from ecological studies with any mathematical bias, whatever be the distribution of the exposure level of the risk factor. In the case of an exponential risk model, it is proved, in general that the relative risk from ecological studies tends to be larger than the estimate computed from individual studies. Further it is proved that the two estimates of relative risk will be identical if, and only if, the exposure level to the risk factor has a normal distribution with a constant variance, that is the same between the exposed and the unexposed groups of individuals. Several probability models for the distribution of exposure are used to illustrate the results.The analogous results. when the individuals in a cohort study are exposed simultaneously to several competing risks are also briefly discussed. A numerical example is given in Section 8.