ARCHIVED - Chronic Diseases in Canada

Volume 29 · Supplement 1 · 2010

Cancer and the environment: Ten topics in environmental cancer epidemiology in Canada

Shirley A. Huchcroft, Yang Mao and Robert Semenciw, Editors

https://doi.org/10.24095/hpcdp.29.S1.01

Basic Concepts in Epidemiology

Several of the exposure-specific chapters that follow deal with methodological issues unique to the exposure being examined. This chapter briefly sets out general principles and strategies in environmental cancer epidemiology and uses specific exposures to illustrate concepts. These concepts provide context for the critiques of the evidence that are contained in the exposure reviews and may be helpful for readers not familiar with epidemiological methods. Detail on methods can be found in publications devoted to environmental epidemiology, medicine and statistics.^1-9a

Types of evidence from epidemiological studies

For both ethical and practical reasons, studies to investigate the effects of environmental exposures on human populations must be observational rather than experimental. Unlike the experiment, where only the agent of interest is manipulated and all other extraneous factors are held constant, the observational study must contend with the difficulty of identifying, measuring and controlling (either by design or analysis) the many factors, other than the one of interest, that could influence outcome. In addition, measuring both the exposure and the disease outcome can be problematic. The ability of an epidemiological study to provide accurate risk estimates depends heavily on the strength of its design and the types of information it uses.

A strong study design uses accurate measurements that result in little misclassification, controls extraneous factors that could confound the results and permits causal inference. General criteria for concluding that a relationship is causal rather than only an association include a strong association, consistency, specificity, a relationship in time, a biological gradient (dose-response effect), biological plausibility and coherence of the evidence.9b Generally speaking, the stronger the association, the less likely it is to have occurred as a result of chance or to be the result of confounding by another factor. An association that has been repeatedly observed by different persons in different places, under different circumstances and over time is considered to be consistent and unlikely the result of some constant error or fallacy that permeates every enquiry. Specificity is a more difficult concept. Suffice it to say that an association that is limited to specific individuals and to particular sites and types of diseases is a strong argument in favour of causation. Demonstration that the suspected causal factor preceded the effect (and is consistent with the known latency of the disease) is further evidence for a causal association, as is a biological gradient or dose-response effect (i.e., the greater the exposure, the higher the risk of disease). Biological plausibility lends further strength to the argument for causation, but often cannot be demonstrated; what is biologically plausible depends upon the biological knowledge of the day. And finally, coherence means that the cause-and-effect interpretation of an association should not seriously conflict with the generally known facts of the natural history and biology of the disease. In his discussion of this criterion, Hill gives the example that the association of lung cancer with cigarette smoking is coherent with the temporal rise in the two variables and with the sex difference in both smoking rates and mortality from lung cancer.9c

In terms of the strength of a design, a continuum extends from descriptive analyses (useful for formulating hypotheses) to the “natural experiment” (in which an exposure has occurred to a defined group of people who can be compared to a similar group of individuals not exposed). In between, lie study designs that yield evidence of varying degrees of strength. These approaches are described below.

Ecological (or cross-sectional) studies relate cancer mortality or incidence rates (usually age- and sex-adjusted) with characteristics of regions. The units of analysis in these studies are populations or groups, rather than individuals. Thus, the ecological design provides no information on the relationship between exposure and disease at the individual level. The measure of association is the correlation coefficient. Values for the correlation coefficient range from −1 through 0 to +1, representing, respectively, a perfect negative correlation, no relationship and a perfect positive relationship. Example of ecological studies could be the rates of bladder cancer of various communities by water supply (chlorinated municipal water as opposed to well water), or skin cancer rates of communities with different average numbers of hours of sunlight per day.

Although this study design can be a useful preliminary step in investigating an association between disease and a suspected causal factor, the evidence it provides for a cause-effect relationship is relatively weak for at least three reasons. First, a relationship that applies with respect to groups of people does not necessarily apply at the individual level. This is referred to as the ecological fallacy. For example, it is conceivable that people who developed bladder cancer used well water rather than chlorinated municipal water even though they lived in communities with a chlorinated water supply. Second, to impute a cause-and-effect relationship, the suspected “cause” must precede the effect. If both “cause” and “effect” are measured at the same time, there is no assurance that the cause has preceded the effect. This is a particular problem in cancer epidemiology where a long latency between exposure and the development of cancer is the norm. Third, there is little opportunity in this type of study design to control for other factors, besides the study factor, that could affect outcome. For example, if the population of the sunnier communities tended to be more prone to skin cancer (e.g., have fair skin), than those of the less sunny communities, an apparent relationship between average daily sunshine exposure and skin cancer could be overestimated.

In case-control studies, individuals with the disease of interest are compared with individuals without the disease on factors being investigated as potential causes. The measure of association in this context is the odds ratio (OR). When the control group is representative of the general population with respect to the suspected causal factor, the odds ratio provides a good estimate of the degree of risk of the disease for persons with the attribute relative to those without it. For example, an odds ratio of two means that the risk of disease for persons with the attribute is approximately twice the risk for those without it. The case-control design is stronger than the ecological design for the three reasons mentioned above: the unit of analysis is the individual, a time interval between exposure and disease onset can be approximated, and information on a variety of other factors can be collected. It may be weaker than other designs discussed below in that measurement of the potential exposure factors can be limited. The case-control study is of particular value for diseases, such as cancer, that are relatively rare.

A cohort is a group of persons who share a common experience within a defined time period. In cohort studies, the disease status of individuals known to be exposed to a particular factor is determined at a later date and compared to the disease status of individuals known not to have been exposed. The measure of association in the cohort study is the relative risk (RR). This is the risk of disease in the exposed group, expressed as a rate, divided by the risk in the unexposed group. The cohort study is often much more costly than the case-control study when the disease is rare because a very large number of people must be included in order to accumulate a sufficiently large number of participants with the outcome of interest. Also, depending upon the time between the exposure and the disease, selective losses to follow-up can be a major weakness.

A nested case-control study is a case-control study conducted within a cohort. For example, workers in a factory (the cohort) who have developed cancer can be compared with those who have not, in terms of their specific jobs and/or exposure to the agent of interest. This study design can benefit from the advantages of both the case-control and cohort approaches in that similar information is collected for both the case and control groups. The nested case-control study is especially useful where biological specimens have been procured in a cohort study, particularly if they can provide data on biological markers of exposure, susceptibility or disease natural history.

The natural experiment is one that arises from the activities of humanity. It is a variant of the cohort design, and one in which a group of individuals exposed to an event that would not normally occur without the actions of humankind—a nuclear accident, for example—are then followed for disease occurrence relative to individuals not exposed. For example, much of the information we know about the effects of radiation exposure has been derived from follow-up studies of individuals exposed to radiation fallout from the atomic bombings of Hiroshima and Nagasaki.

It is not uncommon for the different types of studies of an exposure-disease relationship to yield different results. A cause-effect relationship is increasingly likely if various study designs, executed in different populations, suggest the same relationship (even though the strength of the relationship may differ) and if the association increases with the magnitude of the exposure (i.e., a dose-response relationship is observed).

The results of multiple epidemiogical studies are often aggregated using two techniques: a meta-analysis and a pooled analysis. A meta-analysis produces a weighted average of risk estimates from previously published studies. Studies are often weighted on the basis of the variability of the risk estimates or to reflect in some other fashion the quality of the studies. A pooled analysis combines the original data on individual exposures and outcomes from multiple studies. It is methodologically generally preferred to a meta-analysis.

Measuring outcome

Cancer epidemiology studies can use either incidence or mortality as a measure of outcome. Mortality information is often more readily available because mortality data are part of the vital statistics that most countries collect. Cancer mortality data approximate incidence data for cancers that are highly fatal. Mortality data are less useful in epidemiological studies for cancers where mortality is low since factors, other than those that cause cancer may contribute to death from cancer and, thus, obscure the etiology. Also, use of cancer mortality information usually limits the amount of other information that can be collected, such as occupational and residential histories, and behaviours such as smoking.

One way of obtaining more comprehensive exposure information is through studies using incident cancer cases and personal interviews. This greater detail and precision renders incidence studies better able to detect relationships than mortality studies. Canada is fortunate to have the Canadian Cancer Registry as part of a national cancer registration system to which all provinces and territories contribute information.¹⁰

Measuring exposure

In order to estimate risk, it is important to assess the amount of exposure to the person, group or area being monitored. Exposure assessment can be either direct or indirect. An example of direct exposure measurement is the use of radiation monitors worn by workers. Indirect exposure assessment includes predicting exposure from levels monitored in various media (air, water, food, soil) and reconstructing historical exposure patterns (e.g., by using job classifications and exposures known to be associated with specific jobs).

Examples of exposure indices in order of increasing accuracy are as follows: 1) a binary categorical assessment (when, in fact, there are a range of individual exposures); 2) a matrix of categories associated with a person’s exposure, along with a length of exposure; 3) subject-specific exposure measurements; 4) the effective biological dose received by an individual; and 5) extending the previous index to incorporate information on the genetic susceptibility of the individual to the dose received. Cumulative exposure is a commonly used index, calculated by multiplying exposure intensity by duration of exposure.

Assessment of exposure to any environmental contaminant is difficult because the general population is often not aware of specific exposures and may have difficulty remembering proxy indicators of exposure, such as residential history, drinking water sources and dietary intakes from 10 to 40 or more years ago. Environmental measurements may not be available for the earlier periods. Thus for many studies there has been an element of misclassification. To the extent that this misclassification is non-differential (e.g., random error), elevated risks probably represent an underestimation of the true risk. Where misclassification is systematic (e.g., the tendency of persons with the disease to report exposures of concern more often than those without the disease), overestimation of risk is likely to occur. This is referred to as exposure bias.

Controlling extraneous factors

Controlling extraneous factors that can distort the risk estimate is one of the biggest challenges in epidemiology and various design and analysis strategies have been developed for this purpose. One design approach is to restrict participant inclusion so that the study groups are as homogeneous as possible and sources of variability are reduced. One example is the inclusion of people of one sex and/or within a limited age range. A second approach is matching, whereby controls are selected for inclusion in the study if they match individual cases on certain attributes (e.g., age group and sex). A third approach is to collect as much descriptive information about the study participants as possible so that the study groups can be compared to determine how similar they are on factors other than those of interest. A factor that differs between the comparison groups and is associated with the outcome of interest is a potential confounder which can distort the relationship being studied. Analytic strategies for controlling potentially confounding factors involve mathematical models to adjust the risk estimate for the distorting effects of the confounders. Direct and indirect age-adjustment of rates, logistic regression, multiple linear regression and the Cox proportional hazards model are some of these techniques.

Références

1. ^ Goldsmith JR. Environmental epidemiology: Epidemiological investigation of community environmental health problems. Boca Raton, Florida: CRC Press; 1986.
2. ^ Brooks S. Environmental medicine. St. Louis, Missouri: Mosby; 1995.
3. ^ Breslow NE, Day NE. Statistical methods in cancer research. Volume 1 – The analysis of case-control studies. IARC Scientific Publications No. 32. Lyon: International Agency for Research on Cancer; 1980.
4. ^ Breslow NE, Day NE. Statistical methods in cancer research. Volume II – The design and analysis of cohort studies. IARC Scientific Publications No. 82. Lyon: International Agency for Research on Cancer; 1987.
5. ^ Steenland K, Savitz D, editors. Topics in environmental epidemiology. New York: Oxford University Press; 1997.
6. ^ Aldrich T, Griffith J, editors. Environmental epidemiology and risk assessment. New York: Van Nostrand Reinhold; 1993.
7. ^ Bertollini R, Lebowitz MD, Saracci R, et al., editors. Environmental epidemiology: exposure and disease. Boca Raton: CRC Press; 1996.
8. ^ Kelsey JL, Whittemore AS, Thompson WD, et al. Methods in observational epidemiology. 2nd ed. New York: Oxford University Press; 1996.
9. a,b,c Hill AB. Principles of medical statistics. 9th ed. New York: Oxford University Press; 1971.
10. ^ Band PR, Gaudette LA, Hill GB, et al. The making of the Canadian Cancer Registry: Cancer incidence in Canada and its regions, 1969 to 1988. Ottawa: Minister of Supply and Services Canada; 1993. Catalogue Number C52 42/1992.