The epidemiologic models previously described provide a framework for considering the association among variables in an attempt to determine cause and effect. In the context of epidemiology, causal association can be defined as an association between categories of events or characteristics in which an alteration in the frequency or quality of one category is followed by a change in the other category. To determine whether an observed association between a risk factor and disease is likely causal, it is first necessary to demonstrate a statistical association between the risk factor and disease outcome. Next, it is necessary to rule out that the association is explained by bias (i.e., that the people studied don’t represent the population of interest). Even when a strong, statistically significant, and unbiased association between a risk factor and disease is demonstrated, there is always the possibility that the observed association is noncausal. Two issues, confounding and effect modification, can result in noncausal associations.
Confounding is the confusion of the effect of the presumably causal variable under study with the effects of other extraneous variables so that all or part of the supposed causal effect of the “causal” variable on the dependent variable is actually explainable by the extraneous variables. For example, in men, baldness is associated with the risk of myocardial infarction; however, this is most likely not a causal association. Age increases both the likelihood of baldness and the risk of myocardial infarction; thus, age confounds the association between baldness and myocardial infarction. Confounding takes place when two or more potential risk factors co-occur in a group so that it is not possible to determine the independent effect of either risk factor. A confounder must be associated with both the exposure variable (e.g., physical inactivity) and the health outcome (e.g., disease, injury, or death) and also with the health outcome independently of the exposure variable. Said another way, if a physically active group of people were observed to have a lower rate of disease than a sedentary group but also were younger and smoked less tobacco, it would not be possible to conclude that physical activity protected against disease independently of age and smoking. Rather, the higher rate of disease in the sedentary group might be explainable by older age and smoking, not by the group’s low physical activity. Figure 2.5 illustrates age as a confounder.
Of the 500 active people, 250 have disease and 250 do not. So, the rate is 50%. Of 500 inactive people, 400 have disease. So, their rate is 80%, 1.6 times the rate among the active. However, this difference in rates is confounded by age differences between the active and inactive. Both the diseased and not diseased active people are equally split between young and old. However, 75% of inactive people are old (25 of 100 without disease and 300 of 400 with disease). Hence, the higher rate of disease among the inactive might be explained by their higher age.
The possibility of confounding is higher in observational studies, which are common in physical activity epidemiology, than in randomized controlled trials. This is so because the random assignment of participants to groups in randomized trials reduces the likelihood that the comparison groups (e.g., exposed vs. not exposed) differ with regard to a potential confounding variable.
For example, many observational studies have shown greater physical activity at baseline to be associated with lower risk of all-cause mortality. Does increased physical activity cause the lower mortality risk, or are more-active people generally in better health and therefore at lower risk of dying? Does overall health status confound the observed association between more physical activity and decreased mortality? Most well-conducted observational studies use statistical techniques to adjust, or attempt to adjust, for indicators of health status such as blood pressure, lipid levels, obesity, and cigarette smoking to determine whether the association of physical activity with morbidity or mortality risk has independence from confounders. A simple way to determine this is to divide the number of cases by the suspected confounder. For example, if higher age is positively associated with a higher rate of disease and negatively associated with physical activity, dividing the number of cases of each of the exposed and unexposed groups by their cumulative ages would standardize the rates according to age (i.e., the denominator for the rate becomes person-years of exposure) and remove most of the confounding influence of age.
Case–control studies can minimize the effects of confounding by matching each diseased case with one or more nondiseased persons who have the same scores on risk factors that are possible confounders. Nevertheless, many studies report that the association of physical activity and all-cause mortality persists after adjustment for the measured confounders. The association may still be the result of residual confounding, though as most studies do not measure all potential confounders or do not measure these confounders perfectly or do not determine whether they change across time. For example, other diseases that may affect both physical activity and mortality, such as diabetes or depression, are often not taken into consideration. This points to the difficulty for any single observational study to account for all potential confounders and therefore to provide definitive evidence that the apparent effect of physical activity in reducing risk is wholly independent of other possible explanations for the reduced risk.
The accumulation of results of numerous investigations using different samples and measuring different confounding variables adds weight to a causal association between physical activity and chronic disease risk.
In addition to confounding, effect modification, also called interaction, needs to be considered in any evaluation of the possibility of causal associations. Effect modification in epidemiology refers to a situation in which two or more risk factors modify one another’s effects on an outcome. For dichotomous variables (variables with only two levels or categories, e.g., yes or no, sedentary or active), effect modification means that the effect of the exposure on the outcome depends on the presence of another variable. For example, the effect of the physical activity categories sedentary or active on the presence or absence of CHD might differ depending on some third factor, such as sex; physical activity might reduce the risk of disease among men but not in women. In such a case, sex would be termed an effect modifier. We can evaluate the effect modification of continuous variables by determining the extent to which the effect of exposure on outcome depends on the level of some other variable.
For example, the effect of physical activity assessed by questionnaire on the risk for CHD may depend on the body mass index (BMI, body weight in kilograms divided by the square of height in meters). The risk of disease among the least-active people might be greatest among those having the highest BMI. If so, BMI would be considered an effect modifier in the association between activity and CHD. Figure 2.6 illustrates age as an effect modifier. Mortality decreases linearly with higher levels of physical activity. However, the slope of the decrease is much steeper among older people. The least-active older person has a much higher mortality rate than does the least-active young person. But the mortality rates for older people become increasingly similar to those of young people at higher levels of physical activity. Said another way, low activity is a bigger problem for older people, but high activity protects against mortality regardless of age. Hence, age modifies the effect of physical activity for reducing mortality risk.
Determining the degree to which certain factors act as effect modifiers can provide important information for the development of preventive strategies. For example, if it were demonstrated that individuals with high BMIs are at higher risk of CHD only if they also are relatively inactive compared with those who have normal or low BMIs, the evidence for effect modification might be sufficient to prompt interventions designed specifically to increase physical activity among individuals with high BMIs. The potential for confounding and effect modification increases the difficulty of interpreting epidemiologic studies of the direct effects of physical activity on health. Age can be a confounder because of a direct association between age and increased risk for death and most chronic diseases. Therefore, to examine the association between physical activity and health outcomes, the effect of age needs to be removed. Age can also change the magnitude of risk associated with other variables. Thus, age is also considered an effect modifier. For example, the risk of CHD increases with age, increased blood pressure, and decreased levels of physical activity. But age is also associated with decreased physical activity and increased blood pressure. Therefore, to determine whether there is an association between physical activity and CHD risk, the effect of both age and blood pressure must be controlled. In epidemiologic studies, this type of control is typically done by statistical analysis.
When you read and interpret information about the association of physical activity and health outcomes, it is important to be aware of the potential consequences of confounding (i.e., the confusion of two supposedly causal variables) and effect modification (i.e., the influence of a third variable on the strength or direction of the association between two other variables) and to determine whether these issues have been satisfactorily addressed.