Gordis Epidemiology 6th Edition 2019 PDF Chapter 14

Chapter 14 From Association to Causation: Deriving Inferences From Epidemiologic Studies Not everything that can be counted counts, and not environmental and genetic factors. As we shall see in the everything that counts can be counted. chapter on genetics and environmental...

Chapter 14 From Association to Causation: Deriving Inferences From Epidemiologic Studies Not everything that can be counted counts, and not environmental and genetic factors. As we shall see in the everything that counts can be counted. chapter on genetics and environmental factors, studies —William Bruce Cameron, 19631 of disease etiology generally address the contributions of both genetic and environmental factors and their interactions. Learning Objectives This chapter discusses the derivation of causal To describe a frequent sequence of study inferences in epidemiology. Let us begin by asking, designs used to address questions of etiology “What approaches are available for studying the etiology in human populations. of disease?” To differentiate between real and spurious associations in observational studies. To define the concepts of “necessary” and Approaches for Studying Disease Etiology “sufficient” in the context of causal relationships. If we are interested in whether a certain substance is To present guidelines for judging whether an carcinogenic in human beings, a first step in the study association is causal based on the guidelines of the substance’s effect might be to expose animals to set forth by the US Surgeon General and to the carcinogen in a controlled laboratory environment. discuss the application of these guidelines to Although such animal studies afford us the opportunity broader questions of causal inference. to control the exposure dose and other environmental To describe how the guidelines for causation conditions and genetic factors precisely and to keep originally proposed by the US Surgeon General loss to follow-up to a minimum, at the conclusion of have been modified and used by the US Public Health Service and the US Preventive Services the study we are left with the problem of having to Task Force. extrapolate data across species (i.e., from animal to human populations). Certain diseases seen in humans have neither occurred nor been produced in animals. I n previous chapters, we discussed a variety of It is also difficult to extrapolate animal doses to human designs of epidemiologic studies that are used to doses, and species differ in their responses. Thus, determine whether an association exists between an although such toxicologic studies can be useful, they exposure and a disease outcome (Fig. 14.1A). We still leave a gnawing uncertainty as to whether the then addressed different types of risk measurement animal findings can be generalized to human beings. that are used to quantitatively express an excess in We can also use in vitro systems, such as cell culture risk. If we determine that an exposure is associated or organ culture. However, because these are artificial with a disease, the next question is whether the systems, we are again left with the difficulty of extrapo- observed association reflects a causal relationship lating from artificial systems to intact, whole human (see Fig. 14.1B). organisms. Although Figs. 14.1A and B refer to an envi- In view of these limitations, if we want to be able ronmental exposure, they could just as well have to draw a conclusion as to whether a substance causes specified a genetic characteristic or some other disease in human beings, we need to make observations risk characteristic or a specific combination of in human populations. Because we cannot ethically or 269 270 SECTION II Using Epidemiology to Identify the Cause of Disease Environmental Environmental Exposure or Host Exposure or Host Characteristic Characteristic Is an Is the Observed Association An Association Association Observed? Is Observed Causal? Disease or Disease or Other Health Other Health Outcome Outcome A B Fig. 14.1 (A) Do we observe an association between exposure and disease? (B) Is the observed association between exposure and disease causal? practically randomize human beings to exposure to a Clinical Observations suspected carcinogen, we are dependent on nonran- domized observations, such as those that come from case-control and cohort studies. Available Data APPROACHES TO ETIOLOGY IN HUMAN POPULATIONS Case-Control Studies Epidemiology often capitalizes on what have been called “unplanned” or “natural” experiments. (Some think Cohort Studies that this phrase is a contradiction in terms, in that the word “experiment” implies a planned exposure.) What we mean by unplanned or natural experiments is that Randomized Trials we take advantage of groups of people who have been exposed for nonstudy purposes, such as occupational Fig. 14.2 A frequent sequence of studies in human populations. cohorts in specific industries or persons exposed to toxic chemicals. Examples include people affected by the poison gas leak disaster at a pesticide manufacturing data, the analysis of which might shed light on the plant in Bhopal, India, in 1984 and residents of Hiro- question. We can then carry out new studies such as shima and Nagasaki, Japan, who were exposed to the cohort and case-control studies, as discussed in radiation from the atomic bombs dropped on both prior chapters, which are specifically designed to cities by US forces in 1945. Each of these exposed determine whether there is an association between an groups can be compared with an unexposed group exposure and a disease, and whether a causal relation- (e.g., residents of Chennai, India or Tokyo, Japan) to ship exists. determine whether there is an increased risk of a certain The usual first step in carrying out new studies to adverse effect in persons who have been exposed. explore a relationship is often a case-control study. For In conducting human studies, the sequence shown example, if Ochsner had wanted to further explore his in Fig. 14.2 is frequently followed. The initial step may suggestion that cigarette smoking may be associated consist of clinical observations at the bedside. For with lung cancer, he would have compared the smoking example, when the surgeon Alton Ochsner observed histories of a group of his patients with lung cancer that virtually every patient on whom he operated for with those of a group of patients without lung cancer—a lung cancer gave a history of cigarette smoking, he was case-control study. among the first to suggest a possible causal relationship.2 If a case-control study yields evidence that a certain A second step is to try to identify routinely available exposure is suspect, we might next do a cohort study  14 From Association to Causation: Deriving Inferences From Epidemiologic Studies 271 Fig. 14.3 Another example of association or causation. (DILBERT © 2011 Scott Adams. Used by permission of ANDREWS MCMEEL SYNDICATION. All rights reserved.) (e.g., comparing smokers and nonsmokers and deter- A. Causal B. Due to Confounding mining the rate of lung cancer in each group or compar- Characteristic Characteristic ing workers exposed to an industrial toxin with workers Under Study Under Study Observed Association Observed Association without such an exposure). Although, in theory, a randomized trial might be the next step, as discussed earlier, randomized trials are almost never used to study Factor X the effects of putative toxins or carcinogens and are generally used only for studying potentially beneficial agents. Conceptually, a two-step process is followed in Disease Disease carrying out studies and evaluating evidence. However, in practice, this process often becomes interactive and Fig. 14.4 Types of associations. deviates from a fixed sequence: 1. We determine whether there is an association or correlation between an exposure or characteristic For example, if we designed a study to select controls and the risk of a disease (Fig. 14.3). To do so, in such a way that they tended to be unexposed, we we use: might observe an association of exposure with disease a. Studies of group characteristics: ecologic (i.e., more frequent exposure in cases than in controls). studies (discussed in Chapter 7) This would not be a true association but only a result b. Studies of individual characteristics: cohort, of the study design. Recall that this issue was raised case-control, and other types of studies in Chapter 7 regarding a study of coffee consumption 2. If an association is demonstrated, we determine and cancer of the pancreas. The possibility was suggested whether the observed association is likely to be that the controls selected for the study had a lower a causal one. rate of coffee consumption than was found in the general population. Types of Associations INTERPRETING REAL ASSOCIATIONS REAL OR SPURIOUS ASSOCIATIONS If the observed association is real, is it causal? Fig. 14.4 Let us turn next to the types of associations that we shows two possibilities. Fig. 14.4A shows a causal might observe in a cohort or case-control study. If we association: we observe an association of exposure and observe an association, we start by asking the question, disease, as indicated by the bracket, and the exposure “Is it a true (real) association or a false (spurious) one?” induces development of the disease, as indicated by 272 SECTION II Using Epidemiology to Identify the Cause of Disease the arrow. Fig. 14.4B shows the same observed associa- A. Causal B. Due to Confounding tion of exposure and disease, but they are associated Increased Increased only because they are both linked to a third factor, Coffee Drinking Coffee Drinking Observed Association Observed Association which is called a confounding variable and designated here as factor X. This association is a result of confound- ing and is noncausal. Confounding is discussed in Smoking greater detail in Chapter 15. In Chapter 7 we discussed this issue in relation to McMahon’s study of coffee and cancer of the pancreas. Increased Risk Increased Risk McMahon observed an association of coffee consumption of Pancreatic of Pancreatic Cancer Cancer with risk of pancreatic cancer. Cigarette smoking was known to be associated with pancreatic cancer, and Fig. 14.5 Interpreting an observed association between increased coffee drinking and cigarette smoking are closely coffee drinking and increased risk of pancreatic cancer. associated (few smokers at the time of that report did not drink coffee) (Fig. 14.5). Therefore, was the A. Causal B. Due to Confounding observed association of coffee drinking and cancer of the pancreas likely to be a causal relationship, or could Increased Increased the association be due to the fact that coffee and cigarette Cholesterol Cholesterol Observed Association Observed Association smoking are associated and that cigarette smoking is a known risk factor for cancer of the pancreas? The same issue is exemplified by the observed Factor X association of increased serum cholesterol level and risk of coronary heart disease (CHD) (Fig. 14.6). Is physical inactivity a causal factor for increased risk of Increased Increased colon cancer, or is the observed association due to Risk of CHD Risk of CHD confounding? That is, are we observing an association Fig. 14.6 Interpreting an observed association between increased of physical inactivity and colon cancer because both cholesterol level and increased risk of coronary heart disease (CHD). are associated with a factor X (such as a smoking), which might cause people to have both physical inactiv- ity and an increased risk of colon cancer? Is this distinction really important? What difference does it make? The answer is that it makes a tremendous 10 Nonsmokers difference from both clinical and public health stand- Smokers 8 points. If the relationship is causal, we will succeed in Percent reducing the risk of colon cancer if we promote physical 6 activity, both for the individual but also at the population level. However, if the relationship is due to confounding, 4 then the increased risk of colon cancer is caused by factor X. Therefore increasing physical activity will have 2 no effect on the risk of colon cancer. Thus it is extremely important for us to be able to distinguish between an 0 association due to a causal relationship and an associa- 4 5 6 7 8 9 10 11 BIRTH WEIGHT (SCALE IN POUNDS; INTERVALS OF 4 OZ) tion due to confounding (which is noncausal). Let us look at another example. For many years it Fig. 14.7 Percentage distribution by birth weight of infants of mothers who did not smoke during pregnancy and of those mothers who has been known that cigarette smoking by pregnant smoked 1 pack of cigarettes or more per day. (From US Department of women is associated with low birth weight in their Health, Education, and Welfare. The Health Consequences of Smoking. infants. As seen in Fig. 14.7 the effect is not just the Washington, DC: Public Health Service; 1973:105.)  14 From Association to Causation: Deriving Inferences From Epidemiologic Studies 273 result of the birth of a few low-birth-weight babies in association reflected a causal relation. Others, including this group of women. Rather, the entire weight distribu- a leading statistician, Jacob Yerushalmy, believed the tion curve is shifted to the left in the babies born to association was due to confounding and was not causal. smokers. The reduction in birth weight is also not a He wrote as follows: result of shorter pregnancies. The babies of smokers are smaller than those of nonsmokers at each gestational A comparison of smokers and nonsmokers shows that age (Fig. 14.8). A dose-response relationship is also the two differ markedly along many environmental, seen (Fig. 14.9). The more a woman smokes, the greater behavioral and biologic variables. For example, smokers her risk of having a low-birth-weight baby. For many are less likely to use contraceptives and to plan the years the interpretation of this association was the pregnancy. Smokers are more likely to drink coffee, beer subject of great controversy. Many believed the and whiskey and the nonsmoker, tea, milk and wine. The smoker is more likely than the nonsmoker to indulge in 3,650 these habits to excess. In general, the nonsmokers are 125 revealed to be more moderate than the smokers who are Mean weight in ounces 3,400 Mean weight in grams 115 Non- shown to be more extreme and carefree in their mode smokers 3,150 of life. Some biologic differences are also noted between 105 Smokers 2,900 them: Thus smokers have a higher twinning rate only in whites and their age for menarche is lower than for 95 2,650 nonsmokers.3 85 2,400 In view of these many differences between smokers 75 2,150 and nonsmokers, Yerushalmy believed that it was not 36 37 38 39 40 41 42 43+ Gestation in completed weeks the smoking that caused the low birth weight but rather that the low weight was attributable to other charac- Fig. 14.8 Mean birth weight for week of gestation according to maternal smoking habit. (From US Department of Health, Education, and Welfare. The teristics of the smokers. It is interesting to examine a Health Consequences of Smoking. Washington, DC: Public Health Service; study that Yerushalmy carried out to support his position 1973:104.) at the time (Fig. 14.10).3 14 10 9.5a 8.9b 12.0 Percent of low-birth-weight infants 12 8 10 6.0b Percent