Causality in Epidemiology, Health Policy, and Law

Philip Cole

Editors' Summary: Determining the impact that environmental forces have on human health is an integral part of environmental law and policy. A determination of this impact must, of course, begin with a determination of whether there has been any impact at all. Evaluating the causal relationship between environmental forces and human illness is, therefore, essential. This Article, written by a professor of epidemiology at the University of Alabama at Birmingham's School of Public Health, examines the epidemiologic process for assessing causation, both for purposes of environmental litigation and for purposes of environmental regulation. The Article analyzes three situations in which it is necessary to evaluate causation: an individual research study, an assessment of a general causal hypothesis, and an examination of a specific individual's illness. The Article discusses the criteria for establishing causation and reviews factors, other than causation, that may lead to apparent correlations between exposures and diseases. Finally, the Article suggests ways to approach evaluations of causation to promote more effective use of scientific evidence in environmental law and policymaking.

Dr. Cole received his M.D. from the University of Vermont and his Dr.P.H. from the School of Public Health, Harvard University. He is Professor of Epidemiology, School of Public Health, and Senior Scientist, Comprehensive Cancer Center, at the University of Alabama at Birmingham. Preparation of this Article was supported by an award from the Shell Oil Company Foundation.

The author is grateful to the following persons who offered constructive suggestions on this Article: Dr. Elizabeth Delzell, Mr. Timothy Hardy, Mr. David Oliver, Dr. Brad Rodu, Dr. Kenneth Rothman, Dr. Jeffrey Roseman, Dr. Dimitrios Trichopoulos, and Dr. John Waterbor.

[27 ELR 10279]

Epidemiology is the basic science of public health¹ and, historically, has focused on identifying the environmental causes of disease in human beings. During the last 30 years, epidemiology's findings on disease causation have become a major basis for health-related public policy. Most recently, epidemiology has become central to health-related law, particularly in the resolution of so-called toxic torts. One aspect of epidemiology, the evaluation of possible cause-effect relationships, has been difficult to apply in these newer areas. This Article attempts to lessen that difficulty by presenting an approach to assessing causality that is suitable for science, policymaking, and law.

Background

The earliest criteria of causality for the biological sciences were developed in the mid-nineteenth century to permit the identification of a particular pathogenic microorganism as the cause of a disease. These criteria, known as Koch's, or the Koch-Henle, postulates, related to the isolation and identification of a bacterium and to the production of disease by its transmission in the laboratory.² In 1976, Dr. Alfred Evans revised the Koch-Henle postulates to make them more applicable to emerging infectious diseases, particularly those due to viruses and to deficiencies of the immune system.³ The Koch-Henle-Evans postulates remain useful today for cause-effect relationships that involve a single or dominant causal factor, that are strong, and that have a short induction period (the time between exposure to the cause and the appearance of disease). These features typify the causal relationships of infectious diseases.

The Koch-Henle-Evans postulates are not well suited to the chronic diseases that emerged as major medical and public health problems in the twentieth century. These conditions have causal relationships that usually involve a set of factors, are weak with a low correlation between any one of the factors and the disease, and have an induction period that extends from years to decades.

The 1964 report of the U.S. Surgeon General, Smoking

[27 ELR 10280]

and Health,⁴ advanced criteria of causality for chronic diseases. These criteria (consistency, strength, specificity, and coherence of the association; and appropriate temporal relationship) were included, along with others (biologic gradient, plausibility, experimental evidence, and analogy), by Sir Bradford Hill in his now widely used criteria of causality.⁵ The present approach modifies and extends the Hill criteria. It is based on the perception that the need to evaluate causality arises in three related, but quite different, contexts: the individual research study, the general causal hypothesis, and the cause of the illness of a specific human being.

BASIC CONSIDERATIONS

A causal hypothesis cannot be literally proven or disproven by science. Rather, its credibility is repeatedly modified, up or down, as new evidence becomes available. Thus, at any time and in the mind of each person, a causal hypothesis lies at a point on a spectrum of credibility. This spectrum often is described as if it extended from "not proven" at one extreme, to "proven," at the other (see Figure A). However, the full spectrum of credibility is conceptually twice as broad and extends from "disproven" at one extreme, through "not proven" at the center, to "proven" at the other extreme (see Figure B). The central, "not proven," position is adopted correctly by a person who is unfamiliar with the relevant information or who considers the evidence for and against the hypothesis as about equal. This model of the spectrum of credibility of a causal hypothesis includes the extremes "disproven" and "proven." In actuality, scientists can adopt neither extreme position for they must remain free to alter their assessments on the basis of new evidence. A willingness to modify their assessments characterizes scientists and explains why the extremes of the credibility spectrum are untenable: they represent the indefensible view that meaningful, contrary information could not come into existence.

In science and public health, a causal hypothesis seen as near the center of the spectrum is "worthy of study." As it becomes more credible it reaches a point where a regulatory agency may consider the advantages and disadvantages of acting, or of failing to act, "as if" the hypothesis were true. If its credibility increases further, the hypothesis may become "established" and serve as a foundation for research in related areas. Science can also refute a false hypothesis and so redirect scientific investigation onto a more productive course.⁶

The law is intended to be formal and rigorous in matters of logic, evidence, and proof. It is in law that rules of evidence apply and that structured argument proceeds toward a limited objective. However, the law, like science, uses ambiguous language to describe benchmarks on the spectrum of credibility of a causal hypothesis. This is unfortunate as these points correspond to the level of certainty, or burden of proof, required to establish causation. Further, the burden of proof differs depending on several factors, especially the nature (e.g., civil or criminal) of the case and the court having jurisdiction. In most torts, the plaintiff must show that his or her alleged exposure more likely than not caused his or her particular illness.

[SEE Figures A and B IN ORIGINAL]

The second basic consideration is that epidemiologic, or observational, research can be highly persuasive as to disease causation. It is widely believed that epidemiologic research does not address causation directly because it does not involve the randomization of subjects and the controlled administration of the agent under study. Experiments, usually done with animals, do have these attributes and are considered by some persons as being able to demonstrate cause-effect relationships in a visible, virtually mechanical way. But regardless of the appeal of such demonstrations, no causal hypothesis can be proven absolutely, no matter how much evidence exists in its favor.⁷ Nonetheless, experimental studies are generally more valid than are epidemiologic studies, which are more subject to bias and confounding. These limitations of epidemiologic research are described below. However, experimental studies are also limited as shown by the fact that they are often inconsistent with one another. Furthermore, even consistent experimental findings have a major obstacle to overcome before they address causality in people—the difficulty of generalizing from animals to human beings. This animal-to-human inference, usually termed an extrapolation, is in fact a generalization. This generalization invokes both a quantitative extrapolation over the exposure range (usually several orders of magnitude downward from heavily exposed animals to lightly exposed persons) and a series of

[27 ELR 10281]

implicit assumptions relating to matters such as dose equivalence, pathways, and products of metabolism. The useful approach to interpreting epidemiologic and experimental research lies in recognizing that each has inherent strengths and limitations. The most persuasive causal arguments rest both on experimental findings for biological credibility and on epidemiologic findings for relevance to human beings.

Context One: The Individual Study

The evaluation of causality begins in context one (C1) with a critical assessment of each relevant epidemiologic study. The seven criteria used in C1 (see Table I) are best described with an example. Let us say that we conducted a retrospective follow-up study of lung cancer among male foundry workers. We identified a large group of men who had been employed at a foundry and determined that over several decades they experienced 60 deaths due to lung cancer. We estimated from the group's characteristics (mainly its age composition) and from the length of follow-up that 20 lung cancer deaths would have occurred among its members if they had died from this disease at the same rate as the comparable general population. These findings are usually expressed as a standardized mortality ratio (SMR). The SMR is the ratio of the number of deaths observed to that expected, or 3.0. (Standardization is a process by which demographic and other characteristics of the study group are minimized as explanations of any difference between the observed and expected numbers of deaths. Also, there are epidemiologic research designs other than the retrospective follow-up study in common use. These describe findings with various measures of association other than the SMR. For example, the case-control study uses the "odds ratio" or the "relative risk." These measures are computed differently from the SMR but have a very similar interpretation.)

Table I

The contexts and the criteria of causation.

1. THE INDIVIDUAL STUDY

(a) minimal systematic error

(b) minimal confounding

(d) strength of association

(e) internal consistency

dose-response

uniformity

specificity

(f) temporality

(g) biologic plausibility

2. THE GENERAL CASE

(a) external consistency

replicability

strength

(b) coherence

3. A SPECIFIC PERSON'S ILLNESS

(a) context 2 is met

(b) specific exposure characteristics

(d) absent or minimal alternative cause

The first criterion of causality in C1 is the absence, or near absence, of bias. Here, bias does not mean "prejudice" as in lay usage but "systematically (not randomly) in error." The search for bias usually is the search for a procedural shortcoming that affected only some of the study's subjects. There are general biases that may lower the validity of all epidemiologic studies.⁸ In addition, each study design has types of bias to which it is prone. For example, a problem common to retrospective follow-up studies, such as our study of foundry workers, is the failure to include all of the time during which there was a risk of death among study subjects. This falsely lowers the expected number of deaths and so falsely raises the SMR.

The second criterion in C1 is the near absence of confounding. Confounding occurs when the suspect cause in a study (foundry work) is associated with a real cause of the disease of interest. This real cause produces a false (or amplifies a true) association between the suspect cause and the disease and is termed a confounder. In our example the major confounder is cigarette smoking. If the foundry workers smoked more than the general population of men, then some, possibly all, of their lung cancer excess is due to their smoking and not to agents in the foundry. The epidemiologist will minimize confounding from recognized causes of the disease by employing appropriate strategies in study design and by suitable standardization. The resulting SMR will not reflect the confounder (smoking) or demographic factors. Let us say that the original SMR of 3.0 in the example is reduced to a "residual" SMR of 2.0 by standardization for smoking. This means that one-half of the excess lung cancer of the foundry workers (the excess over a baseline, or "null," SMR of 1.0) was associated with, and presumably caused by, their smoking. However, the residual SMR of 2.0 indicates a persisting association between foundry work and lung cancer and requires evaluation with the remaining criteria of C1.

The third criterion in C1 is the near absence of random or chance effects. In the example, the question becomes: how probable is it that an unconfounded SMR of 2.0 (or greater) would occur by chance alone if, in fact, foundry work and lung cancer are not actually associated? It is unfortunate that chance remains a major consideration in evaluating research for it is rarely consequential and its value long has been challenged.⁹ Today, bias and confounding are recognized as far more important issues than chance in interpreting data. A modern evaluation of causality gives chance little weight in the assessment of any one study and virtually none in the assessment of several studies that address the same question. This position, still surprising to many persons, is best explained when it is clear how the role of chance is described. Chance usually is described by a "p-value" (P). P is the probability that an association (an SMR) as strong as or stronger than that found in a study would arise by chance if the suspect cause and the disease were not, in fact, associated with one another. In the foundry example, P (determined by a test of statistical significance) is 0.03, or 3 percent. This means that if the true SMR relating foundry work to lung cancer is 1.0, our finding of 2.0 would

[27 ELR 10282]

arise by chance in only 3 of 100 studies similar to ours. By convention, a P of 0.05 or less is considered "statistically significant," implying that the related result probably did not arise by chance and has a different explanation. Thus, our finding of 2.0 is unlikely to be due to chance and so is likely to be due to bias, confounding (by something other than smoking), or causation.

The role of chance is also described by a confidence interval (CI). The CI is the range of estimates of the true SMR with which the study result is consistent. Our SMR of 2.0 has a 95 percent CI of 1.5 to 2.6 (determined from tables or various formulas). The interpretation is: from this study the best estimate of the SMR is 2.0 and it is 95 percent likely that the true value of the SMR associating foundry work with lung cancer mortality lies in the range of 1.5 to 2.6. The 95 percent CI is by far the most commonly used. The 99 percent CI (which is wider) and the 90 percent CI (which is narrower) are also used.

In terms of P, the major limitations of chance in the assessment of research results are:

(i) P is a complex measure incorporating both the strength of an association (how much the SMR differs from 1.0) and the number of subjects in a study, among other things. A weak SMR, say 1.5, could have a statistically significant P of 0.05 or less in a large study. Or a strong SMR, say 7.0, could have a nonstatistically significant P, say 0.15, in a small study. Thus, the interpretation of a single P requires considerable judgment. A series of Ps attached to many related associations may be virtually impossible to interpret.

(ii) P has little or no meaning when the SMR is 1.0 or near 1.0 since P describes how probable it is that a positive result would arise by chance. Thus, an asymmetry is produced in that the credibility of a positive result may be supported by a small P but the credibility of a null result is not enhanced by the accompanying large, uninterpretable P.

(iii) Interpretations of P often are accompanied by an implication that a study can be interpreted only in terms of chance or causality. However, there are four, not two, interpretations that may be made of a study: bias, confounding, chance, and causality. A small P does indicate that chance is an unlikely explanation for a positive result but it neither supports nor detracts from any of the three alternatives. Thus, P does not address causality in any direct way. The misperception that it does is common even among expert witnesses who contend erroneously that a small P excludes chance and proves causality (plaintiff's side) or that a large P requires that chance, not causality, be seen as the basis for a study's results (defendant's side).

The difficulty of evaluating chance is often underestimated even by sophisticated scientists, regulatory agencies, and courts. As one student of causal thinking indicated, assessing chance "has so befuddled a substantial contingent of the scientific community…that it has led to wholesale misjudgments of studies and entire subject areas."¹⁰

The fourth criterion of causality in C1 is strength of association. Most epidemiologists consider SMRs below 2.0 as weak and those below 3.0 are considered weak by some.¹¹ SMRs ranging from 2.0 or 3.0 to 5.0 usually are described as moderate and those greater than 5.0 as strong. In practice, few weak associations have become accepted as describing a causal relationship. In any case, scientists will not accept a weak association as reflecting causality unless it is well supported by evidence relating to the other criteria of causality.

The fifth criterion is internal consistency, the extent to which a study's results are consistent with one another. There are three major aspects of internal consistency, the most important being dose-response. A pattern of ever-higher SMRs among persons with ever-higher exposures to the suspect cause supports causality. This is based on common sense and on experience showing that causal relationships without dose-response are rare. The second aspect of internal consistency is the uniformity of findings among different groups of subjects. Disease-causing agents act similarly among persons of different sex, race, etc. An association that is similar among different groups of people is more likely to reflect causality than one that varies, unless there is some reasonable basis for the variation. The third aspect of internal consistency is the restriction of the association to one or a few diseases, sometimes referred to as "specificity." A study that finds an agent associated with many diseases is more likely revealing systematic error than causality.

The sixth criterion of C1, temporality, relates to the suspect cause's preceding the occurrence of disease. This would seem not to warrant mention. However, because epidemiology is nonexperimental and many studies are retrospective (done after the occurrence of disease), it is necessary to ensure that the cause-effect sequence is not reversed. The criterion of temporality can be extended to include other factors such as a suitable induction period.¹² This period will be years to decades for most chronic diseases.

The seventh criterion in C1 is biologic plausibility. This is a description, even conjectural, of a reasonable means by which the suspect cause could produce the disease at issue.

After a report has been assessed with the criteria of C1, the study is interpreted in terms of bias, confounding, chance, and causality. (A fifth interpretation, fraud, is not considered here.) These interpretations are not mutually exclusive and most studies are appropriately interpreted, to some degree, in terms of each of them. Every study harbors some bias for none is perfect. A favorable assessment of a study regarding bias is only that none is identifiable. The effects of known confounders can be minimized but unknown confounders may remain. Thus, while a study may be free of confounding, there is no way to know this until the mechanism of the cause-effect relationship at issue is understood. Chance cannot be excluded from the interpretation of a study because P cannot be zero. The fourth interpretation is "valid" which, because of the orientation of this Article, is equated with "causal." But, more generally, valid means the opposite of biased. So, valid means either valid-causal or valid-noncausal, depending on what is actually true. For clarity, this Article uses valid to mean causal, but provides reminders that noncausal is equivalent as a possible valid outcome. Validity, whether causal or noncausal, is responsible to some degree for the results of nearly every study since few are so poor as to produce results that bear no relationship to reality. Since none of the four contenders can be excluded, the interpretation of a study is actually the allocation of a reviewer's confidence among them.

Perspective is provided by considering how the criteria of C1 pertain to experimental studies. Criterion one, relating to bias, is relatively unimportant in experiments as the random assignment of subjects to an exposed and to an identically treated control group causes errors not to be systematic but to affect all study groups equally. Criterion two, relating to confounding, is usually unimportant in experiments since the specific agent under evaluation as a cause is actually used in the study. Criteria three through five (random error; strength; internal consistency) are of about the same importance for experimental as for observational studies. Criterion six (temporality) is a non-issue in the interpretation of experiments because the exposure precedes the disease by design. Criterion seven (biologic plausibility) is a major issue in experimental work since even results that clearly indicate causality in animals can be generalized to human beings only to a limited degree.

Context Two: The General Case

Context two (C2) assesses causation as a general proposition. Continuing with the example, but going beyond the particular study, the issue has become: to what degree does all available evidence support, or refute, the hypothesis that foundry work causes lung cancer? It might seem that C2 would not be addressed until several studies have been accepted as supporting causality in C1. But this is too restrictive. There may be two or more studies that fail individually to support causality in C1 but that together are reasonably supportive. For example, one study might fail C1 because it appears biased while the other fails because it appears confounded. However, neither bias nor confounding may seem a reasonable explanation for both studies considered together.

The first criterion in C2 is external consistency, the extent to which all epidemiologic studies support causality. External consistency has two major components: replicability addresses the similarity of results among the studies; strength has the same meaning as in C1 but relates to the average of all reported results. When most of a large number of studies report a strong SMR, causation is very likely to be accepted. When the available studies present SMRs that are variable and weak or moderate, causation is unlikely to be accepted. Exceptions to these generalizations arise. For example, the association between environmental tobacco smoke and lung cancer is very weak with an SMR of about 1.3. However, the association is found consistently and has strong biological support. Therefore, Dr. Dimitrios Trichopoulos suggests that the association be considered causal.¹³

External consistency is quite important: unless a positive finding is usually obtained, the association at issue is unlikely to be accepted as causal. Consistency is often evaluated in a critical, but subjective, review of the literature. Objective approaches, collectively termed meta-analysis, have been developed to evaluate external consistency. However, most meta-analyses have assessed only the most obvious questions such as those relating to chance and to dose-response. Fortunately, the procedures of meta-analysis recently have been scrutinized and suggestions offered for improvement.¹⁴

Coherence is the second criterion in C2. A causal proposition that is congruent with all or nearly all relevant biological information is coherent. Coherence is more demanding than the criterion biologic plausibility in C1 as it relates to phenomena for which there is evidence. Nonetheless, coherent information is impressive only insofar as it is specific, for general information is commonplace.

The final criterion in C2 is favorable response to manipulation: remove the suspect cause, or protect against it as with a vaccine, and the incidence rate of the disease in question should decline. This criterion is difficult to apply to chronic diseases whose induction periods span decades and whose causes can be reduced only gradually.

After evaluation in C2, the causal hypothesis is assigned a position on the spectrum of credibility. This may be done informally by an individual scientist or by a small group. It may be done formally by a panel of scientists representing several disciplines and assembled by a scientific agency. The panel may express its assessment by categorizing the hypothesis according to the sponsoring agency's classification system. For example, the Working Groups of the International Agency for Research on Cancer use a five-point scale extending from "carcinogenic to humans" to "probably not carcinogenic to humans."¹⁵ This group approach and the use of designated categories has both advantages and limitations.¹⁶

Causality also may be assessed by a panel convened by a regulatory agency. The panel may consist of external advisors including scientists or it may be comprised solely of the agency's staff. The specific causal question at issue for the agency's purpose should be clear and explicit. A regulatory agency usually is less concerned with the ultimate scientific issue and more with a practical matter: should it act as if the causal relationship in question is true? This is a legitimate question to be posed by an agency that is charged with protecting the public's health. It is especially appropriate if a positive response elicits not a regulation but rather an evaluation of the advantages and disadvantages both of acting and of not acting in any particular way. However, if a regulation ultimately is adopted, it must not be found to be arbitrary or capricious if it is to withstand legal challenge.

The credibility of a causal hypothesis also is assessed in law by a judge, or by a jury acting under a judge's guidance, in tort cases. This is done by taking a sounding of the opinions of the scientific community. This process is quite limited in that the scientific community usually is represented by a few experts who are known to hold one or the other position on causality and who have been selected by attorneys.¹⁷ Another limitation in law is that the commonly used burden of proof for "establishing" general causation in C2 is highly subjective. It is that a causal hypothesis can be accepted as true if it is supported by the "preponderance of the evidence." Much rests on what this is understood to mean. Preponderance can mean simply the "majority." But a majority of unimpressive information is not meaningful. It is suggested here that the meaning of "the preponderance of the evidence" be separated from amounts and types of information. Rather, preponderance could be equated to a specified minimum position on the spectrum of credibility of a causal hypothesis. For example, a court might state that for a plaintiff to prevail, the causal hypothesis at issue would require a credibility of, say, 60 percent or more. This "minimum credibility" approach would place the law and science on the same continuum.

Context Three: A Specific Person's Illness

Context three (C3) is presented as it might appear in a tort. The first criterion in C3 is that the criteria of C2 are met; an agent cannot be considered to cause the illness of a specific person unless it is recognized as a cause of that disease in general.

Before describing the remaining criteria of C3, two matters require mention. The first relates to the burden of proof that applies in C3. Usually, the question is whether the plaintiff's illness "more likely than not" was due to the exposure at issue. First, the plaintiff will have to show that he or she experienced exposure to the agent and was diagnosed with the disease that are causally linked by the preponderance of the evidence, however defined, in C2. To show further that his or her particular case of the disease was "more likely than not" so caused, the plaintiff also will have to establish that the documented exposure produces an attributable proportion (AP)¹⁸ of 50 percent or more and that this applies to the plaintiff's particular case. That is, the plaintiff must show that among persons with his or her exposure and disease, 50 percent or more of cases of the disease are due to that exposure. (This 50 percent or more of cases "so caused" should not be confused with the 50 percent or more of the evidence required to meet the conventional definition of "preponderance of the evidence" in C2.) An AP of 50 percent results from an SMR of 2.0. The benchmark is 2.0 because, among cases of an illness occurring in a group of people with an SMR of 2.0, 50 percent of cases are due to background causes and 50 percent are due to the cause at issue. The AP, the percentage of cases due to the cause in question, is calculated as:

AP = [(SMR - 1)/SMR] x 100%¹⁹

The second matter relates to the misperception that although causality, as a general proposition, can be virtually established, it cannot be established in the case of the illness of a specific person. However, causality can be established as more likely than not in the case of an individual's illness. As indicated, the AP is estimated from an SMR that describes an accepted causal relationship. This AP pertains, on average, to each case of illness among exposed persons. For example, the SMR of lung cancer due to smoking 20 cigarettes per day is 10.0 and the AP is 90 percent. From this we infer, as a first approximation, that for each smoker with lung cancer it is 90 percent likely that the disease was caused by smoking. When an epidemiologic measure is applied to a specific person it may require modification to reflect that individual's circumstances. The remaining criteria of C3 address these possible modifications.

The second criterion of C3 is that the plaintiff, in fact, experienced the exposure under circumstances associated with an SMR of 2.0 or more. This relates to issues such as the magnitude of the exposure and the passage of a reasonable induction period.

The third criterion in C3 is that the plaintiff was diagnosed with the specific illness to which the accepted causal relationship pertains. There are subtleties here as similar-sounding names may be used for diseases with different causes. Also, an organ may be the site of different illnesses, even within a category of diseases. For example, there are two major forms of cancer of the uterus and their causes have little or nothing in common.

The fourth criterion in C3 is the absence, or near absence, in the plaintiff's experience of a cause of the disease other than the alleged cause. The nonsmoking foundry worker who contends that foundry work caused his lung cancer is more likely to prevail than is the smoking worker. However, the history of having experienced an alternative cause should not preclude a plaintiff's prevailing. Even among foundry workers who smoke and who have lung cancer some of the cancers will be due to their foundry work (assuming that foundry work was judged a cause of lung cancer in C2) and not to their smoking. The plaintiff may address this issue by assessing the probability that his or her disease was produced by each of several possible causes. For example, the smoking foundry worker may have probabilities of causation of 60 percent, 30 percent, and 10 percent for smoking, foundry work, and background causes respectively and will not prevail. The nonsmoker may have probabilities of, say, 55 percent for foundry work and 45 percent for background causes and will prevail.

To meet the overall burden of proof the plaintiff will attempt to show that a general cause-effect relationship is "established" (by the preponderance of the evidence) and that he or she experienced the cause in circumstances that make it more likely than not that his or her illness was attributable to this cause. However, there is an unrecognized premise that underlies this reasoning: an SMR of 2.0 is the benchmark of "more likely than not" only if the causal hypothesis is virtually certainly correct. To illustrate, say that the general causal hypothesis has a credibility of 80 percent, high enough to meet any definition of "preponderance" but distinctly not 100 percent. Does an SMR of 2.0 still serve as the threshold of "more likely than not"? Literally, no. An 80 percent credibility level and an SMR of 2.0 equate, as joint probabilities, to a likelihood of disease causation in the individual of only 40 percent (80 percent of an AP of 50 percent). An SMR of about 3.0 would be the required benchmark of "more likely than not" if the credibility of the general hypothesis is 80 percent (80 percent of an AP of 67 percent is 54 percent). It is not suggested that courts use the actual probability of correctness of the general hypothesis in rendering a "more likely than not" judgment. A reasonable approach would be to resolve the general issue at a "preponderance" (or credibility) level of, say 60 percent or more, and then to consider that it had been resolved at the 100 percent level when C3 is considered. This is not to suggest what courts should do but only to make explicit this previously unrecognized premise.

Perspectives

Causality is evaluated by panels of scientists to discern truth, by legislative and regulatory bodies to protect health, and by courts to provide justice. The deliberations of these entities should be facilitated by use of a set of explicit criteria of causality, by a clear description of the burden of proof for the purpose at hand, and by use of a contextual approach to allow all information to be considered in a logical sequence.

In science, causality is assessed primarily by moving from C1 to C2, that is from the specific to the general. The law, however, moves from the general to the specific: it begins with an assessment in C2 and then moves to C3, where it evaluates whether an established cause more likely than not produced a plaintiff's illness. Thus, science and law have a common ground in the general issues of C2 but primary interests in different specific issues. These different specific interests perhaps explain some of the difficulty that scientists and jurists have had in communicating as they try to resolve questions of causality. Difficulties may be reduced by the use of the three contexts to permit all parties to move in the same sequence from one set of issues to the next.

The contextual approach also provides general perspectives on causation. Epidemiologists often are asked: how do causation and association differ? This question lies in C1 and a response is that, rather than thinking of differences, causation should be seen as one of the four possible explanations for associations that occur. Causation becomes the favored explanation when its positive criteria are well met and when the other three explanations (bias, confounding, chance) appear unlikely.

A second general perspective relates to formulating public policy. Policymakers interpret information on causation as do scientists; indeed, they may be scientists. Yet, in C2 they use a low threshold of causation, that "as if" guideline already mentioned. When a regulatory agency decides to act "as if" a causal relationship applies, it may conduct a risk assessment. There are many types of risk assessments²⁰ but those developed by regulatory agencies are usually estimates of the maximum amount of illness that could be produced by a specified degree of exposure to a suspect disease-causing agent. It should be recognized that such risk assessments are exercises in public policy rather than in scientific inference. Although the underlying information is of scientific origin, the major determinants of the outcome—regulatory standards—are policy-driven assumptions. When such risk assessments are recognized as exercises in public policy it becomes clear that they should not, in and of themselves, be the basis for a regulatory standard. Rather, such risk assessment should be only a part of an evaluation that considers all of the advantages and disadvantages of setting a standard at any particular level. Unfortunately, risk assessment has been considered widely as a form of scientific inference and has been used to prescribe standards.

A third perspective transcends contexts. It is that the establishment of a causal relationship is a major scientific achievement. It is also rare. The difficulty of establishing causation is documented in the assessments of scientific bodies that evaluate causality. For example, the National Toxicology Program (NTP) of the United States publishes periodically an Annual Report on Carcinogens. Since 1982, each report has included two listings, one of "substances … that are known to be carcinogenic" and one of "substances … that may reasonably be anticipated to be carcinogens." ("A" and "B" lists, for brevity.) The 1982 report included an A list of 22 agents and a B list of 95.²¹ The most recent report, published in 1994, had an A list of 24 agents including only one of those from the 1982 B list.²² Only 1 of 95 agents moved from the B to the A list in 12 years. Thus, the NTP recognizes the difficulty of establishing causality even after a high level of suspicion has fallen on an agent.

Causality lies in the philosophy of science, a branch of philosophy not of science. This implies that causal judgments require scientific information that has been evaluated in a structured, logical manner. It seems reasonable to expect an agency or a person who contends that a causal relationship does (or does not) exist to state the criteria used in arriving at that judgment and the evidence that meets those criteria.

1. COMMITTEE FOR THE STUDY OF THE FUTURE OF PUBLIC HEALTH, THE FUTURE OF PUBLIC HEALTH (Washington, D.C., National Academy Press 1988).

2. ALFRED EVANS, CAUSATION AND DISEASE: A CHRONOLOGICAL JOURNEY (New York, Plenum Publishing 1993).

3. Alfred Evans, Causation and Disease: The Henle-Koch Postulates Revisited, 49 YALE J. BIOLOGY MED. 175-95 (1976).

4. U.S. DEPARTMENT OF HEALTH, EDUCATION, AND WELFARE, SMOKING AND HEALTH, REPORT OF THE ADVISORY COMMITTEE TO THE SURGEON GENERAL OF THE PUBLIC HEALTH SERVICE, PUBLICATION NO. 1103 (Washington, D.C., U.S. Government Printing Office 1964).

5. Bradford Hill, The Environment and Disease: Association or Causation?, 1965 PROC. ROYAL SOC'Y MED. 293-300.

6. INTERNATIONAL AGENCY FOR RESEARCH ON CANCER, INTERPRETATION OF NEGATIVE EPIDEMIOLOGICAL EVIDENCE FOR CARCINOGENICITY, IARC SCIENTIFIC PUBLICATION NO. 65 (Nicholas Wald & Richard Doll eds., Oxford, Oxford University Press 1985); Philip Cole, Saccharin and Bladder Cancer, in EPIDEMIOLOGY AND HEALTH RISK ASSESSMENT (Leon Gordis ed., New York, Oxford University Press 1988).

7. Philip Cole, The Hypothesis Generating Machine, 4 EPIDEMIOLOGY 271-73 (1993).

8. David Sackett, Bias in Analytic Research, 32 J. CHRONIC DISEASES 51-63 (1979).

9. THE SIGNIFICANCE TEST CONTROVERSY (Denton Morrison & Ramon Henkel eds., Chicago, Aldine Publishing Co. 1970).

10. Kenneth Rothman, The Challenge of Enduring Significance, Address to the Society for Epidemiologic Research (June 22, 1995).

11. Ernst Wynder, Introduction, Workshop on Guidelines to Epidemiology of Weak Associations, 16 PREVENTIVE MED. 139-41 (1987).

12. Robert Hoover & Philip Cole, Temporal Aspects of Occupational Bladder Carcinogenesis, 288 NEW ENG. J. MED. 1040-43 (1973).

13. Dimitrios Trichopoulos, Risk of Lung Cancer From Passive Smoking, in 1994 PRINCIPLES & PRAC. ONCOLOGY 1-8 (Vincent DeVita et al. eds., Philadelphia, J.B. Lippincott Co.).

14. Sander Greenland, A Meta-Analysis of Coffee, Myocardial Infarction, and Coronary Death, 4 EPIDEMIOLOGY 366-74 (1993); Malcolm Maclure, Demonstration of Deductive Meta-Analysis: Ethanol Intake and Risk of Myocardial Infarction, 15 EPIDEMIOLOGIC REVS. 328-51 (1993).

15. INTERNATIONAL AGENCY FOR RESEARCH ON CANCER, OVERALL EVALUATIONS OF CARCINOGENICITY: AN UPDATING OF IARC MONOGRAPHS 1 TO 42, SUPPLEMENT 7 (Geneva, World Health Organization 1987).

16. Noel Weiss, Ambiguities in the IARC Criteria for Evaluation of Carcinogenic Risks to Humans, and a Recommendation, 7 EPIDEMIOLOGY 105-06 (1996).

17. Philip Cole, The Epidemiologist as an Expert Witness, 44 J. CLINICAL EPIDEMIOLOGY 35S-39S (1991).

18. Philip Cole & Brian MacMahon, Attributable Risk Percent in Case-Control Studies, 25 BRIT. J. PREVENTIVE MED. 242-44 (1971).

19. Id.

20. Dennis Paustenbach, The Practice of Health Risk Assessment in the United States (1975-1995): How the U.S. and Other Countries Can Benefit From That Experience, 1 HUM. ECOLOGY RISK ASSESSMENT 29-79 (1995).

21. NATIONAL TOXICOLOGY PROGRAM, PUBLIC HEALTH SERVICE, U.S. DEPARTMENT OF HEALTH AND HUMAN SERVICES, THIRD ANNUAL REPORT ON CARCINOGENS (Washington, D.C., U.S. Government Printing Office 1982).

22. NATIONAL TOXICOLOGY PROGRAM, PUBLIC HEALTH SERVICE, U.S. DEPARTMENT OF HEALTH AND HUMAN SERVICES, SEVENTH ANNUAL REPORT ON CARCINOGENS (Washington D.C., U.S. Government Printing Office 1994).