Ciência & Saúde Coletiva ISSN: 1413-8123 [email protected] Associação Brasileira de Pós-Graduação em Saúde Coletiva Brasil

Kundi, Michael Causality and the interpretation of epidemiologic evidence Ciência & Saúde Coletiva, vol. 12, núm. 2, março-abril, 2007, pp. 419-128 Associação Brasileira de Pós-Graduação em Saúde Coletiva Rio de Janeiro, Brasil

Available in: http://www.redalyc.org/articulo.oa?id=63012215

How to cite Complete issue Scientific Information System More information about this article Network of Scientific Journals from Latin America, the Caribbean, Spain and Portugal Journal's homepage in redalyc.org Non-profit academic project, developed under the open access initiative 419 REVISÃO REVISION Causalidade, Epidemiologia Causalidade, Existe um debate atual sobre como e como sobre atual debate um Existe quando o impacto provocado por agente ou deter- ou agente por uma provocado de impacto causa o como quando interpretado ser pode origi- minante causação”, Hill de “critério Bradford chamado Austin O doença. de seminal trabalho do nário de aplicado frequentementeé Susser, Mervyn e inter- a tendência uma causas há de e esquemáticaevidência forma de falta a erradamente pretar Não causal. relação de falta deevidência como de- no ouevidência de agente avaliação um para critérios existemde propensão à respeito descar- diz para que nem doença, uma causar terminantepro- comentário, Neste causação. de avalia- noção a a tar para dialógica aproximação uma ponho Começando determinante. ou agente um de neces- ção itens quatro epidemiológica, evidênciaassociação, comtemporal, relação abordados: ser sitam é populaci- equivalênciafator e um ambiental equivalênciaa dos contra-argumentos, maioria há Na não Se causação. onal. de potencial o atribuído estu- de partir a evidênciainsuficiente há casos, fa- evidên- num outra casos,confiança Nesses epidemiológicos.a dos diminui ou aumenta en- que No cia usada. ser pode doença à relacionado provisório tor é causação de veredito de qualquerpretexto sob tanto, adiada ser pode não ação uma e estado nosso me- quando evidência,tomar melhor preciso encontrar é que mostrar conhecimento saúde. de da proteção para imediatas didas Resumo Palavras-chave Palavras-chave Causality, Epidemiology Causality, There is an ongoing debate regarding debate ongoing an is There Key words Key Abstract how and when an agent’s or determinant’s impact determinant’s or agent’s an when and how orig- causation, to respectof with causation asinterpreted criteria be can Brad- The Austin disease. Sir target of some work seminal the schemati- from often inatingare Susser, Mervyn tenden- and Hilla ford is there furthermore, and, causa- There applied for callyevidencerelation. of causal lack the a misinterpret of to cy lack for con- evidenceevidence of as assessment tion the for criteria no to are propensity determinant’s or agent’s pro- an dismiss cerning I to criteria there commentary, are nor this disease, a In cause causation. an of of assessment notion the the for approach dialogue epidemiologic a from pose Starting determinant. tempo- addressed: or be agent to need issues four evidence, no equiva- environmentalare association, relation, there ral If equivalence. population andthe attributedlence, is factor a counterarguments, valid there often, More causation. disease of potential used epidemiologic befrom evidenceinsufficient can be willevidence other cases, these In factor studies. a in confidence decreases or increases that though Even disease. a to related causally being action provisional, is causation of verdict every knowledge present health our if postponed for be not measures must immediate demand to appears protection. Causality and the interpretation of epidemiologic evidence * evidence epidemiologic of interpretation the and Causality epidemiológica evidência de interpretação e Causalidade and 1 C S Col S C 114:969–974 Rev . The author declares he declares author The . Institute of Environmental of Institute Health, Center for Public for Center Health, University Medical Health, Kinderspitalgasse Vienna. Austria. of Vienna 1095. 15, michael.kundi@meduniwien. ac.at. 1 (2006). doi:10.1289/ehp. (2006). http:// via available 8297 27 [Online dx.doi.org/ of part is and 2006] March collaboration scientific the between EHP financial competing no has May 8 Received interests. March 27 accepted 2005; 2006. * This article was originally was article This * journal the by published Health Environmental Perspectives Michael Kundi Kundi Michael 420 Kundi, M. Kundi,

The principle of causality, so deeply embedded Defining cause and causality in humans’ minds that it has been thought of as immediately evident, is the very foundation not The most advanced sciences, physics and chem- only of all three monotheistic world religions but istry, have altogether abandoned the concepts of also of the first staggering steps of science [de ni- cause and effect. These terms are no longer used hilo nihil (nothing can be born of nothing) 1. in these sciences. Newton had already replaced Hume 2 was the first to note that there is no logi- cause and effect with functional relationships; cal foundation in the assumption that if in the however, to make himself understood to his con- past every event has had a cause, this will also be temporaries, in the third book of his Principia the case in the future and, furthermore, that what (1726) he spoke about causes (especially to de- we perceive in daily life as well as in science is only fend his position of what can be called a minimal a sequence of events but not cause and effect. Al- sufficient cause). Nevertheless, “cause and effect” though Hume deeply believed in the truth of the remained terms used in physics, somewhat anach- principle of causality, he pointed to the role of ronistically, especially for scholarly purposes un- the human mind in constructing reality and the til the end of the 19th century. Mach 7, alluding to futility of scientifically proving its validity. Kant 3, Hume, stressed the psychological nature of these as he became acquainted with Hume’s thoughts, concepts and pointed out that “in nature there is was awakened from his metaphysical slumber, or no cause and no effect” and that these concepts so he kept saying, and set out to solve the prob- are results of an economical processing of per- lem of how Newton’s physics, which he thought ceptions by the human mind. of as eternally true, could be possible in the face The notion that diseases have natural causes of Hume’s demonstration that it cannot be in- and are not God’s punishments or trials or curses ferred from experience. of malicious beings or results of supernatural The Copernican turn in Kant’s reasoning was forces has not even fully penetrated Western cul- to imply the principle of causality from the as- ture, let alone become the prevailing view world- sumption that it is among the conditions of ev- wide. Despite its metaphysical character, the eti- ery experience. Indeed, if A is a necessary condi- ologic axiom that every disease has an endoge- tion of B, then B is a suffi- cient condition of A. nous and/or exogenous cause was extremely suc- Hence, if for every experience we make (B) it is a cessful and is still the foundation of scientific precondition that everything has a cause (A), then medicine. However, what actually “causes” a dis- from the fact that we do have experiences (B), it ease has from the very beginning been a matter follows that everything has a cause (A). Howev- of controversy. Indeed, a single clinical phenom- er, to make this a logically coherent theory, Kant enon can have quite different “causes,” and one had to sacrifice “objective knowledge”—that is, “cause” can have quite different clinical conse- the Ding an sich (the “thing in itself”) remains quences (Table 1). These facts are not consistent incomprehensible for the human mind. For more with the original concept of causation, which than 100 years, the philosophy of science circled states that a cause is an object that is followed by around either the assumptions or the (untoward) another, and where all objects similar to the first consequences of Kant’s solution. When in 1905 are followed by objects similar to the second 2. Not Einstein published his special theory of relativity even for infectious diseases does this (strong) con- and his theory of the interaction of electrons and cept of causation hold. (Hume gave several “def- light 4,5 , the very foundation of Kant’s philosophy initions” of a cause, among these also what has was called into question: the universal truth of been called the counterfactual approach, dis- Newton’s mechanics 6 and the validity of the de- cussed below.) terministic concept. How, then, should cause and causation be These considerations not only profoundly defined? In a review of definitions of “causation” changed modern science but also resulted in an in epidemiologic literature, Parascandola and open-ended controversy within epistemology. Weed 8 delineated five categories. However, all of And last but not least, epidemiology and the in- these definitions (summarized in Table 1) have terpretation of epidemiologic evidence are deep- severe deficits. Not totally unexpected, the defi- ly connected to these fundamental considerations nitions found in the literature are insufficient to about the nature of human knowledge. provide a basis for the notion of disease causa- 421 Ciência & Saúde Coletiva, 12(2):419-428, 2007

Table 1 Definitions of causation from the epidemiologic literature (modified from Parascandola and Weed 2001).

Definition Main criticism A cause is something that produces or creates an effect. Tautological because “ production” and “creation” are synonyms of “causation” A cause is a condition without which the affect cannot Only very few diseases could then have a cause a occur. A cause is a condition with which the affect must occur. Again, only few diseases could then have a cause b

A cause is made up of several components, no single Introduces unnecessary complexity in cases of simple one of which is sufficient of its own, which taken dose response and in cases of interaction between together must lead to the effect. components

A cause is a condition that increases the probability of Does not distinguish between an association and a occurrence of the affect. “cause” c

A cause is a condition that, if present, makes a Is, in the strict sense, unprovable because there is difference in (the probability of) the outcome. only one world and one cannot observe it twice-once with once without the condition a Many disease definitions already include a cause( e.g., AIDS is a clinical syndrome in the presence of HIV infection of CD4 cells), but this must not be confused with a necessary cause. All clinical symptoms that occur in AIDS patients can have a variety of other” “causes”. b For example, falling from the 27th floor onto the pavement is not necessary cause for breaking the skull because many other processes can lead to this effect; however, it can be seen as a sufficient cause. Except for injuries due to extreme physical or chemical conditions and exposure to extremely contagious infectious agents that lead to death (e.g., rabies)or do not result in immunity(e.g., gonorrhea),there are no sufficient causes in this strict sense. C Following this definition, male sex would be a cause of lung cancer.

tion. As pointed out above for physical phenom- one of abstraction and generalization because no ena, it is also impossible for disease processes to two diseased human beings given the same diag- draw an ontologic demarcation within the indefi- nosis have exactly the same features. nite stream of events between causal and non- In this pragmatic sense, disease cause can be causal associations. defined as follows: Given two or more popula- Consider a human being as a complex input– tions of subjects that are sufficiently similar for output system that is described by a path through the problem under study, a disease cause is a set a state space (of likely very high dimensionality) of mutually exclusive conditions by which these that may or may not explicitly depend on time. populations differ that increase the probability The task is to solve the equations that relate the of the disease. In some cases, the similarity must input stream, the output stream, and the inter- be high, such that only homozygous twins can be nal states to each other. The solution could give studied; in other cases, maybe only sex and age the probability that the human being will be in must be considered, or the state of immunity. To some internal state of disease at some point in avoid encumbering the definition with unneces- time given a set of initial and/or side conditions. sary complexity, we use the term “conditions” and If we were in possession of such a tool, we would the active verb “increase.” What is meant is that a not need the crutch of a concept of causation. number of extrinsic and/or intrinsic factors (i.e., Meanwhile, in a pragmatic sense, it is reasonable conditions) can be discerned that are present be- to stay with this concept but hold in mind that it fore diagnosis of the disease and that prevail at a is just an economical way to organize the other- time and for a duration that is compatible with wise unfathomable stream of events and to take what is known about the natural history of the the necessary steps to counteract or prevent the disease. Hence, this temporal relation is a precon- disease process. The process of diagnosis itself is dition for an agent to be considered a causal fac- 422 Kundi, M. Kundi,

tor. The “conditions” must be mutually exclusive and no evidence of a causal relationship 12 . The (e.g., groups of males characterized by one of the former expresses an important piece of evidence following conditions: smoking or having smoked that may have substantial consequences on steps cigarettes, cigars, pipes only, more than one of taken to prevent health hazards, whereas the lat- these, or none), because otherwise the increase ter simply expresses lack of knowledge. It is, how- in the probability of the disease cannot be ever, often misunderstood as an exculpation of uniquely related to any one of them. the agent in question and is readily misused by This definition is in line with the main de- interested parties to claim that exposure is not signs of epidemiologic studies: the cohort, the associated with adverse health effects. case–control, and the randomized controlled tri- Some examples of such statements illustrate al. It is also in line with the pragmatic definition the point: that assessment of causality affords more than just . A “formal causation analysis based on an the observation of an increased incidence or prev- application of the Hill criteria confirms that there alence in some group or the other. This is the is no causal relationship between diesel exhaust point from which Sir Austin Bradford Hill start- and multiple myeloma” 13 . ed his considerations that led to what are now . “Applying a weight-of-evidence evaluation commonly called the “Bradford Hill criteria” 9. to the PCB [polychlorinated biphenyl] epidemi- ologic studies can only lead to the conclusion that there is no causal relationship between PCB ex- Taking refuge in causality posure and any form of cancer” 14 . . “Results of these studies to date give no con- It seems that the first time causality entered the sistent or convincing evidence of a causal relation discussion on epidemiologic results was during between RF [radiofrequency] exposure and any the tobacco controversy in the late 1950s and early adverse health effect” 15 . 1960s. In particular, the criticism of Fisher 10 con- There are significant differences between cerning the conclusions drawn from the British these statements. The last one claims that there is Doctors Study by Doll and Bradford Hill 11 initi- no “consistent or convincing evidence” (whatev- ated a detailed consideration of the concept of er this may be) of a causal relation. Hence, it causality that led to the famous presidential ad- points mainly to the lack of knowledge accumu- dress by Bradford Hill to the Section of Occupa- lated so far. The second one goes a step further: It tional Medicine of the Royal Society of Medicine claims that risk assessment based on the weight- in 1965. In this talk, Bradford Hill discussed nine of-evidence approach [as applied by the U.S. En- issues that should be addressed when deciding vironmental Protection Agency 16 or the Interna- whether an observed association is a causal rela- tional Agency for Research on Cancer 17 ] leads to tionship. These issues, now called the “Bradford the conclusion of no causal relationship. Howev- Hill criteria”—although they were not intended er, there is no category of this type in the weight- as criteria and not all of them have stood the test of-evidence approaches. Either the category “not of time—are still the starting point of many a likely carcinogenic to humans” 16 or “evidence sug- treatise on the subject today. gesting lack of carcinogenicity” 17 may be used. The Bradford Hill criteria were established Because of the by far higher demands on quality such that, in the case they are met for a specific and size of studies set out to dismiss the assump- factor, this would increase our confidence in this tion of carcinogenicity, there is an inherent im- factor being causally related to the disease. How- balance of classification concerning carcinogenic- ever, they were not intended to dismiss a factor ity and lack of carcinogenicity. The first statement as potentially causing the disease: “None of my goes still further: It claims that an analysisbased nine viewpoints can bring indisputable evidence on the Bradford Hill criteria confirms that there for or against the cause-andeffect hypothesis and is no causal relationship. Because the only Brad- none can be required as a sine qua non ”9. ford Hill criterion that is essential is “temporal Some statements in the past few years about relation,” the only way to confirm— based on the relationship between environmental or oc- these so-called criteria—that there is no causal cupational factors and human health have used relation is to demonstrate that exposure com- the terms “causality” or “causal” in a negative menced after disease onset. All other evidence sense—that is, claiming that there is no evidence may reduce the weight in favor of a causal rela- for a causal relationship. First, one has to discrim- tionship but cannot confirm that there is no caus- inate between evidence for no causal relationship, al relationship. 423 Ciência & Saúde Coletiva, 12(2):419-428, 2007

Are there criteria for causation? been detected already and are subject to an indi- vidual and/or public risk–benefit evaluation. So During the past decades, Bradford Hill’s criteria we are dealing with either less obvious hazards have played almost the same role in occupational or those that occur only rarely or in a small pro- and environmental risk assessment as Koch’s pos- portion of the population. The evidence may stem tulates for microbiology 18 . As was the case with from all kinds of sources, but often we start only Koch’s postulates, which cannot be fulfilled for from the pessimistic assumption that an agent many infectious agents, so Bradford Hill’s criteria either not present in the natural environment or are supportive (for the assumption of a causal re- present only at much lower levels may be harm- lation) only if fulfilled, but cannot be used to dis- ful to health. Or it may be that during routine miss the assumption of a causal relation. It is a surveillance, a high prevalence of a (rare) disease complete misinterpretation of the nine issues con- is observed that coincides with a (rare) environ- sidered by Bradford Hill that they can be a type of mental condition. How should we come to a con- checklist to establish causation. But it may turn clusion whether the suspected environmental out that they owe their popularity, still persisting condition is causing disease? It might be worth- after 40 years, exactly to this misconception. while to stress that there are cases where we do Because the definition of a disease cause giv- not need the verdict of causation before we take en above affords the existence of mutually exclu- action (e.g., a not very important food additive sive conditions, in a strict sense, causation can be may be banned on weak evidence of harmful ef- indicated only by (experimental) production and fects). An important part, and a much ignored control of all (relevant) conditions. This, howev- one, of Bradford Hill’s article deals with such sit- er, leads to ethical problems if the factor is po- uations, as Phillips and Goodman 19 pointed out. tentially debilitating or lethal. And it is practical- Starting from the definition of a disease cause ly impossible if the latency is long, as it is for stated above, it is obvious that three main issues chronic diseases. Resorting to animal experimen- need to be addressed (to simplify the discussion, tation can reduce some of these problems but let us speak of the set of exclusive conditions as introduces new ones, because inference from re- of an agent or determinant A): sults in animals to effects in humans is far from . Is the probability of the disease conditional trivial. Hence, we are often left with a number of on the presence of A higher than in the absence problems that cannot be optimally solved, and of A? (association) therefore there is no set of criteria that, if ful- . Is the set of conditions to which the source filled, would result in attributing a factor as ei- populations are exposed sufficiently similar ex- ther causally related or not. This does not mean cept for A? (environmental equivalence) that we cannot, to the best of our present knowl- . Are the features of the populations that dif- edge, come to a decision concerning the relation- fer with respect to exposure to A such that, for ship of an agent and a disease. Or, as Bradford the problem under investigation, they can be con- Hill 9 said 40 years ago: All scientific work is in- sidered equivalent? (population equivalence). complete—whether it be observational or experi- Association . Although we can to some degree mental. All scientific work is liable to be upset or rely on statistical decision theory concerning an modified by advancing knowledge. That does not observed difference, some problems need to be confer upon us a freedom to ignore the knowledge addressed: First, there are cases where we observe we already have, or to postpone the action that it an incidence only in those exposed to A and con- appears to demand at a given time . trast it to the overall incidence in the population (as was the case with hepatic angiosarcoma in workers exposed to vinyl chloride monomer). If A pragmatic approach the disease is extremely rare in the population, it may not be feasible to do a conventional epide- Concerning a particular chemical or physical fac- miologic study. However, if a plausible mecha- tor, general medical knowledge may suffice to at- nism of action can be delineated, the observation tribute it as harmful and as causing illness or of an unexpectedly high incidence of the disease death (but even in extreme cases such derivations may suffice for a verdict of causation. Second, in may not be altogether valid—e.g., the statement the case–control approach, we estimate not the that it is impossible to climb Mt. Everest without conditional probabilities of the disease but their respiratory aid). But in a developed society, ob- ratio. Furthermore, it is questionable whether viously, hazardous conditions are likely to have statistical decision theory based on random sam- 424 Kundi, M. Kundi,

pling can be applied without further consider- strength of the association between the confound- ation. Typically, all cases of the target disease oc- ing variable and both A and the outcome required curring within a specified Causation in epidemi- to substantially alter inferences about the exist- ology region (or even only those diagnosed in one ence of an association between A and the outcome. or several hospitals) and during a specified peri- These approaches may replace the earlier proce- od of time are intentionally included, and only dures, as already applied by Bradford Hill. controls are sampled (either from the population Population equivalence . The counterfactual or from hospital cases presenting with other than approach to causality (last statement in Table 1), the target disease). To apply statistical decision although of questionable empirical content, has theory, we have to assume that the cases are a ran- great heuristic strengths. A counterfactual cause dom sample from the distribution of all samples is defined as something that leads to a difference related to all time/space intervals. Furthermore, in the disease propensity with respect to the same the population from which the cases and controls target (population). Although, of course, it is then originate has, in general, not been stable during impossible to ever empirically demonstrate such the relevant past. Cases of the target disease that a cause, it points to the importance of consider- occurred before study onset are not included, and ing all features of the populations that are sub- also migration in and out of the target area may stitutes for the target exposed to A or not exposed play an important role, as might deaths from oth- to A, respectively. Ideally, all features of these sub- er and maybe related causes. Because of these cir- stitutes should be equal. However, this would af- cumstances and the additional problem of reli- ford restriction to homozygous twin studies with ably assessing the presence of A retrospectively, twins who shared the same experiences except for case–control studies are often denied the poten- exposure to A. However, for practical purposes, tial to form the basis of a causal interpretation. it will suffice to demonstrate equivalence with However, this is exaggerating the difficulties as- respect to the features that determine suscepti- sociated with this study type. Especially if several bility to A, disposition to develop the target dis- case–control studies from different areas and time ease, and the interaction between disposition and periods are available, a generalization about the susceptibility (i.e., the joint distribution of these ratio of incidences can be made if the different features). sources of bias have been thoroughly addressed. Unfortunately, as a National Cancer Institute Finally, even if the relative risk (whether estimat- workshop has stressed 24 , there is insufficient evi- ed from rate ratios, odds ratios, or hazard ratios) dence to stratify populations based on suscepti- is high, statistical significance may not be reached bility to develop cancer. For other chronic dis- if the number of cases exposed to A is low. eases, such as atherosclerosis, Alzheimer’s disease, Environmental equivalence . Ideally, those ex- and obstructive pulmonary disease, there might posed to A should share the same conditions, be- be even fewer evidence-based criteria for dispo- sides A, with those not exposed to A. If not, all sition and susceptibility. Therefore, a still more relevant conditions that are potentially related to modest approach must be followed that is em- both A and the outcome (i.e., confounding con- bedded in the universal scientific scheme of bold ditions) must be included in the data set to ac- trial-anderror correction. count for them in the analysis. Failing to do so— As a minimum requirement, we must address that is, controlling for some but not others—may the features that are known to be related to dis- increase confounding instead of removing it 20 ; on ease incidence (in most cases, age will be among the other hand, controlling for a variable that is these features); features that indicate early steps downstream of A may remove the effect of A 21 . of the target disease (e.g., polyposis for colon can- Because the number of potentially confounding cer), thereby keeping in mind that agent A may factors is indefinite and judgment about the de- be effective only during certain steps of the patho- gree of similarity between environmental condi- logic process; and features that may determine the tions depends on limited experience, there is al- potential to counteract or aggravate the disease ways the possibility that an observed association (e.g., social class). Scientific discussion may re- is due to confounding. On the other hand, the veal that potentially important features have been mere suspicion that an observed association is due left out. In this case, considerations of the poten- to confounding does not conform to scientific rea- tial bias hereby introduced may reveal that the soning because it cannot be refuted by a finite se- effect of A has been underestimated (e.g., if those quence of empirical tests. Analysis of uncontrolled exposed to A can be considered less prone to de- confounding 22,23 can give an idea about the velop the target disease). If the investigation re- 425 Ciência & Saúde Coletiva, 12(2):419-428, 2007

sulted in a positive association between A and the equivalence condition, then other considerations target disease, we might conclude that no further and evidence can be substituted to support the investigation is needed; if, on the other hand, no assumption of a causal relation (see below). association was revealed, there is indeed a need Temporal relation, association, and environ- for error correction. An analogue procedure fol- mental and population equivalence suffice for a lows from a suspected overestimation of the as- verdict of potential causation. This assertion can sociation. only be refuted by the following: Environmental equivalence and population . Evidence that demonstrates that A is a down- equivalence are usually termed the ceteris pari- stream condition of some other factor B (e.g., bus condition and are often jointly discussed. It Helicobacter pylori infection instead of gastritis is, however, important to discriminate between as a potential causal factor for atherosclerosis) environmental and population characteristics. . Evidence that A is associated with B, the es- Only the former can be targets of change; the lat- sential causal agent (e.g., technical tetrachloroet- ter, although not stationary at all, must be taken hene contaminated with epoxybutane) as side conditions that can be controlled only by . Evidence that essential side conditions have active selection. It is also important to consider been overlooked that need to be present to make self-selection processes in observational studies A effective or to make non-A preventive (e.g., a where features of the environment may determine specific receptor phenotype). to some degree features of the population and vice It is not necessary to demonstrate a mecha- versa. nism of action. Bradford Hill 9 and others point- It goes without saying that all investigations ed to the landmark 1854 study of John Snow, who that are assessed for a causal interpretation must demonstrated that the rate of cholera deaths in be scrutinized for potential biases (especially ex- London was 14 times higher in households sup- posure and outcome misclassification and re- plied with water from the Southwark and Vaux- sponse or observer bias). However, it is insuffi- hall Company compared with households sup- cient merely to point to a potential bias without plied with water from the Lambeth Company 25 . considering the effect this bias may have had on Although Snow suspected a living organism con- the results. For example, in cohort studies, expo- taminating drinking water by proximity to sew- sure misclassification can lead to a bias only in age, another 30 years elapsed before Robert Koch the opposite direction of the reported association. isolated Vibrio cholerae, and more than 100 years Under the precondition that all investigations before the mechanism of action of the cholera have been thoroughly assessed concerning asso- toxin was established. The original observation ciation, environmental equivalence, and popula- of Snow sufficed to state that something in the tion equivalence, and potential biases, and still water supplied by one company potentially caused the following set of statements can be derived, cholera and to take appropriate action (closing then it is reasonable to allocate A among the po- the pump), and there was no need to wait until a tentially causal factors of the target disease: mechanism of action had been demonstrated . The temporal relationship between expo- (thereby probably sacrificing the lives of thou- sure to A and disease onset (or diagnosis) con- sands of people). However, if a mechanism of forms action can be established, the requirements for . There is an association between exposure to epidemiologic evidence outlined above can be A and the target disease. somewhat relaxed. . Environmental characteristics in which ex- Because of difficulties inherent in observa- posed and unexposed populations live can be con- tional studies, it may be impossible to demon- sidered equivalent during the etiologically rele- strate environmental and/or population equiva- vant period except for A. lence to a sufficient degree, and therefore addi- . Characteristics of exposed and unexposed tional evidence and considerations are necessary populations are sufficiently similar to consider to support the notion of a causal relation between them equivalent. agent A and the target disease. There is no possi- Only the first two statements are essential; the ble evidence beyond the three points stated above latter two can be substituted by evidence from that will refute epidemiologic evidence in favor experimental or other research demonstrating a of a causal relation besides more and “better” mechanism of action that does not depend on in- epidemiologic evidence. Stakeholders tend to dividual characteristics or environmental factors. “flood” the scientific literature with inconclusive Furthermore, if it is impossible to demonstrate the (powerless and/or biased) studies in the hope that 426 Kundi, M. Kundi,

the balance of evidence will turn in favor of a less to causal inference. It is assumed that epidemi- strong association between agent A and the tar- ologic evidence has been put forward that is get disease. Assessment of evidence must take this evaluated along the criteria outlined above. A into consideration and make proper use of such scientific dialogue of conjecture and refutation information (which in most cases will result in at first tries to dismiss the notion of a causal re- disregarding it altogether). lation between agent/determinant A and disease There is an extensive literature about “crite- D along the four issues “temporal relation,” “as- ria” for causal inferences in the health sciences, sociation,” “environmental equivalence,” and most of which goes back to the seminal work of “population equivalence.” There are valid and Bradford Hill 9 and Mervyn Susser 26 . Although invalid counterarguments. If the dialogue ends neither author meant to establish a checklist, but without valid counterarguments, no further ev- only to formulate issues that aid in this task, ap- idence for the verdict of causation is necessary. plication has been more or less schematically fol- More often than not, epidemiologic evidence lowing these criteria. However, there is no rule will be insufficient (e.g., due to short duration that can guide the decision. How many of the cri- of exposure). In this case, other evidence may teria must be fulfilled? Is one counting more than support or weaken the assumption of a causal the other? What to do if none is fulfilled? There relation between A and D. The most important is no straightforward answer to these questions, of these arguments favoring or against causa- and every single case merits its own specific line tion are shown in Table 3. Arguments against of argumentation. causation are often not symmetrical to argu- Tables 2 and 3 propose a dialogue approach ments in favor of causation. For example, a long-

Table 2 A pragmatic dialogue approach to causal inferences about an agent or determinant A with respect to a disease D; evidence from epidemiologic studies.

counterarguments In favor of causation Valid Invalid Temporal relation Exposure to A commenced after of D No mechanism of action of A on any or all stages of D has been established

Association A is a downstream factor of agent/ Exposure to A has not been precisely determinant B that has been indicated as a assessed. causal factor of D.A is associated with B There could be exposure that has been indicated as a causal factor misclassification. of D. There is differential bias (response or There is potencial bias (response or observer bias)in the direction of an observer bias) with unknown affect on association between A and D. There has the association between A and D. been differential disease misclassification There has been disease misclassification in cohort studies. in case-control studies but nor associated with exposure.

Environmental Confounding conditions with a combined There a could have been confounding. Equivalence effect exceeding that of agent A have not been considered.

Population Equivalence A is associated with selection into the There is a potential selection bias with study population. unknown effect on the association Risk of A applies only to a subgroup of the between A end D. population. Exposure is associated with a prior risk to develop the disease.

At this stage no further evidence for establishing causation unless valid counterarguments have been put forward 427 Ciência & Saúde Coletiva, 12(2):419-428, 2007

Table 3 A pragmatic dialogue approach to causal inferences about an agent or determinant A with respect to a disease D: evidence increasing or decreasing confidence in a potential causal relation between A and D.

Type of evidence Increasing confidence Decreasing confidence From prior Results conform to predictions from theoretical Although there are sound arguments for specificity knowledge considerations and/ or prior knowledge about of outcome, specificity of type of exposure, or specificity of outcome in different subgroups of specificity regarding the outcome in different the population.Association between A and D is subgroups of the population, data do not conform coherent with biologic knowledge and/or a to these expectations.There is knowledge about plausible mechanistic model of action can be mechanism of action that indicates lack of affect of delineated. A and D

From Strength of association between A and D There are known confounders not considered in Epidemiology exceeds that of potential confounders existing investigations strong enough to explain Association between A and B is consistently the observed affect. observed in different population, with different There is substantial heterogeneity in the effect of A types of studies, or in different time intervals. on D in different populations, different study Manipulating A in the population changes types, or different time intervals. pattern and/ or frequency of D. In the case a Manipulating A in the population does not affect meaningful meter of the “dose” of A can be occurrence of D. defined, there exists a dose –response A meaningful “dose” meter can be defined but the relationship. relationship between “dose” and response is not monotonous.

From animal Long-term animal studies in different species There exist animal models of the diseased, and in studies indicate an association between A and D none of these models A os effective. (or, an analogue of in these species). No promoting or antagonizing effect of A with a A enhances the effect of a known pathogen B. variety of other agents could be found in different In animal experiments, intermediate steps exposure regimes relevant for human exposure. of the pathogenic process can be evoked by In different species that are sensitive to other exposure to A. exposures producing affects expected to be similar to those of A, the latter is ineffective.

From i n vitro Exposure cell or tissues react or get damaged by In cell line or tissues sensitive to exposures similar studies exposure to A consistent with the pathogenic to A, no effect of exposure to A is found. process of D No changes in cellular processes or alterations of Upstream events can be observed by exposure to signaling pathways can be evoked by exposure to A. A that may lead to D in the intact organism. No promoting or antagonizing affect of A with a A enhances the effect of a known cellular variety of other agents could be found. pathogen B

term experiment in animals that results in a higher for chronic diseases such as cancer there must incidence of the target disease in exposed animals be epidemiologic evidence before an extrinsic supports causal inference, whereas a negative re- agent can be ascribed a hazardous potential for sult does not support the assumption of no caus- human health. Considering the long latencies al relation, because the tested species or strain involved in these diseases, there is a need to de- may lack a decisive feature (e.g., an enzyme) that fine procedures that give answers about a po- is present in humans and necessary for A to pro- tential causal relationship in a more rapid fash- duce D. There are, however, cases where a posi- ion. Traditional epidemiologic evidence can be tive result in animal experiments cannot be taken provided only ex post, when the health impair- as evidence for causation because of processes ment has already occurred in a significant frac- not present in humans. tion of the exposed population. There is an ur- Most risk assessment procedures demand that gent need to connect the disciplines of molecu- 428 Kundi, M. Kundi,

lar biology and epidemiology 24 . Such collabora- factors such as electromagnetic fields, knowledge tion should result in a) a better characterization is more limited, and new approaches must be of the study participants with respect to suscep- designed. tibility and b) early markers of responses to the Despite its metaphysical character, the prin- agent in question that can be assessed long be- ciple of causation or, more specifically, the no- fore occurrence of manifest disease. With regard tion that every disease has a cause has been of to such new approaches, it is of paramount im- great heuristic value and likely will govern our portance to investigate the mechanism of inter- future endeavors for better understanding of the action of the extrinsic agent with the organism in relationship between the environment and hu- order to define potential cofactors and sensitive man health until we have accumulated more end points. For chemical substances, in silico knowledge and may describe the process by a sys- methods and structure–activity considerations tem of equations. However, the complexity of the may provide first answers to a potential path of problem may be too great ever to lend itself to action (e.g., binding to a receptor). For physical complete description.

References

1. Lucretius. De Rerum Natura . [On the Nature of the 16. U.S. EPA. Guidelines for Carcinogenic Risk Assessment, Universe]. Trans. RE Latham. London: Penguin Books; Draft. NCEA-F-0644. Washington DC:U.S. Environ- 1951. mental Protection Agency; 1999. 2. Hume D. A Treatise of Human Nature. Book I. Of the 17. IARC. Preamble to the IARC Monographs. Lyon, Understanding . Glasgow: William Collins Sons & Co.; France:International Agency for Research on Cancer; 1962. 2004. [accessed 2005 Feb 5]. Available from: http:// 3. Kant I. Kritik der reinen Vernunft . Darmstadt: Wissen- monographs.iarc.fr schaftliche Buchgesellschaft; 1975. 18. Koch R. Die Aetiologie der Tuberkulose. Berliner Klin 4. Einstein A. Ueber einen die Erzeugung und Verwand- Wochenschr 1882; 19:221–230. lung des Lichtes betreffenden heuristischen Gesich- 19. Phillips CV, Goodman KJ. The missed lessons of Sir tspunkt. Annalen der Physik 1905a; 17:132–148. Austin Bradford Hill. Epidemiol Perspect Innov 2004; 5. Einstein A. Zur Elektrodynamik bewegter Koerper. doi:10.1186/1742-5573-I-3 [Online 10 April 2004]. Annalen der Physik 1905b; 17:891–921. 20. Maldonado G, Greenland S. Estimating causal effects. 6. Newton I. Philosophiae naturalis principia mathemati- Int J Epidemiol 2001; 30:1–8. ca . 3 rd . ed. London: Guil. & Joh. Innys; 1726. 21. Kaufman JS, Poole C. Looking back on “Causal think- 7. Mach E. Die Mechanik: historisch-kritisch dargestellt . ing in the health sciences.” Annu Rev Public Health 2000; 9th . ed. Leipzig: F.A.Brockhaus; 1933. 21:101–119. 8. Parascandola M, Weed DL. Causation in epidemiolo- 22. Greenland S. The impact of prior distributions for un- gy. J Epidemiol Community Health 2001; 55:905–912. controlled confounding and response bias: a case study 9. Bradford Hill A. The environment and disease: associ- of the relation of wire codes and magnetic fields to child- ation or causation? Proc R Soc Med 1965; 58:295–300. hood leukemia. J Am Stat Assoc 2003; 98:47–54. 10. Fisher RA. Cigarettes and cancer. Cent Rev 1959:60–66. 23. Robins JM, Rotnitzky A, Scharfstein DO. Sensitivity 11. Doll R, Bradford Hill A. The mortality of doctors in re- analysis for selection bias and unmeasured confound- lation to their smoking habits. BMJ 1954;1:1451–1455. ing in missing data and causal inference models. In: 12. Altman DG, Bland JM. Absence of evidence is not ev- Halloran ME, Berry DA, eds. Statistical Models in Epi- idence of absence. BMJ 1995; 311:485. demiology . New York: Springer Verlag; 1999. p.1–92. 13. Wong O. Is there a causal relationship between expo- 24. Carbone M, Klein G, Gruber J, Wong M. Modern cri- sure to diesel exhaust and multiple myeloma? Toxicol teria to establish human cancer etiology. Cancer Res Rev 2003; 22:91–102. 2004; 64:5518–5524. 14. Golden R, Doull J, Waddell W, Mandel J. Potential hu- 25. Snow J. On the Mode of Communication of Cholera . 2 nd man cancer risks from exposure to PCBs: a tale of two ed. London: John Churchill; 1855. evaluations. Crit Rev Toxicol 2003; 33:543–580. 26. Susser M. Causal Thinking in the Health Sciences: Con- 15. Ahlbom A, Green A, Kheifets L, Savitz D, Swerdlow A. cepts and Strategies in Epidemiology . New York: Oxford Epidemiology of health effects of radiofrequency ex- University Press; 1973. posure. Environ Health Perspect 2004;112:1741–1754; doi:10.1289/ehp.7306 [online 23 September 2004]. Article presented in 23/10/2006