American Journal of EpidemiologyCopyright 1995 by The Johns Hopkins University School of Hygiene and Public HealthAH rights reserved

Vol. 141, No. 2Printed In U.S-A.

HISTORICAL PAPER

AMERICAN

Journal of EpidemiologyFormer); AMERICAN JOURNAL OF HYGIENE

1976 by The Juhru Hnpkiiu Univtnity School of Hym'ene and Public Health

VOL. 104 DECEMBER, 1976 NO. 6

Reviews and CommentaryCAUSES

KENNETH J. ROTHMAN

The conceptual framework for causespresented here is intended neither as areview nor an expansion of knowledge, butrather as a viewpoint which bridges the gapbetween metaphysical notions of cause andbasic epidemiologic parameters. The focus,then, is neither metaphysics nor epidemiol-ogy, but the gulf between them. In thesame spirit as recent discussion on thesepages about definitions of basic epidemi-ologic terms such as rate (1), commonagreement on the conceptual interrelation-ship of causes may facilitate communica-tion about causes of illness.

A strong motivation for presenting thisscheme is the often-heard confusion of twoimportant but distinct epidemiologicissues: confounding and effect modifica-tion. These two properties of variableshave different areas of relevance (2). Theconfounding property is not an intrinsiccharacteristic of any variable. Confounding,defined as distortion in an effect measureintroduced by an extraneous variate, oc-curs only in the context of a particularstudy, und the same variable which con-founds in one study may not confound thesame association in another study setting.

Department or Epidemiology, Harvard School ofPublic Health. 677 Huntington Avenue. Boston, MA02115.

In fact, the principles of good study designmay call for preventing a potentially con-founding variable from being confoundingfor example, by matching in the selectionof subjects. On the other hand, effect modi-fication, defined as differing values of theeffect measure at different levels of anothervariate, is an inherent characteristic of therelationship between two causes of anillness (effect). This relationship is notgoverned by the particulars of any study; itis an unalterable fact of nature. To be sure,the magnitude of effect modification de-pends on the scale of measurement of theeffectfor example, a risk ratio which isconstant over age generally implies a riskdifference which changes with age. But ithas been proposed that the risk differencescale is the appropriate scale for assessingeffect modification, other scales represent-ing metameters of the risk difference whichdistort to varying degrees the assessment ofeffect modification (3). Effect modifiersmay or may nut be confounders in a givenstudy, or they may confound in somestudies but not in others. Conversely, con-founders may or may not be effect modifi-ers. The following discussion presents ascheme for the interrelationship of causeswhich may provide a useful way for think-ing about effect modification as a descrip-tion of nature.

587

90

Historical Paper 91

588 KENNETH J. ROTHMANTYPES OF CAUSES

A cause is an act or event or a state ofnature which initiates or permits, alone orin conjunction with other causes, asequence of events resulting in an effect. Acause which inevitably produces the effectis sufficient. The inevitability of diseaseafter a sufficient cause calls for qualifi-cation: disease usually requires time tobecome manifest, and during this gesta-tion, while disease may no longer be pre-ventable, it might be fortuitously cured, ordeath might intervene.

Common usage makes no distinction be-tween that constellation of phenomenawhich constitutes a sufficient cause and thecomponents of the constellation which arelikewise referred to as "causes". Anotherqualification for sufficient causes is restric-tion to the minimum number of requiredcomponent causes; this implies that thelack of any component cause renders theremaining component causes insufficient.Thus, measles virus is referred to as thecause of measles, whereas a sufficientcause for contracting measles involveslack of immunity to measles virus andpossibly other factors in addition to ex-posure to measles virus. The term cause,then, does not specify whether the refer-ence is to a sufficient cause or to a compo-nent of a sufficient cause.

Most causes that are of interest in thehealth field are components of sufficientcauses, but are not sufficient in them-selves. Drinking contaminated water is notsufficient to produce cholera, and smokingis not sufficient to produce lung cancer, butboth of these are components of sufficientcauses. Identification of all the compo-nents of a given sufficient cause is unneces-sary for prevention, in that blocking thecausal role of but one component of asufficient cause renders the joint action ofthe other components insufficient, and pre-vents the effect. Even without being able toidentify the other components of the suffi-cient cause for lung cancer, of which smok-ing is ont component, it is possible to

prevent those cases of lung cancer whichwould result from that sufficient cause byremoving smoking from the constellation ofcomponents.

A specific effect may result from a vari-ety of different sufficient causes. The dif-ferent constellations of component causeswhich produce the effect may or may nothave common elements. If there exists acomponent cause which is a member ofevery sufficient cause, such a component istermed a necessary cause. Necessarycauses are often identifiable as part of thedefinition of effect. For example, the pos-session of a vermiform appendix is neces-sary for appendicitis, and infection withthe tubercle bacillus is a necessary causefor tuberculosis. Though sometimes devoidof useful significance, a necessary causecan be a useful component cause to iden-tify. Whereas many different componentcauses have been identified for severaltypes of cancer, the hope exists for identi-fication of a final common pathway repre-senting a necessary cause for cancer of alltypes.

Figure 1 is a schematic illustration of thecausal components of a disease. The dis-ease (effect) has three sufficient causalcomplexes, each having five componentcauses. In this scheme "A" is a necessarycause, since it appears as a member of eachsufficient cause. On the other hand, thecomponent causes "B", "C" and "F",which each appear in more than one suffi-cient cause, are not necessary causes,because they fail to appear in all threesufficient causes.

ETIOLOGIC FRACTION

Causal research of disease focuses oncomponents of sufficient causes, whethernecessary or not. The public health impor-tance of a component cause of disease in aparticular population is determined by thefraction of the disease (the effect) whichresults from the sufficient causes(s) towhich the component cause belongs. Theepidemiologic parameter etiologic fraction

Am J Epidemiol Vol. 141, No. 2, 1995

92 Historical Paper

CAUSES 589

SUFFICIENTCAUSE

ISUFFICIENT

CAUSE1

SUFFICIENTCAUSE

m

FIGURE 1. Conceptual scheme for the causes of a hypothetical disease.

(4) (population attributable risk) measuresthis dimension of a cause-effect relation-ship. Each component of a sufficient causehas, as its etiologic fraction, the fraction ofdisease attributable to that sufficient cause(plus the fraction attributable to any othersufficient causes which contain the samecomponent). Consider, for example, thecauses schematically represented in figure1. If Sufficient Cause I accounts for 50 percent of a disease, Sufficient Cause II30 percent, and Sufficient Cause III 20 per cent,the etiologic fractions for each of the com-ponent causes are: A, 100 per cent; B, 80per cent; C, 70 per cent; D, 50 per cent; E,50 per cent; F, 50 per cent; G, 30 per cent;H, 30 per cent; I, 20 per cent; and J, 20 percent. (The example would have to bemodified slightly if there were individualswith more than one sufficient causebecause for these people blocking one suffi-cient cause would not prevent disease.)This example illustrates that the sum ofetiologic fractions of a set of factors is notlimited above by unity, as is observed withalcohol and tobacco as etiologic agents formouth cancer, each having an etiologicfraction greater than 50 per cent.

RISK

The notion of a risk as a continuousmeasure obscures the concept of risk as

applied to an individual. For an individual,risk for disease properly defined takes ononly two values: zero and unity. The appli-cation of some intermediate value for riskto an individual is only a means of estimat-ing the individual's risk by the mean risk ofmany other presumably similar individu-als. The actual risk for an individual is amatter of whether or not a sufficient causehas been or will be formed, whereas themean risk for a group indicates the propor-tion of individuals for whom sufficientcauses are formed. An individual's risk canbe viewed as a probability statement aboutthe likelihood of a sufficient cause fordisease existing within the appropriatetime frame.

STRENGTH OF A CAUSAL RISK FACTOR

The model proposed also illustrates howcharacterization of risk factors as "strong"or "weak" has no universal basis. A compo-nent cause which requires, to complete thesufficient cause, other components withlow prevalence is thereby a "weak" (com-ponent) cause. The presence of such acomponent cause modifies the probabilityof the outcome only slightly, from zero- toan average value just slightly greater thanzero, reflecting the rarity of the comple-mentary component causes. On the otherhand, a component cause which requires,

Am J Epidemiol Vol. 141, No. 2, 1995

Historical Paper 93

590 KENNETH J. ROTHMAN

to complete the sufficient cause, othercomponents which are nearly ubiquitous isa "strong" (component) cause. Inepidemi-ologic terms, a weak cause confers only asmall increment in disease risk, whereas astrong cause will increase disease risk sub-stantially.

Thus the strength of a causal risk factordepends on the prevalence of the comple-mentary component causes in the samesufficient cause. But this prevalence isoften a matter of custom, circumstance orchance, and is not a scientifically general-izable characteristic. Consider the follow-ing simplified example (5): in a societywhere most people eat high phenylalaninediets, inheritance of the (rare) gene forPKU would appear to be a "strong" riskfactor for phenylketonuric mental retarda-tion, and phenylalanine in the diet wouldappear to be a weak risk factor. In anothersociety, however, in which the gene forPKU is very common and few people eathigh phenylalanine diets, inheritance of thegene would be a weak risk factor and phen-ylalanine in the diet would be a strong riskfactor. Thus, the strength of a causal riskfactor, as it might be measured by the"risk ratio" (relative risk) parameter, isdependent on the distribution in the popu-lation of-the other causal factors in thesame sufficient cause. The term strengthof a causal risk factor retains some meaningas a description of the public health im-portance of a factor. However, the commonepidemiologic parlance about strength oicausal risk factors is devoid of meaning inthe biologic description of disease etiology.

SYNERGY

Synergy, also termed effect modificationor positive interaction, may be defined asthe relationship between factors which ex-hibit a joint effect that exceeds the sum ofthe separate effects, with the effects mea-sured on an appropriate scale (3,6). Syn-ergy implies that two component causesare member-, of the same sufficient cause.Neither of two such causal components of a

sufficient cause can have any effect (aspart of that sufficient cause) without thepresence of the other causal component.Two such causes, and, indeed, all thecomponents of a given sufficient cause, aremutually synergistic. Thus, inheritance ofthe PKU gene and phenylalanine in the dietare synergistic in producing phenylketo-nuric mental retardation. If two causes arecomponents of different sufficient causesfor the same effect, and are not mutualmembers of any other sufficient cause forthat effect, they will exhibit no synergy andare thus considered independent in thebiologic (not statistical) sense. If two com-ponents of a sufficient cause have no ef-fect outside that sufficient cause, they willexhibit complete synergy, in the sense thatno increase in risk can occur from eitherfactor unless both factors are present. It iscommonly observed that causes seem tohave leas than complete synergy with othercauses, because some increase in the prob-ability for the effect is observed even whenthe causes occur in the absence of comple-mentary, synergistic causes. This patternresults from such causes being members ofmore than one sufficient cause. Synergyresults from component causes being mu-tual members of a sufficient cause, but thecomponent causes each may also be mem-bers of other sufficient causes with differ-ent complements, thereby also havingindependent effects and making the over-all interrelationship of two causes one olincomplete synergy. Thus, two factorsmay be mutual members of a sufficientcause, and each may separately be a mem-ber of another sufficient cause. Each factorin the absence of the other has an effectthrough one sufficient cause, but togetherthey have an effect through three sufficientcauses, displaying incomplete synergy.Necessary causes are at least partiallysynergistic with all other causes of thesame effect, inasmuch as a necessary causeis a member of every sufficient cause.

Figure 1 suggests many synergistic rela-tionships. For example, "D" and "E" are

Am J Epidemiol Vol. 141, No. 2, 1995

94 Historical Paper

CAUSES 591

completely synergistic with each other andeach is partially synergistic with "A", "B"and "C". Partial synergy exists between"B" and "C"their effect is dependent ontheir joint presence in one sufficient cause,but each also has independent effects inanother sufficient cause. The extent towhich two factors are synergistic depends,like the strength of a causal risk factor, onthe distribution of other factors in a partic-ular population. Consider synergy betweenfactors "B" and " F " in figure 1. Thissynergy results from both factors' presencein Sufficient Cause II, although "B" exertsa separate effect in Sufficient Cause I, and"F" in Sufficient Cause III. In a populationwhere factor "H" were absent, "B" and"F" would exert their effects only throughSufficient Causes I and III, because theabsence of "H" would eliminate any dis-ease from Sufficient Cause II. Thus, "B"and " F " would be completely independent.In another population in which factor "C"were absent, all disease would result fromSufficient Cause II, and "B" and "F" wouldbe fully synergistic.

An example of incomplete synergy wouldbe the mixture of independent and interac-tive effects that alcohol consumption andsmoking have on risk of mouth and phar-ynx cancer (7). Evidence suggests thateither of these factors will increase cancerrisk in the absence of the other, but thecombined effect of both exceeds the sum ofthe individual effects. This would suggestat least three different causal complexes,one involving alcohol but not smoking, oneinvolving smoking but not alcohol, and oneinvolving both. (Some scientists haveargued that alcohol in the absence of smok-ing has no effect, which leads to a modelwith two sufficient causes, one with smok-ing but not alcohol, and another with bothsmoking and alcohol.)

DISCUSSION

The conceptual scheme in figure 1 couldbe modified u> permit greater complexity.Antagonistic causes might be included as a

part of causal chains which lead to anelement being included in a sufficientcause; the joint action of the components ofa sufficient cause could be considered aneffect with its own constellations of suffi-cient causes. Alternatively, the different"pies" in figure 1 might be constructed as asequence of causal events, with branchingpathways representing those segments ofpies which differ, necessary causes beingcommon to all pathways, etc. The possibil-ities for adapting this framework to virtu-ally any complicated causal relationshipreinforce its utility as an intuitive base forcausal thinking.

The model also provides a conceptualframework for the understanding of latentperiod, the time interval between theaction of a (component) cause and themanifestation of disease. When a compo-nent cause "acts" or occurs, the othercomponents needed to complete a suffi-cient cause may not be on hand. If thecomponent cause of interest has a long-lasting "effect", as time passes the othercomplementary components may add their"effects" and gradually complete the suffi-cient cause. At the point in time at whichthe sufficient cause is completed, the dis-ease process is set in motion, though usu-ally not yet manifest. The latent period isthe interval during which a sufficient causeaccumulates plus the time it takes for thedisease to become manifest. Early recogni-tion could theoretically reduce the latentperiod to coincide with the interval duringwhich the sufficient cause accumulates.

For some diseases, such as rabies, virtu-ally the entire period from internalizing theviral agent to the occurrence of symptomsrepresents time during which the disea.sebecomes manifest, because "internalizing"the viral agent is effectively a sufficientcause (a qualification would be that inter-ventive treatment during the "incubation"period might prevent clinical manifesta-tion). On the other hand, the larger part ofthe 10-20-year latent period for the devel-opment of vaginal cancer after in utero

Am J Epidemiol Vol. 141, No. 2, 1995

Historical Paper 95

592 KENNETH J. ROTHMAN

exposure to diethylstilbesterol probablyrepresents accumulation of a sufficientcause (which might ordinarily be com-pleted at menarcht or during adolescence)with only a few years for the disease to be-come manifest after the sufficient causeis complete. The term incubation periodhas often been applied to the period be-tween the accumulation of a sufficientcause and the time at which disease be-comes manifest. Because reversal of dis-ease may be possible during the preclin-ical development, as in the example ofrabies treatment, or catalysts may act toshorten the incubation period of diseaseswhich might otherwise fester in a preclini-cal state interminably (growth enhancersfor neoplasms are an example), it might bebetter to view the latent period as solelythe accumulation of a sufficient cause,components of which might include devel-opmental catalysts and/or the lack of inter-ventive treatment.

Chronic exposures make it more difficultto quantify, and, indeed, to conceptualizelatent period; different doses of an expo-sure accumulated over time may give riseto different risks. Such situations may beviewed as reflecting a set of different suffi-cient causes, each with a different dose ofthe exposure as a component cause. Smalldoses would presumably require a morecomplex set of complementary componentcauses to complete the sufficient causethan large doses. This extension of thecausal model accommodates the descrip-tion of dose-response relationships, andprovides a basis for the common findingthat for many carcinogens the dose is re-lated directly to risk and inversely to latentperiod.

This presentation does not treat thesubtle issues which arise in arriving at aworkable definition of disease. Such issuesobviously pertain to a full consideration ofcauses, inasmuch as disease definition maybe based on experiential criteria (8), butthese considerations go beyond the scope ofthis paper.

It might be argued that the schemepresented here is superficial because theoccurrence of disease in any individualinvolves a collection of component causeswhich constitute a sufficient cause that isunique, by its complexity. Individuallyunique sufficient causes, however, woulddetract equally from all generalized causalmodels. Furthermore, despite individualdistinctions it seems likely that therewould be broad similarities in the compo-nents of sufficient causes for different indi-viduals.

REFERENCES

1. Elandt-Johnson RC: Definition of rates: Sonicremarks on their use and misuse. Am J Epidemiol102:267-271, 1975

2. Miettinen OS: Confounding and effect modifica-tion. Am J Epidemiol 100:350-363, 1974

3. Rothman KJ: Synergy and antagonism in cause-effect relationships. Am J Epidemiol 99:385-:i88.1974

4. Miettinen OS: Proportion of disease caused orprevented by a given exposure, trait, or interven-tion. Am J Epidemiol 99:325-332, 1974

5. MacMahon B: Gene-envirnnment interaction inhuman disease. J Psychiat Re8 6(supp l):393-40-J.1968

6. Rothman KJ: Estimation of synergy and antago-nism. Am J Epidemiol 103:506-511, 1976

7. Rothman KJ. Keller AZ: The effect of joint expo-sure to alcohol and tobacco on risk of cancer of themouth and pharynx. J. Chronic Dis 25:711-71(>.1972

8. MacMahon B, Pugh TF. Epidemiology. Principlesand Methods. Boston, Little, Brown and Com-pany, 1970

Am J Epidemiol Vol. 141, No. 2, 1995