Tobacco Documents Online

B565 F62-37 P 113 FEIN AI„1, J IMiED SPECIAL ARTICi S~~. Incidence,, Prevaience, and Evidence Scientific Problems in Epi{derrniologic Statistics for the Occurrence of' Cancer ALVAN R. FEINSTE_IN, M.D. New ligven, Connecticut' and W!est'Haven, Connecticut JOHN M. ESDAILE, M.D.- New Haven, Connecticut Fromltie Clinical IEpidemiolbgy Unit and Ithe Rob- ert Wood I Johnson Clinical! Scholars Program. Y'ale University School of Medicine, New Haven, Connecticut, and 1 the Cooperative, Studies Pro; gram Coordinating Center, Veterans Administra» tion Hospitail West Haven, Connecticut. This work was supported in part by the RobertiWood Johnson Foundation Grant 6309 and a grant from the Andrew W: Mellon Foundation. Dr. Esdaile was supported in ipart by the Arthritis Society of Canada, the Frank McGill Fund, and Fonds de Ia Recherche en Sante du! Quebec. Requests for, reprints should be addressed to Dr. Alvan R. Feinstein, Y'ale University School ~ of' Medicine, 333 Cedar Street, New Haven„ Connecticut 06510. Manuscript submitted January 3, 1986, and accepted February 11, 1986. 'Current address: McGill University, Montreal, Quebec, Canada. COPY'R I, ~, F{T YDRKFd `lit=n T.-.. 87 NIEW JFtrt:idy~: •"` '-R(IUP NY Major changes'In policy for health, rxiMtion, and Indastriai safety have been' proplosed because of the "epidemla of apprehension" produced by'statistics for the occurrence rates of cancer: TIFMe atatistical InformaT tion, howevery contains gtass'vioiations of epklemiologic principles and ~ scientific standa;lyds for cnedible evidence. The calculations often de- pend on an Improper linkage~of numeratbrs and denominators; and the calculated ratles, assernblbd from reports of overtly detected cases of cancer, represent neither incWence norprevalbnaer Many of the secu- lar, or regional Ichanges In ratlas are due to correspondin® changes'in the availebility and dissemination of diagnostic tecFxtoiogy; but the tec hno- logic ~ changes have not been adequately evaluated I or Investigated. Improvements will require ctrasticafiy aRered approaches to the use of death certificates, tttmor registries, and clues'from necropsy data. Flardly'a day goes kny wittna.f<'another frightening announcemertt'about'the occurrence rates of cancer. They are regularly repcxtad'as increasing in groups of people demarcated according to age, gender, race, or resi- dence in certain states, cities,,or'ottber geographic regions. The ddnrto- graphie demarcations are often accompanied by statistics showing that the : incxeased rates are associated with diverse features' of envir ~~onrnent or nutrition, various types of cMemibals, certain occupations and I indus- tries, or other suspected etiologic culprits. Receiving these frepuertt announcements, the public has,derelbped what Lewis Thomas [1]I has called an i "epidemic of apprehemsion."' Cancer appears to IUrk in every aspect of daily life: the air we breathe, the water we i drink, tHe food we eat the homes we live in, the substances we touch, and the work we do. The purpose of'this essay, is to examine: the methods with whichithe cancer'statistics are assembled, to note the many problarms'that impair the scientif'Iccredibiiity of thedata, and to suggest rmethods of'improvirtg the defects, DEFINUT!IONS @F' PREVALENCE AND INCIDENCE Like other forms of'epidbrniokxgic data, statistics for the occurrence of: cancer are usually reported as rates of 1 incid®nce or prevalence. These two terms are constantly used to express: distinctions in the rrtary' occurrence rates for which epidemiology is famous; and a krtowiedge of the distinctions is a professional lhallmark of epidemiologists. Prevalence is defined as the relative~freyuertey with which a particular event is noted to be present, at a single point in time, in aparrtictyiar population or group of people. Incidence is defined!as the relative frequency with which the event newly appears as the persons at risk are followed forward diring a subsequertU intervall of time. For example, if we examine a, January 1987 The American Joumal of iMedicine Volume 8?r) 113-13.'1 201.5003313

INCIDENCE, PREVALENCE. AND CANCER EVIDENCE-FEINSTEIN and ESDAILE' community population of 100,000 persons and find 400 cases of a ~ particular cancer, its prewalence : is 400 of' 100,000 or 0.004. During the next year, if 200 new cases of that~cancer develop, its incidencemoukf usually be calculated as 200 of 10'0,000 or, 0.:002. If 300 1per- sons of the population die during the year, the lincidence of morrtality, i.e., the total mortality rate, is 300 of 100;000, or 0.003. If 50 of the deaths are attributed!to the cancer,, the incidence of, cause-specific mortality for the cancer is 50 of 1100,000 = 0.0005. Although ithese definitions are easy to istate and illus- trate; the data needed for accurate calculations ~ of inci- dence and prevalence are:not easy to obtain. SCIEN718FIC PROCESSES IN ACCUISITIOfiIIOF DYKTA Rates of occurrence arecalculated when an appropriate numerator is divided by an appropriate denominator. The statistical activities involve the calculation itself, and the various adjustments and correlations that are made with the results thereafter. The scientific activities involve the basic processes that collect appropriate information to be counted as denominators and numerators, The processes reqwire at least four scientrfic procedures in identification, surveillance, registrat'ion, and Iinkage. The first three pro+ cedures must be carried outto find and count3he individu- al persons who will constitute the denominators and nu- merators that appear as the separate numerical constitu- ents of any occurrence rate: The linkage procedure allows the denominators and inumerators to be joined ifor the actual'calculation. Identification. The itlenYGfication of the persons who are counted in a rate requires a~choics of two conditions- one for the denominator state and the other for the numer- ator state-and the :perforrr>ance of examinations to de- tect the presence or absence of the conditions. For the de ninators of most epidemiologic rates, thee selected condition is sirnple: It is the existence of a person who resides in a specified geographic region. In many occurrence rates, however, the persons in the denomina- tor are further subdivided for additional demographic attri- butes, such as gender;, age, race, occupation, or socio- economic status. The geographic and other demographic attributes are usually determined from information collect- ed in the household surveys perf ormed during the decen- nial census. For the numerators, the rates depend on some other conditi©n; which is often a particular medical evert such as development of a selected disease or death. Although the occurrence ofi death is obvious, the detectioni of' a particular disease is often quite difficult, requiring special activities in clinical examination, technologic testing, and diagnostic reasoning. Surveillance. For the derominator~ and numerator con- ditions to be appropriately identified, the persons who have those conditions must be suitably examined. The surveillance process contains the arrangements that are made for the examinations to ~be conducted with suitable frequency, intensity, scope, and'accuracy. The persons in the denominator are usually identified from the surveillance of a national census; which is con- ducted every 10 years to collect data for everyone: in every region. The scope of the process will be incomplete if' all the persons who live in a i particular region are not located and counted, and' accuracy will be impaired if indiwidVial persons offer erroneous infomation to the cen+ sus taker.. Because the census is conducted only once every 10 years, whereas rates of disease occurrence are usually calculated'each year; the denominators must be estimat- ed in each region for the Intercensal years. The diverse strategies used for these estimations depend! on annual data about deaths; births, and migration in the region. Data about deaths and births are easy to get, but patterms of migration are often difficult to quantify in a highly mobile society. When checked at the end'of each decade, the intercensal estimates are sometimes found to: be quite accurate, but in many instances, the numbers are egre- giously wrong. Doll and Peto [2] have described some of the problems that rnake: "the: available population esti'- rnates... surprisingly unreliable," with, one of the cited examples being the 50 percent increase, despite appar- ently suitable "adjustrnent,"'between the number of Unit- ed States men aged 85 1years'or older who were estimat- ed in 1959, and the ~number later counted in the census of . 1960. Regardless of the quality with which the denominators are counted or estimated, the major scientific problem in surveillance is detection of the numerator event. If'the event is death, its occurrence is readily identified; but, events such as disease cannot be confidently noted as present or absent unless suitable diagnostic examinations have been imposed. Because most persons do not re- ceive routine medical surveillance, and because even the patients who have pieriodic check-ups may not receive all of~ the examinations needed to diagnose: or rule out the presence: ofi every disease, the aillnents identified for citation as numerator events depend on diverse clinical phenomena and patterns of medical I practice, In the ab- sence of a standardized rnedicall surveillance! process, substantial variations in surveillance can occur among different persons in the same geographic region, and among different geographic ! regions" The effects of, these variations will be discussed later. When the medical surveillance process is conducted for numerator events, the existence of a prevalent condi- tion can be determined at the :same time that the denormi- nator, condition is noted. For example, in special noncen-~ sus studies,, the prevalence of' certain diseases may be found when persons are askedlduring househokllsurveys to state their particular clinical conditions or diagnoses. If 114 January 1987 The American Journal of'Medicine Volume 82 2015003314

INCIDENCE;,PREWd4LENCE. ANDiCA'NCER~EVIDENCE-FEINSiIEIN and ESD'A'ILE the numerator evea is an incident condition, however, its occurrence cannot be accurately determined without two special acts of'surveillancie. In the first act; the subjects must be suitably examined I at the start of the follow-up interval to find Ithose whohaveatready, bad the event,,and to remove them from the group "at risk" in the denomina. tor. In the second act of surveillance, which transcends the mere perf ' e of a~single diagnostic examination, the denominator persons who are free of the event of interest must receive the appropriate follow-up examina- tions that will either detect the new event n hitappears in the future, or, demonstrate that it has remained absent. Registration. The information acquired during the pro- cesses of'identification and surveillance does not become statistics until the data have been registered lat a suitablb site. For denominator data, this site is usually the national census bureau; but different agencies are used for the numerator data The agencies include municipal and state health departments, as well as national bu.reaus of, "vital statistics." In the United I States, the municipal and state health departments may anaiyze their information sepa- rateiy; but they also transmit their data i to 1 the National Center for Health Statistics„which then carries out nation- al and regional analyses: In the past few decades, new mechanisms have been developed for the registration of numerator data. Special agencies, such as tumor registries, have been established and maintained to collect routine information about'certain diseases, and certain ~ad hoc registries have been created for special research purposes. The ad hoc registries, as exemplified bytheSEER'(Surveillance; Epidemiology, and End Results) system [3]'for assembling idata about cancer, consist of consortiurns of, laboratories; hospitais, or other medical settings in nine regions of'the United States and Puerto Rico; Each consortium collects data on all l neww cases of cancer reported for persons in the corresponding "catchrnent" region. When the registered data are assembled, the statistical processes begin. By counting what has been registered for the denominator, we find N, the size of the denomina- tor group. By counting the registration of' numerator events, we learn their frequency, fL These two sets of counts produce the occurrence rate, calculated as YN. If the numerator event is a "cross-sectional" condition, determined at the same time as the denominator, fJN indicates prevalence. If the numerator eventis "Iongitudi nall" and noted at a later date, fJN indicates incidence. Linkage. In the foregoing calculation, f and' N were joined by a virgule-the siash ~ rriark that serves both to separate and I link the two i numbers. If i each denominator person has received suitable surveillance, the occurrence (or nonoccurrence) of, the numerator condition will be determined directly, and' it can be easily linked for that person. If appropriate surveillance is not established, and if data are obtained in some other manner, the persons in the numerator must be suitably linked to their status in the denominator. This linkage is not easy for most epidemiologic statis- tics, because the incidence rates of diseases or death aree not'obtained lby'direct~ longitudinal follow-up of the individ- uai subjects counted !in the denominator. The denominator information is collected' at the census bureau,, but thee numerator inf rtion is coiiectediat some othersite: The dataare joined by counting the corresponding values of N I and f for persons who apparently live in the same region, but whose actual status has not been verified in both , denominator and numerator. Thus, although the virgule in, the fVN of an occurrence rate implies that~ the f and I N I values arise from suitable processes of' identification, surveillance, and linkage, these processes are often,inad- equate, particularly for the incideneecaicuiations with which the rest of'this essay is concerned. EVOLUTION OF INCIDENCE RATES In the early deveiopment of epidemiologi'c statistics during the 18th and 19th centuries, the incidence data consisted! of rates of total mortality for a igeographic region, such as alcity; county, or'nation. With this focus, almost all of, the required scientific processes were relatively easy to carryy out. The denominator group was simply defined: as the inhabitants living in a particular region; the: numerator condition was a deatha regardless of cause, in any mem- ber of'that group;, the occurrence of death itself could readily be determinet and;, with, the relatively stable, nonmobile life styles that existed before the 20th century, the linkage problem was trivial. The persons who died in a particular geographic region had almost all been born there and continued to live in~that region thereafter. The main scientific challenges of the earVy era were in the requirements for complete registration. An effective census mechanism was needed for an accurate count of the persons contained in the denominator. For an accu+ rate count of deaths as numerator events; each person who died had I to be cited, according, to law, in a death certificate submitted for registration at an appropriate agency. The arrangemenf and development of these registra- tion processes - particularly for the relatively new work ofi death registration - were fundamental activities of such pioneer figures in viral statistics as John Graurnt; William. Farr, and Lemuel Shattuck. The data obtained with the !death and census registrations could provide the necessary values otif and N; and the resuits could easily be linked to calculate UN as an annual mortality rate in each iselected Iregion. An important distinction of, the f/NlquotieM is that it is structured as a proportion, butthe result was called a rate rather than a proportion (or' percentage), because the result came from two different enumerati'ions, using two different sources of data. The indbiidual members of the January 198, The American Journal of'Medicine Volume 8'2 115' 24150 0i315

INCIDENCE, PREVALENCE. AND CANCER EVIDENCE-FEINSTEIN and ESDAILE denominator group had not received the individual surveil- lance necessary to identify and enumerate each person's fate in a manner that would let UN truly represent a proportion. Although the annual mortality rates for a general popu- lation are really estimates rather than directly counted observations, the improved registration processes~of the 19th century made the information trustworthy enough to be used for three major types of'applications: 1. By comparing the death rates with sirrluftaneously registered birth rates, demographers could plot annual trends of the population's composition in different regions. This information could be usedlfor intercensallestimates, but was particularly irnpor<ant for, the "statists" who stud- ied various aspects of a nation's "political arithmetic"'[4].. 2. By noting',the mortality rates in different age groups,, actuaries could construct the "'life tables"' that would denote remaining life expectancy at different ages. This information could be used both demographically as an index of' national', "heaith"' and also commercially to es- tablish appropriate costs for life insurance policies. 3: By linking the occupations and socio-economic sta- tus of both the numerator and denominator groups, social- minded epidemiologists or demographers could demon- strate disparities in death rates for certain occupations andlsocial classes. Although tlire registered death certificates had a citation of the lethal disease(s), the diagnostic information in the 18th and 19th centuries was generally too unstandardized and' inconsistent for effective routine usage. After efforts were made in the second half of'the 19th century [5]Ito standardize the nomenclature of disease and to select a single disease that' would be regarded as the cause of death, the stage became set for 20th century statistics. The "fmodern"'statistics contain not onlythe total mortali- ty rates„but also calculations of'cause-specific mortality rates for individual diseases. ADVANCES IN THE 20.T}H CEN111UR1f' During i the 20th century, as the registration processes became well established throughout technologically well- developed countries, severall major expansions took place in the use of'epidemiologic statistics for incidence:. 1. The lethal causes cited on death certificates could demonstrate cause-specific mortality rates, but the data became used to show the actual incidence of the cited diseases. These rates of' incidence in different demo- graphic groups then became converted into probabilities that represented' the "risk" for development of each disease in individoai members of, each group. 2. Changes in the annual trend of these incidence rates or risks were associated with sudi concomitant phenom- ena as environmental pollution, chemical composition of water, industrialization, and public health or medical inter= ventions. 3'. The statistical trends and associations then led to major beliefs about the role of these pherwmena, as causes ofidisease„ and to major decisions inpubiic policy for healthi and medical research. Thus, the rising secular trends in the annual! incidence of, cancer have produced the belief that a "cancer epidemic" has occurred in technologically advanced societies; and associations of these trends with various forms of urbanization, industrial- izationa,and otherfypes of information~have led to the idea that most cancers are produced by carcinogens in water,, food, tobacco, chemicals, and other suspected sources outside the human body. SCIENTIFIC PROBLEMS IN!20TH CENTURY DATA. Although a complete scope of registration for births, deaths, and regional populations was generally well es- tablished in technologically well-developed countries dur- ing the 20th century, the concomitant advances in tech- nology began to produce or demonstrate substantial new scientific problems in identification, surveillance, registra- tion, and linkage. The identification of, disease has been aitered I by nosoibgi'c changes in concepts of diagnosis; and by dramatic changes in diagnostic and therapeutic technology; The frequency and intensity, of surveillance have been greatly increased by irnproved access to medi- cal care; and by routine examinations performed for screening„empioyrnent, insurance, or other reasons. The recording,and analysis ofiregistered dAta have been sub- stantially enhanced by computers and other advances in data processing, and the linkage process has been com- plicated l by the migrations encouraged by changes in occupations, transportation, and sociocultural patterns. Some of the new problems affect denorninators, oth- ers affect numerators, and all of them affect the statistics calculated as incidence rates for cancer. Identification of Denominators. With incidence rates being used'to show the "risk" of'developrnent of certain diseases, the denominator groups could no longer be easily assembled from census counts. The groups would have'to be modified to include only those persons who were actually at risk. Persons would have to be removed from the at-risk group if'they did not have the particular anatomic organ in which the disease might deveiop;,or if the disease had already occurred. For example, women cannot be counted as being at risk for the incidence of endometrial cancer if'they have had'a hysterectomy or if the cancer is already presentas'a prevalent condition, The problem of absent anatomic structures was creat- ed because new technology allowed various diseases to be treated I with i a relatively safe and simple surgical' re- moval of'a person's uterus, thyroid, gallbladder, or other organs. The rernovedlsfiuctures were then unavailable to be at risk for the development ofi cancer (or other subse- quent diseases), This problem, when Suitably considered, has usually been managed with~sampie surveys or other 116January 1987 The American Journal of Medicine Volume 82 201Si003316'

INCIDENCE„RREWAUENCE. ANDicANCER EVIDENCE-FEINSTEIN and ESDAILE sources of dataiabout4he prevalence of subjects withthe removed structures. This information could then be used to adjust the ominators appropriately for calculating risk. The problem of identifying denominator subjects who alfeady have a cancer is much more important, however, and its statistical conseq ues have still not been fully resolved. The improved diagnostic technology began to show that almost all cancers have a diverse spectrum of clinical manifestations and courses. The same cancer that might appear overtly, with "characteristic" manifes- tations andl prompt fatality, could also occur silently or lanthanically [6] with no overt symptoms; or it could I progress undetected during the patient's life and be first found at necropsy. If rates of incidence were to be calcu- lated correctly, the prevalence cases must be suitably identified. Without these identifications, the denominators will be too high, because they include piersons who al- ready have the cancen The subsequent numerators will' also~be too high if the tanthaniccases are later discovered and regarded las new incidences of disease. The occurrence of a large!reservoir of lanthanic cancer - which exists without producing prominent symptoms to 1 eliciY medical attention and diagnosis -has become well demonstrated by 20thtcentury necropsy studies. For example, from 10 percent to 90 percent, of the primary cancers found in various structures at necropsy had not been previously diagnosed during i the patient's lifetime [I7',8]. The existence of 9anthanic cancer, has also become the basis for "screening carnpaigns°' that are intended to find and give "early" treatment to cancers that might otherwise escape detection until much later, when they begin to produce overt clinical manifestations. Although the effects of absent structures can'receive relatively easy statistical ladjustments, the effects of lanth- anic disease have not yet been suitably reckoned in statistics concerned with i incidence. Ifi the denominator population is supposed to contain i only persons who are truly at risk for ai particular disease, what shouldlbe done to identify and exclude the persons who already'have it? For example, in the population of' 10'0,000 subjects cited earlier, suppose the initial examination process identified 400 cases of primary, lung cancer, so that the original prevalence was 0:0.04. The number of, persons identified by the process, however, would depend entirely on the intensity of'the examination. If itconsisted solety'of ques- tions about a characteristic symptom, such as cougFeing blood, lung, cancer would not be suspected in persons without symptoms. If the examination included routine chest roentgenography, many lanthanic cases of' lung cancer might be found, but the investigators would' still miss the' cases whose symptomless diagnosis required either Pap smears of sputum, flexible bronchoscopy, or both. The customary justification for the scientific neglect of lanthanic prevalence is that the denominator effects are too minor to warrant major, expensive efforts at detection and adjustment. If we did all the work needed to find and remove the 400 prevalent cases of' lung cancer from the population of' 100,000 persons, the at-risk denominator would become 99,600 persons. When 200 new cases of' lung cancer are later detected, the incidence rate would i be 200 of 99,660 = 0.002008-a result that hardly differ's from the value of 0.002 calculated as 200 of' 100,000. The policy of avoiding massive efforts to identify'preva- lent cases can therefore be ju,stified by its negligible statistical effects on the denominator; but the policy does not deal with subsequent scientific problems in the nu- merator. If death is the incidence event, no problems occur; but if the numerator event is subseqNent detection of the cancer in a living person, the problems can be enormous. Suppose the 100,000' persons in the denomi- nator had received a suitable diagnostic examination. With, those examinations, many of the subsequent 200 ""new" cases of lung cancer might have been detected as a prevalence event, If half of those 200 cases were identified initially, the prevalent cases would rise from 400, to 500'and prevalence would increase by 2'5'percent to a, rate of 500 of 100,000 = 0+005. The later incidence of' "new" lung cancer, however would' n falllfrom 200 to 1100, lowering the incidence rate by 50'percent to 100'of 99,500, = 0.00'1. The failure to identify prevalent cases, although statistically innocuous in the denominator, can thus have major effects when the incidence numerator contains living persons. Identification of Numerators. When identified! in living persons, the numerator events can be distorted both by the denominator difficulties just cited and also by prob- lems to be discussed later in the surveillance process: In most epidemiologic calculations of incidence for cancer (and for many other diseases), however, the numerator events are identified at death, The diagnoses recorded'on death certificates are tabulated to provide rates of cause- specific mortality, which are then regarded as incidence rates for the cited causes. This statistical strategy„which has been used for many' decades in almost all nations throughout the world, is the single greatest and most overwhelming scientific defect in the epidemiologic datw that are called "vital statistics." Of'the many major flaws produced by this defect, only five will be cited here. They arise from difficulties produced when death substitutes for life as a source of data about disease, when diagnostic names are chosen for human ailments, when death is ascribed to a single cause, when irubrics are revised for deathTcertificate coding, and I when changes occur in i no- sologic concepts and diagnostic technology. Dlsease In dead and living patients: If a particular cancer is uniformly lethal, its incidence could regularly be determined from thedeaths iticauses. With modem teich- Januaii The /lwnerdcan Journal of iMediclhe Volume 82 117 2015003317

INCIDENCE. PFREV ALENCE. AND CANCER EWIDENCE --FEINS7EIN ar~f ESmAq,E' nology, however, many cancers have been found I to bee either nonfatal or treated so suacessfully that they are not~ causes of death. ConseqNently, the incidence rates will inevitably be too small'if the numerators depend solely on death-certificate citations. Although used for more than a century, the peculiar scientific strategy of substituting cited Icauses of death for the actual occurrence of diseases in living persons has led to many outstanding errors. One obvious error, to be discussed in a later section, is that the rates of incidencie& will rise fallaciously if the count of occurrences expands to include cancers detected in living patients. Choice of' diagnostic tities for disease: Before a standardized nomenclature was developed in the second half' of' the 19th century ['5], doctors could use a huge variety ofi terms as diagnostic names of disease. Thee development and decennial revision of a standard''interna tional system of nosography have provided a specific set of entities for use in nomenclature, but the taxonomy contains more than a thousand entities [9]. Although thee diagnostic names of' these entities have been standard- ized by international agreement, no criteria have ever been, offered for making the diagnoses. Consequently, massive observer variability can occur when the diagnos- tic terms are applied by individual clinical practitioners. In one study [,10], for example, it was noted that the same cfinicall condition was often called', ernpbyserna in the United States and chronic drorrchitis in the United King- dom, In a different investigation, the diagnostic terms used by practitioners seemed toidepend on the medical school they had attended andltheir date of graduation'[1'1]. In the absence of standardized I criteria for diag.noses i two patients with the same ailment can receive different diagnoses from different clinicians. Nevertheless, the in+ consistent diagnoses assembled' in i death-certificate data have constantly been tabulated as though they represent- ed diseases. Ascribing, death to 1 a single cause: Another conse- quence of modem technology and increased life expec- tancy has been the identification of' multiple diseases in individual patients. With rnodem methods of diagnosis readily available and frequently employed, most patients are reguiariy, found to have several diseases, many of which might have acted individually or collectively as a cause of death. Substantial difficulties and inconsisten- cies are thereby created, even after pos. ..tmortem exami- nations, in deciding which disease among several was uniquely, rresponsible for a person's death. Nevertheless, when incidence'rates are calculated from cause-specific mortality data, each death has been ascribed to a single _- disease. Revision of rubrics and hierarchies in coding: Be- cause so many different diseases can be cited as individu- al causes of death, statistical tabulations of incidence rates for each disease would be difficult to analyze and interpret. Consequentl j+, groups of diseases are regularly consolidated into a smaller mmber, of'coding rubrics that are used for the ultimate enumerations. Thus, the rubric of' respiratory cancer may include primary cancers of the trachea, bronchi, and pulmonary par y,rna. The rubricc of digestive-tract cancer may include primary cancers located anywhere in the gut from the pharynx to the anus, as well as cancers of the pancreas, hepatobiliary system, and retroperitoneal space. One scientific problem in tlfiese coding i rubrics is that thei'r, contents are regularly arOW in different eras, so that putrrronary ederna may be classified as a pulmonary disease inone era, and as a, cardiac disease in another [ 101. For incidence data about cancers, the'greatest scientif- ic problem in the coding rubrics is their use of an arbitrary set of'hierarchical rankings for selecting a single cause of' death, Regardless of the diligence or accuracy with which a clinician may have assembled the medical evidence„ formulated the diagnostic reasoning, and prepared the deathi certificate, the'diagnoses entered on the certificatee are often rearranged when the information is Coded'at thee registration agency. The agencies maintain a list of selec- tion rules, sometimes called "Nosology Guidelines,'" [ 12] that provide an arbitrary set of hierarchical''ranks of'lethal- ity for different diseases. Because these rules are applied to alllthe diseases listed on the death certificate, regard- less of the clinician's cited sequence of causes, the agency's hierarchical guidelines may often select some- thing that alters the clinician's choice of the cause of, death. The hierarchical rules differ in different countries and are regularly changed within the same country. Conse- quently, disease A may be regarded as more lethal than disease B in one era and nation; and their rankings may be reversed at some other time and place. In the two report- ed occasions [13,14] in which the hierarchical'selection process was tested, the disagreements in assigning a single cause of'death were so great among, international agencies that the authors of one study recommended that "a new basis needs to be developed for mortality statis- tics." The recommendation has not been carried out. The arbitrary ranks of the hierarchical system are par - ticulariy important for the incidence of cancer, because lethal priority is often given to a cancer whenever it is listedlanywhere'ona death certificate. For example, the clinicianis original decision was rejected, and cancer of the lung i was coded as the' cause of death in i a patient whose death certificate suggested that he'died of myocar- dial infarction seven years after successful pneumonecto - my [1,12]. Changes In rwsology and!tecFnofogy: Tlhe'scientific distress thatmightbe produced by these many caprices is relatively minimal, however,,compared with the impact of changing nosologic concepts and diagnostic technology. With new nosologic concepts of labeiing disease, certain entities - such as dropsy or ehlorosis - may disappear simply because'they are no longer regarded as diseases 11'8 January 1987 TlheAmerican Journal of Medicine Volume 82 "0J00 33,18

! [ 15]j With new methods of' detecting disease, certain entities-such as systemic lupus erythematosus and di- verse cancers - may appear to increase only because they now become identifiedlduring life. If the new diagnostic methods also lead l to stricter nosologic s ds and' criteria„ certain diseases may appear to decrease because the available evidence does not satisfy the strict diagnostic criteria. For example, regardless of'whatever else'has contributed to the declin- ing incidence rates of pulmonary tuberculosis in the'20th century, one promi ncause has been ~ the demand for suitable microbiologic and roentgenographic evidence: With this d. ~emandi patients having only the clinical mani- festations of fever, hemoptysis, and' suitable types of pulmonary rales would'no longerreceiue the diagnosis of tuberculosis that they would have been given almost routinely 8f)' years ago: Similarly, a patient dying with cachexia and l an abdornirnal mass in whom, formerly, stomach cancer would be' diagnosed, could no longen receive that diagnosis when the nosologic criteria ~ began to demand suitable roentgenographic evidence of the cancer[10]J. Despite the'obvious impact of, changing nosology and improved diagnostic teclnnology, their roles are constantly ignored in the rise or fall l of incidence for many other diseases„ particularly caneer. The problems are further discussed in the next section, Effects of Surveillance. Since special diagnostic tech, nology is required to show that certain diseases are present or absent, the application ofi this technology is a crucial feature in the surveillance process for both preva- lence in denominators and I incidence in numerators. For the population under study to, receive suitable surveil- lance, the mere invention and existence of a technologic procedure are not enough; the procedure must be dis= seminated and widely used. For example, although roentgen rays and~ microbial smears had I been developed in the' 1890s, the diagnosis of pulmonary tuberculosis in 11'9'20.'at New Haven Hospital still depended mainly on ai clinical history and physical findings; only 50percent of the patients had I undergone chest radiography and only 10 percent had'had the tiiber- cie bacillus microscopically demonstrated [16]. The cyto- logic smear used to diagnose cancer in desquamated cells was invented by Papanicolaou in 1945; yet almost a decade later, in 1954, the Pap smear of sputum was not yet being used routinely at New Haven i Hospital to help identify or rule out the existence of lung cancer [ 10]. In the past decade, the sharply rising annual rates of pancreatic cancer have aroused'suggestions that another new can+ cer "epidemic" is occurring,,but little or no attention has been given to the role of dramatic new technology- abdominal ultrasonography, computer-assisted I tomogra+ phy, and endoscopic retrograde cholecystpancreato - graphy-in detecting pancreatic cancers that formerly would have escaped identification during life. INCIDENCE, PREVALENCE, AND CANCER EVIDENCE-FEINSTEIN andIESDAILE Without adeqNate attention to the incidence and preva- lence of surveillance with diagnostic technology, and without suitable, consideration of the technologic effects on what becomes detected or left unidentified, the rates of reported diseases simply represent the rates of diagnostic detection: They cannot be accepted as scientific enumer- ations of' the true incidence and prevalence of those diseases. Problems in Aleglstratkxn. Two of'the main problems in the census-bureau registration of denominators were briefly outlinedlearlier: For one of these problems, when the intercensal estimates of, denominators are found!to be wrong, the census bureau can readily correct its data retroactively. If'the denominator estimates have already been used, however, to calculate myriads of annual inci- dence rates for an enormous number of' diverse geo- graphic regions, the attempts to create retroactive cor- rections wouki be formidable. Accordingly, the originalty calculated rates are usually allowed to stand, with the fervent hope that the errors will not be too striking or misleading. The' second problem, which is more substantial l oc- curs when the scope ofithe census is inadequate, leading, to the under-enumeration of' certain ethnic groups. Al- though the conseqwences are most often i discussed as a political issue in legislative representation, the statistical impact can be striking for rates of i disease. For example, after the 1970 census in the Unfted I States, subsequent studies showed that about 10 percent of black men had not been counted [171. If'we suppose that all deaths from cancer in black men are identified and counted as numer- ator events in the medical settings in which ithe patients die, but that only 90 percentofthe black men are actually counted I in i the census denominator, the incidence of cancer deaths, when f is divided by 0.90N rather than by N, will be 1.11 higher than it'should be. The rate oficancer deaths in black men will' be falsely raised by 111 percent. Regardless of' the size of the incidence-rate errors produced by these two problems in denominator registra- tion, a different set of'major problems has been produced by new modem methods of surveillance and registration for numerator events. In the new mechanisms described earlier,, the data collected and set to a special regionall registry can incliide information about cancers discovered during life as well i as those identified at necropsy. The reports that are newly submitted to, the registries each year are then i used Ias numerators and converted to inci- dence rates when liroked I to the corresponding regional denominators estimated from census data. One obvious result of this new registration mechanism would Ibe a sharp increase in the incidence of cancer. The registered reports will include not only all the cancers cited I in i death certificates, but all the other identified cancers that are not lethal and that were formerly not counted as numerator events. AltFjougFn substantially larg- er than before; the newly calculated I incidence rates will January 1987' The American JouAf~,~e J"1~9 ~9~ 119

INCIDENCE, PREVALENCE, AND CANCER EVIDENCE-FEINSTEIN and ESDAILE still, however, be scientifically untrustworthy. 71he rates will continue to be altered by technologic changes whose incidence and prevalence, not having been determined, will remain unadjusted in the diagnostic reports, Problems in Linkage. When an incidence rate is calcu- lated as YNI the original implication is that alll of the N. subjects in the denominator were initially examined and found to be free of' the numerator event. The further implication is that suitable follow,up examinations were performed, with the numerator event found present in f' members of'the group;, and labsent in the other N-f mem- bers. Although incidence rates are interpreted as though they had these two implications, none of the epidemiologicc rates for the occurrence of cancer (or other diseases) are actually constructed in this manner. As notedlearlier, the linkage of f and IN is a purely statistical activity, using data from two different sources: the census bureau and l the agency that collected the numerator information. None of the N subjects in the denominator are examined initially to see whether they have cancer; and for each of the nine years of intercensal estimates, many of the denominator subjects are notactually counted. Of the f subjects count- ed in the numerator, few (if' any) were speeifically checked I to see whether they were indeed appropriately contained in i the denominator. The policy of'forming these rates by purely statistical linkage is a pragmatic necessity of epidemiology, since too many persons are involved for each person to be suitably examined and followed. Nevertheless, the policy has severalI scientific hazards. One hazard, as noted earlier, is created by frequent geographic migration in a highly mobile society. Many of' the numerator persons who die in a particular region may not have been counted in the denominators; and many of the persons who, were counted'in the denominators may have received medical care and diagnosis in some other region, or they may have moved away and Idied elsewhere. A second hazard,, called the ecologic fallacy, occurss when some external agent -such as changes in water, air,,sales of tobacco, or tactics in health care-is associ- ated with changes in the incidence of a particular disease in the region [ 18]. Because the data for these associations are assembled as a total composite of everything occur- ring in the region, the events that produce the main change in totals for the external agent may occur in one part of the region; but the main alterations in disease may occur in a different part of the region. The two unrelated entities may then be fallaciously associated when the events are linked in the composite totals. Thus, if a major rise in use of vitamins occurs in Albany while cancer of the colon has a, major rise in Brooklyn, the two disparate events will be associated in an ecologic fallacy when increasing colon cancer is correlated with increasingiuse of vitarnins for the entire state of'New York. The existence and Imagnitude of these two hazards are relatively easy to check. The disparities between numera• tor and I denominator persons could be determined from suitable sample surveys. The ecologic fallacy could be tested by examining the associations found lin small parts of the region, rather than in the total composite data. These two simple types of scientific investigations are seldom performed, however,, for the data of vital statis- tics. Census bureaus may investigate the effects of erro- neous intercensal estimations in denominators, but the incidance rates of cancer have not been checked for the combined effects of'numerator and denominator dispari- ties. Sociologists, who christened the ecologic fallacy [18], regularly check for it initheir research, and health- care researchers also investigate it [ 1I9], but correspond- ing investigations hav® seldom been carried out for the incidence rates of disease, despite the ease with which such research could be conducted in an age of comput- ers. Although the hazards of~ migration and the ecologic fallacy will occur whenever a purely statistical linkage iss used for the numerators and denominators, a more sub- stantial scientific problem arises from numerator datal supplied by new mechanisms of registration for cancer in living patients. Although constantly reported and interpret- ed I as incidence rates, the results found with i the new registration methods are not really rates of'irucidence. The calculations represent the rates with which new cancerss have been reported, not the rates with which the cancers have newly appeared in persons who were previously cancer-free. When used'increasingly in clinical practice orin publicc eampaigros, diagnostic screening procedures will reveall many prevalent cases of i lanthanic cancer that may have existed undetected for long or short periods of time : When these prevalent cases are newly reported, however, they will be countedlas incident cases. The calculated "inci- dence rates" will therefore offer a grossly inaccurate account of'true incidence. The calculated rates willl be strongly affected by the incidence with which new diag- nostic technology becomes available and routinely ap- plied to identify cancers that formerly might have escaped detection in the lanthanic reservoir [2D]. CONCLUSIONS AND RECOMMENDATI(fNS' If the modem cancer "epidemic'" might be explained as a statistical artifact, and if' many of' the statistics about cancer rates are egregiously wrong, what cani be done to improve the situation, and to get accurate scientific data on which to base enlightenedidecisions for public policy? Of the six suggestions that follow, the first and fourth can be carried out promptly and inexpensively. The others will require either new research or new approaches in research, Eliminate Cause-Specific Mortality Rates. In the cur= rent mechanisms of, preparattiora and collection„ death certificates are legal instruments required for burial or 120 January 1987 The American Journal of Medicine Volume 82 20150'0 i32G

IMCIDEMCE.,PREVALIENCE. AND QA'rJDER'EVIDENCE-FEINSTEIM and ESDAILE cremation of the dead. The occurrence of'death and the concomitant' demographic data listed on the certificate can continue to be used l successfully to calculate total niortaiity, rrates. These rates can still be effectiveiy ana- lyzed for the'demographiC; actuarial, and socioeconomic purposes for which the data have been used since the 19th century;, and the accxacy 'of the rates can be improved with sample surveys that show the kinds of numerator and denominator adjustments needed for geo- graphic migrations. On the other hand, death certificates cannot be used as scientific instruments for dbnoting diseases or indicatirxg lethal maladies. Almost 50 years' ago, after carefully analyzing the ' Yeiiabitity of statistics of separate causes of death," Raymond IPearll['9] concluded'tfiat "the material is fundamentally of a dubious character." Except fordeaths unequivocally attributable to homicide' or trauma. Pearl suggested that the nosologic data recorded on a death certificate could not be accepted "unreservedly at their face vaiue " Because the situation has not improved' during the past half century,, tabulations and analyses of' cause-specilic mortality rates should be regarded as a statistical I idea whose time has passed: Develop an Additional Death Certificate as a Scirantific : Document. The medical information listed oni death certificates couid I become~ valuable scientific data if the death certificate wereAransf ~~Drmed into a scientific~docu- ment. The transformation wouid require a new system wholly different from the ambiguous, inconsistent proce < dure in igeneral use today. The legal and'scientific needs could I be met separately by preparing, tl+vo death, certifi+ cates; rather than one. lme first would be a simple short form, intended legally to verify the occurrence of: death and to state whether a, coroner's inquest is required. The second document would serve the scientific purposes of indicating what diseases were: beiievedl present at the time of death, what evidence existed to sustain the diag- noses, and whether a, single disease could be uneqwi~ro ~- cally regarded' as the lethal I agent. If the diagnostic and' lethal-cause evidence was equivocai, it would so be noted i and classifiedl Because the scientific evidence couid' probably not be suitably prepared in ia routine manner throughout an entire nation, special' registration regions could be estabiished~-analogous to regional tumor regis- tries-where deliberate surveillance, monitoring, and' supportive personnel could be used to maintain scientific quality in the procedure. Improve the International Taxonomy of Disease. If sci- entific quality can be established in the basic evidence of' death cerrtificates, analogous improvements can then be instituted to repair the nosographic and clinical inconsis- tencies of the international taxonDrnry for nomenciaturee and l classification of disease: Some obvious first, steps would be to discontinue the custom of arbitrary hierarchi- cal rankings for rearranging the diseases listedlon death, certificates, to develop better methods for dealing withi January 1987 The Amerbcan Journai of Medibine Volume 82 rnuitipie nosographic entities, and to establish operational criteria'for making specific diagnoses. Add Scientific Clinical Precision to Analysis of Registry Data. For scientific analysis, the' infmrmatiomcoilected in special tumor registries should have an enlarged ciini'r cai scopeand classification. Because ofithe probiems in determining the previous status'of a newly reported case, the annual results cannot be called inciderr¢e rates, and should receive some other name, such as occurrenice rates or reportedirates. The rates should also be more'preciseiy categorized, from data obtained by careful review of' each patient's past history and the clinical manner in which each cancer was manifested and detected. Certain cancers will be detected because they produced symptoms or other man* ifestations that acted as thei iatrotropic stimulus [6] I in making, the patient' seek medical attention. In other pa- tients, the cancer will be diagnosed while lantfianic [6], having produced no sy,rnptorns; or symptoms that'had not become:innporRant iatrotfopic complaints. In the lanthanic patients, the cancer wouid have escaped detection ex- cept for the special screening or other diagnostic test that was part of'the general examination process. In some instances, such as a routinely scheduled peri- odic x-ray or sputum Pap smear, the screening test is specifically aimed at finding a lanthanic cancer. In other instances, such as a lesion found on chest radiography performed for othen reasons (e.g;, to ruie outtubercuiosis; or to show cardiac size), the lanthanic discovery will be inadvertent. By dividingithe newly reported cancers into complainant, specific lanthanic, and inadvertent lanthanic detections, we can get a better idea of the types. of changes that are occurring in annual rates, and of the diagnostic patterns that are contributing to the changes. IfI the terms ircidemt and prevalent are to be rnain- tained, they can ~be used'more preciseiy'if applied after a review of' previous roentgenographic findings or other appropriate diagnostic information obtained before the cancer was overtly identified. If, absent, in previous suit- abie examinations, the cancer can be unequivocally clas- sified as irciderot. If the patient had no pertinent, previous examinations or complaints, and if the cancer was found during a test, performed routinely for screening or other purposes„ it can be unequivocally classified as prevalent. Suitabte criteria can then be developed for an appropriate intermediate classification of cancers found in other cir- cumstances. These irnproverrnents in classification will require careful attention to data that are now generally neglected lor disdained: the particular iatrotropic stimulus that made the patient seek the! mediaal I examination that led to the eventual detection of the cancer; and the results of previous appropriate examinations of the same body structure in which the cancer has been found. Use Old Medical Records to Determine the Incidence and Prevalence of Technologic.,Charnges. At institu+ tions that still lmaintain both old medical records and index 121 201 500 3321

INCIDENCE, PREVALENCE. AND CANCER EVIDENCE-FEINSTEIN and ESDAIUE files for accessing the diagnoses citedl in the records, appropriate records cart be retrieved and reviewed to note the kind of' information collected and used in diagnostic identifications of the same disease at different eras. (This was the ty,pe of research used to show the frequent absence of chest x-ray and l microbial evidence when pulmonary tuberculosis was diagnosed in patients at a promine4 academic hospital'in 1920 [16]). The mainte- nance of voluminous old l medical records, however, is often regarded as a major administrative nuisance at most institutions; and the records are increasingly destroyed as soon as they lose their value as legal documents: Never- theless, the records have an irreplaceable scientific value because they are the only availabie information that can indicate what has truly happened in time as the result of changinginosoiogic concepts and diagnostic technology. Make Better Epidemiologic Use of Necropsy Data. Another valuable btAoften overlooked source of scientific information in epidemiology is the existence of collectedd reports of necropsy findings at different eras. The necrop- sy results are particularly useful' not for their anticipated revelations about a cause of death in a particular patient, but for their unexpected findings of other diseases[6,20]. The unexpected, unsuspected diseases noted at necropsy can help indicate the size of, the undetectedi lanthanic "reservoir" of disease that~ will become detected and reported as "new" cases when improved diagnostic technology is applied to living persons. Since the subjects who undergo necropsy may not be a representative sample of' all deaths, data to be analyzed should focus on the unexpected conditions foundiat necrop, sy, not on the apparent causes of death. For example, because necropsy may not be performed if lung cancer has been unequivocaify, diagnosed during life; the reiative fre- quency of' previously diagnosed lung cancer at necropsy will not be an interpretable rate. On the other hand, in subjects who died of other causes and in whom necropsy, was performed for other reasons, the necropsy discovery, rate of'coexisting but previously unsuspected lung cancer can, be valuable. The subjects who underwent necropsy without a previous suspicion of' lung cancer will provide a "sampie" of the general population in whom death, due too other causes, allowed necropsy to, be the "compietee screening" examination that reveafedithe existence of the lung cancer. If the necropsies in i all these unsuspected subjects are used as a denominator, and if the lung i cancers found ass unanticipated discoveries are used I as a numerator, the result will approximate the same type of' occurrence rate now being', calculated at tumor registries. If the necropsy, discovery cases are removed from the tumor-registry nu- merators, the tumor registry results will show the occur- rence rate of, cancers that are detected during life. The results in the unsuspected necropsy group will show the occurrence rate of cancers that are undetected during life. By rnotingtheseculartrends in these two sets of occurrence 122 rates, we can get a clear Idea ofi what is really happening in the existence; diagnosis, andlreporting of cancers. The argument is sometimes offered that necropsy dis- covery rates are not comparable to tumor registry rates because the tumor-registry discovery rates represent the risk of'cancer in a particular age group, whereas the nec- ropsy discovery rates represent a subject's "lifetime" risk. This argument seems to emerge from a peculiar double standerdlin statistical calculations. WlJhena living 65-year- old man is'foundlto have lung cancer In a screening exami- nation, the result is counted not as a "lifetime risk," but as the occurrence of a lung i cancer among similarly eligible 65-year-old men, If a dead 65-year-old man is found to have a lung cancer in a, screening examination (i.e., necropsy), the resuitican analogously be counted as the occurrence of a lung cancer in the simiiarly eligible group of 65-year-old men. Another possible objection to the proposed use of the necropsy discovery rate is that it should be viewed not as an occurrence rate, but as a proportionate mortality rate: In the latter rate, the denominator consists of a count of all thee single causes cited'for a series of deaths. The numerator contains a count of the frequency with which a particular disease occurs among the total!number of causes of death:. The proposed necropsy discovery rate, however, is quitee different from a proportionate mortality rate, because thee necropsy rate does not emerge from counting causes of, death, andi because several differeft diseases might be detected in the necropsy of a single subject. Thus, some- one who dies of a stroke would Ihavs the stroke counted as a! cause of death in either a cause-specific mortality rate or a, proportionate mortality rate. The same subject's incidental, unexpected necropsy findings of gallstones, coronary dis- ease, cancer of, the pancreas, and diverticulosis wouldl each be counted as part of, the necropsy discovery rates in those diseases. None of these scientific improvements will be easy to achieve. Some of them-such as the development of operational criteria for making diagnoses-represent fun- damental scientific challenges that have been neglected for more than a century. Others-such as better use of medical records and careful attention to the clinical phe- nomena that preceded diagnoses of disease-will require a resurrection of'scientific attention to activities that were once regarded as irn ; but that have been allowed to languish in the efflorescence of' technologic data. Yet others - such as a more creative epidemioiogic use off necropsy information-will need a renewed respect' for what the deadican teach the living„artd a recognition thatt necropsy can be useful in epidemioiogic research and not just in clinicopathologic conferences. Can all these "tadicai" but old-fashioned I scientific activities be restored or newly instituted?' If we want to determine whether the rise of new technology has caused I cancers as weN as detected them, the scientific contribu- tions of old-fashioned forms of research may be our bestt hope. ... January 1987 The American Journal of'Merlicine Volume 82 201500 33 c2