induced risk review - bmjinduced abortion. however,the original study does not distinguish between...

Journal of Epidemiology and Community Health 1996;50:481-496

Induced abortion as an independent risk factorfor breast cancer: a comprehensive review andmeta-analysis

Joel Brind, Vernon M Chinchilli, Walter B Severs, Joan Summy-Long

AbstractStudy objective - To ascertain, from thepublished reports to date, whether or nota significantly increased risk ofbreast can-cer is specifically attributable to a historyofinduced abortion, independent ofspon-taneous abortion and age at first full termpregnancy (or first live birth); to establishthe relative magnitude of such risk in-crease as may be found, and to ascertainand quantify such risk increases as maypertain to particular subpopulations ofwomen exposed to induced abortion; inparticular, nulliparous women and parouswomen exposed before compared withafter the first full term pregnancy.Included studies - The meta-analysis in-cludes all 28 published reports which in-clude specific data on induced abortionand breast cancer incidence. Since somestudy data are presented in more than onereport, the 28 reports were determined toconstitute 23 independent studies. Overallinduced abortion odds ratios and odds ra-tios for the different subpopulations werecalculated using an average weighted ac-cording to the inverse of the variance. Anoverall unweighted average was also com-puted for comparison. No quality criteriawere imposed, but a narrative review ofall included studies is presented for thereader's use in assessing the quality ofindividual studies.Excludedstudies - All 33 published reportsincluding data on abortion and breast can-cer incidence but either pertaining only tospontaneous abortion or to abortion with-out specification as to whether it was in-duced or spontaneous. These studies arelisted for the reader's information.Results - The overall odds ratio (for anyinduced abortion exposure; n= 21 studies)was 1.3 (95% confidence interval of 1.2,1.4). For comparison, the unweightedoverall odds ratio was 1.4 (1.3,1.6). Theodds ratio for nulliparous women was 1.3(1.0,1.6), that for abortion before the firstterm pregnancy in parous women was 1.5(1.2,1.8), and that for abortion after thefirst term pregnancy was 1.3 (1.1,1.5).Conclusions - The results support the in-clusion of induced abortion among sig-nificant independent risk factors for breastcancer, regardless of parity or timing ofabortion relative to the first term preg-nancy. Although the increase in risk wasrelatively low, the high incidence of both

breast cancer and induced abortion sug-gest a substantial impact of thousands ofexcess cases per year currendy, and a po-tentially much greater impact in the nextcentury, as the first cohort of women ex-posed to legal induced abortion continuesto age.

(J Epidemiol Community Health 1996;50:481-496)

Epidemiological evidence of a positive as-sociation between induced abortion and theincidence of breast cancer was first presentedby Segi et al' in 1957 based on cases diagnosedbetween 1948 and 1952. Experimental evi-dence of a causal association between inducedabortion and breast cancer in rodents was pre-sented by Russo and Russo2 in 1980. Yet,despite the alarmingly high incidence of bothbreast cancer and induced abortion, the lastfour decades have produced neither consensusof opinion within the medical research com-munity nor a sense of urgency to arrive at one.Although a few dozen studies have appearedworldwide, and many of them support a pos-itive association, the potential of induced abor-tion as a breast cancer risk factor continueslargely to be minimised. For example, the re-cent study by Daling et al' which reported asignificant, 50% increase in the overall riskattributable to induced abortion, was publishedin the Journal of the National Cancer Institutewith an accompanying editorial by Rosenberg4which described the results as "far from con-clusive". Similarly, in a more recent study ofwomen in Greece, Lipworth et al' confirmed theoverall findings ofDaling et al' but neverthelessconcluded: "At this stage, perhaps all that canbe definitively stated is that any risk associatedwith induced abortion is at most statisticallymarginal". Previous reviews have also notserved to clarify this issue. The New EnglandJ7ournal of Medicine's extensive, 1992 review ofbreast cancer 6 fails to mention abortion at all,even among potential risk factors. The sameis true for the recent breast cancer review 7

published in The Lancet. Reviewers who haveincluded a discussion of induced abortion asa real or potential risk factor have not beencomprehensive,8-12 and also often fail (as domany ofthe original epidemiological studies) todistinguish between induced and spontaneousabortion.89 Even when the distinction is made,erroneous citations are common. For example,Harlap" cites studies by Hadjimichael et al"and Vessey et al'4 as examples of studies of

Department of NaturalSciences, BaruchCollege, The CityUniversity ofNewYork, 17 LexingtonAvenue, New York,NY 10010, USAJ Brind

Center forBiostatistics andEpidemiology andDepartment ofPharmacology,Pennsylvania StateUniversity, The MiltonS Hershey MedicalCenter, Hershey,PA 17033, USAV M ChinchilliW B SeversJ Summy-Long

Correspondence to:Professor J BrindAccepted for publicationApril 1996

481

on February 17, 2021 by guest. P

rotected by copyright.http://jech.bm

j.com/

J Epidem

iol Com

munity H

ealth: first published as 10.1136/jech.50.5.481 on 1 October 1996. D

ownloaded from

http://jech.bmj.com/

Brind, Chinchilli, Severs, Summy-Long

Table 1 Epidemiological studies on abortion and breast cancer incidence which do not report specific data on inducedabortion

Author(s) Year Nation of study population Overall OR (everlnever) Significant? (yeslno)

Wynder et al" 1960 Japan 1.64 NoStewart et aPF 1966 Israel (Jewish) >1 Yes*Valaoras et at92 1969 Greece 1.22 YestSalber et al" 1969 USA (white) 0.97 NotLowe and MacMahon24 1970 Wales 0.89 NotYuasa and MacMahon25 1970 Japan 1.25 YestMirra et al26 1971 Brazil 1.30 NotRavnihar et at97 1971 Slovenia 0.94 NotLin et at2" 1971 Taiwan 1.30 NotPaymaster and Gangadharan29 1972 India 0.74 YesStavraky and Emmons30 1974 Canada 1.59 NoAbeatici et al3' 1975 Italy 0.36 NoHerity et al12 1975 Ireland Data not shown NoSoini 1977 Finland 1.6 YesChoi et al34 1978 Canada >1 YestToti et al3 1980 Italy <1 NotPaffenbarger et alt6 1980 USA (all races) 0.81 YesKelsey et al3' 1981 USA (all races) 1.7 No§Lubin et al38 1982 Canada 1.0 No§Vessey et al'4 1982 England 0.84 No§*Helmrich et al9 1983 95% USA 1.0 NottEnachescu and Lemneanu40 1984 Romania 2.45 Yes**Talamini et al" 1985 Italy 0.74 NottLevshin and Chepurko'9 1985 Russia 1.6 Yes**Hadjimichael et al"3 1986 USA 3.5 Yes§**#4Kvale et a12 1987 Norway 0.84 NottYuan et a13 1988 China (PRC) 0.89 NoBernstein et at" 1990 USA (white) 1.13 NottSellers et at45 1993 USA (99% white) 1.2 No§#%Gandra et al'7 1993 Portugal 0.5 YesAndrieu et at16 1993 France 1.0 NottRao et at" 1994 India 0.8 NoAndrieu et al" 1994 France 1.4-2.1 No***ttt* Data given in terms of number of pregnancies rather than number of subjects.t Data shown are as presented in 1995 reanalysis of Michels et aP".t Data given in form other than odds ratio or relative risk.§ Abortions stated to be all or mostly all spontaneous.** OR given is for abortion before full term pregnancy only.tt Other paper(s) by same group on same or overlapping study population contain specific data on induced abortion and areincluded in the present meta-analysis.j4 Cohort study.§§ Specific data on induced abortion collected but not shown.* Two or more abortions only.

ttt Abstract only.

induced abortion, whereas the former dealtexclusively, and the latter almost exclusively,with spontaneous abortion. It is therefore thepurpose of the present study to establishwhether or not clear trends exist in the epi-demiological literature specifically about anyoverall relationship between induced abortionand breast cancer. Evaluation of relationshipswithin certain subgroups are included wheresufficient data have been published. This ana-lysis should prove useful in clarifying directionsfor future research, and provide a basis forguidelines governing clinical practice. It is alsohoped that the present work will eliminatethe current confusion regarding spontaneousversus induced abortion vis-a-vis breast cancerrisk, and that it will ultimately help womenconsidering elective abortion to make betterinformed choices.

MethodsSEARCH METHODSPublished studies were located using the Med-line (National Library of Medicine, USA) data-bases back to the earliest available publicationdate (1966), using the subject search terms

"abortion", "breast" and "cancer", and bysearching bibliographies of original studies andreview papers. English translations of studiespublished in Japanese, 1516 Portuguese, 17 andRussian'819 were professionally provided bycontract with the Frank C Farnham Co (Phila-delphia, PA, USA).

STUDIES EXCLUDED FROM THE QUANTITATIVEANALYSISTable 1 lists all published studies to date whichreport on the association between abortion andbreast cancer incidence but which do not reportspecific data on induced abortion. 13l4l7l9 Insome cases, the distinction between inducedand spontaneous abortion was not made duringdata acquisition, and in others, data were col-lected separately and combined for analysis.The 1979 study by Levshin and Chepurko19 isbriefly summarised in English in Remennick's1990 review," in which the "slightly increased"risk of breast cancer in women with abortionbefore first full term pregnancy is ascribed toinduced abortion. However, the original studydoes not distinguish between induced andspontaneous abortion. The studies of Hadii-michael et all' and Sellers et al45 report datafor spontaneous abortion only. In the latterstudy, the authors collected data on inducedabortion, but stated only that, "The reportedfrequency of induced abortions was low". Ves-sey et all4 combined data for induced and spon-taneous abortion before first full term

pregnancy, but induced abortion accounted for"only a handful" of the 113 cases and 127controls in this category. Bernstein et al4 pre-sented only combined data in terms of "in-complete pregnancy" in their continuation ofthe study of Pike et al.49 In the original study,49which we include in the quantitative meta-

analysis, separate data are presented for in-duced and spontaneous abortion. The reports

482



j.com/

J Epidem

iol Com

munity H


ownloaded from


Induced abortion breast cancer meta-analysis

ofHelmrich et al,39 Talamini et al44 and Andrieuet at46 represent studies whose data on inducedabortion and breast cancer risk are presentedin other reports which are included in thequantitative meta-analysis. Of the 33 studiesnot reporting specific data on induced abortion,29 present data in terms of relative risk (RR)or odds ratio (OR), with a range of values from0.36 to 3.5 (table 1). Of these, five studiesreport values at or below 0.8, 12 between 0.8and 1.2, and 12 report values above 1.2. Thus,the lack of a significant trend is obvious.

DESIGN OF THE META-ANALYSISTo date we have located 28 original publishedreports which describe a total of 23 independ-ent studies which report data specifically oninduced abortion and breast cancer. Based onthe particular questions most commonly ad-dressed, the meta-analysis summarises the dataaccording to the following categories:1. Overall breast cancer OR for women with ahistory of: one or more induced abortions: (n=21 studies)2. Breast cancer OR in women with a historyof one or more induced abortions before a firstfull-term pregnancy, who are: either parous ornulliparous: (n = 7 studies)a: nulliparous: (n =7 studies)b: parous: (n = 6 studies)

3. Breast cancer OR in (parous) women witha history of one or more induced abortionsonly after the first full term pregnancy: (n = 6studies)

NARRATIVE REVIEW OF INCLUDED STUDIESSince the detailed "epidemiological study on

cancer in Japan" was published by Segi et allin 1957, before the basic methods of epi-demiological data analysis were standardisedby Mantel and Haenszel,50 their data were ana-lysed differently from all subsequent studies.This would present no difficulty for inclusion

in the meta-analysis if raw data were given interms of numbers of exposed and unexposedpatients and controls. However, Segi et al' re-

ported their data in terms of numbers of preg-nancies of each type - ie, live births, still births,spontaneous abortions, and induced abortions.Thus, the numbers of exposed and unexposedsubjects cannot be ascertained, even thoughthe numbers of patients (n = 432) and controls(n = 1713) are specified. Fortunately, data re-

ported for exposure rates in two other Japanesestudies provide a basis for a reasonable es-timation of the number of exposed patients inthe study of Segi et al.' Published in 1968, thestudy of Watanabe and Hirayama'5 presentsRR calculations for each number of inducedabortions in patients admitted for breast cancer

surgery between 1940 and 1942. Thus, themean numbers of abortions per patient and percontrol who had had at least one inducedabortion are 1.92 and 1.82, respectively. Thesecond Japanese study that can be used isthe 1982 study by Nishiyama'6 of a patientpopulation admitted from 1970 through 1979,and the mean numbers of induced abortionsper patient and control who had had at leastone are 1.82 and 1.65, respectively. The onlyother study of Japanese women giving data oninduced abortion and breast cancer is that ofHirohata et alf, in which only dichotomousdata are given. Since the patients studied bySegi et al' were hospitalised for breast cancerbetween 1948 and 1952, the study populationsof Watanabe and Hirayama'5 and Nishiyama'6bracket them in time in addition to agreeingclosely on the induced abortion exposure rate.Therefore, we have averaged the exposure ratesfor patients (1.87) and controls (1.74) in thesetwo studies in order to estimate the number ofexposed patients (n = 53) and controls (n =86) represented by the numbers of artificiallyaborted pregnancies given by Segi et al 1 Theseassumptions and calculations make it possibleto include the study of Segi et all in the meta-analysis under category 1 (table 2), although

Table 2 Odds ratios and 95% confidence intervals for different categories of exposure to induced abortion in component studies of the meta-analysisOR and 95% confidence interval for exposure catego?y

Induces abortion before FFTPNation of study Any induced Induced abortionRef no Year population abortions Any parity Nulliparous Parous after FFTP only

1 1957 Japan 2.63t (1.85, 3.75) -

15 1968 Japan 1.51t (0.91, 2.53) -

18 1978 Russia 1.71t (0.80, 3.64) - - - -52 1979 Yugoslavia 0.50t (0.33, 0.74) - - - -49 1981 USA - 2.37 (0.85, 6.93) - - -16 1982 Japan 2.52t (1.99, 3.20) -

53 1983 USA 1.2 (0.6, 2.3) 2.2 (0.7, 7.2) 5.5 (0.8, 36.8) 1.34 (0.3, 5.6) 0.89 (0.4, 2.0)54 1984 France 1.32 (0.97, 1.77) -

51 1985 Japan 1.52 (0.93, 2.48)55 1988 Denmark - - 2.91t (0.77, 16.2) -

56 1988 USA 1.2 (1.0, 1.6) 1.1 (0.8, 1.6) 1.3 (0.8, 2.2) 0.9 (0.5, 1.4) 1.4 (1.0, 1.9)57 1989 USA 1.9 (1.2, 3.0) -58, 59 1989, 1990 Sweden, Norway 0.9 (0.5, 1.3) 1.09 (0.71, 1.56) - 0.82t (0.44, 1.51) 0.58 (0.38, 0.84)61, 63 1991, 1993 Italy 0.92 (0.80, 1.06) - 0.8 (0.4, 1.5) -64 1993 USA 1.0 (0.7, 1.4) --65 1993 USA 3.1 (2.0, 4.8) -66 1994 USA 2.44 (1.0, 6.0) ---67 1994 France 1.1 (0.7, 1.8) -

3, 68 1994 USA 1.36 (1.11, 1.67) 1.5 (1.1, 2.0) 1.7 (1.1, 2.6) 1.4 (1.0, 2.0) 1.5 (1.0, 2.2)69 1995 USA 0.99 (0.81, 1.21) - - - -5 1995 Greece 1.51 (1.24, 1.84) 1.68 (1.25, 2.25) 0.98 (0.56, 1.73) 2.06 (1.45, 2.90) 1.59 (1.24, 2.04)70 1995 Netherlands 1.9 (1.2, 3.1) 1.45 (0.76, 2.75) 0.9 (0.4, 2.3) 2.6 (1.0, 6.8) 1.7 (0.9, 3.1)71 1996 USA 1.23 (1.00, 1.51) - - - -

* First full term pregnancy or first live-birth.t Raw odds ratio and confidence interval calculated by StatXact.

483



j.com/

J Epidem

iol Com

munity H


ownloaded from



their analysis was also restricted to parouswomen. Another salient feature of this studyis a non-cancer outpatient control populationslightly older than the patient population (ap-proximate median ages, 53 versus 48, re-spectively).

In addition to describing the oldest data setin the literature on induced abortion and breastcancer, the 1968 study of Watanabe and Hi-rayama'5 is also unique in that the 238 breastcancer patients are compared to 110 stomachcancer patients as controls. Here again, controlswere slightly older than cases (approximatemedian age, 48 versus 43, respectively).The 1978 study by Dvoirin and Medvedev"8

concerned 227 Russian and Kabardin patientsand 500 controls in the former Soviet Union.It was reviewed in English by Remennick,'2who reported RR estimates of 2 for one or twoinduced abortions and 3.4 for three or moreinduced abortions. However, these data aregiven for induced and spontaneous abortioncombined in the original study.'8 Where theoriginal study does show data specifically re-lating to induced abortion, it shows only di-chotomous data for induced abortion inKabardin (RR= 1.4) and Russian (RR= 1.89)women (combined RR= 1.71), but withoutshowing any significance values, confidenceintervals, or raw data. Hence, we developed acomputer program to find all possible ex-posures for a population of 227 patients and500 controls which would yield an OR of 1.71.Out of all the possibilities, the one which yield-ed the widest 95% confidence interval (CI)(0.80,3.64) was taken as the most conservativeestimate for inclusion in the meta-analysis.The 1979 study of Burany52 compares 250

Yugoslavian breast cancer patients with anequal number of healthy controls matched forage, ethnicity, and residence. This study isunique in the high rate of exposure, with 60%of cases and 75% of controls showing a historyof between 1 and 20 induced abortions. It isalso the only case-control study to report astatistically significantly reduced breast cancerrisk, although the difference is almost entirelyrepresented by subjects who reported one in-duced abortion. It is also noteworthy that thisstudy produced atypical risk profiles for re-productive variables, showing neither any tend-ency toward increasing risk with age at first livebirth, nor toward decreasing risk with parity.The 1981 study of 163 young white Amer-

ican breast cancer patients by Pike et al" isperhaps the best known study in the field. Infact, it has often been referred to as the firststudy to report an increased risk of breastcancer in women who had experienced inducedabortion,45 although it appeared almost a quar-ter century after the more powerful study ofSegi et al. The restriction to subjects under 33years ofage at diagnosis is unusual, but justifiedby the fact that induced abortion had onlybeen legalised in the US about the time ofthe beginning of the data collection period(1972-78). The analysis is also restricted toabortion before first full term pregnancy. Caseswere age matched both to healthy "neigh-borhood" and "friend" controls, and, although

RR calculations are presented for spontaneousand induced abortion combined, raw data aregiven, so that the crude OR may be calculatedspecifically for induced abortion. In calculatingthe crude OR, we opted to compare exposedsubjects with those who simply had no exposureto induced abortion, rather than with those whohad no abortion ofeither type, thus generating amore conservative point estimate (2.37 versus2.50). A subsequent study of this populationwith additional patients and controls," did notdifferentiate between induced and spontaneousabortion.The 1982 study of Nishiyama"6 compared

767 radical mastectomy patients from a singleprefecture in Japan with an equal number ofagematched, normal controls identified through amass breast cancer screening programme. Themedian age of patients and controls was ap-proximately 51 years.The 1983 study of Brinton et al53 involved

1362 cases and 1250 healthy control subjects,all identified between 1973 and 1977 througha mass screening programme in 28 centres inthe US. Patients and controls were racematched and age matched within five years,with a median age of approximately 53 years.Despite the large study population, however,the then very recent nature ofinduced abortionlegalisation severely limited the number of ex-posures reported. In fact, although raw datawere incompletely reported, it appears that onlyabout 20 cases had any history of inducedabortion. Curiously, although the calculationof an OR of 2.2 for abortion before first fullterm pregnancy is in close agreement with that(2.37) of Pike et al,49 Brinton et al5 char-acterised their findings as "contrary to Pike etal (1981)".Le et aF54 studied 240 French breast cancer

patients under age 46, diagnosed between 1982and 1984, with the aim of measuring the effectof oral contraceptive use. Patients werematched with hospital controls (± 2 years inage), 22% of whom had non-gynaecologicalmalignancies. Induced abortion history wastaken as one of "nine classical risk char-acteristics", considered by the authors to bepotential confounding variables for oral con-traceptive use.The 1985 paper by Hirohata et al"5 describes

the first third of a cooperative study of breastcancer among Japanese women in Japan, Jap-anese women in Hawaii, and white women inHawaii. The study was designed to examinethe role of dietary and reproductive history, butapparently, data were never collected on eitherspontaneous or induced abortion in the Ha-waiian parts ofthe study. The Japanese patientsnumbered 212, with 212 matched (for agewithin 5 years) hospital controls without canceror breast disease and 212 random neigh-borhood controls. An unusual finding of thisstudy was a null association with family historyof cancer.

In their 1988 study, Ewertz and Duffy" com-pared reproductive histories in 1486 Danishbreast cancer patients diagnosed during 1983and 1984, and who were under 70 years old(median age approximately 53 years), with

484



j.com/

J Epidem

iol Com

munity H


ownloaded from


Induced abortion - breast cancer meta-analysis

1336 control women drawn at random fromthe general population and stratified by age indecades. The only data given which specificallyrelated to induced abortion were for nulliparouswomen. Although this RR was reported as3.85 and significant, since the reference groupconsisted entirely of parous women, this figurerepresents the combined effect of nulliparityand induced abortion. Fortunately, the authorsalso calculated the RR of nulligravid women,compared with parous women. Thus, we wereat least able to calculate a corrected RR of 2.91for nulliparous women (table 2).The 1988 study of Rosenberg et ar6 is a

continuation of the 1983 study of Helmrichet al,39 which did not separate induced andspontaneous abortion data. Although this studyincludes data from four major east coast citiesand includes 3200 cases and 4844 cancer freehospital controls under 70 years of age, it isseverely weakened by a very large age differenceand a consequent cohort difference in inducedabortion exposure rates between cases and con-trols. In particular, since the subjects werecollected between 1978 and 1985, and sincethe median patient age was 52 and the mediancontrol age 40 years, the average patient in thestudy was in her 40's, but the average controlsubject was only about 30 when induced abor-tion was legalised nationwide in 1973. In fact,more than three times the number of controls(49%) as patients (16%) in the study wereunder age 40. Nevertheless, an overall RR of1.2 (with borderline significance) emergeswhen the data for all ages and parities arecombined (table 2).The 1989 study of Howe et aF7 reports data

on all 1451 women from upstate New York(including Long Island) under the age of 40who were diagnosed with breast cancer between1976 and 1980. Since this age matched, neigh-borhood control study was based entirely oncomputerised records, the possibility of recallbias was eliminated, although the possibleeffects of certain variables such as family historycould not be evaluated. Unfortunately, datapresented on abortion before first full termpregnancy did not distinguish between inducedand spontaneous abortion. A particularly note-worthy finding ofthis study is of 10 patients andno controls with a history of two consecutiveinduced abortions.The 1989 study of Harris et a158 is a com-

puterised registry study of the cohort of Swed-ish women who had induced abortions duringthe period 1966-74. Although the prospectivenature of the study precludes the existence ofresponse bias, the study nonetheless suffersfrom serious methodological weaknesses.Firstly, the incidence of breast cancer amongsubjects who had undergone induced abortionwas compared with the expected incidencefrom general population statistics. These stat-istics included the study cohort and were notadjusted for the protective effect of parity, eventhough the nulliparity rate was considerablyhigher among the general population (49%)than the study cohort (41%). Secondly, theauthors inexplicably restricted their study co-hort to those whose abortion occurred before

age 30. This had the effect ofdisproportionatelyeliminating older breast cancer patients fromthe analysis, as the authors' own comparisonof "total cohort" versus "study cohort" datashows. However, a case-control study withoverlapping authorship and most of the studypopulation in common was published byAdamiet ar9 in 1990. This latter study includes 317Swedish patients (with one age matched, non-hospital control each) and 105 Norwegian cases(with two age matched controls each) underthe age of 45 and 40 years, respectively, anddiagnosed during 1984-85. As noted above,the Swedish population is largely included inthe computerised cohort study of Harris et al,58but we have chosen the better designed, case-control study59 for inclusion in the overall ORcalculation of the meta-analysis to representthis population (category 1, table 2), althoughthe point estimates do not differ substantiallybetween the two studies (0.77 versus 0.9, re-spectively). Concerning data pertaining toabortion before first full term pregnancy, Harriset al58 reported data for women who were nul-liparous versus parous at the time of abortion,who are thus included in the meta-analysisunder category nos 2 and 3 (table 2). However,the OR for abortion before first full term preg-nancy among women parous at diagnosis isonly given in the study of Adami et al.59 Un-fortunately, the OR for this statistic given inthe paper (0.6; 95% CI: 0.3,1.5) does notinclude multiple abortions, for which the au-thors did not calculate an OR. We have there-fore recalculated the OR for one or moreabortions using the raw data given. The valuethus obtained (0.82; 95% CI: 0.44,1.51) isincluded in the meta-analysis under category2b (table 2).A continuing case-control study in the

greater Milan area of northern Italy has gen-erated a number of published reports, four ofwhich60-63 have included data specifically oninduced abortion. The most recent, a 1993report by La Vecchia et al3 is a summary ofdata on many types of cancer, including 3048breast cancer cases and 4981 cancer free, hos-pital controls. In this report, the data on breastcancer are limited to overall risk among subjectswith one or two or more induced abortions,which we have combined for category 1 in themeta-analysis (table 2). Data from the 1987,601991,61 and 199262 reports, are superseded bythose of the 1993 paper.63 The 1991 paper byParazzini et alV also reports RRs for abortionbefore first birth, but only distinguishes be-tween induced and spontaneous abortion innulliparous women (440 cases and 449 con-trols). Hence, data from this report are includedin the meta-analysis under category 2a (table2). An unusual feature of this study populationis the lack of a significant overall trend in riskwith respect to parity, with subjects with 1-3children showing raised (1.2-1.4, but not stat-istically significant) risks, and those with fouror more children, slightly (0.8) but significantlyreduced risk.The 1993 study of Moseson et al64 on 370

breast cancer patients and 783 normal controlsfrom a New York City screening clinic is un-

485



j.com/

J Epidem

iol Com

munity H


ownloaded from



usual in several important respects. Firstly, thestudy was published many years after its com-pletion, with patients diagnosed in the halfdecade of 1977-81. Since most cases and con-trols were postmenopausal, very few wouldhave been exposed to induced abortion, whichhad only been legalised in New York in 1970.Secondly, this cohort effect is compounded bysubstantial differences in age between casesand controls. The authors acknowledged thatcases were an average of 3 years older thancontrols; thus more of them were post-menopausal. More importantly, almost twiceas many controls as patients (16.1% versus8.9%, respectively) were in the 22-44 year agestratum, the only age stratum that would havehad any significant exposure to legally inducedabortion. Hence, despite adjustment of the ORfor age in the analysis, considerable un-derestimation of the overall RR for inducedabortion might be expected. This difficulty ap-pears to have been avoidable, since the tablegiving the distribution of subjects according toage indicates a large excess of controls in everystratum, and an age matched control groupcould therefore have been selected. A thirdpotential difficulty is the likelihood of partialoverlap of the study population with that ofthe study ofRosenberg et al,56 whose cases weregathered during 1978-85, and included casesfrom two large prestigious hospitals in NewYork City. Since Moseson et al'4 collected casesfrom the most prominent screening centre inNew York City, it may be assumed that someofthese cases ended up in both studies. Anotherunusual feature of this study is that the numberof induced abortions was ascertained indirectlyby subtracting the number of births and mis-carriages from the number of total pregnancies,since the authors considered induced abortionhistory "too sensitive a question", and ac-knowledged that it "may have resulted in anunderestimate of the abortion rate in the studygroup".The 1993 study by Laing et al'5 is exclusively

on African-American women, specifically, 503cases from the Washington, DC area who werediagnosed between 1978 and 1987, and 539non-cancer hospital controls matched for age(5 year age groups). Most of the cases andcontrols in the study were postmenopausal(mean age: 57.2 and 56.1 years, respectively).Both crude and adjusted (by multiple logisticregression) ORs are presented, with the latter(which are used in the present meta-analysis)limited to 405 cases and 463 controls for whomcomplete data were available. For inducedabortion, the data were reported for three agestrata, and the OR went up with age, reaching4.7 in subjects age 50 and over. In the meta-analysis (category 1) we have combined theORs given for the three age strata (table 2).The 1994 study by Laing et al"' has so far

only been published as an abstract. It also isexclusively on African-American women fromthe Washington, DC area, but the cases werediagnosed between 1989 and 1993. Only over-all ORs (category 1) obtained by conditionallogistic regression analysis are presented for the138 patients who had at least one unaffected

sister, with these sisters serving as paired con-trols. While this novel study design at leastpartially eliminates the confounding effect offamily history, it is likely that the sister controlswere generally younger than the patients, afeature which would tend to inflate the OR. Italso is likely that age was adjusted for in theanalysis, although this is not stated in the ab-stract.The 1994 study of Andrieu et al/7 focused

on the interaction of abortion and family his-tory. The study population is comprised of 495cases, 354 "friend or colleague" controls, and431 non-cancer hospital controls, all obtainedbetween 1983 and 1987 from a study on oralcontraceptive use and breast cancer. The agerange of subjects was 20-56 years, with a meanof approximately 44.5 years for patients andboth control groups, even though they hadbeen only matched to + 5 years. Of particularnote is the interaction of induced abortion andfamily history of breast cancer (mother, sister,grandmother, or aunt). Among subjects re-porting a positive family history and one in-duced abortion, an OR of 1.3 (non-significant)was calculated, which rose to a significant 7.1among subjects reporting two or more inducedabortions.The 1994 studies ofWhite et al'8 and Daling

et al' concern essentially the same white patientpopulation derived from a tumour registry inWashington state - patients aged 45 and underwho were diagnosed between 1983 and 1990.Controls were identified from the general popu-lation through random digit telephone dialing,and appear to be about 2 years younger thanthe patients, on average. The only differencein the patient population of the two studiesis that the former (n=747) was restricted toinvasive cancer, while the latter (n=845) alsoincluded 98 patients with in situ carcinoma.Both studies used the same control group.The former studyy6' was designed primarily toinvestigate the effects of oral contraceptive useon breast cancer risk, while the latter3 focusedon induced abortion. However, due to differ-ences in study design, we have elected to in-clude some of the data from each study in themeta-analysis, for the following reasons. Whiteet al'8 calculated ORs for induced abortionusing the entire patient and control populationsfor the calculation. However, Daling et al' re-stricted their calculations of all ORs concerninginduced abortion to women who were everpregnant (689 cases and 781 controls). Theeffect of thus deleting the nulligravida is toarrive at an estimate of the risk associated withinduced abortion and nulliparity combined.Not surprisingly, therefore, the OR based onwomen who were ever pregnant (1.5)' is slightlyhigher than that based on the entire studypopulation (1.36).68 Therefore, we have electedto use the more conservatively estimated dataofWhite et al"5 in the meta-analysis for category1 (table 2). Data pertaining to abortion beforeversus after first full term pregnancy (categories2, 2a, 2b, and 3; table 2) are only given byDaling et al.'The 1995 study of Brinton et al'9 is focused

on the effect of oral contraceptives on breast

486



j.com/

J Epidem

iol Com

munity H


ownloaded from



cancer risk, and shows only data for one andtwo or more induced abortions in women whowere ever pregnant. The 1648 patients (withinvasive or in situ carcinoma) and 1505 controlswere drawn from three regions of the US:Atlanta, Georgia, central New Jersey, and thesame Seattle, Washington area covered earlierby the studies of White et al68 and Daling etal3, with subject collection in the Seattle areabeginning when that of the previous studiesleft off - ie, mid-1990, and ending with theend of 1992. Patients in the Brinton et al study69also appear to be slightly older than controls.An unusual feature of this study is the ad-justment for race (white, African-American, or"other"), rather than keeping the study un-iracial or matching for race. The authors' cal-culation indicating an OR of 1.20 for African-American women is not surprising, since it isknown that breast cancer incidence is higherin premenopausal African-American womenthan in white American women. However, it isa cause for concern that ORs for other variables,such as induced abortion, are adjusted for thisdifference, since the reason for the racialdifference is unknown, and since African-American women are vastly over representedamong induced abortion patients. Thus it ispossible that adjustment for race, rather thaneliminating the effect of a confounding variable,actually nullifies the effect of the variable understudy. Another question is raised by the factthat all information on control subjects wastruncated at the time of initial screener in-terview, but the authors do not indicate thatthe period during which the controls werescreened for participation is the same as theperiod during which patients in the study werediagnosed. If these periods did not overlapprecisely, any differences would constitute anadditional source of error. It is expected thatBrinton and colleagues will publish a sequel tothis study focussing on induced abortion, atwhich time the results may be more fully evalu-ated.The 1995 study of Greek women by Lip-

worth et ar involved 820 patients, (diagnosedduring the years 1989 through 1991), 795cancer free hospital controls, and 753 "healthyvisitor" controls. Although controls werematched for age and residence, the age match-ing was crude (± 5 years), and age distributiondata were not given. Hence significant agediscrepancies between cases and controls mayexist.The 1995 study by Rookus and van

Leeuwen70 has so far only been presented asan abstract. It is a population based study of918 case-control matched (for age and region)pairs, all under age 55. Cases were diagnosedbetween 1986 and 1989 and originally gatheredfor a study on oral contraceptives and breastcancer risk. Dichotomous data are given foroverall induced abortion, as well as for thetiming of induced abortions relative to first fullterm pregnancy. As in the studies of Daling etalP and Brinton et al,69 ORs are calculatedexclusive of nulligravida - ie, for women whowere ever pregnant.

The 1996 study of American women byNewcomb et alT is a very large study, with6888 patients obtained from tumour registriesand 9529 non-hospital controls, and withsubjects (under 75 years old) drawn fromWisconsin, Massachusetts, Maine, and NewHampshire. However, most of the abortionsreported among cases and controls (97%) werespontaneous. It is also noteworthy that, in thesame manner as in the 1995 study of Brintonet al,69 information on control subjects wastruncated at the time of screener interview,with the timing ofthis collection period (relativeto diagnosis dates) not given. Any deviationsfrom the precise overlap of these periods mayresult, for example, in patients and controlsubjects with identical birth dates being con-sidered as having different ages. A unique fea-ture of this study is the establishment of anarbitrary gestation length of six months, beyondwhich all pregnancies are characterised as fullterm. While excluding late term abortion fromthe analysis would be acceptable, includingthem in a category expected to have an opposite(ie, downward) effect on risk, is questionable.Of particular concern is the exclusion of 66cases and 50 controls for whom the precisetiming of pregnancy termination was notknown with respect to the six month dividingline. Since 66 cases represent, proportionately,twice as many subjects as 50 controls, a sig-nificantly (twofold) raised risk among thesewomen is thus ignored. Although these authorsalso reported having found no statisticallysignificant differences regarding timing ofabortions relative to full term pregnancies (cat-egories 2 and 3) or number of abortions, nodata for these subgroups are given. Finally, thisstudy evidences a trend, similar to that foundby Laing et al5 of increasing risk with age atdiagnosis (divided into five 10 year age strata).Thus, the reported RR of 1.11 for womenunder 40 years of age at diagnosis rises to 2.02for women age 70 and over.

STATISTICAL METHODSDescriptive statistics were presented as ORsor RRs in all studies included in this review,except in the case of Segi et al,' from whichORs were calculated as described previously.Except for five studies which reported onlyraw ORs,'5 16 18 52 57 a multiple logistic regressionanalysis was used to arrive at an estimate ofthe OR, adjusted for age and other prognosticfactors such as parity and age at first full termpregnancy (or age at first live birth). In addition,one of the studies reporting only raw ORs alsostated that the use of the conditional binomialdistribution did not change the ORs.57 Sevenof the studies in the meta-analysis did notreport an overall dichotomous OR (category1), but rather, reported separate ORs on thebasis of single versus multiple exposures,'654 68 69differences in age at diagnosis,65 or differencesin parity.5356 For these studies, the overall di-chotomous OR and 95% CI were calculatedfor each study according to a weighted averageformula using the natural logarithm ofthe given

487



j.com/

J Epidem

iol Com

munity H


ownloaded from


Brind, Chinchilli, Severs, Suninmy-Long

Segi et al1Watanabe & Hirayama 15Dvoirin & Medvedev18Burany 52Nishiyama 16Brinton et al 53Le et al 54

Hirohata et al 51Rosenberg et al 56Howe et al 57Harris/Adami et al 58,59

La Vecchia et a/ 61,63

Moseson et al 64Laing et al 65Laing et al 66Andrieu et al 67White/Daling et al 68,3

Brinton etal69Lipworth et al 5Rookus & van Leeuwen 70

Newcomb et al 71

1957196819781979198219831984198519881989

1989-901991-93199319931994199419941995199519951996

JapanJapanRussiaYugoslaviaJapanUSAFrance

HI '

Japan HUSAUSASweden & Norway v ---

ItalyUSAUSAUSAFranceUSAUSAUSANetherlandsUSA

Weighted average (weight = 1/variance)

Unweighted average

101

0CI--

H

ICI

C

HOHi

0.3 0.5 1 2 3 4 5 6

Odds ratios

Figure I Overall odds ratios (ORs) for any induced abortion history (category 1, table 2). Point estimates and 95%CIs for each conmponent study, and for weighted and unweighted average are plotted on a logarithmic scale.

OR and the inverse of the variance.72 For thosestudies which did not use logistic regression,we calculated exact 95% CIs for the ORs viaStatXact (Cytel Software Corp., Cambridge,MA, USA).For the meta-analysis, a weighted average for

the pooled OR was obtained for each exposurecategory using the log OR and inverse of thevariance as described above. For overall, di-chotomous exposure (category 1), the un-weighted average was also calculated forcomparison (figure 1).

ResultsFigure 1 shows a semi-logarithmic plot of theoverall dichotomous ORs and 95% CIs (errorbars) for induced abortion and breast cancer(category 1) for each of the 21 independentstudies for which such data were presented(representing data published in 26 separatereports) or could be calculated (see Statistical

Table 3 Summary of pooled odds ratios for induced abortion and breast cancer

Exposure Parity at No of Pooled odds (95% confidencecategory Dx studies ratio interval)

Any Any 21 1.3 (1.2, 1.4)Before FFTP Any 7 1.4 (1.2, 1.6)Before FFTP Nulliparous 7 1.3 (1.0, 1.6)Before FFTP Parous 6 1.5 (1.2, 1.8)After FFTP Parous 6 1.3 (1.1, 1.5)

FFTP = first full term pregnancy or first live-birth.

methods). The weighted (1.3, 1.2-1.4) andunweighted (1.4, 1.3-1.6) averages, both ofwhich significantly exceed unity, are alsoshown.

Table 2 lists the ORs and CIs for each ofthe 23 independent studies (representing datapublished in 28 separate reports) included inthe meta-analysis for each category for whichdata were reported. Table 3 summarises theweighted averages and 95% CIs for each cat-egory. All of the averages significantly exceedunity.

Table 4 lists the proportion of studies foreach classification with (a) an OR greater thanunity, (b) a significantly positive OR and (c) asignificantly negative OR. For each clas-sification, a majority of the studies exhibitedan estimated OR greater than unity. Of thesignificant findings for each classification, thepositive results outnumbered the negative res-ults. In particular, there were 10 significantlypositive findings and only one significantly neg-ative finding out of 21 independent studies forcategory 1.

DiscussionINCLUSION, EXCLUSION, AND WEIGHTINGIn the present work, we have endeavored to beas conservative and as inclusive as possible,thus to avoid introducing any subjective biasesof our own through such means as the im-position of quality criteria. Hence, the single

488

I=

q

t30



j.com/

J Epidem

iol Com

munity H


ownloaded from



Table 4 Direction of association between induced abortion and breast cancer in component studies of meta-analysis

Exposure category Parity at Dx Proportion of Proportion of significantly Proportion of significantlyORs >1 (%) positive ORs (%) negative ORs (%)

Any Any 16/21 (76) 10/21 (48) 1/21 (5)Before FFTP Any 7/7 (100) 2/7 (29) 0/7 (0)Before FFTP Nulliparous 4/7 (57) 1/7 (14) 0/7 (0)Before FFTP Parous 4/6 (67) 3/6 (50) 0/6 (0)After FFTP Parous 4/6 (67) 3/6 (50) 1/6 (17)

FFTP =first full term pregnancy or first live-birth.

exclusion criterion is the absence of data re-lating specifically to induced abortion.

Nevertheless, maximal inclusion of pub-lished studies also may introduce error in twoways, namely, by repetition of data due tooverlap of study populations, and by the some-times wide variations in study quality. We haveaddressed the former problem with care toinclude only one data set for studies whereinthe overlap of study populations was nearlycomplete, as in the studies of Harris et al8 andAdami et al" and those of White et al" andDaling et al3, and wherein one study supersededanother as a more recent report of a continuingstudy, as in the studies from northern Italy.6063In the case wherein overlap of a small pro-portion of study subjects was likely - ie, in thestudies of Rosenberg et al" and Moseson et al64- we have chosen to include both as separatestudies. However, it is noteworthy that theoverall ORs of these two studies were similar(1.2 and 1.0, respectively), and the possibleerror (in the direction of underestimation ofthe overall OR) due to the partial overlap wouldnecessarily be slight. Regarding differences instudy quality, we have chosen the most widelyaccepted and objective method of weighting,namely, according to the inverse ofthe reportedvariance of the log OR. For comparison, wehave also calculated the unweighted average ofthe overall OR, and, although its CI is (notsurprisingly) somewhat wider, the point es-timate (1.4) is very close to the weightedaverage (1.3), and both are significant. Ac-knowledging, however, that no statistical for-mula could possibly account for the many largeand small differences in study design and de-scriptive statistical presentation in the variousreports, we also have opted to include a ratherdetailed narrative review of the individual in-cluded studies as well as the individual dataentered into the quantitative meta-analysis. Bythis method, we have aimed to provide thereader with as complete as possible a qualitativeas well as quantitative review of the extantliterature. Ideally, a meta-analysis would bebased on a compilation of the raw data (in-cluding data on other prognostic variables)from each subject from each component study.A logistic regression analysis could then beapplied to the master data base to get a morereliable estimate of the overall OR. With sucha database it might even be possible to performmore sophisticated statistical analyses than lo-gistic regression, such as proportional hazardsregression of the age at time of breast cancerdiagnosis.Another general limitation to the present

meta-analysis is the observational nature of

studies on abortion and breast cancer, sinceobservational studies inherently contain morebias than randomised trials. Recent discussionsin the literature7"77 have addressed the con-cerns that arise with claims of causality whenrelatively small ORs are reported, whether ina single study or in a set of studies. Given therelatively small magnitude of the cumulativeORs (table 3) we have calculated, the questionarises as to whether these are real effects orartifacts of the biases that occur within ob-servational studies. One attempt to distinguishartifact from reality is to look for consistencyacross the independent studies. Table 4 il-lustrates the clear consistency that emerges inthe present meta-analysis, with the over-whelming majority of the studies favouring apositive association.

THE "FILE DRAWER" PROBLEMIn any meta-analysis, the "file drawer" ar-gument may be invoked, particularly if themagnitude of both the individual and cumul-ative ORs (tables 2 and 3) is small. That isto say, if there is an underlying bias against thepublication of negative data, the significantlyelevated ORs generated by the present meta-analysis may be artefactual. However, sinceinduced abortion is an unusual surgical pro-cedure which is politically and legally, as wellas personally, sensitive, there is indirect evi-dence to suggest the opposite trend in bias,that is, against the publication of data whichreflect a positive association with breast cancerincidence.

It is certainly understandable that the firstobservation of increased breast cancer risk withinduced abortion should have been interpretedwith caution. For example, Segi et all backin 1957, having observed that, "The rate ofartificial interruption of pregnancy is sig-nificantly larger in all the subgroups amongthe cancer cases than the control cases", werenonetheless "rather hesitant . . . in inducingsome definite conclusions". However, it is pe-culiar that almost 40 years and over 20, mostlypositive studies later, the most recent in-vestigators should report their own, sig-nificantly positive data with extraordinaryreluctance. Witness the literal bottom line ofthe recent report ofNewcomb et al": "our datasuggest that the risk of breast cancer associatedwith any pregnancy termination is likely to besmall, if it exists at all".Perhaps the most widely known study which

reported a positive association between inducedabortion and breast cancer is that of Pike etat" in 1981, on young women in California.

489



j.com/

J Epidem

iol Com

munity H


ownloaded from



The following year, Vessey et al,'4 in their ownstudy at Oxford, called the findings of Pike etal "provocative and worrying" and offered theirown (slightly and insignificantly negative res-ults) as "entirely reassuring", even though theirstudy population contained "only a handful"of subjects with induced abortion, and wastherefore inadequate to address the issue ofinduced abortion and breast cancer risk at all.Finally, the conspicuous absence of any men-tion of induced abortion relative to breast can-cer risk in prominent medical journal reviews(eg, the New England Jrournal of Medicine6 andLancet7) may be seen against the conspicuousclaim, by the American Medical Association inits own Journal,78 as recently as December of1992, that the risk of maternal death fromchildbirth is, at "a conservative estimate" (of4.7 deaths per 100 000 live births) "nearly 12times greater than the legal abortion mortalityratio of 0.4". Lifetime breast cancer risk iscurrently estimated to be approximately 12%in the US, for example, where induced abortionis a very common exposure (approximately 1.6million per year). Thus, it is easily seen thatany demonstrable risk increase due to inducedabortion would make this elective procedurefar more risky than live birth, at least in thelong term, as the risk of immediate maternaldeath is vanishingly small for any pregnancyoutcome. Therefore, while we are aware of nospecific cases wherein positive data have beenwithheld from publication, indirect evidencesuggests that any bias against publication ofdata concerning induced abortion and breastcancer would be in the direction of keepingpositive rather than negative data "in the filedrawer".

RECALL OR RESPONSE BIASThe possibility of bias due to differential recalland/or reporting by patients versus control sub-jects merits serious consideration in any retro-spective questionnaire or interview based study.It looms still larger as a possible confoundingvariable in any association of low magnitude,particularly when such a sensitive exposure asabortion is in question.With regard to abortion and breast cancer,

the issue of recall bias was raised by Harris etal58 in 1989 as an explanation for the alreadyclear trend in worldwide data: "Most of theearlier epidemiological studies showed in-creased risk among women who had had anabortion; one reason for this could be recallbias". In particular, these authors58 postulatedthat recall bias would be in the direction ofexaggerating RR, on the supposition that, "Awomen with cancer is perhaps more likely toremember and report a previous abortion thana healthy control".A test of this hypothesis was subsequently

published by the same group in 199. 79 Inthis study, the authors compared prospective,computerised data reported in their 1989study58 with data on the same Swedish studypopulation that had been gathered by retro-spective interview for an earlier (1986)80 studyon oral contraceptives and breast cancer. As

evidence of response bias, the authors reporteda differential discordance between computerregistry based data and interview based data,specifically, that an excess of cases relative tocontrols (7 versus 1, respectively) had "overreported" induced abortions, and that an excessof controls relative to cases (16 versus 5, re-spectively) had "under reported" induced abor-tions.79 From these discrepancies, the authorscalculated that the OR for induced abortionbased on the interview data (0.95) was sig-nificantly inflated compared with that based onthe computer registry data (0.63; ratio of theORs= 1.5, 95% CI 1.1,2.1).79With regard to the issue of "over reporting"

in this study,79 we do not hesitate to concur withDaling et al,3, who commented, "we believe itis reasonable to assume that virtually no womenwho truly did not have an abortion would claimto have had one". Daling et aP went on furtherto recalculate the OR inflation reported byLindefors-Harris et aF9 with all positive reportsof induced abortion history (whether by in-terview or computer) taken as true, and theyshowed that the spurious risk increase wentdown from a significant, 50% to a non-sig-nificant 16%, attributable to the "under re-porting" among controls.Even closer scrutiny reveals the claim of

"under reporting"79 to be on no firmer groundthan that of "over reporting".79 The 1986 study,from which the interview based data weregathered,80 contained no abortion data, whichwere instead reported in the 1990 study ofAdami et al.59 There was, however, a singledifference in the study populations between the1986 and 1990 reports: The latter59 had onecontrol subject, obtained from a populationregister, for each of the 317 Swedish patients,including the 196 who were under 40 years oldat diagnosis, but the former study80 had hadan additional control selected from a fertilityregister for each of the cases under 40. Theonly segment, it turned out, of the 1991 studypopulation to show a substantial discordanceattributable to under reporting, was the controlpopulation under 40 years old, with 12 controlsubjects (compared with only four patients)reporting no induced abortions, but for whomabortions were listed on the computer registry.It is not possible to determine from the pub-lished data if most (or even all) of these underreporting control subjects were from the extra,fertility register control group80 which was notused in the 1990 study.59 However, it seemsreasonable to postulate the existence of re-sponse bias between subjects solicited from atumour registry or a population registry versusthose solicited from a fertility registry. That is,women whose names were obtained from afertility register, who are thereby identified asmothers by the interviewers, might be morereluctant to admit having had any abortionsthan women identified merely as women oras citizens. Unfortunately, the deletion of thefertility register controls80 from the 1990 reportwas not explained.59 In any event, it is clearthat the 1991 paper79 does not provide credibleevidence of response bias between cases andcontrols.

490



j.com/

J Epidem

iol Com

munity H


ownloaded from



Daling et al3 also offered, as further evidenceagainst the response bias argument, their ownfinding of a null association of cervical cancerand induced abortion among 214 cases and321 controls gathered and interviewed in thesame manner as those in their breast cancerstudy. Nevertheless, despite the compellingcase made by Daling et aP3 against a responsebias interpretation of their own data, Ro-senberg, in her accompanying editorial,4 main-tained that "the possibility of reporting bias"was "a major concern" in the study, with noacknowledgment whatsoever that Daling et alhad indeed addressed the issue.

Outside of the study of Harris et al,58 theonly other computer registry based study ofinduced abortion and breast cancer is that ofHowe et al,57 which contains direct evidenceagainst the response bias hypothesis. In in-terpreting their finding of significantly raisedrisk (table 2), these authors noted "under re-porting" and "inconsistent reporting" on thefetal death certificates. However, they foundno evidence of bias, with instances of suchmisreporting having "occurred similarly amongthe cases and the controls".57More recently, Lipworth et al5 suggested that

their own study on women in Greece, with its"permissive social environment with respect toinduced abortion", might therefore "provide auseful complementary insight" in order to testhypotheses that "have been invoked to explain,in noncausal terms, the reported associationbetween induced abortion and breast cancer".They concluded that their own data (in ex-cellent agreement with those of Daling et a13and Howe et al57; table 2) did not result fromresponse bias, since "healthy women in Greecereport reliably their history of induced abor-tion" .

Despite the overwhelming evidence that theassociation ofinduced abortion and breast can-cer does not result from reporting bias, theresponse bias argument continues to be ad-vanced with vigour. Recently, Rookus and vanLeeuwen70 attributed their significantly higherOR obtained from a more rural and tra-ditionally religious region of The Netherlands(compared with a highly urbanised region) to"differential misclassification bias". Surely thisis but one of many possible explanations fordifferent results between two regions with sub-stantial differences in many variables, includingethnicity and a host of lifestyle factors. It isalso noteworthy that the exposure rates forboth regions are very low, and that both showpositive overall associations between inducedabortion and breast cancer. Also recently, New-comb et al' have claimed that their results"suggest a bias in reporting", since the RR isslightly higher among American women withinduced abortions before versus after legal-isation in the US in 1973 (1.35, 95% CI 1.01,1.80 versus 1.12; 0.84,1.49, respectively).However, their data speak for themselves: Eachof these point estimates falls well within theother's CI, thus providing no support for thesuggestion of reporting bias. On the contrary,the almost significant (p = 0.09) trend they re-port for RR to increase with age at diagnosis

is continuous, and it shows up even when onlypost-1973 abortions are included.7'

SPECIFIC EFFECT OF INDUCED ABORTION VERSUSDELAYED CHILDBIRTHA crucial consideration in the assessment of thereal magnitude ofbreast cancer risk attributablespecifically to induced abortion is the ability todistinguish this from the known increased riskattributable to a delay in the first full termpregnancy by any means.8' From the point ofview of women considering abortion, parouswomen would be subject only to the in-dependent effect of induced abortion, whereasnulliparous women (about half of Americanabortion clients), would be subject to both riskenhancing effects of the abortion, dependingon their age at time of abortion and if andwhen they subsequently have any children.From the point of view of breast cancer

aetiology, delay of first full term pregnancy isone of only two risk factors (the other beingionising radiation) known to influence primarycarcinogenesis. Presumably, delaying the firstcomplete pregnancy increases the time periodduring which undifferentiated breast tissue canaccumulate potentially tumourigenic mut-ations. Induced abortion, however, may in-dependently increase risk via the tumourpromoting effect of the considerably raised oe-stradiol concentrations of early pregnancy,while denying a woman the benefit of thedifferentiating effect of the hormonal milieu oflate pregnancy. This differentiating effect ispresumably the mechanism by which an early,completed pregnancy confers permanent pro-tection against breast cancer.28'82 In addition,induced abortion may enhance the oestrogenmediated proliferation of normal but primitivecells, resulting in the presence of more cellswhich are vulnerable to subsequent primarycarcinogenesis.From the point of view of epidemiology, the

differential effects of delaying the first full termpregnancy and artificially terminating a preg-nancy in progress have been resolved in twoways. Firstly, by assessing the risk of breastcancer specifically in populations ofnulliparouswomen, the specific effect of induced abortioncan be measured, providing the controls in-clude the nulligravida. In the present meta-analysis, only seven studies assessed risk innulliparous women. 3 5 53 55 56 6 70 Six of the sevenused nulliparous women (all but one3 includingthe nulligravida) as controls. Only Ewertz andDuffy55 used only parous women as controls,but since they also provided data on the riskof nulliparity per se, we were able to subtractout this effect in order to arrive at the net RRattributable specifically to induced abortion innulliparous women in their study (table 2).The resulting pooled OR in the meta-analysis(category 2a, table 3) is the same as that of theoverall OR (category 1).The second method of arriving at the specific

overall effect (ie, in parous and nulliparouswomen; category 1) of induced abortion is toinclude a term in the calculation of the OR forthe effect of age at first live birth or first full

491



j.com/

J Epidem

iol Com

munity H


ownloaded from



term pregnancy. Thirteen of the 21 studies inwhich an overall OR was reported (table 2,category 1) calculated the OR by multiple lo-gistic regression."55 53 54 56 59 63 64 67 69-71 Two stud-ies which did not5768 reported that it made nodifference, five' 15161852 studies only reportedcrude ORs, and one65 had insufficient dataavailable on age at first full term pregnancy.Thus there are 15 studies for which the overalleffect of induced abortion has been measuredwith the possible confounding effect of age atfull term pregnancy in parous women ac-counted for. Recalculation of the pooled ORusing only these studies slightly reduces thepooled OR (to 1.2), which is still significant(95% CI 1.1,1.3).The same is true for other potential con-

founding variables for which terms were gen-erally included in the multivariate analyses -namely, parity, age at menarche, oral con-traceptive use, and some measure of socio-economic status (usually, educational level).A few studies also adjusted for other factorssuspected of influencing breast cancer risk,such as alcohol and fat consumption, althoughnone of these studies reported any significanteffects of these variables.

INDUCED ABORTION BEFORE VERSUS AFTER FIRSTFULL TERM PREGNANCYWith regard to the question ofinduced abortionbefore versus after first full term pregnancy inparous women, (categories 2b versus 3; table3), the aggregate OR is slightly but not sig-nificantly higher for the former (1.5 versus 1.3).Since only six studies addressed both thesequestions3 5 53 56 59 70 and reported adequate datafor the meta-analysis, a firm conclusion cannotbe drawn at this time. However, if furtherresearch verifies this trend, it would provideevidence that induced abortion specifically in-creases breast cancer risk both by amplificationof previously transformed, potentially can-cerous cells, and of the number of normal, butprimitive, cells (much more numerous beforefirst full term pregnancy) vulnerable to sub-sequent mutagenesis.

It is also noteworthy that four5535970 of thesix studies so far published reported higherORs for induced abortion for before versusafter first full term pregnancy, and the findingsof the two which did not are explicable interms consistent with the presently proposedmechanisms. Specifically, Daling et a!5 alsofound no influence of age at first full termpregnancy as an independent risk factor. Thus,their finding of no added effect of abortionbefore versus after first full term pregnancysupports the concept that both of these typesof exposure (ie, delayed first full term preg-nancy and induced abortion before versus afterfirst full term pregnancy) operate via the samemechanism (ie, by increasing the opportunityfor primary carcinogenesis in normal but prim-itive cells). In contrast, Rosenberg et al6 founda higher OR for abortion after compared withbefore first full term pregnancy (table 2). How-ever, this can be ascribed to the very largecohort effect in their study. Specifically, since

the average patient was over age 40 while theaverage control was only about 30 when in-duced abortion was legalised in the US, thepotential exposure of patients to induced abor-tion before first full term pregnancy was un-doubtedly much lower than that of controlsubjects.

Unfortunately, the finding of no differentialrisk increase for abortion before comparedwith after the first full term pregnancy in onestudy3 has been interpreted as conflicting withprevious animal data. Specifically, Daling et a3refer to this finding in their own study as being,"not completely in accord with the results inexperimental animals". Rosenberg4 called thesame finding "a striking inconsistency with themodel". Such conclusions are unwarranted,since the animal model to which these authorsreferred2 did not include testing the effect ofinduced abortion after full term pregnancy.Rather, Russo and Russo2 compared breastcancer incidence in rats whose first pregnancywas aborted (via hysterectomy) before exposureto the chemical carcinogen, dimethylbenz-anthracene, with that of rats who carried thepregnancy to term and to that of rats whonever mated. The aborted group had a highmammary tumour incidence rate (78%), asdid the virgin rats (71%) compared with themarked protective effect of carrying the firstpregnancy to term (6% tumour incidence).Furthermore, histological examination ofbreast tissue from these animals revealed in-complete differentiation of primitive structuresin the virgin and aborted rats, compared withthose allowed to bear pups. While these findingsprovide excellent experimental evidence of themechanism responsible for the protective effectofearly first full term pregnancy (the abrogationofwhich is one way induced abortion increasesbreast cancer risk), this animal model system(wherein abortion precedes carcinogen ex-posure and wherein the effect of abortion onparous animals is not measured) does not fullyaddress the question of the independent effectof induced abortion, which we largely ascribeto the oestradiol mediated promotion of thegrowth of previously transformed cells.

EFFECT OF INDUCED VERSUS SPONTANEOUSABORTIONWhatever the details of the mechanism(s) bywhich induced abortion may independently in-crease the breast cancer risk, the fact that thefirst trimester of pregnancy is characterised byhigh levels of ovarian oestradiol makes this riskfactor consistent with most others (eg, earlymenarche, late menopause, postmenopausalobesity), which are also associated with someform of oestrogen excess. However, the overalllack of association found with spontaneousabortion raises the important question of whyany early termination of a pregnancy, whethernatural or artificial, does not have the sameeffect. Various hypotheses have been offered toexplain this apparent paradox, ranging from"the inherent difficulty in detecting" spon-taneous abortion5 to the possibility that "therelatively short gestational length" of spon-

492



j.com/

J Epidem

iol Com

munity H


ownloaded from



taneously aborted pregnancies might makethem less likely to raise breast cancer risk.3Lipworth et aP have even suggested that thisdiscrepancy might provide a reason to dismissthe association of induced abortion and breastcancer altogether, as an artifact "generated bysubtle information bias".However, consideration ofthe endocrinology

ofnormal compared with threatened early preg-nancy provides a straightforward explanation:The first trimester of most pregnancies whichend in miscarriage is characterised by sub-normal oestradiol secretion. As early as 1976,Kunz and Keller83 found subnormal maternaloestradiol to be the most reliable predictor offirst trimester miscarriage. In their 1990 studyof 221 pregnancies, Witt et al84 observed thatmaternal oestradiol in women with apparentlynormal pregnancies of 11 weeks' gestation orless (from last menstrual period) averaged onethird lower in pregnancies that ended in afirst trimester miscarriage. More strikingly, theyobserved that in pregnant women with threat-ening symptoms (significant vaginal bleeding),oestradiol averaged only one sixth the averagenormal pregnancy level in pregnancies whichwent on to miscarry in the first trimester.84Recently, Stewart et al85 performed daily lon-gitudinal hormone measurements on 24 normalwomen of proven fertility. They detected stat-istically significantly higher maternal oestradiollevels within six days after the luteinizing hor-mone peak in conceptive cycles (n = 14) thatresulted in viable pregnancy. In contrast, con-ceptive cycles that ended in spontaneous abor-tion (n= 9) showed a subnormal oestradiol risethat did not significantly exceed non-con-ceptive levels until the 10th day after the peak,by which time oestradiol begins to decline in anon-pregnant cycle.

EFFECT OF SINGLE VERSUS MULTIPLE INDUCEDABORTIONSTen of the studies in the present meta-analysispresent overall ORs for two or more inducedabortions5 16525456596367-69. However, these 10studies represent a subset in which the overallOR for one or more induced abortions is lower(1.1; 95% CI: 1.0,1.3) than that obtained forall 21 studies providing this statistic (1.3; 1.2,1.4; table 2, category 1). Thus, the finding thatthe OR for two or more abortions is identicalto that calculated for one or more ([.1; 1.0,1.3) may not be representative. In fact, sevenofthe 10 studies reporting the multiple abortionOR report slightly (though not significantly)higher ORs for two or more, as opposed toone abortion. 5 16 52 54 59 68 69 The extant data aretherefore insufficient to draw any firm con-clusions about any overall dose effect of in-duced abortion at the present time.A particularly important reason to refrain

from dismissing the apparent lack of a doseeffect of induced abortion is given in the studyof Howe et al,57 whose multiple abortion dataset was excluded from the calculations abovebecause it appears to describe only a specialcase of multiple abortion - ie, two consecutive

induced abortions with no live birth in-tervening. Since this history pertained to 10cases and no controls (out of 1451 matchedpairs), the OR could not be calculated. If theprincipal mechanism of risk elevation by in-duced abortion is the oestrogenic growth pro-motion of existing abnormal cells or cloneswhich would otherwise be eliminated (or atleast inhibited) by the completion of the preg-nancy, then one would predict a much greaterdose effect if two (or more) artificially in-terrupted pregnancies followed consecutively.Thus, it would be particularly useful if theprospective data base used by Howe et al,57which has been growing since 1980 (when thatstudy was terminated), were followed up toverify this trend.

EFFECT OF GESTATIONAL AGEMost studies did not specify the gestational ageof the fetus at the time of induced abortion,with the exception of the studies of Pike et al49(<12 weeks), Howe et aF7 (20 weeks or less),Daling et al' and Rookus and van Leeuwen70(1-8 weeks and 9-12 weeks, calculated sep-arately). However, since the overwhelming ma-jority of induced abortions occur in the firsttrimester, and almost all the rest in the secondtrimester (which would still be expected toincrease risk, as reported by Daling et al'), it ishighly unlikely that the overall results reportedwould be materially affected by third trimesterabortions. Indeed, Howe et al7 found thatinclusion of third trimester abortions did notaffect the results in their study. Of the twostudies which divided the analysis according toearly and late first trimester abortions, one3found the later abortions (9-12 weeks) to beassociated with a slightly (but insignificantly)higher OR (1.9; 95% CI 1.3,2.9) than theearlier abortions (1-8 weeks: OR= 1.4; 1.0,1.8), and the other study70 found the reverse(1-8 weeks: OR= 2.1; 1.1,4.2; >8 weeks: OR=1.6; 0.8,3.5). Thus, there is no reason to sus-pect that new technologies, (such as mi-fepristone/misoprostol) that would result ingenerally earlier terminations, would not alsobe associated with increased breast cancer risk.

EFFECT OF AGE AT FIRST (OR ONLY) ABORTIONTwo studies have examined this question,namely, those of Rosenberg et at56 and Dalinget al.3 The former study only considered nul-liparous women in this regard, and reportedno significant differences in risk, with adjustedORs ranging from 1.0 to 1.5 over the age rangeof under 20 years to 30 years and over. Dalinget al' also reported no significant differences,but they noted a trend toward increased riskin women with first induced abortion under 18years old and over 29 years old, which theycorrelated with the histological data from thehuman biopsy study of Russo et al.82 In notingthat the rate of cell proliferation is likely to behighest in the youngest subjects, Daling et al'have prudently suggested that the greater eleva-tion in risk for women under 18 at the time of

493



j.com/

J Epidem

iol Com

munity H


ownloaded from



their first (or only) abortion may be real andshould be further investigated.

IMPORTANCE OF AGE AT DIAGNOSISBecause the incidence of breast cancer in thewestern world rises with age throughout thelifespan86 (whereas in Japan, incidence levels offin the fifth decade and actually drops somewhatlater on86), the range of ages at diagnosis inany epidemiological study is the most crucialdeterminant of the practical significance, interms of excess cases expected, of the RR de-termined in the study. Many of the studiespublished thus far have been restricted toyounger women - ie, under age 33,49 40,45,354596869 or 57,67 generally because onlyyounger women would have been exposed tolegalised abortion. As we have already dis-cussed in the present report, most studies thatinclude older women are not only weakenedby the lack of exposure of the older women,but also by the cohort effect of having thecontrols younger than the patients, which tendsartificially to lower the calculated ORs. In broadterms, even if the overall weighted pooled ORof 1.3 (±0.1) were to be applicable only towomen up to age 50, in whom the incidenceof breast cancer is about 2%, and this 30%odds increase were to be applied only to theapproximately 800 000 patients having theirfirst induced abortion each year in the US, forexample, the calculated excess incidence ofbreast cancer would be 4700 (± 1600) casesper year in the US. As abortion has been legalin the US for up to a quarter century, an excessincidence of this magnitude should already beoccurring. Since over 30 000 cases are alreadydiagnosed in women under age 50 each year,an excess incidence of 4700 might well escapeour notice.

Yet, as significant a public health tragedy asthis figure suggests, there is reason to believethat it may seriously underestimate the mag-nitude of the present and future problem. Forexample, the recent study of Rookus and vanLeeuwen70 reports a significant, overall OR inpatients under age 55 diagnosed between 1986and 1989, of 1.9, which is identical to thatreported by Howe et al7 for patients under age40 a decade earlier. Another recent example isthe study of Lipworth et al,5 which (althoughthe age distribution of subjects was not given)had no age restriction and which reported anoverall OR of 1.5 1 forwomen in Greece, where,according to the authors, "Even before theirlegalisation, induced abortions were practicedwith widespread social acceptance". In the US,the recent study of Laing et at65 on African-American women had no age restriction onpatients, and 62% of those patients with ahistory of induced abortion were age 50 andover. The results are particularly troubling sincethe OR was found to increase with age, up to4.7 in the 50 and over stratum. Newcomb etalf' reported a similar trend of increasing riskwith age at diagnosis.Thus, the available evidence so far suggests

that the 30% (± 10%) increased risk calculatedin the present meta-analysis will probably apply,

at a minimum, to incidence rates at advancedages, where such rates are much higher. At acurrently estimated lifetime risk in US womenof 12%, the 800 000 first abortions performedeach year would thus generate 24 500 (± 7800)excess cases each year, once the first cohortexposed to legal abortion reaches their ninthdecade, in the fourth decade ofthe 21 st century.Furthermore, it is worthy ofemphasis that eventhis forbidding figure does not reflect the non-specific effect of induced abortion in delayingfirst full term pregnancy, which has been dis-cussed in the present review, but was explicitlyeliminated from the quantitative meta-analysis.This effect would apply variably to the ap-proximately 800 000 first abortion patientseach year, and it could raise the estimate ofexcess breast cancer incidence which may beattributable to induced abortion considerably.

EFFECTS OF INTERACTION WITH OTHERVARIABLESA few investigators have begun to explore thepossible interaction of induced abortion withat least one risk factor other than age at firstfull term pregnancy, namely, family history.Thus, Parazzini et al62 found no interaction atall, although their numbers were small, andthey also found, contrary to most other reports,no overall effect of reproductive risk factors inwomen with positive family history, reportingno "strong or significant effect of the bestrecognized factors for breast cancer risk, andseveral of the observed trends were in the op-posite direction". Only two other studies ad-dressed the interaction of family history andinduced abortion. Andrieu et at67 calculated anOR of 7.1 for women with two or more inducedabortions and a family history, but the numberof subjects (nine patients and four controls)was very low. Daling et aP found only a slightlyhigher OR for women with a positive first- orsecond degree family history (1.8 versus 1.5overall), but they found much stronger as-sociations when they also figured in the effectof age at first (or only) induced abortion. Thus,the OR went up to 3.7 for women whose firstinduced abortion was over age 30 (14 casesand 3 controls), and it was incalculable forwomen whose first abortion was under age 18,since such family history and induced abortionhistory applied to 12 cases and no controls.

ConclusionsWe believe that the present review and meta-analysis summarises a literature that documentsa remarkably consistent, significant positive as-sociation between induced abortion and breastcancer incidence, independent of the effect aninduced abortion has in delaying first full termpregnancy. Moreover, the increased risk is seenin both prospective and retrospective studiesfrom around the world, in populations with thewidest imaginable differences in ethnicity, diet,socioeconomic and lifestyle factors and socialmorays, and which also differ widely in sizeand in many aspects of design, and whose dataextend over more than half a century in time.

494



j.com/

J Epidem

iol Com

munity H


ownloaded from



We are convinced that such a broad base ofstatistical agreement rules out any reasonablepossibility that the association is the result ofbias or any other confounding variable. Fur-thermore, this consistent statistical associationis fully compatible with existing knowledgeof human biology, oncology and reproductiveendocrinology, and supported by a coherent(albeit incomplete) body of laboratory data as

well as epidemiological data on other risk fac-tors involving oestrogen excess, all of whichtogether point to a plausible and likely mech-anism by which the surging oestradiol of thefirst trimester of any normal pregnancy, if it isaborted, may add significantly to a woman'sbreast cancer risk.

Finally, it should be acknowledged that in-duced abortion is the most common electivesurgical procedure currently performed in theUS. While other elective, risk enhancing mat-ters of choice, such as cigarette smoking, re-

quire thousands of exposures to producedetectable increases in cancer incidence, theinduced abortion patient's risk of breast cancer

later in life is measurably increased after a singleexposure. Therefore, while the need for furtherresearch cannot be denied, especially giventhe existence of prospective data57 that can bestudied with minimal cost, there exists the morepresent need for those in clinical practice to

inform their patients fully about what is alreadyknown.

The authors wish to thank Alicia Fisher for running the com-

puter analyses, and Cheryl Gates for her assistance in preparingthe manuscript. This work was supported in part by a LegislativeInitiative Grant, Contract nol175169, from the Department ofEducation of the Commonwealth of Pennsylvania JS-L), andby a Fellowship Leave Award (JB) from Baruch College of theCity University of New York.

1 Segi M, Fukushima I, Fujisaku S, et al. An epidemiologicalstudy on cancer in Japan. GANN 1957;48(Suppl):1-63.

2 Russo J, Russo IH. Susceptibility of the mammary gland to

carcinogenesis. Am 7 Pathol 1980;100:497-512.3 Daling JR, Malone KE, Voigt LF, White E, Weiss NS. Risk

of breast cancer among young women: relationship toinduced abortion. .7 Natl Cancer Inst 1994;86:1584-92.

4 Rosenberg L. Induced abortion and breast cancer: more

scientific data are needed (editorial). _7 Natl Cancer Inst1994;86: 1569-70.

5 Lipworth L, Katsouyanni K, Ekbom A, Michels KB, Tri-chopoulos D. Abortion and the risk of breast cancer: a

case-control study in Greece. Int3 Cancer 1995;61:181-4.6 Harris JR, Lippman ME, Veronesi U, Willett W. Breast

cancer. (first of three parts). New Engl Med 1992;327:319-28.

7 Hulka BS, Stark AT. Breast cancer: cause and prevention.Lancet 1995;346:883-7.

8 Kvale G. Reproductive factors in breast cancer epi-demiology. Acta Oncol 1992;31:187-94.

9 Michels KB, Hsieh C, Trichopoulos D, Willett W. Abortionand breast cancer risk in seven countries. Cancer CauseControl 1995;6:75-82.

10 Kelsey JL, Gammon MD, John EM. Reproductive factorsand breast cancer. Epidemiol Rev 1993;15:36-47.

11 Harlap S. Oral contraceptives and breast cancer: cause andeffect? 7 Reprod Med 199 1;36:374-95.

12 Remennick LI. Induced abortion as cancer risk factor: a

review of epidemiological evidence.3Epidemiol CommunityHealth 1990;44:259-64.

13 Hadjimichael OC, Boyle CA, Meigs JW. Abortion beforefirst live birth and risk of breast cancer. Br3r Cancer 1986;53:281-4.

14 Vessey MP, McPherson K, Yeates D, Doll R. Oral con-

traceptive use and abortion before first term pregnancy inrelation to breast cancer risk. BrJ7 Cancer 1982;45:327-3 1.

15 Watanabe H, Hirayama T. Epidemiology and clinical aspectsof breast cancer. Nippon Rinsho 1968;26: 1853-9 (in Jap-anese).

16 Nishiyama F. The epidemiology of breast cancer in To-kushima prefecture. Shikoku Ichi 1982;38:333-43 (in Jap-anese).

17 Gandra L, Barros H, Moreira A, Calheiros F, MagalhaesN. Risk factors for breast cancer. Acta Med Portuguesa1993;6:129-33 (in Portuguese).

18 Dvoirin VV, Medvedev AB. Role of women's reproductivestatus in the development of breast cancer. In: Methods

and progress in breast cancer epidemiology research, Tallin,1978. Moscow: Oncology Science Center of the USSRAcademy of Sciences, 1978;53-63 (in Russian).

19 Levshin VF, Chepurko AD. Reproductive history and breastcancer. Soviet Medicine 1986;6: 15-21 (in Russian).

20 Wynder EL, Bross IJ, Hirayama T. A study of the epi-demiology of cancer of the breast. Cancer 1960;13:559-601.

21 Stewart HL, Dunham LJ, Casper J, et al. Epidemiology ofcancers of uterine cervix and corpus, breast and ovary inIsrael and New York City. Jf Natl Cancer Inst 1966;37:1-95.

22 Valaoras VG, MacMahon B, Trichopoulos D, Po-lychronopoulou A. Lactation and reproductive historiesof breast cancer patients in greater Athens, 1965-67. Int_7 Cancer 1969;4:350-63.

23 Salber EJ, Trichopoulos D, MacMahon B. Lactation andreproductive histories of breast cancer patients in Boston,1965-66. Jf Natl Cancer Inst 1969;43:1013-24.

24 Lowe CR, MacMahon B. Breast cancer and reproductivehistory of women in South Wales. Lancet 1970;1(639):153-7.

25 Yuasa S, MacMahon B. Lactation and reproductive historiesof breast cancer patients in Tokyo, Japan. Bull World HealthOrgan 1970;42: 195-204.

26 Mirra AP, Cole P, MacMahon B. Breast cancer in an areaof high parity: Sao Paulo, Brazil. Cancer Res 1971;31:77-83.

27 Ravnihar B, MacMahon B, Lindtner J. Epidemiologic fea-tures of breast cancer in Slovenia, 1965-1967. Eur _7Cancer 1971;7:295-306.

28 Lin TM, Chen KP, MacMahon B. Epidemiologic char-acteristics of cancer of the breast in Taiwan. Cancer 1971;27:1497-504.

29 Paymaster JC, Gangadharan P. Some observations on theepidemiology of cancer of the breast in women of westernIndia. IntrJ Cancer 1972;10:443-50.

30 Stavraky K, Emmons S. Breast cancer in premenopausaland postmenopausal women. _7 Natl Cancer Inst 1974;53:647-54.

31 Abeatici S, Raso ANL Merlo G, Lepora P, Dorigo M, MussaA. Computerized study of biological and environmentalcharacteristics of patients with breast cancer. Minerva Med1975;66:2295-305 (in Italian).

32 Herity BA, O'Halloran MJ, Bourke GJ, Wilson-Davis K. Astudy of breast cancer in Irish women. Br.7 Prev Soc Med1975;29: 178-81.

33 Soini I. Risk factors of breast cancer in Finland. Int .7Epidemiol 1977;6:365-73.

34 Choi NW, Howe GR, Miller AB, et al. An epidemiologicstudy of breast cancer. Am _7 Epidemiol 1978;107:510-21.

35 Toti A, Piffanelli A, Pavanelli T, et al. Possible indication ofbreast cancer risk through discriminant functions. Cancer1980;46: 1280-5.

36 Paffenbarger RS Jr, Kampert JB, Chang HG. Characteristicsthat predict risk of breast cancer before and after themenopause. Am _7 Epidemiol 1980;112:258-68.

37 Kelsey JL, Fischer DB, Holford TR, et al. Exogenous es-trogens and other factors in the epidemiology of breastcancer. _7 Natl Cancer Inst 198 1;67:327-33.

38 Lubin JH, Burns PE, Blot WJ, et al. Risk factors for breastcancer in women in northern Alberta, Canada, as relatedto age at diagnosis. _7 Natl Cancer Inst 1982;68:211-17.

39 Helmrich SP, Shapiro S, Rosenberg L, et al. Risk factorsfor breast cancer. Am .7 Epidemiol 1983;117:35-45.

40 Enachescu D, Lemneanu S. Associated mammary neoplasmrisk factors considered by means of retrospective ep-idemiologic investigations. Sante Publique 1984;27:225-31.

41 Talamini R, La Vecchia C, Franceschi S, et al. Reproductiveand hormonal factors and breast cancer in a northernItalian population. Intc Epidemiol 1985;14:70-4.

42 Kvale G, Heuch I, Eide GE. A prospective study of re-productive factors and breast cancer. Am_7Epidemiol 1987;126:831-41.

43 Yuan J-M Yu MC, Ross RK, Gao Y-T, Henderson BE. Riskfactors for breast cancer in Chinese women in Shanghai.Cancer Res 1988;48:1949-53.

44 Bernstein L, Pike MC, Krailo M, et al. Update of the LosAngeles study of oral contraceptives and breast cancer:1981 and 1983. In: Mann RD, ed. Oral contraceptives andbreast cancer The implications of the present findings forinformed consent and informed choice. Park Ridge, NJ: Pan-theon, 1990;169-81.

45 Sellers TA, Potter JD, Severson RK, et al. Difficulty be-coming pregnant and family history as interactive riskfactors for postmenopausal breast cancer: the Iowa wo-men's health study. Cancer Cause Control 1993;4:21-8.

46 Andrieu N, Clavel F, Auquier A, et al. Variations in the riskof breast cancer associated with a family history of breastcancer according to age at onset and reproductive factors..7 Clin Epidemiol 1993;46:973-80.

47 Rao DN, Ganesh B, Desai PB. Role of reproductive factorsin breast cancer in a low-risk area: a case-control study.BrJ Cancer 1994;70:129-32.

48 Andrieu N, Demenais F. Role of genetic and reproductivefactors in breast cancer. Genet Epidemiol 1994;11:A285.

49 Pike MC, Henderson BE, Casagrande JT, Rosario I, GrayGE. Oral contraceptive use and early abortion as riskfactors for breast cancer in young women. Br _7 Cancer1981;43:72-6.

50 Mantel N, Haenszel W. Statistical aspects of the analysis ofdata from retrospective studies of disease. _7 Natl CancerInst 1959;22:719-48.

495



j.com/

J Epidem

iol Com

munity H


ownloaded from



51 Hirohata T, Shigematsu T, Nomura AMY, Nomura Y, HorieA, Hirohata I. Occurrence of breast cancer in relation todiet and reproductive history: A case-control study inFukuoka, Japan. Natl Cancer Inst Monogr 1985;69: 187-90.

52 Burany B. Gestational characteristics in women with breastcancer. J7ugosl Ginekol Opstet 1979;19:237-47 (in Serbo-Croatian).

53 Brinton LA, Hoover R, FraumeniJFJr. Reproductive factorsin the aetiology of breast cancer. Br J Cancer 1983;47:757-62.

54 Le M-G, Bachelot A, Doyon F, Kramar A, Hill C. Oralcontraceptive use and breast or cervical cancer: Pre-liminary results of a French case-control study. In: WolffJ-P, Scott JS, eds. Hormones and sexual factors in humancancer aetiology. Amsterdam: Elsevier, 1984:139-47.

55 Ewertz M, Duffy SW. Risk of breast cancer in relation toreproductive factors in Denmark. Br J Cancer 1988;58:99-104.

56 Rosenberg L, Palmer JR, Kaufman DW, Strom BL, Sch-ottenfeld D, Shapiro S. Breast cancer in relation to theoccurrence and time ofinduced and spontaneous abortion.Am J Epidemiol 1988;127:981-9.

57 Howe HL, Senie RT, Bzduch H, Herzfeld P. Early abortionand breast cancer risk among women under age 40. Int JEpidemiol 1989;18:300-4.

58 Harris B-M L, Eklund G, Meirik 0, Rutqvist LE, WiklundK. Risk of cancer of the breast after legal abortion duringfirst trimester: a Swedish register study. BMJ 1989;299:1430-2.

59 Adami H-O, Bergstrom R, Lund E, Meirik 0. Absence ofassociation between reproductive variables and the risk ofbreast cancer in young women in Sweden and Norway.BrJI Cancer 1990;62:122-6.

60 La Vecchia C, Decarli A, Parazzini F, et al. General epi-demiology ofbreast cancer in northern Italy. IntjEpidemiol1987;16:347-55.

61 Parazzini F, La Vecchia C, Negri E. Spontaneous andinduced abortions and risk of breast cancer. Int J Cancer1991;48:816-20.

62 Parazzini F, La Vecchia C, Negri E, Franceschi S, BoccioloneL. Menstrual and reproductive factors and breast cancerin women with family history of the disease. Int J Cancer1992;51:677-81.

63 La Vecchia C, Negri E, Franceschi S, Parazzini F. Long-term impact of reproductive factors on cancer risk. IntCancer 1993;53:215-9.

64 Moseson M, Koenig KL, Shore RE, Pasternack BS. Theinfluence of medical conditions associated with hormoneson the risk of breast cancer. Int J Epidemiol 1993;22:1000-9.

65 Laing AE, Demenais FM, Williams R, Kissling G, ChenVW, Bonney GE. Breast cancer risk factors in African-American women: The Howard University Tumor Re-gistry experience. J Natl Med Assoc 1993;85:931-9.

66 Laing AE, Bonney GE, Adams-Campbell L, et al. Re-productive and lifestyle factors for breast cancer in African-American women. Genet Epidemiol 1994;11:A300.

67 Andrieu N, Clavel F, Gairard B, et al. Familial risk of breastcancer and abortion. Cancer Detect Prev 1994;18:51-5.

68 White E, Malone KE, Weiss NS, Daling JR. Breast canceramong young US women in relation to oral contraceptiveuse. J Natl Cancer Inst 1994;86:505-14.

69 Brinton LA, Daling JR, Liff JM, et al. Oral contraceptivesand breast cancer risk among younger women. J NatlCancer Inst 1995;87:827-35.

70 Rookus MA, van Leeuwen FE. Breast cancer risk after,induced abortion, a Dutch case-control study. Am J Ep-idemiol 1995;141:S54 (abstract 214).

71 Newcomb PA, Storer BE, Longnecker MP, Mittendorf R,Greenberg ER, Willett WC. Pregnancy termination inrelation to risk of breast cancer. JAMA 1996;275:283-7.

72 Schmid JE, Koch GG, La Vange LM. An overview ofstatistical issues and methods of meta-analysis. J BiopharmStat 1991;1:103-20.

73 Shapiro S. Meta-analysis/schmeta-analysis. Am J Epidemiol1995;140:771-8.

74 Petitti DB. Of babies and bathwater. Am J Epidemiol 1995;140:779-82.

75 Greenland S. Can meta-analysis be salvaged?AmjEpidemiol1995;140:783-7.

76 Shapiro S. Is there or ain't there no baby? Am J Epidemiol1995;140:788-91.

77 Taubes G. Epidemiology faces its limits. Science 1995;269:164-9.

78 Gans JE, Coble YD, Estes EH Jr, et al (Council on ScientificAffairs), American Medical Association Council Report:Induced termination of pregnancy before and after Roe vWade. JAMA 1992;268:3231-9.

79 Lindefors-Harris B-M, Eklund G, Adami H-O, Meirik 0.Response bias in a case-control study: Analysis utilizingcomparative data concerning legal abortions from twoindependent Swedish studies. Am J Epidemiol 199 1;134:1003-8.

80 Meirik 0, Lund E, Adami H-O, Bergstrom R, ChristoffersenT, Bergsjo P. Oral contraceptive use and breast cancer inyoung women. Lancet 1986;ii:650-4.

81 MacMahon B, Cole P, Lin TK et aL Age at first birthand breast cancer risk. Bull World Health Organ 1970;43:209-21.

82 Russo J, Rivera R, Russo IH. Influence of age and parityon the development of the human breast. Breast CancerRes Treat 1992;23:211-18.

83 Kunz J, Keller PJ. HCG, HPL, oestradiol, progesteroneand AFP in serum in patients with threatened abortion.BrJ3 Obstet Gynaecol 1976;83:640-4.

84 Witt BR, Wolf GC, Wainwright CJ, Johnston PD, Thor-neycroft IH. Relaxin, CA-125, progesterone, estradiol,Schwangerschaft protein, and human chorionic gonad-otropin as predictors of outcome in threatened and non-threatened pregnancies. Fertil Steril 1990;53:1029-36.

85 Stewart DR, Overstreet JW, Nakajima ST, Lasley BL. En-hanced ovarian steroid secretion before implantation inearly human pregnancy. J Clin Endocrinol Metab 1993;76:1470-6.

86 Thomas DB. Epidemiologic and related studies of breastcancer etiology. In: Lilienfeld AM, ed. Reviews in cancerepidemiology. New York: Elsevier-North Holland, 1980:154-217.

496



j.com/

J Epidem

iol Com

munity H


ownloaded from


induced risk review - bmjinduced abortion. however,the original study does not distinguish between...

Documents