Seventeen-year evaluation of breast cancer screening: the DOM project, The Netherlands. Diagnostisch Onderzoek (investigation) Mammacarcinoom.

The DOM project is a non-randomized population-based breast cancer screening programme in Utrecht which started in 1974-75. The 17-year effect has been evaluated by a case-control study of breast cancer deaths during the period 1975-92 in women living in the city of Utrecht, born between 1911 and 1925, whose breast cancers were diagnosed after the initiation of the DOM project. Controls (three for each case) were defined as women having the same year of birth as the case, living in the city of Utrecht at the time the case died, and having had the opportunity of screening in the DOM project. Screening in the period 1975-92 indicated a breast cancer mortality reduction of 46% (odds ratio of 0.54, 95% confidence interval 0.37-0.79). The strongest protective effect was found at a screening interval of 2 years or less (mortality reduction of 62%, odds ratio of 0.38), and for the highest number of screens (mortality reduction of 68%, odds ratio of 0.32 for more than four screens). Exclusion of breast cancer deaths that occurred within 1 year of diagnosis, to allow for 'lead-time' bias, gave an odds ratio of 0.61. Early diagnosis of breast cancer by screening reduces breast cancer mortality in the long term. Bias due to the study design may slightly overestimate the protective effect. A screening programme with a 2-yearly, or smaller, interval between successive screens will improve the protection of screening.

The purpose of the present paper is to evaluate long-term benefits of the breast cancer screening in the DOM project in The Netherlands. bv means of a nested case-control studv. Ex aluation Awas made of t-o particular forms of bias to which attention has recent1s been drawn (Hosek et al. 1996;Weiss and Lazoxich. 1996).

SUBJECTS AND METHODS
The DOM [Diacnostisch Onderzoek (investigation) Mammacarcinoom] project started in December 1974 in the city of Utrecht. Initiallv. the screening A-as limited to A-omen aged 50-64 at intake (birth cohort 191 1-25). and thev A-ere screened bv mammography. Of 20 555 eligible A-omen. 14 697 (72%7c) attended for screening. The intervals between successive screening examinations (screens) w-ere of different length. namelv 1. 1'/,. 2 and 4 x-ears. At the first examination. both mediolateral and craniocaudal projections A-ere obtained: in subsequent examinations. mammographx w as restricted to the mediolateral projection. A w-oman who did not participate in the first screeningv was not invited for the second screening and so on. Fix e examinations had been completed by 1984. A breast cancer registry was set up and cooperation w-ith general practitioners. local authorities and the Central Bureau for Statistics (CBS) ensured the followi-up of the invited A-omen. More detailed information about the screening design has been described prexiously (Collette et al. 1984. 1992 onwards. the DOM project was graduallv integrated in the nationviide breast cancer screeninc programme in The Netherlands. This programme invites woiomen aged 50-69. at 2-vear intervals and independently of preceding participation. to be screened.
In the present study. cases u-ere defined as breast cancer deaths in the period 1975-92 in woiomen living in the city of Utrecht w-ho v-ere born betvieen 1911 and 1925 and whose breast cancers were diagnosed betuveen 1975 and 1992. Information about tumour size. axillarv status and mode of detection of the breast cancers at diagnosis w-ere extracted from the DONI project breast cancer registry. Causes of death wiere provided by gYeneral practitioners or hospitals and checked agrainst the breast cancer regristrv. Controls ere defined as w omen lix ing in the cityv of Utrecht at the time the case died. and havinr the same year of birth. Aae matchingr >-as done because response rate to the screening and breast cancer mortalitx are age dependent. For each case. three controls were selected at random from the screening invitation file. i.e. from among all women in the 1911-25 cohort viho were resident in Utrecht in

1974.
For all cases and controls. the screening historv was taken for the time up to and including the date of diarnosis of the case. To evaluate the bias of 'misclassification of exposure' ) Hosek et al. 1996) due to inclusion of the diagnostic screening of screendetected cases that artificiallx restricts the chances of controls haxving undergone screening. the analy-sis wvas also done excluding these screens from the screening historx. A second form of bias. 'lead-time' bias (Weiss and Lazovich. 1996). was ealuated by excluding breast cancer deaths wvith a short follovi-up period after diacnosis (i.e. deaths of patients wvho viere less likely to hax-e been screened). because their inclusion wiould gixe the impression of a disproportionately large number of deaths from breast cancer in unscreened wvomen. Maximum likelihood estimation of the odds ratio (OR) associated with breast cancer screeningr was obtained bv means of a conditional logistic regression analvsis for matched sets (Breslow and Dav. 1980). This measure of effect can be considered as an estimate of the relatixe risk. The anaix ses u-ere performed w-ith the statistical packace EGRET (Egret. 1990).

RESULTS
Betu-een 1975 and 1992. a total of 846 breast cancer cases u-ere diacnosed in Utrecht. and for 177 of them it had been the cause of death. Of these 177 breast cancer cases. 13%/, (n = 23) w-ere detected by screening. 12%/ (n = 21 ) uere detected in the interval betmeen two successive screens and 75%7c (n = 133) were not screen detected (62 were among nexer-attenders). All 177 tumours of the cases uwho had died x-ere invasixe at diagnosis. Most tumours were < 20 mm: 61%. 62% and 36%7L respectivelxr in the three detection groups.
This indicates a 46% reduction in breast cancer mortality reduction among participants of the screening project. ORs from previous analses of the DOM project (Collette et al. 1984. 1992: de Uaard et al. 1986) are also presented in Table 1.
The results of screening on breast cancer mortalitx in the period 1975-92 stratified bx birth cohort are presented in Table 2. The strongest protective effect wvas found in the eldest birth cohort 11911-15). and this decreased in the tu-o vounger birth cohorts.
However. the confidence intervals of the ORs are broad and these differences are not statisticallI significant.
In Table 3. ORs are shou-n for different intervals betu-een the last screenina examination and diagnosis of the case. the number of screens. and participation in all offered screens before or at diacnosis of the case. W'omen w-ho never participated (87 cases aThe number of offered screens before or at diagnosis of the case depends on the date of diagnosis. tWomen who were screened at least once but who did not take up all screens offered before or at diagnosis of the case. Women who took all offered screens before or at diagnosis of the case. and 193 controls) represented the reference group. The strongest protectix e effect was found for the interval of 2 years or less betuA-een the last screen and diagnosis of the case (OR 0.38. 95%'confidence interval 0.18-0.77). The protection of screeninee decreased u-ith increasincg interxal periods since the last screen. Writh reaard to the number of screens. the ORs decreased with increasincg number of screens before diagnosis of the case. The OR for women who participated in all offered screens was lower than the OR for women without full compliance. In the present case-control study. all screening xisits up to. and includina. the diaenosis were counted as a positix e x-isit. For the screen-detected cases (n = 23). the screening, examination. from which the diagnosis was made. was included as part of the screening, histor-. To evaluate the possible bias due to including this examination. an analy sis of the same data (177 matched case-control pairs) was performed. excluding the diaonostic screening, examination. This gave an OR of 0.38 (95%e confidence interval 0.26-0.56). suggesting a higher protectix e effect of screening. To evaluate lead-time-bias'. Table 4 show-s the influence of excluding cases with successively longer follow -up times betmeen diagnosis and dying. If this period is less than 1 year. the effect of screening seems to be overestimated.

DISCUSSION
This case-control approach indicates a 46%e reduction in breast cancer mortalitv after 17 years of follow--up for participants of the screening programme. The strongest protective effect of screenine is found in the oldest birth cohort. women of 60-64 years at the start of the project. A higher lexel of protection in older women has been described earlier in the DOM project (Collette et al. 1984(Collette et al. . 1985(Collette et al. . 1992 and in other studies (Tabar et al. 1995). Cancers detected at screening wxould be expected to be of lower malignant  potential than cancers that were not screen detected. In the present studv. 61 '7c of all screen-detected breast cancers were small (< 20 mm). a percentage that is in reasonable accord with results of the Dutch national screening programme and the Finnish study (Hakama et al. 1995: Koning et al. 1995a). Because the present study concerns breast cancer deaths. it is not surprising that a high percentage (74%-) of the screen-detected tumours were axillarv positive at diagnosis. For both cases and controls. participation in the DONI project was low. However. some controls. of course. responded to the screening after the pseudo-diagnosis: resulting in higher true attendance rates of 68%-. 74%c and 78%/c. respectively. for the three 5-year birth cohorts show in Table 2. Selection bias due to a 'healthv screenee effect' cannot be excluded in this case-control study. because both the number of screenings before diagnosis of the case and compliance show protective effect (Table  3). Two other forms of bias in a case-control desien may also be relevant. 'Misclassification of exposure' bias due to includine the screening examination. from which the diaenosis was made. in the screenin2 history of the screen-detected cases (Hosek et al. 1996) appears to be present in the current study. Its effect is reflected in an OR of 0.54 (with inclusion of the diagnostic screening: 95C% confidence interval 0.37-0.79) and 0.38 (without screening: 95%7s confidence interval 0.26-0.56). The unbiased OR may be expected to lie between the two estimates. because systematic exclusion of the screening examination can cause bias in the opposite direction to that of its inclusion (Hosek et al. 1996). The other form of bias. lead-time bias'. seems also to be present in this study. Too short a follow-up time of incident cases leads to an artificially large number of deaths from breast cancer in unscreened cases (in w-hich lead time is absent). resultina in an overestimation of the protective effect (Weiss and Lazovich. 1996 . The present study indicates a reduction of breast cancer mortalitv of 46%because of screenino. Exclusion of cases with a folloW-up time of less than 1 -ear reduces this figure to 39%c. which is in reasonable agreement with new results from the Swedish trials: a 34%lmortalityreduction for women aaed 50-74 y-ears (Tab.r et al. 1995) and an expected reduction of between 24% and 32%c for women aaed 50-69 y-ears at trial entry (Koning et al. 1995b).
The estimates of the protective effect of the screening on breast cancer mortalityv in the long term have decreased from a 70% reduction after 6 years' follow-up. to a 48%reduction after 12 vears' follow-up. and finally to a 46%7c reduction (95%c confidence interval 21-631 after 17 vears follow-up (Table 1). This decrease max be due to the followinc: first. the screeninr progammme could have had a positive influence on the whole population. includinc women who were never screened. by promoting awareness of breast cancer and so increasing self-examination and readiness to seek early medical help. Another explanation for the observed decrease in the screening effect could be an improved therapy or a change in aggressiveness of the disease in time. A sliaht improvement of the prognosis of breast cancer during 1970-74 to 1980-84 in The Netherlands has been reported (Nab. 1995). Furthermore. lead-time' bias could have had a larger influence in the shorter follow-up periods of the previously published results of the DOM project. leading to greater overestimation of the protective effect in the earlier periods. Finally. it might be that in the short term the screening partly postpones breast cancer death and partly prevents it (Collette et al. 1984(Collette et al. . 1985. However. after 12 and 17 years of follow-up. the protective effect has stabilized. suggaesting a real reduction of breast cancer mortality in the long term. At the moment. the nationwide breast cancer screeninc programme in The Netherlands invites wuomen at 2-year inten-als (Koning et al. 1995a). whereas in the United Kingdom women are invited everv 3 years (Patnick et al. 1995: Asburv et al. 1996. On the basis of the results of the present study (Table 3). a 2-yearly screening programme seems preferable to a 3-yearly proaramme. However. these results should be interpreted with caution because of the small number of cases and controls with an interval of 2-3 years between last screen and diaognosis of the case.
In the present study. protective effects of screenincy were found in shorter screening intervals. after more screens. in older u-omen. and in women who are willing to participate. It was not possible to evaluate a possible interaction betmeen these factors. because of small numbers and their inter-correlation. It is not likely that the protective effect of screening! in the oldest 5-year birth cohort is fully attributable to attendance at more screening examinations. because this birth cohort saw the lowest percentage of cases and controls who were screened two or more times.
In conclusion. early diagnosis of breast cancer by screenin2 reduces breast cancer mortalityin the longyer term. Tw-o forms of bias due to the case-control study design seem to influence the protective effect in different directions: the overall bias probably results in a small overestimation of the overall protective effect.
The choice of a 2-yearly interval in the nationwide Dutch screening programme is supported by the results of the present study.