The evaluation of organized mammographic service screening programs is a major challenge in public health. In particular, there is a need to evaluate the effect of the screening program on the mortality of breast carcinoma, uncontaminated in the screening epoch by mortality from 1) cases diagnosed in the prescreening period and 2) cases diagnosed among unscreened women (i.e., nonattenders) after the initiation of organized screening.
In the current study, the authors ascertained breast carcinoma deaths in the prescreening and screening epochs in 7 Swedish counties from tumors diagnosed in these epochs and in the age group 40–69 years in 6 counties and 50–69 years in 1 county. Data regarding deaths were obtained from the Uppsala Regional Oncologic Center in conjunction with the National Cause of Death Register. The total number of women in the eligible age range living in each county was obtained from the annual population data of Statistics Sweden. Detailed screening data were provided by the screening centers in the seven counties, including the number of invited, the number attended, and whether each individual breast carcinoma case was exposed (screen-detected and interval cases combined) or unexposed (not-invited or nonattenders) to mammographic screening. There were 2044 breast carcinoma deaths from 14,092 incident tumors diagnosed in the prescreening and screening epochs, and the total number of person-years was 7.5 million. Data were analyzed using Poisson regression with corrections for self-selection bias and lead-time bias when appropriate.
The mortality reduction for breast carcinoma in all 7 counties combined for women actually exposed to screening compared with the prescreening period was 44% (relative risk [RR] = 0.56; 95% confidence interval [95% CI], 0.50–0.62). When all incident tumors were considered, both those exposed and those unexposed to screening combined, counties with > 10 years of screening were found to demonstrate a significant 32% mortality reduction (RR = 0.68; 95% CI, 0.60–0.77) and counties with ≤ 10 years of screening showed a significant 18% reduction in breast carcinoma mortality (RR = 0.82; 95% CI, 0.72–0.94) for the screening epoch compared with the prescreening epoch. Within the screening epoch, after adjustment for self-selection bias, a 39% mortality reduction (RR = 0.61; 95%CI, 0.55–0.68) was observed in association with invitation to screening.
Several randomized mammographic screening trials have demonstrated the ability of screening to reduce mortality from breast carcinoma.1, 2 Service screening has been introduced in many countries in recent years. However, the evaluation of population service screening for breast carcinoma remains a complex issue. Because the objective of a breast screening program is to reduce mortality from the disease, the obvious strategy is to compare breast carcinoma mortality before and after the introduction of screening. However, one must ensure that the nominally prescreening and screening periods of interest actually do correspond to no invitation and invitation to screening, respectively, in the geographic areas and age groups studied. One study of breast carcinoma mortality in Sweden3 was shown to have misclassified exposure to screening by as much as 50%, rendering meaningless the comparison of mortality rates between those women nominally exposed and those nominally unexposed to screening.4
In the before-and-after comparison, the effects of screening on breast carcinoma mortality should be separated from other effects on breast carcinoma mortality, such as changes in public awareness of the disease, in the availability of diagnostic and treatment facilities, and in therapy and patient management. The study by Blanks et al.5 from the U.K. addressed this issue by comparing changes in mortality over time in invited and uninvited cohorts, but the inability to identify the date of breast carcinoma diagnosis in the cohorts did not permit an estimate of the full benefit of exposure to screening. Many deaths from breast carcinoma that occur after the introduction of screening in age cohorts invited to screening result from tumors diagnosed before the inception of screening. Figure 1 shows the proportion of breast carcinoma deaths in Dalarna County (formerly known as Kopparberg) between 1978–1987 from tumors diagnosed before and during that period. In the first 5 years, the vast majority of deaths were from tumors diagnosed before 1978, demonstrating that a short-term evaluation period would be influenced significantly by cases diagnosed before the introduction of screening. The proportion of deaths from tumors diagnosed between 1978–1987 tended to increase as the time since the start of the period increased, although in the period as a whole only 43% of the breast carcinoma deaths were from tumors diagnosed within the period. Thus, in this example, if screening became policy in 1978, < 50% of the deaths in the following decade would be from incident cases that might have been affected by the new policy. Furthermore, because it would be unrealistic for the entire eligible population to be screened immediately at the start of the program, the actual proportion of deaths in women exposed to screening would be even smaller.
Blanks et al. acknowledged this observation in their article.5 This problem also was addressed by Tabar et al.,6 who analyzed breast carcinoma mortality in three periods, but in each period only deaths from tumors diagnosed during that period were used.
In the current study, we report on an analysis of breast carcinoma mortality in 7 counties in Sweden, covering approximately 33% of the eligible Swedish female population. We compared mortality from breast tumors diagnosed in the prescreening and screening epochs and in the age range eligible for screening (i.e., in both epochs, deaths from tumors diagnosed before the start of the epoch were excluded). We also compared mortality in screened and unscreened women within the screening epochs. Corrections were made for lead time and self-selection bias when appropriate. Within each county, we use prescreening and screening epochs of the same duration to avoid bias due to the varying death hazard of breast carcinoma by time since diagnosis. For example, tumors diagnosed within a 5-year epoch and followed only within that 5-year epoch would have an average follow-up period of only the first 2.5 years after diagnosis. Within a 10-year epoch, they would have an average follow-up of 5 years. Because the risk of death is greater in the first few years after diagnosis, the shorter period might artificially display a higher average death rate compared with the longer period.
MATERIALS AND METHODS
We used data from seven Swedish counties comprising the Uppsala administrative region: Dalarna, Gävleborg, Södermanland, Uppsala, Värmland, Västmanland, and Örebro (Fig. 2). For each county, we first ascertained the date of inception of the mammographic screening program and the coverage rates in each year. We then assigned a nominal starting date for each county. Clearly, it is not possible to have 0% coverage before the nominal starting date of a screening program and rapidly reach 100% attendance thereafter because it may take several years to complete a single screen of the target population. Even after 100% of the target population has been invited, a proportion will decide not to attend screening. Therefore, for each county, we had to select a starting date with as little screening as possible taking place before it, and high screening coverage prevailing as soon as possible after that date. The length of the periods of observation before and after screening was determined by two criteria. 1) The need to have as much observation time as possible in the screening epoch because it takes several years for the full benefit of screening to emerge.7 In addition, in view of the relatively low proportion of breast carcinoma deaths from tumors diagnosed within a short period, it is advantageous to have relatively long periods of observation both before and during the screening epoch. 2) The periods before and after screening should be of the same length because of the nonconstant death hazard from breast carcinoma mentioned earlier (the average risk of breast carcinoma death in, for example, the first 3 years since diagnosis being greater than the average risk in the first 10 years).
Data regarding breast carcinoma deaths were obtained from the Regional Oncology Center in Uppsala and confirmed by data from the National Cause of Death Register during the years 1979-1998 for Södermanland, Uppsala, Värmland, Västmanland, and Örebro counties; 1958-1997 for Dalarna county; and 1969-1998 for Gävleborg. Because mammographic screening began earlier in Dalarna and Gävleborg counties, earlier breast carcinoma mortality data were required to achieve balance in observational periods in the prescreening and postscreening epochs. Nominal starting dates, corresponding contamination rates (percent of person-years of exposure to screening prior to the starting date, percent of person-years of nonexposure to screening thereafter), and details of the screening regime are shown in Table 1.
Table 1. Nominal Starting Dates, Age Groups Screened, Screening Interval, and Contamination Rates in the Seven Counties
Nominal starting date
Contamination before (% screened)
Contamination after (% not screened)
Age group screened (yrs)
Screening interval (yrs)
Some sporadic screening with very long intervals took place prior to 1984.
Contamination rates varied from county to county, mainly due to varying attendance rates. In addition, in the first 7 years of the screening epoch in Dalarna (1978-1984), a randomized trial was in progress, during which approximately 33% of the relevant population was randomized to receive no invitation to screening.1 The periods of observation chosen to represent before and after screening are shown in Table 2. In each county and in each period, the breast carcinoma deaths of interest were from tumors diagnosed within the 2 periods (referred to as incident tumors below) and within the age range targeted for screening (50-69 years for Värmland County and 40-69 years for the other 6 counties). This approach minimizes contamination of the screening period with deaths from tumors diagnosed before screening was initiated or at ages not included in the screening program. Although several counties also invited women ages 70-74 years, these individuals were not included in this analysis due to low uptake rates in this age group.
Table 2. Overview of the Pre- and Post-Screening Study Periods in the Seven Swedish Countries
The total number of women in the eligible age range living in each county was obtained from the annual population data of the Statistics Sweden to calculate person-years. The screening centers provided data regarding screening attendance in the population and mode of detection of the tumors (diagnosed at screening, in the interval between screens, or in an unscreened woman) within the screening epoch. This information concerning mode of detection was used to subdivide deaths from breast carcinoma and the population at risk (and therefore the person-years) into whether women were exposed to screening. For breast carcinoma deaths, a tumor diagnosed at screening or in the interval between screens was counted as exposed to screening, and a tumor diagnosed in a nonattender or in a person not yet invited to undergo screening was counted as unexposed. Uninvited women in the screening epoch are those women in Dalarna County who belonged to the control group of the randomized trial and those women in any county who developed breast carcinoma before they received their letter of invitation.
In a given year in the prescreening epoch, the contamination rate was calculated as the proportion of the eligible population screened in that year. The person-years subject to contamination were calculated as the number of women screened within that year. The total person-years of contamination were calculated by summing over all years in the prescreening epoch. This total then was divided by the total person-years of observation in the prescreening epoch to give the overall contamination rate.
For the screening epoch the contamination rate was the proportion not screened. For the latter part of the screening epoch this essentially was the proportion that did not attend, but in the first few years of the screening program it also included some women not yet invited. The person-years of contamination for a given year was calculated as the product of the contamination rate and the population in that year. As in the prescreening epoch, the total person-years of contamination was summed over the whole epoch and divided by the total person-years of observation in that epoch to provide the overall contamination rate.
Mortality data were analyzed by Poisson (log-linear) regression.8 We compared death rates in the screening epoch with those in the prescreening epoch, and contemporaneous rates in exposed and unexposed women within the screening epoch. For the screening epoch in Västmanland and Dalarna (last decade only) counties, a correction for lead time was applied because there was appreciable screening in these counties before these periods. The correction for lead time is necessary to account for the possibility that some breast carcinoma deaths that might appear to be avoided in the later periods still in fact occurred, but were not counted in the later period because they were diagnosed in the earlier period because of screening. An example of how this might occur is given in Appendix 1, with the method and full rationale for the correction for lead time. For the other counties, virtually no screening took place before the screening epoch, so no such correction was necessary.
Changes in breast carcinoma mortality over time may be influenced by multiple factors. It therefore was desirable to obtain an estimate of the reduction in mortality over time that was the result of screening alone. This was done in several ways. First, as stated earlier, we compared contemporaneous death rates from breast carcinoma between women exposed to screening and women not exposed within the screening epoch. Because this essentially is a comparison of those women who chose to attend screening with those women who chose not to do so, a correction for self-selection bias was necessary. For this we used the method of Duffy et al.,9 which uses additional risk of breast carcinoma death observed in nonattenders in the randomized trials of screening to produce a more conservative estimate: that of the relative risk (RR) associated with invitation rather than actually receiving screening.
Second, we calculated the changes in death rates over time before screening was initiated. These changes are restricted to 1989 and before, with the exception of Värmland, in which screening began in 1994. Thus, these rates may not reflect more recent changes in therapy, management, or awareness that have affected breast carcinoma mortality rates in the 1990s. Therefore we also calculated the changes between the prescreening epoch and the unscreened part of the population in the screening epoch, also applying a correction for self-selection bias. We therefore scaled down the death rate in the unscreened group in the screening epoch (i.e., the nonattenders) by a factor of 1.36, which is the RR of breast carcinoma death in the nonattenders compared with the uninvited controls in the randomized trials.9
There were 5728 incident breast carcinoma cases in the prescreening epoch and 8364 in the screening epoch. The corresponding numbers of breast carcinoma deaths were 1169 and 875, respectively.
Figure 3 shows three cumulative breast carcinoma mortality rates from incident tumors in the 5 counties with ≤ 10 years of screening: 1) the rates in the prescreening epoch, 2) the rates for those actually exposed to screening in the screening epoch, and 3) the rates for those not exposed to screening in the screening epoch. The last group is referred to as nonattenders, although for the majority of counties they also included a small minority of tumors diagnosed in women not yet invited in the early years of the screening epoch. The mortality reductions in those women who actually received screening in the screening epoch range from 26-54% compared with the prescreening epoch. For all 5 counties with > 10 years of screening, a 45% reduction in breast carcinoma mortality was observed (RR = 0.55; 95% confidence interval [95% CI], 0.46-0.65). Figures 3a and 3b) show the corresponding cumulative mortality rates for the 2 counties with > 10 years of screening. A 34% reduction was observed in Gävleborg County and a 60% reduction was noted in Dalarna County. The combined mortality reduction in these 2 counties was 43% (RR = 0.57; 95% CI, 0.49-0.67). In all 7 counties combined, the breast carcinoma mortality reduction was 44% (RR = 0.56; 95% CI, 0.50-0.62).
Table 3 shows the breast carcinoma deaths from incident tumors by epoch and county, with person-years and mortality rates. RRs for the screening epoch compared with the prescreening epoch also are shown, inclusive of all tumors combined (i.e., exposed and unexposed to screening). All counties except Värmland, which had only 5 years of screening, demonstrated a reduction in mortality in the screening epoch compared with the prescreening epoch. For the two counties with the longest periods of screening, Gävleborg and Dalarna, the reduction was statistically significant, and for Södermanland, Västmanland, Uppsala, and Örebro the reduction was strongly suggestive. Combining the 2 counties with > 10 years of screening, there was a significant mortality reduction of 32% (RR = 0.68; 95% CI, 0.60-0.77). For counties with ≤ 10 years of screening, there was a significant mortality reduction of 18% (RR = 0.82; 95% CI, 0.72-0.94). Note that these are the reductions in mortality associated with the policy of offering screening, not of actually receiving screening. The combined mortality reduction for all women in the screening epoch (exposed and unexposed combined) was greater for counties with > 10 years of screening.
Table 3. Comparison of Breast Carcinoma Mortality from Incident Tumors between the Prescreening and Screening Epochs (Exposed and Unexposed Women Combined in the Latter Group)
However, the mortality reductions in those women who actually received screening did not appear to differ according to the duration of the screening epochs. In addition, the reductions associated with screening shown in Figures 3 and 4 are greater than those observed in Table 3, in which both exposed and unexposed women were combined during the screening epoch. This finding is as expected, but there may be an overestimation of the benefit due to other changes over time (awareness, improvements in therapy, etc.) and to self-selection bias. Table 4 shows the comparison of contemporaneous mortality rates between exposed and unexposed women in the screening epochs, corrected for self-selection bias and lead time. All counties showed a reduction in breast carcinoma mortality associated with attendance at screening, the reduction being significant in all counties except one. Overall, the combined reduction was 39% (RR = 0.61; 95% CI, 0.55-0.68).
Table 4. Comparison of Breast Carcinoma Mortality between Screened and Unscreened Women within the Screening Epochs, with RRs and 95% CIs
Table 5 shows the breast carcinoma mortality rates from incident tumors in the prescreening epoch in each county. In only one county, Gävleborg, was there a significant decrease in the breast carcinoma mortality in the prescreening epoch. The overall breast carcinoma mortality reduction in the prescreening epoch in all 7 counties was 1% per year, which was not statistically significant. When compared with these results, the cumulative 32% reduction in breast carcinoma mortality during the screening epoch suggests that the observed reductions in mortality from incident tumors are due largely to screening and not to other changes over time.
Table 5. Changes in Breast Carcinoma Mortality from Incident Tumors in the Prescreening Epochs
% change per year (95% CI)
95% CI: 95% confidence interval.
+1 (−5, +6)
−2 (−11, +7)
+6 (−4, +18)
+9 (−1, +21)
+1 (−8, +11)
−7 (−12, −3)
−1 (−5, +4)
Overall % change per year
−1 (−4, +1)
The higher cumulative rate of breast carcinoma mortality in unscreened women in the screening epoch compared with women in the prescreening epoch is due largely to the fact that the nonattenders were a self-selected population, often with a higher mortality rate than the general population. After correcting for this self-selection bias, the percentage change in the breast carcinoma mortality rate from that observed among all women in the prescreening epoch to unscreened women in the screening epoch ranged from a 26% reduction in breast carcinoma mortality to a 108% increase. Overall, there was a nonsignificant 12% increase in mortality in unscreened women in the screening epoch compared with the prescreening epoch (RR = 1.12; 95% CI, 0.97-1.30).
Now that several published randomized trials have demonstrated the efficacy of breast carcinoma screening,1, 2, 10, 11 it is important to measure the effect of service screening on the breast carcinoma mortality rate and to distinguish the mortality benefit among women who actually participate in screening from the benefit observed at the population level associated with the policy of offering breast carcinoma screening.6 We found significant reductions in breast carcinoma mortality from incident tumors in the screening epochs compared with the prescreening epoch based on an analysis of 2044 breast carcinoma deaths reported over several decades in 7 counties. In women who actually underwent screening, the mortality reduction was approximately 45%. The mortality reduction among women who actually attend screening is the more accurate measure to use when communicating to women about the benefits of screening, or when calculating programmatic measures such as “number needed to screen.”12 Cost-effectiveness analysis might be similarly based on the mortality reduction among those who actually attend screening, although some argue that the cost-effectiveness ratio should pertain to intention to treat (or, in this case, to screen). Cost-effectiveness analyses previously have shown that mammography screening is cost-effective in comparison with other medical interventions.13, 14 Because data from the current study confirm those obtained in the randomized trials, service screening appears to be highly cost-effective.
The overall breast carcinoma mortality reductions (inevitably diluted by combining exposed and unexposed women in the screening epoch) were greater in counties in which screening had been taking place for > 10 years (approximately 30%) compared with those countries in which screening had been taking place for ≤ 10 years (approximately 20%). These findings are consistent with findings from the randomized trials demonstrating that longer durations of follow-up reveal greater reductions in mortality.
During the screening epoch, women exposed to screening had a breast carcinoma mortality rate that was 39% lower than that in women unexposed to screening after correction for self-selection bias (RR = 0.61; 95% CI, 0.55-0.68). The correction for self-selection bias takes into account the fact that those who choose not to attend screening have a higher breast carcinoma mortality rate compared with a population that is not invited to screening at all.9 The 39% lower breast carcinoma mortality estimates the effect of being invited to screening rather than the effect of actually undergoing screening and therefore underestimates the true benefit. This comparison is based on contemporaneously diagnosed and treated breast carcinoma cases and therefore is not affected by changes over time in therapy or other factors. Thus, the 39% mortality reduction can be attributed only to the earlier detection of breast carcinoma.
A 1% reduction in breast carcinoma mortality per year was observed prior to the beginning of the screening epochs. This can be considered to be the combined effect of increased public awareness resulting in more rapid reporting of breast symptoms, changes in the availability of diagnostic and treatment facilities, and improvements in therapy and patient management. If this trend continues unabated, it would indicate a 12% reduction in mortality between the prescreening and screening epochs for reasons other than screening, and therefore it could be concluded that the majority of the 45% mortality reduction in screened women was because of the screening. This conclusion is consistent with the considerably greater curative effect of therapy on early-stage tumors compared with later-stage tumors. Because mortality is a product of both disease incidence and case fatality, the lack of a reduction in breast carcinoma mortality between the prescreening and screening epochs independent of screening may be the result of increasing incidence. For example, in Dalarna County, breast carcinoma incidence rates per 1000 for the 4 decades 1958-1967, 1968-1977, 1978-1987, and 1988-1997 were 1.17, 1.38, 2.07, and 2.19, respectively. These represent a 2% increase in incidence per year, occurring in both the prescreening and screening epochs.
The results of the current study are consistent with those of Tabar et al.6 The Dalarna data overlap with those of Tabar et al., and demonstrate a similarly large reduction in mortality, possibly because of the long period of screening exposure in Dalarna.
The efficacy of mammographic screening in reducing breast carcinoma mortality already has been demonstrated by randomized trials.1, 2 Gøtzsche and Olsen15, 16 have expressed doubts regarding the quality of the trials. These doubts have been shown to be misplaced,17-20 not least by the publication of more detailed information from the overview of Swedish mammography trials.21 Olsen and Gøtzsche,16 echoed by Black et al.,22 also have claimed that breast carcinoma mortality as an endpoint is biased in favor of screening, despite long-published data demonstrating no evidence of such bias.23 Tabar et al. have shown that in association with invitation to screening, death rates from all causes among breast carcinoma cases (because one cannot expect screening to influence mortality rates among those who do not develop breast carcinoma) are reduced significantly (unpublished observations). The use of breast carcinoma mortality as an outcome therefore is valid and, consequently, the results of the randomized trials are reliable. Therefore, the results of the current study, based on breast carcinoma mortality, also are valid.
Now that the randomized trials of mammographic screening have been completed and are in the long-term follow-up phase, the challenge for researchers is to devise strategies for the evaluation of the service screening programs that currently are in development around the world. Our evaluation used deaths from incident tumors. Although this tends to mean shorter follow-up periods, it also minimizes contamination of screening-exposed cohorts with unexposed cohorts, and vice versa. We also had data regarding attendance at screening and detection mode of the tumors, which enabled us to estimate the reduction in mortality conferred by actually receiving screening, as well as that conferred by belonging to a cohort that was invited to undergo screening. We incorporated corrections for self-selection bias in comparisons involving screening attenders or nonattenders and adjusted for lead time when necessary. The correction for self-selection bias is substantial because we used the estimate from the randomized trials of 1.36 for the RR of breast carcinoma death in nonattenders compared with uninvited controls9. When translated to a comparison with attenders at screening, the RR representing self-selection bias is even larger. Because a number (admittedly small) of the unexposed women in the screening epoch have not yet been invited rather than being nonattenders, this correction is likely to be conservative.
Potential limitations of the current study could be that the assumptions behind our adjustments for lead time and self-selection bias are not valid. For example, there may have been changes since the randomized trials in average lead time or in the RR between nonattenders and an unscreened population. However, to our knowledge there are no studies published to date indicating that these kinds of changes have taken place.
The results of the current study apply to the range of screening intervals prevailing in these seven counties. Variation from county to county in the screening interval is not substantial, with all counties using average intervals of close to 2 years. Greater mortality reductions would be likely with shorter intervals, and lesser reductions would be likely with longer intervals.
The results of the current study, based on 2044 breast carcinoma deaths in 7 counties over a maximum of 40 years, demonstrate a 40-45% reduction that is associated with actually receiving screening and a 30% reduction in mortality in association with the policy of offering screening. The results of the current study, which evaluated the results of service screening in 33% of the population of Sweden, demonstrate that the results from the randomized trials are reproducible when large-volume, well organized screening programs are performed in dedicated mammography screening centers. Organized service screening in Sweden is achieving overall population-based breast carcinoma mortality reductions at least as high as those observed in the randomized controlled trials, and mortality reductions among women who actually attend screening are higher.
The authors thank the staff at the screening centers and the women who attended screening. This work would not have been possible without the cooperation of large numbers of health and informatics services personnel in the seven counties studied.
Appendix 1. Correction for Lead Time
Because we are only using deaths from incident tumors, there is a potential for lead-time bias in the mortality rate observed in a later period if any substantial amount of screening took place in the earlier period. Consider the case of Västmanland, in which there was some screening in the prescreening epoch. Suppose that in the absence of screening a particular tumor would have been diagnosed in 1990 and caused death in 1995. However, suppose further that it was detected early at screening, in 1988, but still caused death in 1995. In our methodology, which was designed to avoid mixing screened and unscreened cohorts, this death would not be counted in the screening epoch because it was not diagnosed within that epoch, despite the fact that screening did not prevent this death.
We therefore need to add some of the deaths from the screen-detected tumors in the earlier period to the later period, or reduce the person-years in the later period to reflect the removal of tumors due to lead time. The average sojourn time in women ages 40-69 years has been estimated as 3.4 years.1 In 1988 in Västmanland, 20,037 women were screened. Therefore, we might expect to lose 3.4 × 20,037 = 68,126 person-years worth of tumors in the second period due to lead time. Therefore we subtract 68,126 from the total of 470,961 given in Table 3 as the person-years in the screening epoch, using 402,835 as our denominator for the screened period. This gives a RR of 0.84, as shown in Table 3, rather than the 0.72 that would have resulted from an analysis that was not corrected for lead-time. A similar correction was made to the Dalarna County person-years in the period 1988-1997 because a significant amount of screening activity took place in the decade immediately preceding.