Explaining the Better Prognosis of Screening-Exposed Breast Cancers: Influence of Tumor Characteristics and Treatment

Since the late 1970s, UK breast cancer mortality rates have been rising in all age groups alongside incidence rates until around 1990 when mortality rates started falling, including for women outside the screening age range (1, 2). This steady decrease is believed to be the result of (i) the implementation and improvement of the national screening program, and (ii) improvements in cancer staging and treatment including the use of adjuvant systemic therapies (3–6). In addition, the inception of a screening service may have resulted in increased breast awareness (earlier response to self-detected breast symptoms) and coincided with better organization of cancer care, including the implementation of multidisciplinary care in 1996 (7), possibly resulting in more effective treatment (8). The Cancer Intervention and Surveillance Modeling Network collaborators estimated that screening service delivery and advances in adjuvant treatment made similar large contributions to decreased mortality in the United States (9). In Sweden, the majority of recent reductions in mortality in the screening age range were estimated to be due to screening (10).

Screening is concerned with the detection of disease at an early detectable stage and can only reduce the rate of death when followed by effective treatment. The expectation is that treatment will be more effective if begun earlier in the disease process. If improvements in treatment are such that better survival is achieved regardless of disease stage at diagnosis, then identifying cancer in early stages becomes less critical, rendering breast screening less relevant. Screening may also lead to overdiagnosis, the detection and treatment of tumors which would not have come to clinical attention during the woman's lifetime had screening not taken place (11).

In England, population mammographic screening has been offered to women for over 20 years. It is therefore important to assess the role of screening on fatality from breast cancer in the context of potential risks as well as benefits, and its current contribution to survival from breast cancer, independent of treatment.

Here, we report on the results from a case–control study with an innovative design which allows us to compare cases who died from invasive primary breast cancer with controls who received a similar diagnosis but who survived the disease up to, and beyond the case's death, with respect to attendance at breast screening prior to diagnosis. The impact of disease attributes and cancer treatment on the relative effect of attendance at breast screening was investigated.

The purpose of this study was not to estimate the impact of screening on breast cancer mortality but to investigate sources of the difference in survival conferred by screening, with corrections for lead time, overdiagnosis, and self-selection bias. The impact of screening on breast cancer mortality was assessed in a companion study (12).

We targeted women residing in the London region, who had been invited to participate in the NHS BSP from 1988 onward, and who had not expressed dissent to their records being used for evaluation purposes. All women who died of primary breast cancer (as stated in part 1 of the death certificate) ages 47 to 89 between January 1, 2008 and December 31, 2009 and who had been first diagnosed with primary breast cancer (invasive) ages 47 to 89 and since 1990 were selected as cases. Cases identified as “Death Certificate Only” were excluded. Each case was matched to one control who was alive at the case date of death, was born within 4 months either side of the case date of birth, and was registered in the same cancer network as the case, at the case date of first diagnosis. The control had a first diagnosis of primary breast cancer (invasive) within 6 months prior to and including the case date of first diagnosis. All cases and controls had been invited to take part in the NHS BSP at least once prior to the date of their first diagnosis.

The OR for breast cancer death associated with ever attending breast screening was postulated to be equal to (or less than) the estimated 0.65 based on matched comparisons of breast cancer deaths with general population controls (13). With one control per case, 650 fatal cancers would provide 80% power to detect such an effect size at the 5% significance level using a two-sided test (14).

Tumor characteristics and treatment data were extracted from the National Cancer Data Repository (NCDR) by the National Cancer Intelligence Network (NCIN) London. Area-based deprivation quintiles based on the 2010 Income Domain of the English Indices of Multiple Deprivation (IMD) were also obtained (15).

Screening history data were traced on the National Health Applications and Infrastructure Services (NHAIS) system of the Health & Social Care Information Center (HSCIC) database. Only breast screens with invitation dates which occurred after January 1, 1988, within age range 47 to 73 years at invitation, and prior to and including date of first diagnosis were included in the analysis. For cervical screening (used in adjustment for self-selection bias, see below), only screen dates which occurred after January 1, 1998, within age range 20 to 70 years at screen, and prior to and including date of first diagnosis were included in the analysis (cervical screening invitation data are not available on the NHAIS system).

Unusually, for a matched case–control study, the primary analysis had to be by unconditional logistic regression, with broad adjustment for the matching factors (cancer network registration and age at first diagnosis, and year of first diagnosis), as the correction for overdiagnosis (see below) entailed removal of large numbers of controls which would have lost the case information in a conditional analysis. After adjustment for overdiagnosis/lead time and self-selection bias, we adjusted for tumor characteristics (histopathologic size, lymph node status, tumor stage as classified by the registry and dichotomized to localized (1) or extending beyond the organ (2–4), histologic grade [Bloom-Scarff-Richardson (BRS), histologic type, and laterality], and for treatment, whether a woman received surgery [mastectomy or breast conservative surgery (BCS)], radiotherapy, chemotherapy, and/or endocrine therapy within 6 months of first diagnosis.

Unknown tumor characteristics and treatment data were treated as separate categories. A sensitivity analysis was performed where relative effects of chemotherapy, radiotherapy, and endocrine therapy compared with no treatment, from randomized trials, were included as offsets in the regression, based on data from the Early Breast Cancer Trialists' Collaborative Group (16). For each treatment (except surgery), offsets were set to 1 if the treatment was not received and to the relevant ratio if treatment was received, that is, 0.79 (95% CI, 0.72–0.85) for polychemotherapy (anthracycline-based regimen; ref. 17); 0.83 (95% CI, 0.73–0.95), 0.79 (95% CI, 0.65–0.95), and 0.82 (95% CI, 0.75–0.90) for radiotherapy of node-negative, node-positive, and node status unknown cancers, respectively (18); and 0.76 (95% CI, 0.70–0.82) for endocrine therapy (19).

Collinearity between treatment and tumor characteristics was assessed using the inflation observed in the estimated SEs of the regression coefficients and was judged to be negligible (20).

Lead time, the amount of time by which the date of diagnosis has been advanced by screening as opposed to symptomatic detection, increases the apparent follow-up time of screen-detected cancers (defined in this study as cancers diagnosed in women having had their last breast screen within 90 days of first diagnosis). This confers a survival bias in favor of screening.

The individual lead time (t) for the preclinical but screen-detectable phase was estimated for each screen-detected cancer assuming an exponential distribution with mean |$\frac{1}{\lambda}$|⁠, where for the rate of transition to symptomatic disease (λ), we used 0.25 (mean sojourn time = 4 years) as estimated by Tabar and colleagues (21).

For a screen-detected breast cancer patient who did not die of breast cancer (i.e., screen-detected controls), we sampled an unconditional random variable from the exponential distribution:

Overdiagnosis corresponds to the detection (and associated treatment) of tumors, which would not have come to clinical attention during the woman's lifetime had screening not taken place. In our study design, the potential for overdiagnosis only applies to controls as by definition, cases were diagnosed with progressive (fatal) disease, and may confer a bias in favor of screening.

Overdiagnosis can therefore be seen as an extreme form of lead-time bias among screen-detected cancers, in which the lead time exceeds the future lifetime. In our analysis, we adjusted for overdiagnosis by excluding control women with estimated lead time extending beyond the case's date of death, and consequently we had to perform adjusted unconditional logistic regression analyses to avoid losing the corresponding case information. This is a highly conservative correction for overdiagnosis, because a control's lead time exceeding the corresponding case's time to death does not necessarily imply that the latter would exceed that control's time to death.

Self-selection, the voluntary compliance with the invitation to screening may bias the estimates of the effect of screening, as factors related to both the decision to attend screening and the underlying risk of getting and/or dying from breast cancer may confound the relationship between exposure to screening and disease outcome.

The OR (ψ) was corrected for self-selection bias using the formula derived by Duffy and colleagues (22) where a correction factor “D_r” is defined as the relative risk of breast cancer death for nonattenders compared with the not invited:

where p is the proportion of control women who attend the screening invitation. “D_r” was estimated using the relative risk of breast cancer death in nonattenders to the cervical screening program compared with the general population, adjusted for confounding of cervical screening attendance with breast screening attendance (see details in ref. 23).

For analyses of time since last screen, the logistic regression was adjusted for deprivation, thought to be the major driver of self-selection bias in this context, and contemporary attendance at cervical screening prior to diagnosis using a three-category variable to partially account for self-selection: “Never screened,” “Formerly screened (>60 months),” and “Currently screened (0–60 months).”

All statistical analyses were performed using the statistical software R version 2.13.0 (The R Foundation for Statistical Computing, http://www.r-project.org/foundation).

This study is part of a protocol for the ongoing evaluation of the English NHS BSP and has received all relevant approvals (details published elsewhere; ref. 24).

A total of 1,493 breast cancer deaths were registered in London during 2008–09, and 1,192 of these were able to be matched to diseased controls and traced for screening history (Fig. 1). Among these 1,192 pairs, 22 were excluded due to a first diagnosis being an in situ disease (8 controls and 14 cases) and 491 for not having been invited to the NHS BSP at least once prior to their first diagnosis (391 controls and 417 cases), leaving 679 pairs in the dataset for the main analysis.

Over 80% of women in our dataset selected were diagnosed from the year 2000 onward (Table 1). As per study design, cases and controls had the same median age at diagnosis (63 years old). Median age at death for the cases was 68 years old.

Median age at first NHS BSP invitation was 52 years for both groups. Among compliers with breast screening, median age at first breast screen was very similar (53 years old), whereas the proportion of women who never attended screening was larger for cases than for controls (26.4% vs. 13.7%), and was mirrored by a larger proportion of control women having attended screening more than once (61.7% vs. 48.1%). In addition, time since last breast screen was over a year longer for the case population (1.2 vs. 2.4 years) and the number of screen-detected cancers, defined as cancers diagnosed in women having had their last breast screen within 90 days of first diagnosis, was higher among controls than cases (37.1% vs. 18.7%). Controls also attended their last or penultimate invitation in larger proportion (41.7% vs. 33.3%, and 57.7% vs. 47.3%, respectively; Table 1).

When assessing the effect of having attended at least one breast screen prior to first diagnosis on case fatality from breast cancer, the conditional OR (cOR) and the unconditional OR (uOR) adjusted for matching factors were found to be very similar (cOR, 0.42; 95% CI, 0.31–0.56 and uOR, 0.44; 95% CI, 0.33–0.58, respectively, Table 2).

To account for overdiagnosis, the unconditional analysis was repeated after excluding screen-detected controls with an estimated lead time extending beyond the date of death of their matched case (uOR, 0.55; 95% CI, 0.42–0.73, Table 2).

The treatment profile of early stage (localized) screen-detected cancers among women diagnosed ages 47 to 74 years, potentially associated with overdiagnosis/overtreatment, was found to be less aggressive than that of symptomatic cancers (Table 3): only 9% of screen-detected women received chemotherapy and 12% had mastectomy. The corresponding figures for symptomatic cancers were 19% and 29%.

Adjustment for deprivation had minimal impact on the relative effect of screening (uOR, 0.56; 95% CI, 0.42–0.75). Regular attendance at cervical screening (as measured by attendance in the last 5 years compared with never screened), but not deprivation, was an inverse predictor of primary breast cancer outcome (uOR, 0.75; 95% CI, 0.54–1.03; P = 0.08). Adjusting for self-selection using a “D_r” estimated on the basis of attendance at the cervical screening program (D_r = 1.19; 95% CI, 1.05–1.36; see details in ref. 23) increased the uOR to 0.69 (95% CI, 0.50–0.94, Table 2).

After excluding screen-detected cancers from the analysis, the unconditional OR was 0.60 (95% CI, 0.45–0.81); it increased to 0.81 (95% CI, 0.59–1.12) after adjustment for deprivation and correction for self-selection which was found to be larger in this subgroup of screen-exposed symptomatic women compared with the never screened (D_r = 1.24; 95% CI, 1.09–1.40, see details in ref. 23 and Table 2).

The risk of fatality reversed slightly with time since last screen compared with no screening with a conditional OR of 0.68 (95% CI, 0.47–0.98) for last attendance over 5 years prior to first diagnosis, down to 0.59 (95% CI, 0.39–0.89) for time since last screen of between 3 months and 2 years (Table 4). The better prognosis observed for screen-detected cancers in this analysis is strongly affected by overdiagnosis, lead time, and length biases, so no inferences are made from this estimate.

Tumor size, tumor stage, and surgery were the most single influential factors on the OR of the effect of screening: from uOR = 0.56 (95% CI, 0.42–0.75) after adjustment for overdiagnosis/lead time and deprivation, to uOR = 0.70 (95% CI, 0.52–0.94), uOR = 0.63 (95% CI, 0.47–0.85), and uOR = 0.68 (95% CI, 0.51–0.92), respectively.

The model was adjusted for tumor characteristics at diagnosis, which included tumor size, lymph node status, extension beyond organ (registry-based measure of stage), histologic grade, histologic type, and laterality (Table 1): the unconditional OR of attending at least one routine breast screen prior to first diagnosis increased from 0.69 (95% CI, 0.50–0.94, after correction for self-selection) to 1.00 (95% CI, 0.71–1.40; Table 2), the tumor characteristics accounting entirely for the effect of screening.

Additional adjustment for treatment factors (surgery, radiotherapy, chemotherapy, and endocrine therapy) had little impact on the OR (uOR, 1.02; 95% CI, 0.72–1.45). Using treatment effect estimates based on RCT data resulted in similar adjustment in the OR (uOR, 1.04; 95% CI, 0.73–1.49; Table 2).

Similar observations were made regarding the mediation, through disease attributes and treatment, of the effect of ever attending breast screening for symptomatic cancers (Table 2), and in terms of time since last screen in all cancers (Table 4).

We estimated the impact of breast screening attendance on breast cancer fatality and investigated the current mediators of that effect. The study was carried out using an innovative case–control design in an urban region (London) with relatively low rates of participation in screening.

Attendance in the NHS BSP reduced fatality risk by 45% after accounting for overdiagnosis and lead time in screen-detected cancers, and by 34% after additionally accounting for deprivation and self-selection. We found the major intermediaries between screening and survival to be the traditional tumor characteristics, in particular smaller size and extension beyond the breast, rather than differential treatment, in agreement with previously findings (25–29). We also observed increased fatality in left-sided or bilateral tumors. Reduction in size and likelihood of spread beyond the organ are considered a necessary effect of a screening program (5). Our results are in agreement with recent reports comparing parallel cohorts in Denmark and Norway, which observed a residual effect of screening after introduction of multidisciplinary medical care in breast cancer management (30, 31).

After adjusting for these disease attributes, and for treatment (in addition to overdiagnosis and lead time), most of the screening effect (screen and interval detection) was accounted for, leaving little scope for other factors to explain the better prognosis of screening.

A beneficial effect of attendance at breast screening was also observed among screen-exposed women diagnosed with symptomatic cancers (40% for interval detection) compared with never screened, although this effect was no longer significant at the 5% level after correcting for self-selection (19% reduction). The residual benefit of screening was also explained by the characteristics of their tumor as previously noted by Day and colleagues (32). Interval tumors have previously been found to be smaller (and more likely to be estrogen receptor-positive) and associated with better survival than tumors in the unscreened population (33–35). We and others have noted that the survival benefit in interval cancers diminishes with time since last screen (36, 37), suggesting that the effect of attendance at breast screening is not entirely due to self-selection, whereas others have not (36).

Women who participated in the screening program but whose cancer was not screen-detected may have developed increased self-awareness as a consequence of the information provided during the screening process. Tyndel and colleagues (38) have reported that women who had a recall had become more aware of their risk of developing breast cancer.

Deprivation is believed to be the main confounding factor between the decision to attend screening and cancer survival; in England, deprivation is associated with lower attendance at breast screening after adjustment for other sociodemographic factors, such as urbanization and the percentage of ethnic minorities (39), and cancer survival is lowest in the most deprived areas (40). Deprivation adjustment had no impact on the OR pertaining to attendance at breast screening in spite of having been shown to impact both attendance at breast screening and survival from breast cancer (41), possibly because cases and controls were matched for diagnosis in the same cancer network, imposing a degree of residential proximity. A similar observation was made for the matched case–control study nested within the UK Age Trial which assessed the effect of the NHS BSP in a cohort of women ages 40 to 49 years (42).

To address self-selection, we chose to use a new approach based on contemporary attendance at cervical screening to estimate the underlying risk of breast cancer death in the different screening groups compared (23). Our results suggest that self-selection bias among this group of women at higher risk of breast cancer is larger than among the general population (12).

Screening is likely to pick up cancers with best and worst prognosis. The fact that the effect of screening is minimal once the regression has been adjusted for tumor characteristics suggests that screening downstages tumors. Further investigation into the impact of tumor characteristics and treatment was limited by the absence of information on receptor status and other biological attributes (43), or type of endocrine therapy. Recent studies have questioned the relevance of tumor staging in treatment decisions, whereas histologic and molecular characterization of tumors and surrounding tissues have been shown to predict clinical outcomes (44–47).

There has been limited scope for the clear distinction of fast- versus slow-growing or steady-state tumors (48, 49), and correspondingly, it has been considered necessary to intervene once a tumor has been diagnosed. Both invasive and in situ lesions have been found to have the potential to recur and metastasize, or to remain symptomless (indolent disease; refs. 50, 51). An estimated 7% (range 3%–15%) of 40- to 80-year-old women of various ethnical backgrounds are found to have undiagnosed, mainly in situ, breast cancer at autopsy (reviewed in ref. 52). Given the upper limit provided by the autopsy results, and the limitations of mammographic detection, an average of 1.16 cm-diameter threshold for women ages 50 to 69 years (53), we believe that our approach which excluded approximately 15% of screen-detected controls accounted adequately for overdiagnosis and lead time.

In addition, our data suggest that the treatments associated with potential overdiagnosis, that is, among screen-detected women diagnosed with early-stage cancers after attending an invitation, tend to be less aggressive compared with the treatment received by early-stage cancers detected symptomatically, that is, they rarely, if ever, include mastectomy or chemotherapy.

In 2013, the UK Independent Review panel of mammographic screening felt that the benefits of screening and those of improved treatments could reasonably be considered independent (6). If the recent improvements in treatment indeed rendered early detection redundant or less effective, we might have expected to see attenuation of the screening effect on fatality by adjustment for treatment factors. This was not the case. We found that most of the improved fatality associated with screening exposure was due to tumor characteristics, in common with earlier studies (25–29).

It is worth observing that effects of treatment in a nonrandomized setting are themselves subject to selection bias, in that those given the most aggressive treatments tend to be those cases at most risk of poor outcome (54, 55). This was the case in this study with chemotherapy (data not shown). Accordingly, in addition to conventional adjustment, we also adjusted for the effects of therapies as found in meta-analyses of the randomized trials (16). The results were the same, in that the major mediating factors for the lower fatality of screening-exposed cases were the tumor characteristics rather than treatment.

In this study, over 80% of the women included received cancer treatment spanning over the last decade (2000–2009). In addition, the study design was such that epoch of diagnosis was confounded with the survival of the cases. It was therefore not possible to assess any change over time in the relative contribution of treatment on the screening effect, and hence on the extent to which time trends in mortality are due to screening or treatment.

Extension of this case–control study design to other regions of England will provide insight into regional self-selection factors. In addition, one of the companion case–control studies, designed alongside this one, will allow us to quantify the rate of overdiagnosis within the current English Breast Screening Program. In the meantime, our results suggest a substantial reduction in case fatality in screening-exposed breast cancer cases. This is largely attributable to more favorable tumor characteristics, consistent with the effect of early detection.

No potential conflicts of interest were disclosed.

Conception and design: N.J. Massat, P.D. Sasieni, S.W. Duffy

Development of methodology: N.J. Massat, P.D. Sasieni, J. Cuzick, S.W. Duffy

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): D. Parmar

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): N.J. Massat, P.D. Sasieni, J. Cuzick, S.W. Duffy

Writing, review, and/or revision of the manuscript: N.J. Massat, P.D. Sasieni, D. Tataru, J. Cuzick, S.W. Duffy

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D. Tataru, D. Parmar

Other (data extraction and preparation, and guidance on treatment and tumor characteristics data): D. Tataru

This work is part of the program of the Policy Research Unit in Cancer Awareness, Screening, and Early Diagnosis (PRU). The PRU receives funding for a research program from the Department of Health (DH) Policy Research Program. This is a collaboration between researchers from seven institutions (Queen Mary University of London, University College London, King's College London, London School of Hygiene and Tropical Medicine, Hull York Medical School, Durham University, and Peninsula Medical School).

The authors thank Dave Graham (Senior Project Manager, Systems and Service Delivery, HSCIC, Newcastle, England) for extracting screening histories and Amanda Dibden (Statistician, Centre for Cancer Prevention, Queen Mary University of London, London, England) for performing initial data checks.

This study was funded by a grant from the UK Department of Health (grant no. 106/0001 to S.W. Duffy).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.