In the present study high breast density was positively associated with death from breast cancer, in a dose–response pattern. Moreover, this association was most pronounced among symptomatic women. Breast density did not substantially contribute to the difference in survival between women with non-symptomatic vs. symptomatic cancers.
Some methodological issues have to be considered. All Swedish residents are given a unique civil registration number at birth which facilitates record-linkage. The Swedish Cause of Death Registry offers complete data with a coverage of 97.3% in 2008  and the cause of death for malignant tumours has been shown to be correct in 90% of cases . Hence, information on cause of death in the present study is expected to have a high completeness and correctness.
Breast density in this study was estimated qualitatively by several radiologists. Both qualitative and quantitative methods of measuring mammographic density have shown an association between density and breast cancer risk . Quantitative measurements are thought to be more exact and reliable . However, there is no consensus which quantitative measure of breast density to use , which is why we consider it relevant to use a readily available qualitative mode of assessment such as a modified BI-RADS. Known determinants for breast density (BMI, HRT, age and menopausal status) were distributed as expected in relation to different categories of density in this study, indicating a valid measurement of breast density. No formal assessment of inter-observer variability was performed at the initial estimation of breast density in this study, which is a limitation. However, in a not yet published study, 1200 recent screening mammograms were prospectively double-read by the same observers as in the present investigation. When applying the BIRADS classification with the modification described above, a kappa coefficient of 0.60 (0.55-0.65, 95% confidence interval) and a quadratic weighted kappa of 0.66 (0.62-0.70, 95% confidence interval) were found. This provides support of a substantial inter-observer agreement between the radiologists.
There was a change from analogue to digital mammography at our institution in 2004. Difference in acquisition method has been reported not to be of great importance when using a qualitative density measure such as BI-RADS .
BMI was used for adjustment as a potential confounder of the possibility to detect a tumour clinically or at screening mammography. It would have been more accurate to adjust for BMI at diagnosis but this information was not available. Weight changes over time cannot be excluded, which could have resulted in residual confounding. Weight gain and a consequent rise in BMI is probably the most likely change to occur with increasing age [36, 37]. We believe this non-differential misclassification could have attenuated the effect seen among women with fatty breasts.
Women with dense breasts, participating in the general screening program in Malmö, were invited at 18 instead of 24 months intervals. This may explain why the interval cancers were only slightly more common among women with dense breasts, which would otherwise have been expected according to data from previous studies [38, 39]. This weak association between interval cancers and high breast density could also have attenuated a potential adverse effect of density on survival in women with non-symptomatic cancers. Indeed, the strongest effect of density was seen in symptomatic cases. Interval cancers, which might have been missed at screening due to a masking effect at mammography, were included in this group, although the results may also indicate an independent effect of density on survival. There is however a problem with small numbers of breast cancer deaths in the stratified analyses, resulting in wide confidence intervals and consequently poor precision, why the results must be interpreted with caution. However, our main result is not a difference between symptomatic and non-symptomatic women but instead an over-all association between breast density and survival.
The increased mortality seen among women with dense breasts could be explained by competing mortality from other causes than breast cancer in women with fatty breasts. This seems unlikely however, as fewer women died from other causes than from breast cancer, and women who died from other causes died at a higher age. Moreover, although fatty breasts were associated with an increased risk of death from other causes, after adjustments for age, BMI, HRT and diagnostic period, these results were considerably attenuated and not statistically significant. It is also necessary to consider potential effects of lead time but to completely adjust for lead time bias is a challenging task. We adjusted our results for age, tumour size, axillary lymph node involvement, grade, hormone-receptor status, use of HRT and BMI. All this factors are likely to be associated with lead time and we believe these adjustments would have diminished potential effects of lead time bias. We are however aware that there could be residual confounding regarding lead time, which may have influenced our results through earlier tumour detected in women with fatty breasts. If so, this would then have led to a spuriously better survival in women with fatty breasts.
It is well known that younger women with breast cancer tend to have more aggressive tumours and younger women tend to have denser breasts. However, the proportion of younger women in this cohort is relatively low; mean age at diagnosis was 64.4 years (SD 7.5). All analysis were adjusted for age, tumour size, axillary lymph node involvement, grade and hormone-receptor status, which we believe would have diminished potential confounding by more aggressive tumours in younger women.
Results from previous studies on breast density and breast cancer specific survival are difficult to compare due to differences in methodology. The present and three previous studies [4, 5, 14] used different qualitative classifications of breast density. Two studies used quantitative estimation of density by a computer-assisted method but one of these studies included only 27 breast cancer deaths . In a recent study from Sweden, evaluating breast density quantitatively, comparing 1115 screening-detected breast cancers and 285 interval cancers in postmenopausal women, women with interval cancers had generally more aggressive tumour characteristics and worse 5 year survival, although these differences were most pronounced among women with non-dense breasts . There are however several differences between their and our study. Firstly, the study by Eriksson and al. used quantitative, computer assisted measures of breast density while we estimated breast density qualitatively. Secondly, symptomatic women not diagnosed as interval cancers were excluded from the study by Eriksson et al. It was not possible to study interval cancer separately in our analysis due to small numbers, which would indeed have been interesting. Thirdly, the end point in the study by Eriksson et al. was 5 year breast cancer specific survival, while most women in our study had a considerably longer follow-up, median 7.8 years (0.5-19.1). The limited number of events in our study must also be considered although the number of events was not given in the study by Eriksson et al.
Two investigations both used the BI-RADS classification of breast density in four categories [15, 16] but differed in size and design. Moreover, Chui et al.  estimated breast density at baseline in a screening program instead of at diagnosis, which may to some extent have attenuated their findings given potential changes in breast density over time. The results by Chiu et al., also based on a Swedish population, were similar to ours with an increased risk of breast cancer death in relation to density and with an even more increased risk in adjusted analyses. A problem in the majority of previous, and the present study, was the low number of included events, ranging from 26  to 127 . Despite a mean follow-up time of 7.8 years, the present study only included 76 events in the main analysis. On the other hand, the study by Gierach et al., based on data from the US Breast Cancer Surveillance Consortium, included 9232 women out of whom 889 died from breast cancer . Their results showed no association between breast density and death from breast cancer. In that study nearly fifteen percent of women were treated with breast conserving surgery and not treated with radiotherapy, which most women in Sweden would have been. Women with BI-RADS category 1 were also to a lesser extent treated with chemotherapy than women in BI-RADS category 4, despite women with fatty breasts having more adverse tumour characteristics. Their results were indeed stratified for stage and adjusted for therapy but it could be that other population related differences, e.g. lack of population based mammography screening in the United States and differences regarding health care financing make results difficult to compare.
The mechanism underlying the association between breast density and breast cancer is not clear. Several studies have addressed the issue of an association between breast density and prognostic tumour characteristics but in most of these studies, no association or conflicting results have been seen between density and tumour size, ALNI or hormone receptor status . Genetic factors could be of importance and there is evidence of common genetic polymorphisms related to both breast density and breast cancer development . Apart from underlying genetics, breast density could biologically reflect hormonal factors [12, 43, 44], and factors related to oxidative stress [45, 46], which could affect epithelial and/or stroma related processes [45–49]. It is possible to hypothesise that such factors could also influence breast cancer risk and/or outcome.