All carriers participated in clinical or research studies at the host institutions which have been approved by local ethics committees (list provided in Additional file 1 Table S1). Informed consent was obtained from all study participants. Subjects were BRCA1 and BRCA2 mutation carriers recruited by 40 study centres in 28 countries through CIMBA (Additional file 1 Table 2). The majority of carriers (97.58%) were recruited through cancer genetics clinics offering genetic testing, and enrolled into national or regional studies. Some carriers were identified by population-based sampling of cases (2.38%), and some by community recruitment (0.04%). Eligibility to participate in CIMBA is restricted to female carriers of pathogenic BRCA1 or BRCA2 mutations age 18 years old or older at recruitment. Information collected included the year of birth; mutation description, including nucleotide position and base change; self reported ethnic ancestry, age at last follow-up; ages at breast or ovarian cancer diagnoses; and age or date at bilateral prophylactic mastectomy and oophorectomy. Related individuals were identified through a unique family identifier. Women were included in the analysis if they carried mutations that were pathogenic according to generally recognized criteria 18 . Further details on CIMBA can be found elsewhere 19 .

Women who carried pathogenic mutations in both BRCA1 and BRCA2 were excluded from the current analysis. The primary analysis was restricted to women self-reported as "white of European ancestry". The number of mutation carriers of non-white ancestry was too small to allow separate analysis. We investigated possible overlap of carriers between studies by comparing the year of birth, exact mutation description and the reported ages, to identify potential duplicate individuals. Where possible we also used other genotype data on SNPs genotyped in the current round (at least 26 SNPs), in previous genotyping rounds or as part of GWAS to find hidden duplicates. When a potential duplicate was identified, we contacted the relevant centres for further information about these individuals, in a manner that protected the identity of the individuals in question, in order to determine precisely the extent of true overlap in subjects and families appearing more than once in the data set. Duplicate mutation carriers were included only once in the analysis. When in doubt, and when centres could not clarify a potential duplication, one of the samples was excluded from the analysis.

DNA samples (in almost all cases, obtained from blood) were genotyped using the iPLEX Mass Array platform at four genotyping centres (Additional file 1 Table S2); the iPLEX included 26 SNPs as part of a larger study. All centres included at least 2% of the samples in duplicate, no template controls in every plate, and a random mixture of affected and unaffected carriers. Samples that failed for ≥ 20% of all the SNPs typed (that is, five or more) were excluded from the analysis. A study was included in the analysis only if the call rate was over 95%, after samples that failed at multiple SNPs had been excluded. For each study, genotypes for at least 98% of the duplicate samples had to be concordant. To assess the accuracy of genotyping across genotyping centres, the four centres genotyped 95 DNA samples from a standard test plate (Coriell Institute) for all SNPs. If the genotyping was inconsistent for more than one sample in the test plate, all the studies genotyped at the centre were excluded from the analysis of that SNP. No SNPs failed this criterion. The present study included eight SNPs: rs1011970 (9p21, near CDKN2A/B), rs10995190 (10q21, near ZNF365), rs704010 (10q22, near ZMIZ1), rs2380205 (10p15), rs614367 (11q13), rs1292011 (12q24), rs10771399 (12p11 near PTHLH) and rs865686 (9q31.2). Based on the quality control criteria, 4 studies were excluded from the analysis of rs2380205 (one due to low duplicate concordance, 3 due to low call rate), 2 studies were excluded from the analysis of rs704010 (low call rate) and 13 studies were excluded from the analysis of rs1292011 (all due to low call rates). As an additional genotyping quality-control check, we also evaluated the deviation from Hardy-Weinberg equilibrium (HWE) for unrelated subjects separately for each SNP and study. Nine studies had HWE P-values in the range 0.005 to 0.05 (two studies for rs10995190, two studies for rs704010, one study for rs10771399, two for rs1292011 and two for rs865686). Upon examination of the cluster plots for these studies and SNPs, none revealed any unusual patterns and these studies were included in all the analyses. After the above exclusions, a total of 19,731 unique mutation carriers (12,599 BRCA1 and 7,132 BRCA2) from 40 studies had an observed genotype for at least 1 of the SNPs and were included in the primary analysis.

The aim of the primary analysis was to evaluate the association between each genotype and breast cancer risk within a survival analysis framework. The time variable for each individual was defined to be the time to breast cancer diagnosis. Each individual was followed until the first breast cancer diagnosis, ovarian cancer diagnosis, or bilateral prophylactic mastectomy or the age at last observation. Only those with a first breast cancer diagnosis were considered as affected in the analysis. Mutation carriers censored at ovarian cancer diagnosis were considered unaffected. Analysis was conducted by modelling the retrospective likelihood of the observed genotypes conditional on the disease phenotypes as previously described 18 . The effect of each SNP was modelled either as a per-allele hazard ratio (HR) (multiplicative model) or as separate HRs for heterozygotes and homozygotes, and these were estimated on the logarithmic scale. The HRs were assumed to be independent of age (that is, we used a Cox proportional-hazards model). The assumption of proportional hazards was tested by adding a "genotype x age" interaction term to the model in order to fit models in which the HR changed with age. Analyses were carried out with the pedigree-analysis software MENDEL 20 ; details of this approach have been described previously 18 21 . We examined between-study heterogeneity by comparing the models that allowed for study-specific log-hazard ratios against models in which the same log-hazard ratio was assumed to apply to all studies.

To investigate whether our results were influenced by any of our assumptions we performed additional sensitivity analyses. If a SNP is associated with disease survival, the inclusion of prevalent cases may influence the HR estimates. Current data indicate that five-year survival after a breast cancer diagnosis is over 80% (Cancer Research - UK, Breast cancer survival statistics) and studies have suggested no difference in survival between mutation carriers and non-carriers 22 . We, therefore, repeated our analysis by excluding mutation carriers diagnosed more than five years prior to recruitment into the study. To examine whether SNP associations differed by type of mutation, we classified BRCA1 mutations according to their potential functional effect 23 24 25 26 . Class 1 mutations were those likely to lead to the absence of protein expression due to i) reduced transcript level and/or degradation or instability of truncated proteins, or ii) absence of transcription. Class 1 mutations comprise truncating mutations expected to trigger nonsense-mediated mRNA decay (NMD) or translation re-initiation but no production of stable protein, and deletion of transcription regulatory regions. Class 2 mutations were those likely to generate stable mutant proteins with partial or total loss of function that might also have dominant negative effect. Class 2 mutations include missense substitutions, in-frame deletions and insertions, as well as truncating mutations with premature stop codons occurring in the last exon. Mutations whose consequences at transcript or protein level could not be inferred were not considered for this classification. These were mainly mutations located in splice sites but not characterised for their effect at the transcript level, or large deletions or insertions with undetermined boundaries.

The associations of these SNPs with ovarian cancer risk were evaluated within a competing risk analysis framework 8 9 21 , by estimating HRs simultaneously for breast and ovarian cancers. In this model, each individual was at risk of developing either breast or ovarian cancer, by assuming that the probabilities of developing each disease were independent conditional on the underlying genotype. A different censoring process was used for the competing risk analysis, whereby individuals were followed up to the age of the first breast or ovarian cancer diagnosis and were considered to have developed the corresponding disease. No follow-up was considered after the first cancer diagnosis. Individuals were censored for breast cancer at the age of bilateral prophylactic mastectomy and for ovarian cancer at the age of bilateral oophorectomy and were assumed to be unaffected for the corresponding disease. The remaining individuals were censored at the age at last observation and were assumed to be unaffected for both diseases.

We further evaluated the associations of these SNPs with breast cancer subtypes defined by the estrogen receptor (ER) status of the tumours in BRCA1 and BRCA2 mutation carriers. The analysis was carried out by an extension of the retrospective likelihood approach to model the simultaneous effect of each SNP on more than one tumour subtype 14 . Briefly, this involves modelling the conditional likelihood of the observed SNP genotypes and tumour subtypes, given the disease phenotypes. Within this framework it is possible to estimate simultaneously the HRs for each tumour subtype and test for heterogeneity in the associations. Only studies that provided tumour pathology information and had genotype information were included in the analysis. To maximise the available information, genotyped mutation carriers that were missing information on tumour characteristics (within each study) were included in the analysis, and their disease subtype was assumed to be missing at random 14 . This is a reasonable assumption given that more than 90% of mutation carriers in our sample were recruited prior to 2007, when it was uncommon to use tumour pathology in selecting individuals for BRCA1 and BRCA2 mutation screening.

To ensure a sufficiently large number of mutation carriers within each stratum, we grouped studies from the same country. All analyses were stratified by country of residence and used calendar-year- and cohort-specific cancer incidences for BRCA1 and BRCA2 27 . For sensitivity analyses, strata with small numbers of mutation carriers were grouped. We used a robust variance-estimation approach to allow for the non-independence among related carriers 28 .

Of the eight SNPs, only rs10771399 in PTHLH was associated with breast cancer risk for BRCA1 mutation carriers (P-trend = 3 × 10^-4, Table 2). The association was consistent with a multiplicative model in which each copy of the minor allele was estimated to confer a HR of 0.87 (95% CI: 0.81 to 0.94). There was no evidence of heterogeneity in the HR estimates across studies (P-het = 0.24, Additional file 1 Supplementary Figure 1). There was no evidence that the HRs varied with age (P = 0.68). The association remained significant, with a similar HR estimate (HR = 0.85, 95% CI: 0.77 to 0.93, P-trend = 6 × 10^-4, Table 3), when long-term survivors were excluded from the analysis, suggesting no evidence of survival bias. Interestingly, the association was restricted to BRCA1 carriers of Class 1 mutations (HR = 0.82, 95% CI: 0.74 to 0.90, P-trend = 3 × 10^-5, Table 3) with no evidence of association for Class 2 mutation carriers (HR = 1.00, 0.87 to 1.15, P-trend = 0.99, P-difference between Class 1 and Class 2 = 0.03).

We found no evidence of association between breast cancer risk for BRCA1 mutation carriers and any of the other SNPs under the trend models (P-trend > 0.15). There was, however, some suggestion of an association under the genotype specific model for rs865686 (2df P = 0.06, Table 2), reflecting a lower HR for heterozygous carriers than for either homozygote genotype. There was marginal evidence of heterogeneity in the HRs across countries for rs704010 and rs865686 (P-het = 0.04 for both), but examination of the forest plots revealed that in each case this was mainly due to a single study/country of relatively small sample size, with the majority of the HR estimates being close to 1 (Additional file 1 Supplementary Figure 1). There was no evidence that the HRs varied by age for any of the SNPs (P > 0.08 for all).

We further evaluated the SNP associations with breast and ovarian cancer risk simultaneously (Table 4). The associations with breast cancer risk remained essentially unchanged in the competing risk analysis, with only the PTHLH SNP rs10771399 being significantly associated with breast cancer risk. There was some suggestion of a possible association between this SNP and ovarian cancer risk for BRCA1 mutation carriers with risk in the opposite direction (HR for ovarian cancer = 1.14, 95% CI: 1.00 to 1.30, P-trend = 0.06) especially among rare homozygotes (ovarian cancer HR for GG = 1.67, 95% CI: 1.05 to 2.64, P-homozygotes = 0.03). This analysis also provided some weak evidence for an association between SNP rs614367 at 11q13 and ovarian cancer risk for BRCA1 mutation carriers under the genotype-specific model (2df P-value = 0.03). There was no evidence that any of the other SNPs are associated with ovarian cancer risk for BRCA1 mutation carriers.

There was evidence of association with breast cancer risk for BRCA2 mutation carriers for four SNPs (Table 2). The minor allele of rs10995190 in ZNF365 was associated with a reduced risk of breast cancer, where each copy of allele "A" was estimated to confer a HR of 0.90 (95% CI: 0.82 to 0.98, P-trend = 0.015). There was also some marginal evidence that the minor allele of rs1011970 near CDKN2A/CDKN2B was associated with increased breast cancer risk (HR = 1.09, 95% CI: 1.00 to 1.18, P-trend = 0.048). None of the other polymorphisms was associated with breast cancer risk for BRCA2 mutation carriers under the multiplicative model. However, SNPs rs865686 and rs1292011 were associated with risk under the genotype specific model (2df-P = 0.007 and 0.03 respectively, Table 2). There was some evidence of heterogeneity in the HRs across countries for rs1011970 (P-het = 0.005). This appeared to be mainly due to the USA stratum. The heterogeneity was no longer significant after removal of that stratum (P-het = 0.42) and the HR estimate for the association with breast cancer risk increased to 1.20 (95% CI: 1.09 to 1.32, P-trend = 1 × 10^-4). There was no heterogeneity for any of the other SNPs (P-het > 0.12 for all, Additional file 1 Supplementary Figure 2). The HR estimates for the four associated SNPs were similar when long-term survivors were excluded from the analysis (Table 3). Consistent with the results of the main analysis, rs10995190 in ZNF365 and rs1011970 near CDKN2A/CDKN2B provided marginal evidence of association using the trend-test statistic (P-trend = 0.05 for both) and SNPs rs865686 was associated with breast cancer risk under the genotype specific model (2df-P = 0.03). SNP rs1292011 was not associated with breast cancer risk in this analysis. A somewhat smaller HR estimate was obtained for the PTHLH SNP rs10771399 compared to the main analysis (per-allele HR = 0.89, 95% CI: 0.78 to 1.00, P-trend = 0.06). The attenuation of the association in the overall analysis could have occurred if the SNP is also associated with prognosis. However, the difference in the HRs was small. The results for the remaining SNPs were similar and non-significant. None of SNPs were associated with ovarian cancer risk for BRCA2 mutation carriers (Table 4).

Table 5 summarises the associations of the eight SNPs with breast cancer ER status in BRCA1 and BRCA2 mutation carriers. Only the PTHLH SNP rs10771399 was associated with ER-negative breast cancer for BRCA1 mutation carriers (ER-negative HR = 0.81, 95% CI: 0.74 to 0.90, P-trend = 3.8 × 10^-5). There was also marginal evidence that SNP rs704010 near ZMIZ1 was associated with ER-positive breast cancer for BRCA1 mutation carriers (ER-positive HR = 1.12, 95% CI: 1.00 to 1.26, P-trend = 0.046). However, the associations between ER-negative and ER-positive breast cancer among BRCA1 mutation carriers were only significantly different for SNP rs1292011 at 12q24 (P-heterogeneity = 0.045).

Despite the small number of BRCA2 ER-negative breast cancers, there was a suggestion that the minor allele of the PTHLH SNP rs10771399 is protective for ER-negative breast cancer for BRCA2 mutation carriers (HR for ER-negative = 0.78, 95% CI: 0.62 to 1.00, P-trend = 0.049), but there was no association with ER-positive breast cancer. There was evidence that SNPs rs10995190 near ZNF365, rs865686 at 9q31.2 and rs1011970 near CDKN2A/B are associated with ER-positive breast cancer for BRCA2 mutation carriers (P-trend = 0.043, 0.028 and 0.05 respectively). However, the HR estimates were not significantly different from those for ER-negative breast cancer.