Correspondence to: Madeleine MA Tilanus-Linthorst, Department of Surgery, Erasmus University Medical Center, Rotterdam, The Netherlands, Tel: 0031-10-7041161, E-mail: firstname.lastname@example.org
Women from high-risk families consider preventive measures for breast cancer including screening. Guidelines on screening differ considerably regarding starting age. We investigated whether age at diagnosis in affected relatives is predictive for age at diagnosis. We analyzed the age of breast cancer detection of 1,304 first- and second-degree relatives of 314 BRCA1, 164 BRCA2 and 244 high-risk participants of the Dutch MRI-SCreening study. The within- and between-family variance in the relative's age at diagnosis was analyzed with a random effect linear regression model. We compared the starting age of screening based on risk-group (25 years for BRCA1, 30 years for BRCA2 and 35 years for familial risk), on family history, and on the model, which combines both. The findings were validated in 63 families from the UK-MARIBS study. Mean age at diagnosis in the relatives varied between families; 95% range of mean family ages was 35–55 in BRCA1-, 41–57 in BRCA2- and 44–60 in high-risk families. In all, 14% of the variance in age at diagnosis, in BRCA1 even 23%, was explained by family history, 7% by risk group. Determining start of screening based on the model and on risk-group gave similar results in terms of cancers missed and years of screening. The approach based on familial history only, missed more cancers and required more screening years in both the Dutch and the United Kingdom data sets. Age at breast cancer diagnosis is partly dependent on family history which may assist planning starting age for preventive measures.
The average risk of detecting breast cancer between 50 and 70 years of age is 5–6% for women in the United States and Europe. A woman's risk of developing breast cancer increases when a first-degree relative has had breast cancer progressively with decreasing age at onset in the relative and with an increasing number of family members with breast and/or ovarian cancer.[1-3] Very high risk and young age at onset are seen in women with a BRCA1 or BRCA2 mutation, who have a cumulative lifetime risk of 45–80%.[2-5] Screening with MRI or other preventive measures such as preventive mastectomy are considered appropriate in these groups from 25 to 30 years onward. However, in about 80% of families that meet the criteria for BRCA testing no causative gene mutation can be detected currently. These women are considered as having an increased familial risk.
Screening and preventive surgery should ideally be tailored to the expected age of breast cancer onset. The estimated risk at a given age differs considerably between studies, both for BRCA1/2 mutation carriers and for women at familial risk, possibly caused by the different populations and inclusion criteria in the studies.[2-7]
Few studies investigated predictiveness of age at onset in familial breast cancer; one was based on different familial syndromes, another on age at onset in one first-degree relative, a third one included BRCA1 or BRCA2 mutation carriers only.[8-10]
National guidelines that advise screening for high-risk groups rely for the starting age mainly on expert opinion, and differ between countries. Most guidelines advise either starting at a fixed age depending only on risk group (e.g., 25 or 30 years in BRCA1 or BRCA2 carriers),[11-13] or take into account the age of the youngest affected relative.[14, 15]
Aside from financial burden, risk reduction strategies have documented side effects, such as false-positive screening results and anxiety. Also, more than 50% of BRCA1 and BRCA2 mutation carriers who opt for preventive mastectomy at 25–30 years of age would not have developed breast cancer before the age of 50 years.[2-5]
The aim of our study was to determine whether the age of initiation of risk reducing interventions could be optimized by combining information on mean age of onset in the risk group with family history.
Material and Methods
Between 1999 and 2007, the Dutch MRI breast-screening study MRISC was performed in women with a BRCA1 or BRCA2 mutation or >15% lifetime risk based on family history. Women were included if aged between 25 and 70 years. Further inclusion and exclusion criteria, methodology and results of the MRISC-study have been reported previously. Of 722 participating women who were DNA tested for BRCA1 and BRCA2 mutations, pedigree information was obtained from the clinical genetic departments. Apart from the mutation status, genetic family identification number and age at onset of breast cancer in first- and second-degree relatives of the index woman were extracted. Parents and siblings are defined as first-degree relatives, and grandparents, aunts and half-siblings as second-degree relatives. If the participant was a BRCA1 or BRCA2 mutation carrier, age at onset was extracted only from the side of the family harboring the mutation. In participants with no BRCA1/2 mutation detected in the family, either the ages on the paternal or the maternal side were used, using the side in which DNA testing had been performed. At least, the four youngest ages at onset and the oldest were registered, as well as the age of any contralateral breast cancer. All index-women were unaffected with breast cancer when entering the MRISC-study. In total, 77 of the 722 index-women had a cancer diagnosed during the MRISC study (index patients) and their age at breast cancer detection was also registered. Index-women with missing ages at onset for breast cancer for first- or second-degree relatives were excluded (five pedigrees).
To test whether age of breast cancer diagnosis was determined by family, we used a random effects linear regression model. Random effects models, in contrast to commonly used fixed effects models, take into account the clustering of women within families and distinguish within-family variation from between-family variation.[18, 19] The random effects model estimates the between-family variance τ2 and the residual variance, which can be interpreted as the within-family variance. Based on these numbers, the proportion of the total variance that is owing to between-family variance (the intraclass correlation) can be calculated. A large between-family variation compared to a small residual variance indicates that age of breast cancer is family determined. We expressed the between-family variance τ2 as a 95% range of the mean age of breast cancer onset per family 95% range=overall mean age ±1.96 τ.
We fitted the random effects model including family ages at breast cancer diagnosis as a random effect and mutation type (BRCA1, BRCA2 or high risk) as a fixed effect. In addition to the within-family and between-family variation, the random effects model estimates an expected age of breast cancer diagnosis for each family. A feature of the random effects model is that if many women in a family have breast cancer, the expected age at diagnosis will be close to the mean family age at onset. If little information on family members is available, the expected age will be close to the overall average age at diagnosis of the mutation type.
We used these expected ages from the random effects model to develop a prediction model for the age at breast cancer diagnosis, based on both mutation type and number of relatives with breast cancer and their age at diagnosis. This approximating model has the following form, where the predicted age at diagnosis in family i from risk group j, yij is given by:
where zij is the predicted difference between-age at diagnosis in family i and mean age at diagnosis in risk group j, nij is the number of affected relatives in the family, tj is the mean age at diagnosis in risk group j and tij is the mean age at diagnosis in family i from risk group j. Thus, Δ(tij,tj) is the difference between the mean age at diagnosis in risk group and mean age at diagnosis in family and is always positive.
β1 and β2 are the coefficients from the fixed effects linear regression analysis.
To ensure early detection of tumors, screening should start before age of onset. In some guidelines based on family history, this is reflected by starting screening 5 or 10 years younger than the age of the youngest affected relative. As the model is calibrated to the mean observed age, with a distribution of the predictions around this observed age, a subtraction is needed to get the advised starting age of screening. Based on the guidelines, subtracting 5–10 years from the youngest family member, we subtract 15 years from the mean (predicted) family age. The advised starting age for screening based on the model is yij−15.
The model was developed in all 1,304 first- and second-degree relatives with breast cancer and applied to the 77 index-patients detected with breast cancer in the MRISC study, to avoid influence of age at entry of the participants in the study.
Comparing three strategies
To determine the value of our model for optimizing the starting age of screening, we compared it to two recommendations from the guidelines, either a fixed start of screening (25 years for BRCA1, 30 years for BRCA2 and 35 years for high risk), or a start based on the youngest age at diagnosis in a relative (5 years before age of diagnosis) solely based on family history. The three approaches were compared, examining the number of cancers that would have been missed, and the average number of screening years to detection.
We validated the performance of the three approaches in data from the United Kingdom; the MARIBS study in which women with a high familial risk were screened with MRI. The age at detection was obtained in 63 women: 37 were diagnosed with breast cancer during the study between August 1997 and May 2004, and 26 BRCA1/2 carriers (to allow for the two cases attributed to BRCA1 on the basis of being in a family with a known BRCA1-mutation without having been tested themselves for the family mutation) thereafter in an extended follow-up study of 3.5 years. Women could be included in the MARIBS study from 35 to 49 years of age. Further inclusion and exclusion criteria, methodology and results of the MARIBS- and the extended follow-up study have been reported.[22, 23] Age at onset in the first- and second-degree relatives of these MARIBS index-patients, their DNA test results and family pedigree-number were obtained.
For each index-patient from the United Kingdom data set, we predicted age of breast cancer with the model developed in the Dutch data, and compared the predicted to the observed ages at diagnosis.
Statistical analyses were performed with SPSS 15.0 and R statistical software 10.1.
The study population
The population consisted of 1,304 first- and second-degree relatives with breast cancer of the 722 MRISC-participant from 474 families, 203 with a BRCA1 mutation, 105 with a BRCA2 mutation and 166 high risk families (Table 1).
Table 1. Description of patients and families in the Dutch MRISC-study and United Kingdom MARIBS-study per risk group
Number of families
Number of relatives with breast cancer
Number of index patients (MRISC detected)
Mean age at diagnosis relatives (95% CI)
Mean age at diagnosis index patients (95% CI)
Mean age youngest relative at diagnosis (95% CI)
Mean age at diagnosis per family (95% range)
Part of the variance in age at onset explained by the family (%)
Number of families
Number of relatives with breast cancer
Number of index patients (MARIBS detected)
Mean age at diagnosis relatives (95% CI)
42 years (34–56)
44 years (35–53)
44 years (37–49)
Women in the families with a BRCA1 mutation were on average youngest when diagnosed (45 years). The number of cancers detected during MRI (and mammography) screening, their median age at detection and median age of the youngest family member is given by risk-group. Age at onset of the 77 index-patients was on average 2.7 years older than the youngest age at onset in a first- or second-degree relative. On seven occasions, there was no breast cancer in a relative.
We could validate our model in 26 BRCA1, 22 BRCA2 and 15 high-risk index patients detected in the United Kingdom MARIBS and its follow-up study.[22, 23] Age at diagnosis in the three risk groups for these patients is listed in Table 1.
Family influence on age at onset
The within- and between-family variation of age at onset was estimated with the random effects model in the overall Dutch data and within each risk group. The between-family variance was large. In the BRCA1 group, the 95% range of mean family ages at diagnosis was 35–55 (τ2=28), indicating that the mean expected age for the families varied between 35 and 55 years, even when the 2.5% oldest and 2.5% youngest families are disregarded. In the BRCA2 group, the 95% range of mean family ages was 41–57 years (τ2=16) and in the high-risk families it was 44–60 years of age (τ2=17). In the overall Dutch data, the between-family variance explained 14% of the total variance between patients in age at onset. Risk group explained an additional 7% of the variance. The rest of the variance remained unexplained. Within the individual risk groups, the part of the total variance that was attributed to between-family variance was largest in BRCA1 (23%) and smaller in BRCA2 and in the high-risk families (both 11%). Figure 1 shows ages at diagnosis in each family, ordered from families with the lowest to highest mean age on the x-axes per risk group.
The random effects model estimates the expected age of onset for each MRISC family. Figure 2 shows the mean age per risk group, the observed mean family ages and the age estimated by the random effect model. Larger families have an estimated age closer to their observed mean, whereas smaller families have an estimated age closer to the overall mean. An example of a small family is the BRCAI family with the highest mean age at diagnosis (75 years). Although this mean family age is high, the number of family members is so small that the predicted age is close to the overall mean. An example of a large family is the biggest dot in the high-risk families. This family has a mean family age at diagnosis of 60 years, which is above the overall mean. The number of family members is so large that their predicted age at diagnosis is the family mean (−15 years).
The predicted age at diagnosis in family i from risk group j, can be estimated as:
For example, the “oldest” family in the development data has a BRCA2 mutation and consists of three women diagnosed with breast cancer at 80, 81 and 82 years of ages. The mean age in this family (tij) is 81. The mean age in the BRCA2 risk group (tj) is 49. This makes Δ(tij,tj) 81−49=32. The number of diagnosed family members (nij) is 3. Thus, zij=−1.04+0.47 × 3+0.26 × 32=8.69. As the mean family age is higher than the mean risk group age (tij>tj), yij=49+8.69=57.69 which is the predicted age at diagnosis for a new family member. This corresponds to the highest predicted age in the middle panel in Figure 2.
Both the effect of the number of family members and the difference between the mean age in the family and the overall mean age in the risk group had a significant effect on the predicted age at diagnosis (β1=0.47, p<0.001 and β2=0.26, p < 0.001). The advised starting age for screening in the model is yij−15, in the example 42 years.
Comparative effectiveness of different starting ages
We compared three different strategies to determine the optimal age to start screening in the development (MRISC) and validation (United Kingdom) data set (Supporting Information Table 2). For each approach, we used the observed age of breast cancer diagnosis and the starting age of screening that would be advised with the particular strategy. For the model approach in the validation data, we predicted age at detection for each woman based on the βs estimated in the development data, but we used the mean age in each risk group from the validation data.
Table 2. Comparison of numbers of cancers missed and years screened till diagnosis in index patients in MRISC study and the United Kingdom MARIBS study, when start of screening is based on either the model or the guidelines advising 25 years old for BRCA1, 30 years for BRCA2 and 35 for high risk, or based on the age of the youngest diagnosed family member
Model prediction minus 15 years
Fixed age per risk group
Youngest relative minus 5 years
United Kingdom data
United Kingdom data
United Kingdom data
United Kingdom data
Average years between starting age and detection
United Kingdom data
United Kingdom data
United Kingdom data
United Kingdom data
Supporting Information Figure 3 shows the predicted and observed ages for the three strategies in both data sets.
Table 2 summarizes the cancers missed, and the average number of years between starting age of screening and detection.
Evaluating the different approaches in the Dutch index-patients, the model-based approach resulted in an average of 12 years of negative (“unnecessary”) screening, and would have resulted in screening starting too late in two women (Table 2). The approach of a fixed age per risk group gave comparable results with the same number of negative screening years, but starting too late in only one women. In the validation data, the model and the fixed age per risk group performed similar, with an average of 14 negative screening years and no cancers diagnosed before screening started. In both data sets, the approach of screening based on only the youngest relative would have missed most cancers (4 in the Dutch data, 13 in the United Kingdom data).
The age at diagnosis of breast cancer was in part determined by the risk-group, but within each risk group, age of breast cancer diagnosis is further significantly influenced by the age at onset in relatives, as shown by the substantial “between-family variation.” In the BRCA1 group, the family explained even 23% of the variance in age at onset. The results were confirmed in the United Kingdom MARIBS data, using the two largest data sets published on screening for breast cancer in BRCA1-, BRCA2-mutation carriers and women with familial risk. We do not yet know the extent to which this family influence on age at onset is caused by inherited factors such as the specific BRCA1 or 2 mutations in that family or single-nucleotide polymorphisms (SNPs) that also influence cancer risk in carriers, by lifestyle or by other factors.[24, 25]
Overall, determining starting age of screening based on our model or starting at 25 for BRCA1, 30 for BRCA2 and 35 for women with familial risk, as suggested in several guidelines, resulted in optimal balance between missed tumors and “unnecessary” screening years. Per risk group, however, there were differences between the two approaches. The model performed better in the Dutch BRCA2- and United Kingdom BRCA1-group, whereas the fixed starting age was better in the Dutch BRCA1 group. These differences are mainly caused by the variation in unnecessary screening years of the guidelines approach between risk groups, for example 17 in BRCA1 and only 8 in high familial risk in the United Kingdom data, whereas the number of screening years based on the model was quite constant between the risk groups.
However, the predictiveness of the model when applied to our study index-patients is limited by the range at which a woman entered healthy in the study and the follow-up (8 years). Especially, in the MARIBS study the age range was small (35–49 years). As shown in Figure 1b, it is clear that also in the United Kingdom breast cancer occurs in BRCA1/2 mutation carriers regularly before 35 years of age. The model may perform better in a general high-risk population. Therefore, further external validation is required to test the performance approach of our model in other populations.
This external validation would also be preferable in registry data as women do not only participate in screening studies according to the guidelines, but may also be alerted by breast cancer in a family member and her age at onset. This may result in a later starting age of screening than advised by the guidelines and therefore an underestimation of the healthy years of screening in our MRISC and MARIBS validation studies. This underestimation would, however, influence our validation of the three strategies to determine screening age equally and therefore will not have greatly influenced the comparison.
Neither the American Cancer Society, the United Kingdom NICE guideline, nor the Dutch guidelines uses age at onset in the family. This may be related to the absence of clear evidence at the time of guideline development.[11-13] The American Cancer Society 2007 guidelines advise screening by MRI for women with a lifetime risk of more than 20%, but do not indicate from what age in the absence of a BRCA1 or BRCA2 mutation.
In the Dutch guidelines, the starting age for screening is based on risk-group alone, 25 years for a woman at ≥50% risk of being a BRCA1 or BRCA2 mutation, 35 years for >30% lifetime risk and 40 years for 20–30% lifetime risk. From our data, these recommendations appear safe but may add years of unnecessary screening, if many ages at onset in the family are known and are high. For BRCA1 and BRCA2 mutation carriers who now opt for risk reducing mastectomy at age 25–30 years, safely postponing risk reducing surgery by 10 or more years could influence many aspects of their life. Therefore, screening and other preventive measures such as surgery should be optimally tailored to the expected age of cancer.
In some guidelines, the family history is already applied. The American College of Radiologists (ACR) and the European Society of Breast Imaging (EUSOBI) combine a woman's hereditary risk with age at diagnosis in the family in their screening advice although in different ways.[14, 15] The ACR advises screening for all three risk-groups from 30 or 10 years of age earlier than the age at diagnosis of the youngest affected relative, whichever is later. The EUSOBI advises screening from 30 years of age, and 5 years younger than the youngest relative. This guideline appeared safe in all three risk groups in our data.
Several existing guidelines are thus completely based on the overall risk in a particular risk group, whereas others combine this with family history but in different ways. Our model can be particularly helpful to tailor preventive advice when the age at diagnosis is known in many relatives.
In our data, the family explained 14% of the total variance between patients in age at onset and risk group 7% and therefore a large part of the variance in age at onset between individuals remained unexplained. This finding is compatible with the large amount of unexplained variance in age at onset in the normal risk population and the numerous lifestyle factors and SNPs that influence age-related penetrance in BRCA1/2 families.[24, 25] The prediction of age of onset might be further improved by indentifying individual factors that contribute to the development of breast cancer.
One limitation of our study is that in both validation groups the women were screened with MRI, which may have reduced the age at cancer detection by several years for many women.[17, 26, 27] We do not know whether any of the relative's cancers were detected by screening. If high-risk women are screened with mammography instead of MRI, then the number of years to subtract in the model may have to be modified.
A further limitation is that we could not study cohort effects regarding the age at onset of breast cancer as it has been described in BRCA mutation-related cancers. De Bock et al. investigated the relationship between risk for breast cancer before 30 or 50 years of age and family history in a population-based setting but could not find a cohort effect. It could be argued that breast screening or other preventive measures should start earlier than any cancer is likely to be detected. This means for a great majority of women a considerable loss of healthy years without intervention.
Furthermore, most women are well aware that screening with mammography, especially at a young age, may also induce breast cancers.[30, 31]
The findings of several studies,[8, 9, 29, 32] one in BRCA1/2 carriers, are consistent with our results, which means that age at onset is in part determined by family history.
In conclusion, age at onset is determined by both family history (14%) and risk group (7%); therefore, age at diagnosis in the family might assist when determining the starting age of screening or other preventive measures for women with high breast cancer risk. Our model can be used to combine information on risk group and on family history and may help to optimize the balance between the disadvantages of starting screening unnecessarily early and the risk of starting too late.
The authors thank Roeline van der Eijk, Ada J van Eekelen and Petra Bos for data management. Professor R Eeles received an educational grant from Vista Diagnostics.