Evidence for familial clustering in breast cancer age of onset

Abstract

Background

Familial clustering of age at onset would have implications for both personalized screening and aetiology, but has not been studied for breast cancer.

Methods

We prospectively studied a cohort of 23 145 sisters to explore whether their breast cancer risk changed near the age at diagnosis of a previously affected older sister. Using an age-time-dependent variable in a Cox regression model, we estimated hazard ratios for breast cancer when participants were near their sister’s diagnosis age, relative to similarly aged women whose sister was diagnosed at a very different age. To rule out a correlation driven by young-onset familial cancer, we separately investigated women who had enrolled at age 50 or older.

Results

Of the 23 145 women, 1412 developed breast cancer during follow-up (median 9.5 years). The estimated hazard ratio was 1.80 (95% confidence interval: 1.18, 2.74) at their sister’s age at diagnosis, suggesting a substantial increase in risk compared with women of the same age but whose sister was diagnosed at a very different age. Restriction to women who enrolled at or after age 50 produced similar results.

Conclusions

This familial clustering suggests that there may be important genetic and/or early environmental risk factors that influence the timing of breast cancer, even when onset is late in life. Personalized screening might need to account for the age at which a sister was earlier diagnosed with breast cancer.

Key Messages

• Previous studies have found elevated risk of breast cancer following the first incident cancer within a sistership, i.e. evidence for heritability.

• In a large prospective assessment of familial correlation of breast cancer age of diagnosis, we found evidence that the highest breast cancer relative hazard is around the age of diagnosis of a previously affected sister.

• To rule out that this correlation might be driven by young-onset familial cancer, we separately investigated women who had enrolled at age 50 or older, and the estimate was little altered.

• This evidence suggests that there may be important genetic and/or early environmental risk factors that influence the timing of breast cancer, even when onset is late in life.

Introduction

Genetics and environment contribute to breast cancer risk, and sisters of affected women have on average twice the risk of the general population.^1–3 Several studies have examined the age- and time-specific associations in women with an affected sister,^4–6 suggesting patterns of elevated risk over time following the first incident cancer within a sistership. One hypothesis is that family-specific absolute risk increases to a threshold⁴ at the first age of onset within the family and then remains constant. Subsequent work done with Swedish registry data^5,6 considered both time and age, reporting a higher risk within families, with no evident reduction in the relative risk across age for women with a first-degree family history.

We use data from the Sister Study cohort⁷ to examine the pattern in risk over time for sisters of cases as they approached and then passed the age at which their sister had been diagnosed with breast cancer. By using age as the primary time scale, we make comparisons among women who are all at the same observed age, and all have elevated risk due to their first-degree family history, but differ in their proximity to the age at which their sister was diagnosed. We reasoned that if there is familial clustering in age at onset for breast cancer, then hazard ratios should be highest at ages nearest a previously diagnosed sister’s age at diagnosis. Such clustering would have implications both for personalized screening and for understanding aetiology.

Methods

This study was reviewed and overseen by the institutional review boards of the National Institute of Environmental Health Sciences and the Copernicus Group.

 

Study population

Between 2003 and 2009, the Sister Study enrolled 50 884 women who were 35 to 74 years old, resided in the USA (including Puerto Rico) and gave informed consent to be followed for health outcomes.⁷ Eligible women had at least one full or half-sister who had been diagnosed with breast cancer, but had never had breast cancer themselves.⁷ We refer to the previously affected sister as the ‘proband’. We considered data from 40 236 participants from data release 7.2, with certain exclusions (Supplementary Figure S1, available as Supplementary data at IJE online): three women withdrew, 57 had breast cancer diagnosed before completing baseline activities and six had unknown diagnosis or timing of onset. To simplify the analysis, we excluded 6821 participants with more than one full sister who had been diagnosed with breast cancer and 53 with proband’s age of diagnosis unknown. To ensure that participants could be diagnosed at the same age as the proband, we excluded 3696 sisters with a proband age of diagnosis before age 36 or after age 75. We further excluded 12 participants with no follow-up time. Of the remaining 40 236 women, we then restricted attention to the 23 145 women whose affected sister was older, a group that provided follow-up times both before and after the sister’s age at diagnosis (Figure 1). By contrast, participants with a younger sister proband can never be diagnosed with breast cancer at or before their sister’s age of diagnosis. Excluding participants with a younger proband necessitated further exclusion of 33 with undetermined birth ordering and 734 whose affected sister was their twin.

Figure 1.

Counts of participants by participant age of entry and proband age of diagnosis

Outcome definition

We defined breast cancer as either ductal carcinoma in situ or invasive breast cancer. Participants were censored at loss to follow-up, end of follow-up (September 2018; data release 7.0) or death (3.1%, n = 706). Age was considered the primary time scale, with left-truncation at enrolment. Incident breast cancers were ascertained through annual self-report updates and, when possible, confirmed via medical record retrieval (82%) following consent. Based on cases with an available medical report, self-reports were highly accurate, with more than 99% positive predictive value.⁸

Statistical methods

As summary descriptive statistics, we report medians and interquartile ranges (IQR) for continuous variables and frequencies with percentages for categorical variables.

The Cox model and exposure

Closeness covariate 

The risk factor of interest was proximity of a participant’s time-dependent age to her proband sister’s age at diagnosis. Our primary approach to model this risk factor was based on a continuous time-dependent covariate. Using each participant’s age during follow-up, we constructed a ‘closeness covariate’, similar to one described by Rabinowitz⁹ 2002, as our measure of proximity to the sister’s age of diagnosis. Specifically, in a Cox model where age is the primary time scale, this time-dependent covariate is a Gaussian function of the difference between the participant’s increasing age, $t$, and the affected sister’s age at diagnosis, $d$, calculated as $Z\left(t\right)=\text{exp}(-0.5(t-d{)}^2/{\sigma }^2)$. This $Z\left(t\right)$ is 1 when t equals d and decreases toward 0, i.e. is constrained to have neglible effect, for t long before or after $d$. The scaling factor, ${\sigma }^2$, in Z(t) controls how quickly the closeness covariate declines. Smaller values of ${\sigma }^2$ result in a steeper curve, and that parameter is estimated through the model fitting process. Because $Z\left(t\right)$ equals 1 when t equals d, the exponentiated coefficient of $Z\left(t\right)$ is interpretable as the estimated maximum (or minimum) hazard ratio for the participant at her sister’s age of diagnosis. We estimated the parameters, including the scaling factor, σ, by maximum likelihood. Figure 2 shows the closeness covariate values over time. For example, for σ equal to 200, by 10 years after the sister’s diagnosis the closeness covariate has declined by a factor of 0.78.

Figure 2.

Our Cox model includes a time-dependent covariate, which is exemplified here for a hypothetical participant whose sister was diagnosed at age 57 (= d). This covariate, Z(t) = exp(-0.5*((t - d )/σ)2), is 1 when the age of the ith participant, t, is the same as the proband age of diagnosis, d, and allows the relative hazard in a Cox model to increase up to d and then smoothly decrease for t beyond that. The ‘scaling factor’, σ, determines the steepness of ascent and descent and is itself a parameter to be fitted to the data. The coefficient for Z(t) in a Cox regression model sets the maximum ln(hazard ratio) in relation to proximity to the proband’s age at diagnosis

As noted, the restriction to participants whose proband was an older sister (59%, n = 23 145) enables consideration of person-time both before and after the proband’s age at diagnosis, as shown in Figure 1. Note that participants with younger affected sisters could only enrol when they were older than the proband’s age at diagnosis. Among the participants with an older-sister proband, 1412 developed breast cancer during follow-up.

Additional sensitivity analyses were performed. The closeness covariate imposes symmetry and unimodality on the effect of interest, so we checked the appropriateness of these approximations by also using two less restrictive modelling approaches, as follows.

Step function

Instead of the closeness covariate exemplified in Figure 2, we modelled closeness with a set of time-dependent indicator variables corresponding to intervals based on the current age minus the proband age at diagnosis, t-d. The referent age-time difference was the interval that included zero. Nine additional intervals, representing 3 to 4 years except at the ends, included age differences on both sides of the referent interval, enabling estimation of a non-symmetrical step function by fitting a constant value within each time interval.

Cubic spline

To further relax the shape and symmetry constraints for the time-dependent closeness covariate, we used a restricted cubic spline to model the age-time-dependent breast cancer hazard as a smooth function of the difference between the participant’s current age and the proband’s age at diagnosis, t-d. Specifically, we used a ‘clamped’ spline, constraining the slopes to be 0 beyond the extreme values, which we fixed at the boundary knots of the 2nd and 98th percentiles of age differences. Additional knot points were assigned at the 5th, 25th, 50th, 75th and 95th percentiles of the time-dependent age differences.

Cox model parameters were estimated using the ‘survival’ package¹⁰ in R (version 3.6.0),¹¹ with robust standard errors to account for multiple sisters from the same family, and clamped splines were estimated with the ‘Epi’ package in R.¹²Two-sided P-values were used with a 0.05 level of significance.

Results

Age at menarche and examiner-measured body mass index (BMI, kg/m²) were similar across the two proband subcategories (Table 1). Participant median age at baseline [IQR] was younger for the older-proband group [54 (47, 60)] and 60% of them were postmenopausal at baseline. A majority of participants (59%) entered the study at an age exceeding their proband older sister’s age of diagnosis (Figure 1).

Table 1.

Baseline descriptive statistics by older proband status

Characteristica,b	Proband age relative to participant age
	Younger	Older	Total
Sample size (no.)	16 324	23 145	39 469
Age at menarche	13 [12, 13]	13 [12, 13]	13 [12, 13]
Body mass index (kg/m2)	27 [23, 31]	26 [23, 31]	26 [23, 31]
Participant age at baseline	59 [53, 65]	54 [47, 60]	56 [50, 62]
Proband age at diagnosis	47 [42, 53]	50 [45, 57]	49 [44, 55]
Participant baseline age >proband age at diagnosis, no. (%)	16 324 (100)	13 670 (59)	29 994 (76)
Postmenopausal at baseline, no. (%)	13 034 (80)	13 821 (60)	26 855 (68)
Age at first birth, no. (%)
Nulliparous	2794 (17)	4489 (19)	7283 (18)
≤20	3235 (20)	4008 (17)	7243 (18)
21-24	4489 (28)	5436 (24)	9925 (25)
25-29	3740 (23)	5427 (23)	9167 (23)
30-55	2041 (13)	3747 (16)	5788 (15)
Proband age of onset, no. (%)
<40	2731 (17)	2453 (11)	5184 (13)
40-49	8212 (50)	9253 (40)	17 465 (44)
50-59	4455 (27)	7844 (34)	12 299 (31)
60-69	915 (6)	3159 (14)	4074 (10)
≥70 years	11 (0)	436 (2)	447 (1)

^aAll age variables are in year units, and continuous variables are characterized by median [interquartile range].

^bThe number of missing values for age at menarche, BMI and age at first birth are 37, 11 and 63, respectively.

Closeness covariate

At the proband age of diagnosis, the estimated hazard ratio was 1.80 (Table 2), indicating a 80% [95% confidence interval (CI) = 18%, 174%] larger relative hazard when the participant’s current age equalled the proband’s age at diagnosis, compared with women of that same age but with very different proband ages at diagnosis, as shown in Figure 3. For example, with ${\sigma }^2$ = 477, a participant with a proband age of diagnosis 30 years before the participant’s current age would have a 30% reduction in risk.

Table 2.

Breast cancer hazard ratios (and 95% confidence intervals) for step function and closeness covariate models

	Null modela		Categorical		Closeness coefficient
	Unadjusted	Adjustedb	Unadjusted	Adjustedb	Unadjusted	Adjustedb
Step function for t − d (years)c
−25 to -5 years			0.67 (0.48, 0.93)	0.67 (0.48, 0.93)
>-5 to -2 years			0.81 (0.62, 1.06)	0.80 (0.61, 1.05)
>-2 to 2			Referent
>2 to 4 years			0.88 (0.72, 1.08)	0.88 (0.72, 1.08)
>4 to 7 years			0.88 (0.74, 1.06)	0.88 (0.74, 1.06)
>7 to 10 years			0.92 (0.76, 1.11)	0.90 (0.75, 1.09)
>10 to 13 years			0.90 (0.73, 1.10)	0.89 (0.73, 1.10)
>13 to 16 years			0.93 (0.74, 1.17)	0.93 (0.74, 1.16)
>16 to 19 years			0.78 (0.59, 1.02)	0.78 (0.59, 1.02)
>19 to 51 years Closeness covariate			0.70 (0.55, 0.89)	0.70 (0.55, 0.89)
exp(−0.5(t − d)2/477)d					1.80 (1.18, 2.74)	1.79 (1.18, 2.74)
Observations	23 145	23 087	23 145	23 087	23 145	23 087
Log likelihood	−12 214.05	−12 148.13	−12 206.46	−12 140.57	−12 210.27	−12 144.38
Chi square (df)			15.19 (df = 9)	15.18 (df = 9)	7.57 (df = 1)	7.50 (df = 1)

^aThe null model is a model fit without the closeness covariate.

^bThe adjusted models included age at menarche, age at first live birth, time-dependent menopause status, body mass index (BMI) and a product term between BMI and menopause status.

^c t is the participant’s time-dependent age, and d is the affected sister’s age at diagnosis.

^d477 = the scaling factor, σ², estimated by maximum likelihood.

Figure 3.

Estimated breast cancer hazard ratios by difference between participant’s age and proband age at diagnosis for three fitted models. The dotted line shows the fitted curve for the hazard ratio based on the covariate from Figure 2. The solid line shows a step function of the hazard ratio based on categories of closeness of participant’s age to the proband’s age at diagnosis. The dash-dotted line shows a fit based on a cubic spline of the age difference on the horizontal axis

Cubic spline

The clamped cubic spline model improves the fit over the null model [chi square of 17.2 (df = 7), P = 0.02], supporting an age-dependent hazard ratio with a peak near the proband age of diagnosis (Figure 3).

Step function

The step function model also supports a hazard ratio that peaks near the proband age at diagnosis. Hazard ratios for the non-referent intervals are all below 1 (Table 2). According to the step function analysis, the estimated relative effect fluctuates across a range of observed ages compared with the proband age of diagnosis. Since these estimates are relative to the selected referent category, the hazard ratio estimate for that referent interval in Figure 3 was scaled upward to enable meaningful comparison. The step function appears to align well with the spline function and both change quickly as the age difference comes close to zero, declining more slowly as the participant ages past the proband age of diagnosis (Figure 3). The step function analysis also enables a range of comparisons. For ages more than 5 years less than the proband age at diagnosis, the estimated hazard was 0.67 (95% CI: 0.48, 0.93) times that of the referent group. For a participant with a proband diagnosed when she was more than 19 years older than the participant’s current age, the estimated relative hazard was 0.70 (95% CI: 0.55, 0.89).

Sensitivity analyses

Increased vigilance around the age when a sister had formerly been diagnosed could have caused bias. We evaluated this behavioural bias by including, in addition to the closeness covariate, a separate covariate indicating whether the participant’s current age was within 2 years of the proband’s age at diagnosis (P  >0.69). Even with this adjustment, the coefficient of the closeness covariate changed little [hazard ratio (HR) = 1.77]. To further assess screening bias, we explored patterns in the relative proportions by ductal carcinomia in situ (DCIS) status and stage for diagnosed participants relative to the difference between their age at diagnosis and that of their proband (Supplementary Tables S1 and S2, available as Supplementary data at IJE online). Two-sided Cochran-Armitage trend tests did not suggest there was earlier detection when participant age at diagnosis was closer to the proband age at diagnosis.

We also fitted Cox models that included known risk factors for breast cancer. Adjustment variables included BMI, age at first birth, age at menarche, time-dependent menopause status, and the product of BMI and time-dependent menopause. The model fit dramatically improved, but those adjustments did not change either the estimated effect of the closeness covariate or its standard error (Table 2).

To assess whether an association was present in families with older-onset breast cancer and not just related to young-onset, as potentially driven by high-penetrance genetic risk factors, we analysed the subgroup of women who joined the Sister Study when at least 50 years old. The hazard ratio remained comparable to that seen with the full sample (HR = 1.71, 95% CI: 1.20, 2.43).

Discussion

We studied a large cohort of women who each had an older sister previously diagnosed with breast cancer, to examine the extent to which risk transiently increases near that affected sister’s age at diagnosis. We estimated an increase of 80% (95% CI: 18%, 174%) in the relative hazard when the participant reached the proband’s age at diagnosis, relative to women of the same age but whose sister was diagnosed at a very different age. Using a less restrictive approach based on step functions, the pattern of risk supported age-time-dependent declines in the relative risk as the participant was observed farther before or after her sister’s age of diagnosis. A third approach, based on cubic splines, also supported a peak near the age of proband diagnosis.

No published earlier analysis has modelled the age-time-dependent relative hazard as potentially influenced by the age separation between a woman’s current age and the age at which her sibling had earlier been diagnosed. Lee et al.⁶ instead considered incidence rate ratios across both age and time scales for sisters of probands as an ‘exposed’ group, using person-time after the first proband diagnosis in the family (but not necessarily only for ages beyond the proband age at diagnosis). An age-dependent incidence, with more than a doubling of the rate seen in the unexposed, was sustained across time for all ages following (in calendar time) a sister’s breast cancer diagnosis.

Peto and Mack reported an intriguing observation that the absolute yearly incidence of breast cancer appeared to remain fixed for at least 20 years after a twin sister was diagnosed, both for monozygotic and for dizygotic twins (Peto and Mack⁴ 2000) and that risk also remained fairly constant (though lower) for non-twin first-degree female relatives.

The seemingly paradoxical almost-fixed incidence rates in sisters, despite their subsequent ageing, could be more understandable in light of our findings. Correlated ages at onset would tend to pull down risk in a woman as she advances beyond her sister’s age at diagnosis, working against the age-based natural increase in risk. An age-at-onset correlation in sisters could thus attenuate the natural age-based increase, resulting in an elevated but fairly constant absolute incidence.

In our Sister Study data, different models, including the spline, step function and closeness covariate, supported a peak near the proband age at diagnosis. However, the estimated spline and step function, which impose no symmetry assumption on shape around the proband age at diagnosis, appeared to show some asymmetry around the point at which the observed age equals the proband age at diagnosis.

The evident asymmetry and lag, which appear in Figure 3, could largely be an artefact related to a kind of length bias in the sampling of the Sister Study cohort, as follows. To be eligible for participation in the Sister Study, a sister needed to have never had breast cancer herself; the sister of a case would become ineligible at the time when she was also diagnosed with breast cancer. That implies that families were more likely to be represented in the Sister Study if the second sister was destined to be diagnosed years later, meaning that she would have remained eligible to participate for years. As an extreme example, if one sister was diagnosed on the same day as the other, they would not at any point have been eligible for the study. In this way, participation was length-biased by the time between the two sister diagnoses. One consequence is that paired occurrences of breast cancer were much more likely to be observed under our design if they occurred at very different ages than if they occurred at similar ages. This could explain the evident lag in the risk for participants. Also, because diagnoses at similar ages are under-represented in the Sister Study, the peak in the pattern shown in Figure 3 has been pruned and our estimate of a 80% increase could be too small. Future analyses based on considering the difference between current age and a sister’s age at diagnosis as a time-dependent predictor, and making good use of consortial or registry data, could clarify whether the relationship we see in the Sister Study is replicable.

The drivers that would underlie a correlation in age at onset for sisters remain to be elucidated. Studies of associations between genetic polymorphisms and breast cancer age at onset are sparse,^13,14 and more research in this area could highlight loci of interest. Many genetic variants and early environmental factors are shared by full sisters, perhaps including early environmental factors from the prenatal and pubertal periods.^15–18 Hypothetically, some factor(s) in the prenatal environment could set a life trajectory for breast maturation that influences age at onset for breast cancer for sisters. Across the life course, early environment is more likely to be shared between sisters than later environments and may be an important contributor to later cancer.¹⁹ For example, a study of women who were exposed to the atomic bombing of Hiroshima found the highest relative risk in those who had been near puberty at the time of the bomb, although their breast cancer typically developed much later in life.²⁰ Furthermore, although speculative, genetic factors could interact with exposures that sisters tended to have in common as they transitioned through puberty, and could set a correlated trajectory for a particular form of breast cancer.

To rule out the possibility that the observed evidence for correlated ages of onset within families is simply due to early-onset breast cancer in families carrying particularly deleterious, high-penetrance genetic risk factors, we carried out a sensitivity analysis that fitted the same model but was restricted to women who enrolled when they were at least 50 years old. The correlation remained, with a 71% increase in the relative hazard (95% CI: 20%, 143%) when the sister reached the age at which her sister was diagnosed. Thus the correlation was evident even for disease with onset later in life.

If confirmed, then there may be practical implications of these findings for counselling and personalized screening. Current recommendations do include more screening for women with a first-degree family history, such as the Sister Study participants, especially at earlier ages,²¹ but the age at diagnosis of the affected family member(s) is usually not considered. Screening might need to be intensified for women when they near the age at which a sister had previously been diagnosed. Also, family history ascertainment should routinely include ages at diagnosis.

Despite the limitations associated with length-biased sampling, the Sister Study offered several advantages in assessing this research question. First, this cohort study yielded more than 1400 incident diagnoses among sisters with an affected older sister—the largest of its kind. Second, its family-based study design, in which sisters of at least one affected sister have been followed for an average of 9 years, is unprecedented.

In summary, we found evidence that the risk of breast cancer increases at ages near the age at diagnosis of a previously diagnosed sister. Future investigations could identify shared early life risk factors and tumour types that drive this clustering. Replication of this evidence for correlation of ages at onset in an independent sample would support a public health recommendation that screening vigilance should be increased around the age at which a sister had previously been diagnosed with breast cancer. The data underlying this article can be shared upon reasonable request at the following site: [https://sisterstudy.niehs.nih.gov/English/data-requests.htm].

Supplementary data

Supplementary data are available at IJE online.

Funding

This work was supported by the Division of Intramural Research at the National Institute of Environmental Health Sciences, in the National Institutes of Health, under projects Z01-ES044005 and Z01-ES102245.