Latent class models of early-life trauma and incident breast cancer

Abstract

Background:

Psychosocial trauma has been hypothesized to influence breast cancer risk, but little is known about how co-occurring traumas—particularly during early life—may impact incidence. We examine the relationship between multiple measures of early-life trauma and incident breast cancer.

Methods:

The Sister Study is a prospective cohort study of U.S. women (N=50,884; enrollment 2003–2009; ages 35–74). Of 45,961 eligible participants, 3,070 developed invasive breast cancer or ductal carcinoma in situ through 2017. We assessed trauma before age 18 using previously studied measures (cumulative score, individual trauma type, and substantive domain) and a six-class latent variable to evaluate co-occurring traumas. We accounted for missing data using multiple imputation and estimated hazard ratios (HR) and 95% confidence intervals (CI) using Cox proportional hazards models.

Results:

Approximately 49% of participants reported early-life trauma. Using the latent class variable approach, breast cancer hazard was higher among participants who had sexual trauma or household dysfunction (HR=1.1; CI:0.93, 1.3), or moderate (HR=1.2; CI:0.99, 1.4) but not high trauma (HR=0.66; CI:0.44, 0.99) compared to low trauma. Breast cancer HRs associated with sexual early-life trauma or household dysfunction were elevated for pre- and post-menopausal breast cancer and by estrogen receptor status. We found no effect modification by race–ethnicity. Estimated effects were attenuated with report of constant childhood social support.

Conclusions:

Breast cancer incidence varied by latent patterns of co-occurring early-life trauma. Models capturing childhood social support and trauma patterning, rather than cumulative or discrete indicators, may be more meaningful in breast cancer risk assessment.

Background

Breast cancer is one of the most commonly diagnosed cancers among United States (U.S.) women,^1,2 affecting approximately 12% of women during their lifetime.³ Psychosocial stress has been associated with breast cancer development^4–6 but prior studies of stress and breast cancer risk have primarily focused on cumulative life course adversity⁴ or stressors occurring in the 1 to 5 years preceding breast cancer diagnosis.^5–9

Stressors experienced during potentially vulnerable periods may differentially affect breast cancer risk via different biological and behavioral pathways including modification of the hypothalamus–pituitary–adrenal (HPA) axis, one of the primary stress response mechanisms. Persistent activation of the HPA axis can result in chronic dysregulation of stress management mechanisms, contributing to greater allostatic load and disease susceptibility related to immune function and inflammatory processes.^10–14 Early-life trauma may also affect breast cancer risk via alteration of circulating hormones that influence pubertal changes, early breast development, and subsequent reproductive behaviors. For example, younger age at menarche, an indicator of accelerated breast development, has been associated with increased breast cancer risk.¹⁵ Childhood and adolescence, in particular, represent potential sensitive periods of breast development during which breast tissue is more susceptible to pre-tumorigenic mechanisms, including rapid proliferation of terminal end buds, ductal elongation, and elevated estrogen sensitivity.¹⁶

Early-life stress, which represents a wide range of psychosocial stressors (e.g., abuse, neglect, and indicators of household dysfunction) and traumatic experiences during childhood or adolescence, is typically assessed via summed scores of adverse childhood experiences (ACEs),^8,17 singular traumatic events (e.g., death of a parent),¹⁸ or substantive domains-such as childhood sexual abuse.^19,20 While these types of measures of early-life trauma (ELT) have been associated with reported breast cancer risk factors,^21–23 including early menarche and menopause onset,^24–27 obesity,²⁸ and alcohol use.^28,29 Traumatic experiences may co-occur and less is known regarding whether specific patterns of concurrent traumas in early-life differentially affect breast cancer risk and whether these associations vary for certain racial–ethnic groups who may also experience greater historical or multigenerational trauma or may vary by the degree of social support experienced during childhood. Using data from the Sister Study, we investigated the relationship between early-life trauma and incident breast cancer, using previously studied measures of early-life trauma—characterized by cumulative number, specific type, or domain, as well as a novel latent class measure of co-occurring trauma.

Methods

Study population

The Sister Study is a prospective cohort study designed to assess environmental and genetic risk factors for breast cancer and other conditions.³⁰ Participants are 50,884 women from the U.S.—including Puerto Rico, ages 35 to 74 years, with a biologic sister diagnosed with breast cancer and no prior breast cancer themselves at enrollment (2003–2009). Baseline and follow-up data collection protocols are described elsewhere.³⁰ We use data release 7.1 (follow-up through 15 September 2017). The Sister Study is overseen by the Institutional Review Board of the National Institutes of Health. All participants provided written informed consent at time of enrollment in the Sister Study. The current analyses used de-identified data and was considered exempt from human subject research by the University of Wisconsin-Milwaukee Institutional Review Board.

Our analytic sample consists of the 47,138 participants who participated in the first biennial follow-up, which represented the start of person time accrual. Characteristics of participants included in the analytic sample are described by participation status (i.e., completed at least one trauma question) of the stand-alone Stress and Coping questionnaire in eTable 1. In brief, participants who did not participate in the trauma section of the Stress and Coping questionnaire were more likely to report Black or non-Black Hispanic/Latina race–ethnicity (26 vs 12%) and food insecurity as a child (14 vs 9.0%).

Early-life trauma

We collected data on early-life traumas during the first follow-up questionnaire (approximately 2 years after enrollment). We adapted trauma-related questions from the 14-item Brief Betrayal Trauma Survey³¹ and addressed 26 types of trauma. Given our focus on childhood and adolescence, we excluded items related to a participant’s children and/or spouses when assessing measures of early-life trauma. Participants indicated whether they experienced each type of trauma before age 13 (childhood) and/or between ages 13 and 17 (adolescence) (eTable 2). We dichotomized each of the included 20 trauma types (yes/no) based on report during childhood and adolescence and 1) summed them as a cumulative score, 2) assessed them individually, and 3) grouped them into seven substantive domains as depicted in Table 1.

Table 1.

Commonly studied trauma domains and types of early-life trauma assessed in the Sister Study

Trauma Domain	Types of Early-life Trauma
Natural Disasters	Natural Disaster – major fire, flood, or other natural disaster
Major Accidents	Major Accident – that resulted in serious injury to yourself or the fear of your own death, or serious injury or death of someone with whom you were very close
Household Dysfunction	Witness attack of a family member by another family member Major issues in personal relationship Serious financial/legal troubles Serious family drug, alcohol,or mental health issues
Sexual Trauma	Unwanted sexual contact by someone CLOSE Unwanted sexual contact by someone NOT CLOSE
Physical Trauma	Hit or attacked by someone CLOSE Hit or attacked by someone NOT CLOSE
Emotional or Psychological Trauma	Emotionally/Psychologically mistreated by someone CLOSE Emotionally/Psychologically mistreated by someone NOT CLOSE Witness suicide/attack of someone CLOSE Witness suicide/attack of someone NOT CLOSE Death of sibling Death of a parent Death of a close friend Major illness (not breast cancer) in someone CLOSE
Personal Illness	Major illness before age 18

We used latent class analysis (LCA) to group Sister Study participants by co-occurring types of early-life trauma³² using Mplus version 8.3 (Muthén & Muthén, Los Angeles, CA).³³ Using the 20 dichotomous early-life trauma types, we estimated five LCA models with 2 to 6 latent clusters with no a priori assumptions regarding the number of latent classes required to best capture the co-occurrence of trauma. We compared LCA model fit using Bayesian Information criterion, Lo–Mendell–Rubin-adjusted likelihood ratio test (p≤0.05), entropy (ideally ≥0.8), and substantive groupings of traumatic experiences (eTable 3). We exported class memberships for analyses.

Incident breast cancer

We ascertained data on incident invasive breast cancer and ductal carcinoma in situ (DCIS) via annual follow-up questionnaires or participant communication with Sister Study staff. Follow-up questionnaire completion rates are approximately 90%. Agreement between self-reported breast cancer and medical record documentation is high, with a positive predictive value of 99% for self-reported breast cancer and 99% for self-reported estrogen receptor (ER) positivity.³⁴ Medical records were obtained for over 82% of participants with breast cancer and self-reported pathology information was used when medical records were unavailable. Due to heterogeneity in breast cancer pathogenesis and prognosis by ER status, we independently assessed breast cancer incidence by ER status.

Covariates

We determined covariate selection a priori based on the literature. Potential confounders included self-identified race–ethnicity (non-Hispanic white, Black, non-Black Hispanic/Latina, other)—a proxy of experienced racism and historical trauma—and measures of early-life socioeconomic position (SEP): highest household education level at age 13 (less than high school, high school graduate, some college, college degree or higher), a subjective measure of family income while growing up (poor, low income, middle income, well off), and as a child, whether there were times when they did not have enough food to eat (yes, no). Although food insecurity can also represent a measure of early-life trauma, in this study, we focus on the association between the traumas noted above derived from the Brief Betrayal Trauma Survey independent of indicators of childhood SEP. Models did not adjust for breast cancer risk factors that post-date childhood and/or adolescence (e.g., age at menarche, parity, age at first birth, etc.) since these factors represent potential mediators of the early-life trauma–breast cancer pathway.

Potential modifiers

We assessed potential effect measure modification by 1) self-identified race–ethnicity (non-Hispanic white and Black only), 2) early-life SEP (using subjective household income – poor/low income vs. middle income/well off), and 3) childhood social support—based on participant report of someone in their family believing in them/wanting them to succeed and/or making them feel important or special (none of the time, a little of the time, some of the time, most of the time, all of the time). A composite variable for constant childhood social support was characterized as reporting “all of the time” for at least one of the two survey items and less than constant social support otherwise.

Statistical analysis

Although fewer than 4.4% of eligible participants were missing data for one or more covariates of interest, we used multiple imputation by chained equations to obtain 20 imputed datasets to better account for missingness when constructing the latent class variables, which combined responses from several categorical variables. This imputation included the 1,436 participants (3.1% of eligible) who were missing all early-life trauma questions. We estimated hazard ratios (HR) and 95% confidence intervals (CI) using Cox proportional hazards regression with age as the time scale (i.e., age at completion of the stress questionnaire or equivalent date for first follow-up if the stress questionnaire was not completed and exit age at breast cancer diagnosis or censoring due to death, end of follow-up, or loss to follow-up). We pooled estimated hazard ratios from imputed data sets using Rubin’s rules.³⁵ We used separate models to assess 1) cumulative trauma scores, 2) each individual trauma, 3) each trauma domain, and 4) the early-life trauma LCA variable as exposures of interest. We adjusted all models for the aforementioned covariates. In ER subtype-stratified analyses, we censored participants with competing (i.e., ER− breast cancer when assessing ER+ breast cancer incidence) or missing ER status at breast cancer diagnosis. We conducted multiple imputation and additional analyses using SAS version 9.4 (SAS Institute Inc., Cary, NC).

We assessed the proportional hazards assumption using a Wald test of the interaction between age and the exposure of interest (p<0.05) for each imputed dataset, which were then assessed across datasets using the pooling of chi-square statistics procedure proposed by Rubin and Li et al.^36–38 To further evaluate observed violations, we stratified models by time-varying menopause status and reassessed potential violations of the proportional hazards assumption using age-exposure interactions. Pre-menopausal models included time from first follow-up to menopause onset, breast cancer diagnosis, or loss to follow-up, whichever occurred first. Post-menopausal models included time from first follow-up or menopause onset, whichever occurred second, through breast cancer diagnosis or end of follow-up.

We tested for potential effect measure modification by 1) race–ethnicity, 2) childhood social support, and 3) childhood SEP in stratified analyses, using Wald tests to evaluate heterogeneity (p<0.05). We also assessed potential racial–ethnic differences in trauma experience and reporting by evaluating early-life trauma models among Black participants using a race–ethnicity specific LCA model. In sensitivity analyses, we assessed overall breast cancer risk among participants who completed the Stress and Coping questionnaire. Secondly, we restricted the sample to participants reporting no trauma in the 12 months prior to stress questionnaire completion to limit the impact of recent trauma on reporting of early-life trauma.

Results

Participants (N=47,138) contributed approximately 318,256 person–years of follow-up [mean (SD): 6.8(1.6) years], including 2,315 who developed breast cancer (n=1,827), DCIS (n=486), or tumors of unknown invasiveness (n=2). Participant characteristics by early-life trauma experience are presented in Table 2. Approximately 50% of participants reported at least one early-life trauma. These participants were slightly younger and a greater proportion reported growing up in poor or low-income households and not having enough food to eat.

Table 2.

Characteristics of the 47,138 Sister Study participants included in the analytic sample^a by experience of early-life trauma before age 18

	No Early-life Trauma	Any Early-life Trauma
	N=23,365 (50%)	N=22,337 (50%)
Age at first follow-up, mean (SD)	59.4 (9.0)	57.8 (8.7)
Follow-up time in years, mean (SD)	6.8 (1.6)	6.8 (1.6)
Age at menarche, mean (SD)	12.7 (1.5)	12.6 (1.5)
Missing	24	18
Parity, mean (SD)	2.0 (1.4)	1.9 (1.4)
Missing	9	25
Age at first live birthb, mean (SD)	24.9 (5.1)	24.6 (5.4)
Missing	34	41
Age at menopause onsetc, mean (SD)	50.0 (5.9)	49.5 (6.3)
Race-Ethnicity, N (%)
Non-Hispanic white	20,010 (86)	18,824 (84)
Black	1,843 (7.9)	1,879 (8.4)
Non-Black Hispanic/Latina	1,021 (4.4)	953 (4.3)
Other	485 (2.1)	676 (3.0)
Missing	6	5
Highest level of education in the household at age 13, N (%)
Less than high school	4,038 (17)	4,178 (19)
High school graduate	8,286 (36)	7,879 (36)
Some college	4,287 (19)	4,275 (19)
College degree or higher	6,533 (28)	5,722 (26)
Missing	221	283
Family income while growing up, N (%)
Poor	1,329 (5.7)	2,202 (10)
Low income	5,472 (23)	6,308 (28)
Middle income	14,975 (64)	12,391 (56)
Well off	1,537 (6.6)	1,389 (6.2)
Missing	52	47
As a child, were there times when your family did not have enough food to eat?, N (%)
Yes	1,310 (5.6)	2,825 (13)
No	22,046 (94)	19,494 (87)
Missing	9	18
Alcohol status at baseline, N (%)
Current drinker	19,102 (82)	18,043 (81)
Former drinker	3,222 (14)	3,578 (16)
Never drank	1,030 (4.4)	710 (3.2)
Missing or Unknown	6	4
Smoking status at baseline, N (%)
Current smoker	1,534 (6.6)	1,982 (8.99)
Past smoker	7,926 (34)	8,370 (37)
Never smoked	13,899 (60)	11,979 (54)
Missing	6	6
Menopausal status at first follow-up, N (%)
Pre-menopausal	4,898 (21)	5,552 (25)
Post-Menopausal	18,454 (79)	16,757 (75)

SD: Standard deviation

^a1,436 missing early life trauma data

^bAmong those parous at baseline (n=38,034)

^cAmong those post-menopausal at first follow-up (n=35,211)

The distribution of early-life trauma events and domains by age period and breast cancer status are reported in eTable 2. Almost one third of participants reported early-life psychological or emotional trauma; 17% of all participants reported emotional or psychological mistreatment by someone close before age 18. Commonly reported domains included sexual trauma (16% overall; perpetrator was someone close (9.5%) versus not close (8.2%)) and household dysfunction (17%), which was commonly reported as serious family problems related to alcohol, drug, or other substance abuse, or mental illness (12%).

We observed no consistent associations between incident breast cancer and (1) cumulative early-life trauma score, (2) most individual traumas and (3) specific early-life trauma domains (Table 3). Given violations of the proportional hazards assumption for cumulative early-life trauma, as well as several trauma domains and individual traumas, we stratified by time-varying menopausal status (eTable 4).

Table 3.

Association between early-life trauma and incident breast cancer (N = 45,702)

	Cases	PY	Age and SEP-adjusted HR (95% CI)a
Early-life trauma cumulative score b	2,237	151,605	1.0 (0.97, 1.0)c
Natural disaster	76	10,276	1.1 (0.86, 1.46)
Major accident	71	10,125	0.97 (0.76, 1.2)
Household Dysfunction	384	53,378	1.1 (0.93, 1.2)c
Witness attack of a family member by another family member	175	23,855	1.1 (0.93, 1.3)
Major issue in personal relationship	13	3,184	0.61 (0.36, 1.1)
Serious financial or legal issues	16	1,994	1.2 (0.71, 1.9)
Serious family drug, alcohol, and mental health issues	258	36,919	0.99 (0.87, 1.1)c
Sexual Trauma	356	50,449	1.0 (0.89, 1.1)
Unwanted sexual contact by someone close	206	30,302	0.97 (0.83, 1.1)
Unwanted sexual contact by someone not close	193	26,424	1.0 (0.89, 1.2)
Physical Trauma	136	17,990	1.1 (0.93, 1.3)c
Hit/attacked by someone close	115	15,414	1.1 (0.90, 1.3)c
Hit/attacked by someone not close	28	4,032	1.1 (0.73, 1.5)c
Psychological Trauma	690	97,400	1.0 (0.92, 1.1)c
Emotionally/psychologically mistreated by someone close	384	53,395	1.0 (0.92, 1.2)
Emotionally/psychologically mistreated by someone not close	52	7,696	0.94 (0.71, 1.2)
Witness suicide/attack of someone close	68	10,533	0.92 (0.72, 1.2)
Witness suicide/attack of someone not close	13	2,763	0.71 (0.41, 1.2)
Death of a sibling	78	9,914	1.1 (0.87, 1.4)
Death of a parent	148	21,694	0.94 (0.79, 1.1)
Death of a close friend	69	11,620	0.87 (0.68, 1.1)
Major illness (not BC) in someone close	90	13,953	0.92 (0.74, 1.1)
Major illness before age 18	49	7,023	0.96 (0.73, 1.3)

BC: Breast cancer; CI: Confidence interval; HR: Hazard ratio; PY: Person-years; SEP: Socioeconomic position

^aAdjusting for age (time scale), race-ethnicity, childhood household education, childhood household income, and childhood food security.

^bEarly-life trauma cumulative score is a continuous measure. All other variables are dichotomous (yes/no).

^cProportional hazards assumption violation observed (p<0.05) for the Wald test of the trauma-time interaction.

The six-class LCA model of early-life trauma provided the best model fit while maintaining a moderate level of certainty that participants were correctly classified (entropy=0.72; eTable 3). eFigure 1a shows the probability plots of the six identified substantive classes: (1) Low early-life trauma (referent; mean probability of reporting each trauma=1.7%), (2) Family health issues, (3) Sexual trauma and family drug/alcohol/mental health issues, (4) High betrayal trauma (i.e., perpetrator was someone close to the participant) and family drug/alcohol/mental health issues, (5) Moderate early-life trauma (mean probability of reporting each trauma=17%), and 6) High early-life trauma (mean probability of reporting each trauma=32%).

Table 4 contains model results assessing early-life trauma via the six-class latent variable. Breast cancer hazards ranged from 1.2 (CI:0.99, 1.4) for participants reporting moderate early-life trauma to 1.1 (CI:0.93, 1.3) for those reporting early-life sexual trauma and family drug, alcohol, and/or mental health issues to 0.66 (CI:0.44, 0.99) for those reporting high trauma compared to participants with low trauma in fully adjusted models. The breast cancer HR for sexual trauma and family drug/alcohol/mental health issues was 1.3 (CI:0.88, 1.8) among pre-menopausal participants and 1.1 (CI:0.88, 1.3) among post-menopausal participants (Table 4). Conversely, the post-menopausal breast cancer hazard for moderate early-life trauma was 1.3 (CI:1.0, 1.5) with an inverse association observed among pre-menopausal women (HR=0.89; CI:0.53, 1.5). Menopause-stratified models demonstrated no effect measure modification (p-value for test of homogeneity=0.59). The HRs for the remaining early-life trauma latent classes were also higher among post-menopausal than those for pre-menopausal participants. Findings for ER+ breast cancer approximated those among post-menopausal women while results for ER− breast cancer were similar to pre-menopausal results (eTable 5).

Table 4.

Association between latent classes of early life trauma and incident breast cancer risk by pre/post-menopausal breast cancer and ER tumor status

		Total Breast Cancer (n=2,315)		Pre-menopausal Breast Cancer (n=337)		Post-menopausal Breast Cancer (n=1,978)
	PY Mean (SE)	Cases Mean (SE)	HR (95% CI)a	Cases Mean (SE)	HR (95% CI)a,b	Cases Mean (SE)	HR (95% CI)a,b
Low early-life traumac	244,025 (24)	1766 (1.2)	1.0 (Ref)	250 (0.24)	1.0 (Ref)	1516 (1.1)	1.0 (Ref)
Family health issues	8,614 (8.8)	61 (0.27)	1.0 (0.78, 1.3)	8 (0.10)	0.63 (0.30, 1.3)	52 (0.26)	1.1 (0.82, 1.5)
Sexual trauma and family drug, alcohol, mental health issues	20,045 (14)	154 (0.47)	1.1 (0.93, 1.3)	34 (0.17)	1.3 (0.88, 1.8)	120 (0.47)	1.1 (0.88, 1.3)
High betrayal trauma and family drug, alcohol, and mental health issues	23,351 (13)	175 (0.64)	1.1 (0.90, 1.2)	24 (0.20)	0.91 (0.59, 1.4)	151 (0.62)	1.1 (0.91, 1.3)
Moderate early-life traumad	16,178 (14)	133 (0.61)	1.2 (0.99, 1.4)	17 (0.15)	0.89 (0.53, 1.5)	116 (0.63)	1.3 (1.0, 1.5)
High early-life traumae	6,043 (11)	27 (0.32)	0.66 (0.44, 0.99)	4 (0.082)	0.55 (0.20, 1.5)	23 (0.34)	0.66 (0.42, 1.0)

CI: Confidence interval; ER: Estrogen receptor; HR: Hazard ratio; PY: Person-years; SE: Standard error

^aAdjusting for age (time scale), race-ethnicity, childhood household education, childhood household income, and childhood food security

^bWald test for differences across strata of menopause status; p-value = 0.59

^cMean probability of reporting each trauma=1.7%.

^dMean probability of reporting each trauma=17%.

^eMean probability of reporting each trauma=32%.

In race–ethnicity-stratified models, breast cancer HR was elevated for moderate early-life trauma among Black participants (HR=1.5; CI:0.85, 2.6; Table 5) compared to low trauma. When we restricted the LCA to Black participants, 85% remained classified within the same substantive latent class despite modified class probabilities of reporting individual traumas (eFigure 1b). We observed no evidence of effect modification by race–ethnicity (p-value for test of homogeneity =0.96). Furthermore, due to the high correlation between race–ethnicity and SEP in the US and the relatively low prevalence of Black participants, we tested potential effect measure modification by childhood SEP, using household income while growing up as a proxy. Breast cancer hazards were higher for early-life trauma latent classes related to family health issues, high betrayal trauma, and moderate compared to low trauma among participants reporting low versus high childhood SEP (p-value for test of homogeneity =0.62; see eTable 6).

Table 5.

Association between latent class measures of early life trauma and incident breast cancer risk by race-ethnicity

	Combined LCA						Black Participant-Specific LCA
	Non-Hispanic white Participants (n=39,852)			Black Participants (n=3,998)			Black Participants (n=3,998)
	PY Mean (SE)	Cases Mean (SE)	HR (95% CI)a,b	PY Mean (SE)	Cases Mean (SE)	HR (95% CI)a,b	PY Mean (SE)	Cases Mean (SE)	HR (95% CI)b
Low early life-traumac	210,782 (92)	1,555 (0.93)	1.0 (Ref)	18,170 (56)	128 (0.42)	1.0 (Ref)	18,260 (65)	128 (0.63)	1.0 (Ref)
Family health issues	7,457 (39)	53 (0.28)	1.0 (0.76, 1.3)	621 (17)	4.4 (0.15)	1.1 (0.40, 3.0)	1,751 (43)	14 (0.24)	1.2 (0.70, 2.2)
Sexual trauma and family drug, alcohol, and mental health issues	16,427 (64)	124 (0.46)	1.1 (0.88, 1.3)	1,979 (37)	13 (0.20)	0.93 (0.51, 1.7)	1,104 (27)	13 (0.35)	1.7 (0.88, 3.3)
High betrayal trauma and family drug, alcohol, and mental health issues	20,464 (55)	152 (0.42)	1.0 (0.87, 1.2)	1,389 (25)	10 (0.22)	1.0 (0.53, 2.0)	1,662 (38)	11 (0.30)	0.90 (0.45, 1.8)
Moderate early-life traumad	13,470 (49)	112 (0.46)	1.2 (0.97, 1.4)	1,462 (30)	15 (0.24)	1.5 (0.85, 2.6)	1,101 (7.7)	6.0 (0.00)	0.81 (0.36, 1.8)
High early-life traumae	4,794 (38)	22 (0.24)	0.67 (0.43, 1.0)	602 (18)	2.3 (0.13)	0.55 (0.13, 2.3)	338 (23)	1.2 (0.09)	0.53 (0.07, 3.8)

CI: Confidence interval; HR: Hazard ratio; LCA: Latent class analysis; PY: Person-years; SE: Standard error

^aWald test for differences across strata; p-value = 0.96.

^bAdjusting for age (time scale), childhood household education, childhood household income, and childhood food security

^cMean probability of reporting each trauma=1.7%.

^dMean probability of reporting each trauma=17%.

^eMean probability of reporting each trauma=32%.

Finally, we assessed the potential modifying effect of childhood social support on the early-life trauma–breast cancer relationship. Although we found no evidence of effect modification (p-value for test of homogeneity =0.48), trends suggested lower breast cancer risk for all early-life trauma latent classes among participants reporting constant childhood social support compared to those without constant childhood support (Table 6). We addressed violations of the proportional hazards assumption among participants reporting less than constant childhood social support in menopause-stratified models; HRs for post-menopausal breast cancer were greater than those for pre-menopausal breast cancer, except for sexual trauma and household drug/alcohol/mental health issues and high betrayal trauma (eTable 4).

Table 6.

Association between latent class measures of early-life trauma and incident breast cancer risk by level of childhood support. (N = 45,606)^a

	Support as a child not all of the time (n = 19,939)			Support as a child all of the time (n = 25,677)
	PY Mean (SE)	Cases Mean (SE)	HR (95% CI)b,c,d	PY Mean (SE)	Cases Mean (SE)	HR (95% CI)b,d
Low early-life traumae	89,585 (12)	644 (0.44)	1.0 (Ref)	147,579 (12)	1,059 (0.35)	1.0 (Ref)
Family health issues	3,582 (4.8)	21 (0.12)	0.88 (0.57, 1.4)	4,783 (3.6)	37 (0.13)	1.1 (0.81, 1.6)
Sexual trauma and family drug. alcohol, and mental health issues	9,634 (5.5)	78 (0.38)	1.2 (0.93, 1.5)	9,845 (7.0)	69 (0.20)	1.0 (0.80, 1.3)
High betrayal trauma and family drug, alcohol, and mental health issues	16,753 (7.5)	132 (0.26)	1.1 (0.93, 1.4)	5,946 (6.3)	37 (0.21)	0.90 (0.64, 1.3)
Moderate early-life traumaf	11,019 (6.5)	92 (0.24)	1.2 (0.99, 1.5)	4,690 (5.7)	37 (0.17)	1.1 (0.81, 1.6)
High early-life traumag	4,406 (5.9)	21 (0.11)	0.75 (0.48, 1.2)	1,420 (4.6)	3.4 (0.11)	0.35 (0.12, 1.1)

CI: Confidence interval; HR: Hazard ratio; PY: Person years; SE: Standard error

^a1,522 participants missing data for childhood support.

^bAdjusting for age (time scale), childhood household education, childhood household income, and childhood food security.

^cProportional hazards assumption violation observed (p<0.05) for the Wald test of the 5 latent trauma-time interactions.

^dWald test for differences across strata; p-value = 0.48.

^eMean probability of reporting each trauma=1.7%.

^fMean probability of reporting each trauma=17%.

^gMean probability of reporting each trauma=32%.

In sensitivity analyses, we limited analyses to participants who completed at least one early-life trauma question included in the Stress and Coping questionnaire (N=45,091); HRs were similar to those observed using the imputed datasets (eTable 7). Secondly, trauma is not limited to early life and can occur across the life course. Less than 2% (n=620) of eligible participants reported experiencing no trauma over the life course and 99% (n=22,607) of participants who reported early-life trauma also reported trauma after age 18. To limit the potential impact of recent trauma on our results, we restricted the sample to participants who reported no trauma during the 12 months prior to stress questionnaire completion (N=27,618) and observed trends comparable to those for the overall cohort (eTable 8). We observed no violations of the proportional hazards assumption in these analyses.

Discussion

To our knowledge this is the first study to comprehensively assess the relationship between multiple measures of early-life trauma and incident breast cancer. We observed no consistent association between early trauma and incident breast cancer based on cumulative scores, individual trauma types, or substantive trauma domains. We further evaluated whether the co-occurrence of individual traumas may differentially affect incident breast cancer. We identified six latent classes of early-life trauma in the Sister Study cohort and observed patterns suggesting that, compared to low trauma, joint sexual trauma and family drug/alcohol/mental health issues may be associated with increased rate of breast cancer overall, regardless of ER or menopausal status. Furthermore, moderate, not high, levels of early-life trauma, were suggestive of increased post-menopausal and ER+ breast cancer risk when compared to low trauma. We saw these trends among participants with less than constant childhood social support, but not among those with constant support. Our results support a complex relationship between patterns of co-occurring trauma, support, and resilience in early life and breast cancer risk in adulthood.

Evaluating the relationship specifically between early-life stress and trauma and breast cancer risk is particularly challenging due in part to the long latency period between exposures in early life and breast cancer development as well as the presence of mediating breast cancer risk factors (e.g., reproductive factors) that may mask persistent underlying effects of early-life trauma. Previous studies have reported an association between childhood sexual abuse and earlier age at menarche^24,26, a suspected breast cancer risk factor¹⁵, but less is known about the potential relationship between early-life trauma and other known breast cancer risk factors, including delayed age at first birth, parity, and time between pregnancies.^39–41

Previous studies have reported positive,^8,42 negative,⁴³ and null^6,17,19 associations between stress and incident breast cancer risk. Variation in the nature of the early-life trauma–breast cancer relationship may also be partially due to breast cancer’s heterogeneous nature, the experience of trauma over the life course, as well as methodologic differences in the measures used to characterize psychosocial stress.^17,19,42,44 Similar to Wise et al., who reported no associations between childhood sexual or physical abuse victimization and breast cancer risk in the Black Women’s Health Study cohort,¹⁹ and Surtees et al., who reported no association between number of childhood difficulties and incident breast cancer in the European Prospective Investigation into Cancer cohort,¹⁷ we observed no clear associations between cumulative early-life trauma scores or individual trauma types and rate of breast cancer. Notably, neither study addressed how clustering of specific trauma types may affect this association.

Cumulative scores and individual traumas may inadequately capture the full impact of experienced early-life trauma, which may vary by race–ethnicity due to persistent social and structural systems that facilitate and perpetuate trauma exposure in early life.⁴⁵ Although substantive classification of latent early-life trauma classes remained largely unchanged in a Black participant-specific LCA, the observed probabilities of reporting individual trauma types varied within given latent classes, suggesting that Black individuals may experience or be susceptible to unique patterns of co-occurring trauma in early life compared with white counterparts. Future studies should carefully consider how differences in ecological context and social disadvantage may contribute to latent measures of early-life trauma.

Across several analyses, high early-life trauma appeared inversely associated with incident breast cancer. The high trauma group represents less than 2% of the cohort, which should be considered when interpreting results. While high early-life trauma may be less common in general, it is also possible that women who would have been classified as high trauma self-excluded from participation in the Sister Study due to competing health complications.⁴⁶

A primary limitation of this study is potential error in the measurement of early-life trauma. Some traumas have discrete initial exposure periods (e.g., natural disasters) whereas others are more chronic (e.g., emotional mistreatment), however both can have persistent health effects. Although we capture different types of early-life trauma, we did not ascertain the frequency, duration, and degree of impact of each type of trauma. Second, we exported latent classes for subsequent analyses, which can increase potential error as analytic models cannot account for error related to probability-based class assignment.^47,48 Future studies should attempt to replicate the utility of latent class measures of early-life trauma and their association with breast cancer incidence. Another limitation of the study was the higher rate of non-response for the Stress and Coping questionnaire by Black and non-Black Hispanic/Latina participants. We used multiple imputation to impute missing values to minimize selection bias due to potential differential non-response. These results were comparable to estimates from analyses limited to participants who completed the Stress and Coping questionnaire, suggesting that selection bias may play a minimal role in the observed breast cancer hazards in the analytic sample

Finally, the Sister Study represents an enriched sample of women with at least one biologic sister diagnosed with breast cancer, a design intended to maximize the likelihood of identifying potential breast cancer risk factors. The Sister Study population is generally older, more educated, and more likely to be healthy and economically advantaged than the general U.S. population, which may contribute to underestimation of the effects of early-life trauma associated with low socioeconomic position. However, the distribution of breast cancer risk factors resembles that of the general U.S. population, suggesting that observed results may still be generalizable and internally valid.^30,49

Strengths of this study include the prospective design of the Sister Study and the comprehensive trauma data. Early-life trauma was assessed prior to breast cancer diagnosis for participants included in the analytic sample. Sensitivity analyses to account for potential recall bias based on recent trauma resulted in similar trends, however, breast cancer hazards were slightly higher across all latent early-life trauma classes among those who reported no recent trauma compared to the overall sample. Furthermore, we assess early-life trauma as the occurrence of trauma during childhood and adolescence whereas prior studies have focused primarily on childhood adversity and breast cancer risk. Studies that limit the scope of experienced trauma to childhood, without considering the effects of subsequent trauma, may overlook the impact of trauma during puberty on reproductive pathways associated with future breast cancer risk. Finally, although other studies have used different methods to evaluate adverse childhood experiences,⁵⁰ this is the first study to our knowledge to use and compare measures of early-life trauma characterized by summative scores, individual traumas, and discreet domains as well as a latent class method that captures co-occurring traumas in the relationship between early-life trauma and incident breast cancer in a single cohort.

In conclusion, we did not observe any consistent associations between early life trauma and incident breast cancer in a large cohort of U.S. women when using cumulative trauma scores, singular trauma types, and trauma domains. However, when using a latent class measure of co-occurring early-life traumas, we observed trends that support an association between two patterns of traumas and incident breast cancer, specifically (1) sexual trauma and family drug/alcohol/mental health issues and (2) moderate ELT. This suggests that nuanced measures that incorporate the co-occurrence of different types of trauma may be more meaningful when assessing incident breast cancer risk.

Availability of data and code:

Data and code are available for replication upon request (www.sisterstudy.niehs.nih.gov)