The Black Women’s Health Study is an ongoing U.S. prospective cohort study of 59,000 African-American women aged 21–69 at entry (27
). In 1995, Essence
magazine subscribers (93.6%), black members of two professional societies (2.5%), and friends and relatives of early respondents (3.9%) responded to a mailed invitation to enroll in a long-term health study by completing a self-administered baseline questionnaire. Every two years, cohort members update their exposure and medical histories by questionnaire; follow-up of the baseline cohort has averaged over 80% across seven cycles of follow-up. Study participants reside in more than 17 states, with the majority residing in New York, California, Illinois, Michigan, Georgia, and New Jersey. The institutional review board of Boston University Medical Center approved the study protocol.
Assessment of outcome
On the 1999 and 2001 follow-up questionnaires, women reported whether they had been diagnosed with “uterine fibroids” in the previous two-year interval, the calendar year in which they were first diagnosed, and whether their diagnosis was confirmed by “pelvic exam” and/or by “ultrasound/hysterectomy.” On the 2003, 2005, 2007, and 2009 follow-up questionnaires, “hysterectomy” was replaced by “surgery (e.g., hysterectomy)” to capture women with other surgeries (e.g., myomectomy), and “ultrasound” and “surgery” were asked as two separate questions. Cases were classified as “surgically-confirmed” if they reported a diagnosis by “ultrasound/hysterectomy” (<2003) or “surgery” (≥2003 questionnaires) and also reported “hysterectomy” under a separate question on that respective questionnaire.
We included cases diagnosed by ultrasound and surgery because surgical cases represent only a fraction of all cases and studies of surgical cases may spuriously identify risk factors associated with disease severity or treatment preference (28
). Ultrasound has high sensitivity (99%) and specificity (91%) relative to histologic evidence (29
). To maximize specificity, pelvic exam cases (N=542) were treated as non-cases (31
Assessment of intrauterine and early life exposures and covariates
In 1997, women reported their country of birth and that of their parents, their birth weight (<4 lbs, 4lbs–5 lbs 8ozs, >5 lbs 8 ozs; or exact birth weight, if known), whether they were born preterm (“3 or more weeks early”), and whether they were a twin or triplet (identical vs. fraternal). Women also reported whether they were in the same room with a smoker for ≥ 1 hour/day for ≥1 years at home (age 0–10). In 2005, women reported the age at which their mothers gave birth to them and their state of residence at birth and age 15. In 2007, women reported how many brothers and sisters (half or full) they grew up with (sibship size) and how many were older than them. In 2009, women reported the highest level of education completed by their mother and father (<12th grade, high school degree or GED, some college or vocational school, college graduate or higher, don’t know/not applicable), whether their home was rented or owned in childhood (age 0–11), whether they were breastfed as an infant (no, yes, don’t know), duration of breastfeeding (months), and whether their mother smoked cigarettes while pregnant with them (no, yes, don’t know).
The baseline (1995) and biennial follow-up questionnaires collected data on several known or suspected risk factors for UL, including reproductive and contraceptive history; weight and height; lifestyle factors (smoking, alcohol, physical activity); geographic region of residence; socioeconomic correlates (education, marital status, occupation, household income); medical history; and gynecologic surveillance (recency of pelvic exam, ultrasound). Reproductive factors, weight (to estimate body mass index (BMI), kg/m2), smoking, marital status, physical activity, and region were updated on follow-up questionnaires and were analyzed as time-dependent covariates.
We assessed the accuracy of self-reported UL in a random sample of 248 cases diagnosed by ultrasound or surgery. Cases were mailed supplemental questionnaires regarding their initial date of diagnosis, method(s) of confirmation, symptoms, and treatment, and were asked for permission to review their medical records. We obtained medical records from 127 of the 128 women who gave us permission and corroborated the self-report in 122 (96%). Among the 188 (76%) providing supplemental survey data, 71% reported UL-related symptoms prior to diagnosis and 87% reported that their condition came to clinical attention because they sought treatment for symptoms or a tumor was palpable during a routine pelvic exam. There were no appreciable differences in UL risk factors by release of medical records (32
Early life factors
We validated self-reported data on infant birth weight, birth order, maternal age, and parental education against birth certificate data from the Massachusetts Department of Public Health among the 637 BWHS participants born in Massachusetts. Spearman correlation coefficients comparing self-reported vs. registry-supplied data were 0.99 for maternal age, 0.87 for birth weight, 0.78 for birth order, 0.66 for maternal education, and 0.68 for paternal education. Agreement for parental education may be greater than estimated because education could have increased since the participant’s birth.
We also assessed the reliability of early life data among women who returned duplicate questionnaires in a given follow-up cycle. High agreement was found for being breastfed (kappa (k)=0.93), months of breastfeeding (Spearman correlation (r)=0.77), maternal age (r=0.96), sibship size (r=0.93), birth order (r=0.94), maternal smoking (k=0.89), passive smoking in childhood (k=0.73), maternal education (k=0.92), paternal education (k=0.93), preterm birth (k=0.86); and birth weight (categorical: k=0.86; exact: r=0.96). The time that elapsed between receipt of duplicate questionnaires ranged from 0 to 557 days (median=61 days), and there were no appreciable differences in correlations by time (data not shown).
Of the 53,126 respondents to the 1997 questionnaire, we excluded women who were ≥50 years (N=11,475) or postmenopausal (N=6,580), women diagnosed with UL <1997 (N=9,846), those lost to follow-up >1997 (N=915), cases without a diagnosis year (N=139) or detection method (N=112), and women with missing covariate data (N=554), leaving 23,505 women for analysis. Those excluded had lower educational attainment than those included, but were similar with respect to parity, age at menarche, and other risk factors for UL.
Person-years were calculated from March 1997 until UL diagnosis, menopause, death, loss to follow-up, or the end of follow-up (March 2009), whichever came first. Age- and time period-stratified Cox regression models were used to estimate incidence rate ratios (IRR) and 95% confidence intervals (CI) for the associations of interest. Exposure variables were categorized in their original form or according to their frequency distribution in the analytic sample. We constructed two multivariable models: the first controlled for age (1-year intervals) and questionnaire cycle, while the second additionally controlled for potential early life confounders, including nativity (born outside the U.S., native born but ≥1 parent born outside U.S., native born with neither parent born outside of U.S.), birth weight (<2,000, 2,000–2,499, ≥2,500 grams), birth order (first-born, later-born), maternal age at participant’s birth (<20, 20–24, 25–29, ≥30 years), and parental education (highest educational attainment by either parent: <12, 12, 13–15, ≥16 years). Secondary analysis involved additional control for adult risk factors that potentially mediate the associations, including age at menarche (years), parity (0, ≥1 births), age at first birth (years), years since last birth (<5, 5–9, 10–14, ≥15), age at first oral contraceptive use (years), history of oral contraceptive use (ever, never), BMI (<20, 20–24, 25–29, ≥30 kg/m2), smoking (current, former, never), current alcohol intake (<1, 1–6, ≥7 drinks/week), participant’s education (≤12, 13–15, 16, ≥17 years), marital status (married/living as married, divorced/separated/widowed, single), occupation (white collar, non-white collar, unemployed), annual household income (≤$25,000, $25,001–50,000, $50,001–100,000, >$100,000, missing), region of residence (South, Northeast, Midwest, West). Because results were largely consistent with the childhood-adjusted model, we did not present mediator-adjusted results.
Tests for trend were conducted by modeling the ordinal categorical version of the exposure and evaluating the associated Wald test statistic (33
). P-values from interaction tests were obtained using the likelihood ratio test comparing models with and without cross-product terms between covariate and exposure variables. Departures from the proportional hazards assumption were tested by comparing models with and without cross-product terms between each exposure, age (<30, ≥30), and questionnaire cycle (1997–2003, 2003–2009). Analyses were performed using SAS statistical software version 9.2 (34
Missing data ranged from 1% (country of birth for respondent and parents) to 36% (maternal age), with the latter variable’s missingness attributable to its omission from the shortened 2005 questionnaire mailed to late-responders. Given the large proportion of women with missing data on at least one early life factor, secondary analyses were also conducted using multiple imputation (35
). This involved using PROC MI in SAS to create five imputed datasets—including all known or suspected risk factors for UL in the imputation—and then combining results across imputed datasets using PROC MIANALYZE (34
). Because both methods produced similar results (available upon request), we present the missing indicator method as our primary analysis.
To increase the sensitivity of disease classification, we repeated analyses after restricting non-cases to women with a recent ultrasound (<5 years ago). We also conducted secondary analyses that used the NIEHS Sister Study’s UL case definition for black women (17
): diagnosis before age 30, including prevalent and incident diagnoses regardless of detection method (N=6,642). Cases diagnosed ≥30 years were excluded from analysis. We used log-binomial regression to estimate prevalence ratios (PR) for the association between early life factors and UL.