Associations between Personal Care Product Use Patterns and Breast Cancer Risk among White and Black Women in the Sister Study

Participants reported breast cancer diagnoses on annual and biennial health questionnaires or by calling the Sister Study helpline. Women who reported an incident breast cancer during follow-up were asked to authorize release of pertinent medical records. Response rates were $>94\%$ over follow-up (Nichols et al. 2013). Among participants in our sample population, 2,326 breast cancers were reported for 304,034 person-years (average follow-up $\sim 5.4\text{years}$). At the time of the present analysis, pathology reports or medical records had been obtained for $>80\%$ of these cases ($n=\mathrm{1,923}$). Confirmation of self-reported breast cancer diagnoses by medical record was very high [$\text{positive}\text{predictive}\text{value}\left(\text{PPV}\right)=99.5\%$] (NIEHS 2010). After medical record review, self-reported ER status information was confirmed for 99% of ER-positive cases and 85% of ER-negative cases. Because agreement between self-reported and medically abstracted data was high, we used self-reported tumor information when medical records were not available.

Information on the use of 48 personal care products was self-reported during the enrollment phase of the study (see Table S1) by inquiring about frequency of use (five-level response option) during the previous 12 months. The five response options in the questionnaire varied according to intended use of the product. For example, the response options for a product intended to be used regularly (e.g., hand lotion) included: a) did not use, b) used less than once a month, c) used 1–3 times per month, d) 1–5 times per week, e) $>5$ times per week. Response options for products that are used less often (e.g., hair dye) included: a) did not use, b) 1–2 times a year, c) every 3–4 months, d) every 5–8 weeks, e) once a month or more. Because PROC LCA requires the same number of response options for each item (i.e., each personal care product), and to ensure that each of the response options had an adequate sample size ($\ge 10\%$ of the sample population), the five original response options were condensed into three (rarely/never used, moderate use, and frequent use); this was done for each product individually (see Figure S1) and was based on the distribution of participants who fell into each frequency-of-use option. Products were then categorized as beauty ($n=14$), hair ($n=15$), or skincare ($n=19$) products, and separate latent classes were defined for each product category. As described in more detail in our previous paper (Taylor et al. 2017), we performed latent class analyses (LCAs) using PROC LCA (Lanza et al. 2015) and SAS statistical software (version 9.3; SAS Institute Inc.). PROC LCA is an SAS procedure for latent class analysis (LCA) developed by the Methodology Center at Penn State (https://methodology.psu.edu/downloads/proclcalta). It allows the user to preprocess data, fit a variety of latent class models, and postprocess the results, all within SAS.

Latent classes within each product category were defined by item-response probabilities (Dean and Raftery 2010) for the products driving each class (Lanza et al. 2007). To identify the classes, we fit a sequence of LCA models starting with two classes and increasing up to six for each model, and we used Akaike’s information criterion (AIC), the Bayesian information criterion (BIC), and entropy (Lanza et al. 2007) to select the optimum number of classes. We used a common classify-analyze approach (the maximum-probability assignment rule) to assign each participant to the class in which she had the highest posterior probability of membership (Bray et al. 2012). To reduce dimensionality and improve interpretability, classification, and precision, each product category included in our model was limited to the personal care products that were most useful for distinguishing between latent classes (i.e., $\ge 10\%$ difference in posterior probabilities between classes), as described by Dean and Raftery (2010). Within each product category, latent classes were described and considered as exposure groups. Women with missing data for any of the individual products included in the corresponding product category (i.e., beauty products, hair products, and skincare products) were classified as missing for a latent class assignment in that product category and were not included in statistical analyses.

The present analysis was limited to non-Hispanic white ($n=\mathrm{42,447}$, 91%) and non-Hispanic black ($n=\mathrm{4,450}$, 9%) women (Table 1). Multivariable Cox proportional hazards models were used to estimate adjusted hazard ratios (adjHRs) and 95% confidence intervals (CIs) for associations between the personal care product latent classes and breast cancer risk. Statistical models used age as the time scale, where participants entered the analysis at their enrollment age (left-truncation) and accrued person-time until they exited at their cancer diagnosis or were administratively censored at their age at last follow-up. Women who reported that they had undergone natural menopause, bilateral oophorectomy, irradiation to the ovaries, or otherwise reported cessation of menstruation were classified as postmenopausal; women who reported that they were still cycling were classified as premenopausal. In analyses investigating associations by menopausal status at the time of breast cancer diagnosis, the person-time of women who became postmenopausal during the follow-up period was counted as premenopausal time at risk up until menopause (after which it was censored for the premenopausal analysis), and subsequent person-time after menopause was counted as postmenopausal person-time at risk. The proportional hazards assumption was visually assessed using ln-ln survival plots; there was no suggestion of time-variant associations.

Models were stratified by race, menopausal status, or both at time of diagnosis or follow-up. The following covariates, measured at baseline, were included in adjusted models: menopausal status (premenopausal or postmenopausal), age at menarche ($<12\mathrm{y}$ or $\ge 12\mathrm{y}$), age at first birth (nulliparous, $<26\mathrm{y}$, or $\ge 26\mathrm{y}$), parity (nulliparous, 1–2 children, or $\ge 3\text{children}$), duration of breastfeeding ($<12\text{mo}$ or $\ge 12\text{mo}$), oral contraceptive (OC) use (ever or never), hormone therapy (HT) use (ever or never), education ($<\text{high}\text{school}$ or $\ge \text{high}\text{school}$), alcohol consumption (never drinker, former drinker, currently drink $<1\text{drink}/\mathrm{d}$, or currently drink $\ge 1\text{drink}/\mathrm{d}$), adult body mass index (BMI) ($<25{\mathrm{kg}/\mathrm{m}}^2$, 25 to $<30{\mathrm{kg}/\mathrm{m}}^2$, or $\ge 30{\mathrm{kg}/\mathrm{m}}^2$), family history (having one sister with breast cancer or $\ge 1$ sister and/or a mother with breast cancer), smoking status (never smoker, former smoker, current smoker), and current region of residence (West, South, North, East). Interaction by menopausal status was tested by adding an interaction term for menopausal status and latent class into the Cox proportional hazard models stratified by race. In sensitivity analyses among postmenopausal white women only (the only subgroup with sufficient numbers for this analysis), models were stratified by ER status (ER positive or ER negative according to the clinical record) and breast cancer type (in situ or invasive).

For all analyses, results are presented only for the personal care product latent classes that included $\ge 20$ exposed breast cancer cases.