Latent class analysis suggests four distinct classes of complementary medicine users among women with breast cancer

Data for this study were obtained from the Long Island Breast Cancer Study Project, a federally mandated case–control study to investigate the high incidence of breast cancer in Nassau and Suffolk counties on Long Island, New York [20]. Cases were identified and interviewed shortly after diagnosis in 1996–1997 and, as part of a continuing prospective follow-up study, a second interview among cases was conducted in 2002–2003, which included a questionnaire assessing CAM use. This study was approved by the institutional review boards of Columbia University Medical Center, University of North Carolina, and other collaborating institutions.

Female residents of Nassau and Suffolk counties, New York, newly diagnosed with a first primary in situ or invasive breast cancer were identified through a rapid reporting system between August 1996 and July 1997, as previously described [20]. Physician consent was obtained for 1837 cases, 1508 women (82.1 %) completed the baseline questionnaire, and 1414 agreed to subsequent contact. The follow-up interview was conducted with cases or their proxies during 2002 and 2003 [21]. Only data from the 764 patients who personally completed the full follow-up questionnaire were analyzed in this study. Non-responders to the full follow-up questionnaire were on average older, had less education and lower household incomes, were less likely to be non-Hispanic white, more likely to have invasive disease, and less likely to have had a recent mammogram, than responders, as previously reported [10]. In the current sample, 94 % of participants were non-Hispanic and white, and 82 % had invasive (vs in situ) breast cancer [10].

The baseline interview, conducted by trained interviewers roughly three months after diagnosis (mean 96 days), included information on demographics, lifestyle, and breast cancer risk factors. The follow-up interview included questions about the first course of treatment for the first primary breast cancer, and CAM use before and after diagnosis and during treatment. Signed medical records release forms were obtained from case women to abstract data relating to tumor characteristics and treatment during the baseline and follow-up interviews. Informed consent was obtained from all participants before each interview.

The CAM section of the follow-up questionnaire included detailed questions about 194 modalities in 7 domains that were developed after a review of the literature: 1) vitamin/mineral supplements; 2) herbs/botanicals; 3) non-vitamin/mineral non-herbal over-the-counter (OTC) health products; 4) mind-body medicine techniques; 5) special treatments (including biofeedback, colon cleansing and others); 6) diet change; and 7) practitioner-based therapies. After an affirmative response to ever use of a particular product, participants were asked specifically about the time following diagnosis through the question, “How many total years have you taken this product since diagnosis?” Frequency of using each endorsed modality during treatment for breast cancer was then assessed at the day, week or month level. Although some studies have included prayer as a CAM modality, we excluded it because a high proportion of participants reported using it and because a separate study had shown that prayer fit into a latent construct different from mind-body medicine techniques [14]. We also excluded special treatments such as biofeedback, colon cleansing, and bioelectromagnetic-based therapy due to very low prevalence of use (1.2 %).

We computed a “cumulative dose index” of each CAM therapy by multiplying the reported number of years used since diagnosis by the reported number of times used per week or month. The cumulative dose index value for vitamins/mineral supplements, herbs, and over-the-counter health products was the number of times taken per day multiplied by years taken; the value computed for mind-body practices and CAM practitioners was the number of times used per month multiplied by number of years used; the value for diet was the number of years used since diagnosis.

Covariates assessed by questionnaire at baseline included age at diagnosis, education, annual household income, marriage history, religion, and race/ethnicity. Known breast cancer risk factors assessed include use of hormone replacement therapy or oral contraceptives, and any first degree family history of breast cancer. Medical history and mammography were also assessed at baseline, in addition to physical activity between menarche and diagnosis, fruit and vegetable intake assessed through a modified Block food frequency questionnaire, life course alcohol use and cigarette smoking history. An unweighted comorbidity index adapted from the Charlson Comorbidity Index [22] was created. An affirmative response at the baseline interview to a history of each of the following conditions contributed 1 point to the index: diabetes, asthma, myocardial infarction, stroke, gallbladder disease, and previous cancer (0, 1, 2 or more). At the follow-up interview, information was collected on weight and height at diagnosis to estimate body mass index (BMI = kg/m²), as well as the first course of treatment (chemotherapy/radiation/hormone therapy) for the first primary breast cancer.

Medical records were abstracted at baseline and again at follow-up to ascertain information on tumor staging (in situ vs. invasive), tumor estrogen/progesterone receptor (ER/PR) status, and first course of treatment for the first primary breast cancer. Concordance between the treatment data collected by interview and from medical records was exceptionally high (kappa >90 %) [21], and thus the information assessed by interview is used here.

The analysis included the 23 modalities used since diagnosis at any dose by at least 10 % of women. Modalities used by fewer than 10 % of women were aggregated into six composite variables according to specified CAM domains. Further, in order to increase the precision of the latent class analysis, factor analysis was performed as a data reduction technique prior to conducting the LCA. A two-step process was used for the factor analysis. Because factor analysis of skewed dichotomous data using Pearson correlation coefficients has been shown to underestimate factor loadings [23], we first computed tetrachoric correlations between dichotomous CAM variables [24]. We then conducted the factor analysis on the tetrachoric correlation matrix using principal axis factoring extraction with oblique factor rotation, and determined the number of factors to be retained using the scree test [25]. Modalities that loaded on a single factor with a factor loading of 0.3 or higher on common factors were grouped together. Each CAM grouping generated through factor analysis represented a subset of modalities with correlated use. The cumulative dose of each grouping was calculated by summing the cumulative dose of all modalities making up the respective grouping. Those not loading on any single factor were retained as separate modalities for use in the LCA.

The summed cumulative dose index for each individual modality or CAM grouping was then categorized into three levels with a cutoff at the median dose among those reporting any use: (1) no use since diagnosis; (2) use since diagnosis at a cumulative dose below the median; and (3) use since diagnosis at a cumulative dose above the median. Latent class analysis was performed with SAS PROC LCA [26] using the consolidated 3-level CAM grouping variables. Models specified to contain one through ten latent classes were evaluated based on Akaike’s Information Criterion (AIC), Bayesian Information Criterion (BIC), and the sample-size adjusted BIC. We also considered the reproducibility of the model, as defined by the percentage of iterations yielding the optimal fit, as well as interpretability of the latent classes, when determining the number of latent classes [27].

Using Bayes’ theorem, SAS PROC LCA computes each individual’s posterior probability of membership in each latent class. Subsequent to determining the best fitting number of classes, participants were assigned to classes based on maximum posterior probability. Bivariate associations of demographic, clinical and behavior variables with assigned latent class were examined using Pearson’s chi-squared tests or analysis of variance (ANOVA) for categorical and continuous variables, respectively. Analyses were performed with SAS v.9.3 (SAS Institute Inc., Cary, NC) using two-sided significance tests.