Author's response to reviews Title: Body fatness and breast cancer risk in women of African ancestry Authors: Elisa V Bandera (elisa.bandera@rutgers.edu) Urmila Chandran (chandrur@cinj.rutgers.edu) Gary Zirpoli (Gary.Zirpoli@RoswellPark.org) Zhigong Gong (Zhihong.Gong@RoswellPark.org) Susan E McCann (Susan.McCann@RoswellPark.org) Chi-Chen Hong (Chi-Chen.Hong@RoswellPark.org) Gregory Ciupak (Gregory.Ciupak@RoswellPark.org) Karen Pawlish (Karen.Pawlish@doh.state.nj.us) Christine B Ambrosone (Christine.Ambrosone@RoswellPark.org) Version: 5 Date: 25 September 2013 Author's response to reviews: see over
To: Prof. Steinar Tretli BMC Cancer From: Elisa V. Bandera, MD, PhD Associate Professor, Department of Medicine Rutgers Cancer Institute of New Jersey elisa.bandera@rutgers.edu 732-235-9874 Date: September 20, 2013 RE: Manuscript MS: 2038651531929437-Second revision Dear Professor Tretli: Thank you for allowing us to respond to further comments by the second reviewer, Dr. Patricia Sheean, on our manuscript Body fatness and breast cancer risk in women of African ancestry. Our point-by-point responses to her comments are shown below. We appreciate the opportunity to further improve our manuscript to journal readers. Best regards, Elisa V. Bandera, MD, PhD Associate Professor of Epidemiology Division of Medical Oncology/Population Science Program The Cancer Institute of New Jersey UMDNJ-Robert Wood Johnson Medical School 732-235-9874 elisa.bandera@rutgers.edu
RESPONSE TO REVIEWER'S REPORT Title: Body fatness and breast cancer risk in women of African ancestry Reviewer: Dr. Patricia Sheean Overall, the revised manuscript is improved. The language is clearer, tighter and reflects significant efforts on behalf of the authors. However, I have several remaining concerns, largely focused on methodologies and objective limitations. Major compulsory revisions: Statistical methodologies: The rationale for why investigators adjust for BMI while looking at WC (and vice versa) is understood. Regardless of the authors response, this approach reflects a statistical overcorrection and as a result, the findings are a bit misleading. It is comparable to examining education, adjusting for income. These two variables, similar to BMI and WC, are typically highly collinear. To this end, it would be helpful for the authors to include comments regarding how they addressed multicollinearity in their regression modeling within the statistical analyses, data presentation and data interpretation. I disagree that interpretations should be left to the reader. Response: We understand Dr. Sheean s concern about potential multicollinearity from including both BMI and waist circumference in the same model, and the potential for this approach to inadvertently introduce bias. To address this, we have carried out in depth additional analyses to examine the impact of assessing BMI simultaneously with WC, and determine if issues with multicollinearity are impacting our risk estimates. We note that a high correlation between variables indicates potential multicollinearity, but also strong confounding and the need to adjust for it. Since the outcome is binary, the options for testing multicollinearity such as variance inflation factor (VIF) and tolerance do not directly apply. These are available only with the PROC REG option. However, a general recommendation to check multicollinearity in logistic models is to run a similar model in linear regression and obtain collinearity statistics. When we obtained tolerance and VIF statistics using this approach, the VIFs for BMI and waist circumference were over 4, indicating that the respective standard errors of these variables are twice as inflated relative to the absence of multicollinearity. However, when we checked the standard errors of these two variables in the logistic model, they were 0.01 for waist circumference and 0.02 for BMI, indicating that there was no substantial inflation, and suggesting that multicollinearity, if any, is not a severe issue in this model. Another sign of multicollinearity is if the global model is significant but none of the individual covariates are significant. In our model that included waist circumference and BMI, the coefficients of both variables were statistically significant (p 0.05). Lastly, we obtained significant associations both when waist circumference was treated in linear and non-linear forms. Put together, these results point to an issue of confounding rather than moderate or severe multicollinearity. We think that it is important to evaluate the independent effects of WC and BMI since WC is increasingly recognized as an independent risk factor from overall obesity, and this approach is now fairly standard within observational epidemiologic studies. This topic is also addressed in Dr. Walter Willett s textbook Nutritional Epidemiology, third edition. He explains that when WC or WHR are included in the model as independent variables with BMI as a covariate, BMI becomes more of an index of lean mass than fatness in this multivariate model because the component of weight due to fatness is accounted for by abdominal circumference. Therefore, the evaluation of WC adjusted for BMI may be a better assessment of the association with
adiposity. Given this, it makes sense that the effect of adjusting for BMI was more pronounced in premenopausal women, who tend to have more lean mass. As mentioned in the manuscript, our findings are in agreement with findings from other studies that have evaluated WC adjusted for BMI (the Carolina Breast Cancer Study and the Nigeria Breast Cancer Study). We have added this text to the manuscript, as requested. Sources: 1. Willett W. Nutritional Epidemiology, Third Edition, Oxford University Press 2013. 2. Kleinbaum, Kuper, Muller. Applied Regression Analysis and Other Multivariable Methods. Second Edition. PWS-Kent 1988. 3. Allison, PD. Logistic Regression using SAS. Theory and application. Second Edition. SAS Institute Inc 2012. BIA methodologies: Again, it is clear why the BIA methodologies were used in this study and I certainly appreciate the reluctance to provide participants with a list of to do s prior to measurement and interview. However, the potential to produce highly variable results are substantial when using a methodology that relies on hydration and may even vary by race/ethnicity. Please provide a more objective view of these data, including the likelihood that the body composition relationships (Table 4) may have been masked by using an imprecise method of measurement. The verbiage on page 13, paragraph 1 is insufficient. Need to consider and include: What biases might have been introduced as a result of using BIA? How could these imprecise measures impact your findings? Response: We have added that potential random misclassification of body composition measures may have affected our ability to detect an association in that paragraph. Minor essential revisions Pg 7, P2: Understanding there is great variability is WC cut-points, please provide a reference for using 2 cm as your cut point for taking a third measure of waist circumference. These sorts of references are important for unifying this methodology. Response: We have added references, as suggested. Pg 14, P2: The lack of data on co-morbid condition should be mentioned in the limitations, especially in light of the associations with BC risk and WC in this study population. Response: We made this change. Conclusive remarks: Seems like the conclusions actually start on pg 14 with the paragraph starting, Our study contributes to the... Suggesting integrating this material with the concepts currently under Conclusions. Response: We made this change.
Last sentence of the current conclusions is too long; suggest shortening and revising. Response: We revised, as suggested. Conclusive remarks within the abstract should be revised to be more objective in light of the methodologic limitations. Response: We revised the conclusions in the abstract, as suggested.