Data collection and study population
We used data from the Sister Study, a nationwide volunteer cohort of 50,884 U.S (including Puerto Rico) women aged 35–74 years with a sister with breast cancer; enrollment occurred September 2003 to March 2009. This analysis examines early enrollees who completed baseline activities by September 21, 2007 (n=31,409). To avoid errors influenced by eating disorders(16–18)
, participants who reported ever having anorexia or bulimia were excluded (n=1,066). Pregnant women delayed baseline activities until at least three months after the end of pregnancy.
Study participants reported weight (pounds) and height (feet-inches) in a computer-assisted telephone interview (CATI) and separately on a self-administered scannable diet questionnaire. During a home visit, trained examiners used digital self-calibrating scales to measure weight and metal tape measures to measure height. The order of completing the CATI, questionnaire, and home visit varied; self reports could be completed before or after the home visit. All measurements were taken three times without shoes. Measurements were rounded to the nearest whole pound for weight and quarter inch for height. Other variables examined from the baseline CATI were weight cycling (frequency of losing and then gaining ≥ 20 pounds), lowest weight since age 20, heaviest non-pregnant/breastfeeding weight, age, race, education level, perceived health status, marital status, household income, smoking, alcohol, physical activity, gravidity, regular multi-vitamin intake, recency of last medical exam, history of depression, and use of anti-depressant medications.
BMI was categorized using Centers for Disease Control and Prevention definitions(19)
. Lifetime weight difference was calculated by subtracting lowest weight since age 20 from heaviest non-pregnant/breastfeeding weight. All statistical analyses were performed using STATA/IC 10.1(20)
Accuracy of telephone interview (CATI)
To assess the accuracy of self-reported weight and height, we first compared CATI-reported values with examiner measures among women who completed the CATI within 30 days of the home visit (n=18,639). The primary source of Sister Study data is the telephone interview, which had less missing data and fewer structural errors (see below) for height and weight. For this analysis, examiner measures were treated as the true value. Percent agreement and weighted kappa statistics were calculated for each variable of interest. Kappa statistics were weighted according to a standard weight in STATA to account for the degree of disagreement. Polytomous logistic regression was used to calculate odds ratios (ORs) and 95% confidence intervals (95% CIs) for reporting accuracy by age, race, education level, perceived health status, marital status, and measured BMI.
To be consistent with the existing literature, we first examined the absolute difference between self-reported and measured weight. Differences between measured and self-reported weight were categorized as under-reporting ≥ 7 pounds, under-reporting 4 to 6 pounds, reporting within 3 pounds, and over-reporting ≥ 4 pounds. Because the relative impact of a specific weight difference will be greater in smaller than larger women, we also calculated the percentage of weight mis-reported; self-reports that differed by less than 5% from measured weights were the referent category. Polytomous logistic regression models explored the effects of measured BMI, weight cycling, lifetime weight difference, and current anti-depressant use on under- and over-reporting, adjusting for age, race, education, perceived health status, and marital status as potential confounders. Models examining weight cycling, lifetime weight difference, or current anti-depressant use also adjusted for measured BMI. Differences between measured and self-reported height were categorized as under-reporting >1 inch, reporting within 1 inch, and over-reporting >1 inch.
To determine the effect of misreporting on BMI categories, we compared categories calculated from CATI-reported data with categories based on examiner-measured data using percent agreement and weighted kappa statistics for all women and stratified by categories of age, race, education level, perceived health status, and marital status. We also determined the sensitivity and specificity of self-reported overweight/obese classification relative to examiner-measured data. To further explore the potential for bias in BMI we stratified on measured BMI and examined the percentage of CATI-determined BMI values that over- or underestimated BMI calculated from examiner measured values.
We carried out additional analyses stratifying by or adjusting for which measure came first, the home visit or CATI.
Accuracy of self-completed questionnaire
Using data from the subset of women with CATI and questionnaire completed within 30 days of the home visit (n=13,985), we carried out similar analyses to assess the accuracy of weight and height reported in the self-completed questionnaire compared with examiner measured data. We then compared the accuracy of the two self-report measures by calculating ratios of OR from models assessing reporting by CATI or questionnaire versus measured data. An analysis including all women (n=21,935) completing the diet questionnaire within 30 days of the home visit had similar results and is not shown.
Reliability of self-reported weight and height was assessed using percent agreement and weighted kappa statistics to compare self-reported data from the CATI and diet questionnaires. Analyses were limited to women who completed the CATI within 30 days of submitting their questionnaire (n=13,316) and had non-missing questionnaire data for weight (n=11,585) and height (n=11,885). Similar to the accuracy analysis, we stratified and adjusted analyses by reporting order with respect to each other and with respect to the examiner measurement.
Correcting structural errors
Prior to analyses, we identified and corrected several problems inherent to the reporting method. Both random and systematic errors occurred with the self-administered diet questionnaire. About 1% of respondents appeared to make frameshift bubbling errors for weight and/or height by mistaking the bubbles in one or more columns as starting at 1 instead of 0. shows a frameshift error in which the respondent filled in the wrong value for weight in the tens place and the wrong values for height in the feet and inches columns. Frameshift errors occurred frequently in the hundreds place of weight, which were detected when an unreasonable weight (<100 pounds) was marked (e.g., 34 pounds instead of 134). We corrected obvious frameshift errors (0.7% of weight values and 0.1% of height values) when questionnaire values differed from both the CATI and examiner reports by >60 pounds or 11 inches.
Example of frameshift errors on self-administered diet questionnaire
Some errors were related to the choice of unit. In the diet questionnaires, a small percentage of respondents appeared to report height in total inches rather than feet and inches as instructed. For example, instead of 5 feet–4 inches, a respondent marked the total inch equivalent (64 inches) which was then mistakenly interpreted as 6 feet–4 inches. We corrected these unit errors in about 0.8% of all responses by checking suspiciously high reports and confirming corrections with CATI and examiner reports. Although these errors occurred for units (inches, pounds) used in the U.S., similar errors could occur for those used in other countries (e.g., meters, kilograms).
There were considerable missing values for weight (13%), height (11%) or both (8%) in the self-administered diet questionnaires. Non-response did not substantially vary by age or BMI category. Missing weight and height were uncommon in the CATI (<1%).
There seemed to be a tendency to round to 0 or 5 when reporting weight in the CATI (59%) and questionnaires (52%), whereas an end digit of 0 or 5 occurred in 27% of examiner measures. We did not correct for this apparent rounding.
We detected infrequent random reporting errors for all modes of reporting. In self-administered questionnaires, random bubbling errors such as pencil smudges were sensitive to the questionnaire scanner. For the CATI, there were occasional data entry errors by interviewers and for examiners, some inconsistencies following measurement protocols. We corrected CATI values if they greatly differed from both examiner and questionnaire values (≥100 pounds for weight; ≥11 inches for height).