Diabetes in US women on the rise independent of increasing BMI and other risk factors; a trend investigation of serial cross-sections

The analyses utilized data from the Behavioral Risk Factor Surveillance System (BRFSS) for the years 2006–2010. For over two decades, the BRFSS datasets, collected via random digit telephone surveys of more than 400,000 households by the health departments of all 50 states, the District of Columbia, Guam, Puerto Rico and the US Virgin Islands have been endorsed as one of the most generalizable primary data sources for investigating numerous health conditions that contribute to morbidity and mortality in the Nation [14]. The BRFSS data sets have been widely utilized in a milieu of reports and articles. Of recent, these data have been used in studies associated with cardiovascular heart diseases [15], physical activity [16], cancer screenings [17], mental health [18] obesity [19], diabetes mellitus [20], and multiple preventive health behaviors [21, 22].

A population-based female sample of 1,291,803 (62%) from the BRFSS annual survey data 2006–2010, were used for these analyses (males were excluded for these analyses due to the focus on females). The sample population included 18–64 year olds across racial categories. To assess diagnoses of diabetes over the 5 year period, we concatenated datasets by variables and year of survey. All male respondents were excluded, as well as female respondents with “missing” and “refused” responses. Of the total 1,291,803 selected in this population there were 123, 385 (10%) excluded due to missing or refused responses. Though proportionally small, due to the size of the missing data there were significant differences between those with complete data and those with incomplete data. However, the proportional differences remained small. The investigators sought National University Institutional Review Board review and were determined to be exempt based on secondary data analyses on public use data [23].

The outcome of interest was a self-reported diagnosis of diabetes. Participants in this study were indicated with diabetes with affirmation of the question, “had ever been told by a doctor that they had diabetes”. Participants who had been told that they had diabetes only during pregnancy and those reporting prediabetes or borderline diabetes were classified as not having diabetes.

The variable age was represented by the age groups in years: 18–24, 25–34, 35–44, 45–54, 55–64, and 65 and over. Exercise was measured as “physical activity or exercise in the last 30 days other than their regular job” indicated as “yes” or “no”. Body mass index was defined as weight in kilograms divided by the square of the respondent’s height in meters (kg/m²) and grouped into adults with body mass index at or below 25 (BMI ≤ 25) and adults with body mass index greater than 25 (BMI > 25). Race was categorized as white non-Hispanic, black non-Hispanic, Hispanic, and other non-Hispanic ethnicities. The variable year was created by indicating which year of the BRFSS the data point was associated with (2006–2010).

Univariate analyses including chi-square tests were used to test for an association between body mass index (BMI), age, race and year of survey response. Multivariable logistic regression was used to estimate the odds of association between BMI and diagnoses of diabetes, independent of other variables. Logistic regression was also used to determine a net-effect of multiple variables on diabetes. Statistical significance was measured with 95% confidence intervals and p-values < 0.05. SAS software, version 9.2, was used for data management and statistical analyses (SAS Institute, Inc., Cary, North Carolina).

Several limitations to this study should be noted. First, diabetes was determined using self-reported data and was not confirmed by physician diagnosis or medical record review. While self-report of physician diagnosed diabetes has been shown to be ideally reliable in population studies where fasting plasma glucose levels are mostly unavailable [43], there is potential for misclassification of disease which may lead to a misrepresentation of disease in this population because the diagnoses of diabetes did not distinguish between type 1 diabetes and type 2 diabetes [44]. However, type 2 diabetes accounts for more than 90% of diabetes cases in the United States [30] Also, the age group 18–24 might be more representative of a population age group with type 1 diabetes. This potential misclassification would likely be non-differential and bias any of the current findings towards the null. Second, findings may not be generalizable to certain states that did not measure diabetes in the BRFSS survey. For instance, the state of California only measured diabetes diagnoses in years 2006 & 2007. Thus, we cannot conclusively articulate actual increases of odds of diabetes in California. Third, the BRFSS excludes certain populations, including those without landline telephones and those residing in institutions and on military bases, and thus might not be representative of the entire U.S. population. Fourth, response rates were not exclusively calculated for the population subset (females) included in this analysis. The average overall response rates for the BRFSS surveys 2006 – 2010 was 33.6% (not excluding the female population) thus representing only one in three of the targeted population. This may impact generalizability despite widespread use of the BRFSS data sets by notable studies/researchers. Fifth, the survey design requires use of software capable of addressing design characteristics (unequal selection, stratification, non-response and demographic variations among populations) that may introduce bias when using statistical tools that do not take these factors into account. Although, we utilized software compatible with the BRFSS data, and were careful to exclude missing responses during the analyses, we did not weight these data to the inverse of the sampling or response design due to our focus on relevant increases in odds. Lastly, interview responses are potentially subject to recall and social desirability biases. More so, higher refusal rates found in telephone surveys may increase bias due to non-response [14]. This factor, enabled by language barriers (non-English and limited-Spanish proficiency), may account for the greater proportion of non-Hispanic whites (over 75%) compared to the low percentage rates of other ethnic minority respondents, as questionnaires were only developed using these languages. The BRFSS addresses these challenges by keeping phone (landline and cell phone) interviews to a reasonable length and increasing the number of adults interviewed in each state, so as to produce a more representative sample and higher quality data amidst changes in communications technology, societal behaviors, and population diversity [14]. More importantly, some segments of the population (e.g., the poor and uninsured) may be misrepresented in the study and there is the slight possibility that subjects may have been selected to participate in multiple years of the cross sectional assessments. Lastly, there were statistically significant differences between those with and without complete data. However, the practical differences of the missingness was minimal.

There are also notable strengths to these data. Most importantly, the BRFSS enables researchers to assess trends that provide preliminary insight into disease (e.g. diabetes) epidemic for populations. This study is unique in capturing characteristics of several influencing lifestyle factors including, physical activity in a large cohort of women residing in the United States. Having multiple years of standard questions allowed for an investigation of odds in the population over a 5-year period while accounting for important risk factors. Interactions were investigated for and not found to be significant. The use of the years was an attempt to control for surrogate variation over the time period and to give us an estimate of the projection of the increase in diabetes independent of these important factors.

A dataset containing public use data files was used for data analysis. This dataset contained aggregate responses which made seeking patient informed consent non-applicable as the data sets used does not involve access to identifiable private information about the persons from/about whom the data were collected.