Data source and sample population
Data collected as part of the National Health and Nutrition Examination Surveys (NHANES) conducted in 2003–2004 and 2005–2006 were used to complete this cross-sectional study. NHANES is a continuous survey based on a complex multistage probability sample designed to provide nationally representative nutrition and health data and prevalence estimates for nutrition and health status measures in the United States [16, 17]. Approval for NHANES data collection was provided by the National Center for Health Statistics Research Ethics Review Board. The current study was limited to adults age 19 years or above, excluding pregnant and lactating females, with food frequency questionnaire (FFQ) responses to questions on candy consumption and two complete 24-hour dietary recalls.
Frequency of candy consumption (primary exposure)
In NHANES 2003–2004 and 2005–2006, a FFQ component was included to gather information on the frequency of foods consumed over the previous 12 months [18]. More recent releases of NHANES did not include the FFQ component and therefore could not be used for this analysis. The FFQ was developed by the National Cancer Institute (NCI) and was based on the NCI Diet History Questionnaire [19]. The 151 items on the FFQ represent a slight modification of the original NCI instrument. Portion size information was not collected with the FFQ. Printed FFQ questionnaires were mailed to the homes of English or Spanish-speaking survey participants 2 years of age and older who provided at least one complete 24-hour dietary recall interview.
The FFQ included two questions on candy consumption over the previous 12 months, namely: (1) how often did you eat chocolate candy; and (2) how often did you eat other candy. Definitions of chocolate and other candy were not provided with the FFQ, and the FFQ did not include a specific question regarding chewing gum. Based on their own interpretation of the two candy categories, survey respondents reported their typical frequency of consumption of each of the two types of candy as one of eleven specified frequency categories ranging from “never” to “2 or more times per day”. The FFQ data files were processed by the National Center for Health Statistics (NCHS) using the Diet*Calc software to provide daily frequencies of consumption, i.e., eating occasions per day (EO/d) by candy type for each respondent based on the categorical response. The Diet*Calc algorithm converted the highest frequency category of “2 or more times per day” to 2 EO/day. Previous studies [20] have found that subjects who eat some foods most frequently tend to eat more of these foods when they consume them, and consequently for some foods there is a positive correlation between the frequency of consumption and the amount of food consumed per eating occasion. The association between gram per eating occasion (g/EO) (from 24-hour recalls) and number of EO/day (from the FFQ) by type of candy consumed (chocolate and non-chocolate) was examined in a preliminary analysis using linear regression analysis for the subset of individuals who reported consumption of candy on one or both of the two days of dietary recall. Portion size of candy by eating occasion (g/EO) was not associated with the frequency of candy consumption in linear regression models of chocolate candy g/EO versus frequency of chocolate candy consumption and non-chocolate candy g/EO versus frequency of other candy consumption (R2 < 1%; data not shown). Adults who reported consuming one candy type in the previous 12 months were more likely to report consumption of the other candy type in the previous 12 months (Pearson chi-square test, p-value < 0.001), and the frequency of chocolate candy consumption was significantly associated with the frequency of non-chocolate candy consumption (p-value < 0.001, linear regression). Frequency of candy consumption as reported was therefore assumed to be representative of total candy intake. Previous assessments in adults have not shown negative effects on cardiovascular risk factors by type of candy [4, 5], therefore daily frequencies of chocolate and other candy were combined into one variable for the present study to derive an estimate of daily frequency of total candy consumption. Based on summed responses in the FFQ for typical frequency of chocolate and non-chocolate candy consumption, adults were divided into three groups based on typical frequency of candy consumption:The ranges within each of the frequency categories were subjectively selected to provide roughly similar proportions of the population in each group while also corresponding to easily interpretable frequencies of consumption over the course of a month or week.
-
Infrequent: ≤ 3 EO/month (daily frequency ≤ 0.09 eating occasions per day);
-
Moderate: > 3 EO/month and ≤ 3.5 EO/week (daily frequency > 0.09 and ≤ 0.5 eating occasions per day); or
-
Frequent: > 3.5 EO/week (daily frequency > 0.5 eating occasions per day).
Physiologic parameters (primary outcomes)
During the examination component of NHANES, participants underwent a physical examination. As part of the examination, body measurement data were collected by trained health technicians following NHANES anthropometry protocols [21]. Body mass index (BMI, kg/m2) was calculated as body weight divided by height squared. Skinfold calipers were used to measure sub-scapular and triceps skinfolds. Sub-scapular and triceps skinfold measurements coded in the data release as “exceeds capacity” (i.e., the amount of adipose tissue exceeded the limits of the caliper) were assigned values of 44 and 45 mm, respectively, as these values correspond to the highest reported value by type of measurement. The sub-scapular and triceps skinfold measurements were then summed per individual. Approximately 5% of the summed values included an assigned maximum value.
Up to three measurements each of systolic and diastolic blood pressure were taken from subjects after resting quietly in a sitting position for 5 minutes [22]. Fasting or non-fasting blood samples were collected for lipid profile and insulin levels. Triglycerides, glucose and insulin were measured only in individuals randomly selected to participate in the morning session after an overnight fast. Triglycerides were measured enzymatically in serum using a series of coupled reactions in which triglycerides are hydrolyzed to produce glycerol [23]. Total and HDL cholesterol were measured in samples from all individuals. HDL cholesterol was measured colorimetrically after precipitation of the apoB containing lipoproteins [24]. LDL cholesterol levels provided in the data release were calculated using the Friedewald calculation: [LDL-cholesterol] = [total cholesterol] – [HDL-cholesterol] – [triglycerides/5] (if triglycerides was 400 or less), where all values are expressed in mg/dL [23]. Plasma fasting glucose was determined by a hexokinase method and insulin was measured with an immunoassay method [25]. For subjects with fasting plasma glucose and serum insulin data, we calculated insulin sensitivity using the quantitative insulin sensitivity check index (QUICKI), defined as 1/[log(Io) + log(Go)] [26].
Demographics and other participant characteristics
Demographic characteristics including age, sex, race (categorized as non-Hispanic white, non-Hispanic black, Mexican American, other Hispanic, or other), and education level (categorized as less than high school graduate, high school graduate/GED or equivalent, or beyond high school) were self-reported by participants [27]. The family poverty income ratio (PIR) was derived from family income data; the data were released as a continuous variable, with all PIRs of 5 or more coded as 5.
The in-home questionnaire included separate questions about participation in vigorous physical activity and participation in moderate physical activity over the previous 30 days for at least 10 minutes; participants provided a yes or no response to each question [28]. For this analysis, adults were categorized into one of three physical activity groups based on the highest level of activity reported: none, moderate or vigorous. Participants also were asked to quantify the average number of hours per day of sitting and watching television (TV) or videos during the previous 30 days; responses were coded as 0 (representing anything less than 1 h), or 1, 2, 3, 4, or 5 or more hours/day [28]. Self-reported use of medication was captured during the in-home questionnaire including use of insulin or diabetic pills for control of diabetes, medication for elevated blood pressure, and medication to lower cholesterol levels [22, 29]. Cotinine concentrations were derived in serum samples collected from survey participants [30]. For this analysis, participants with blood cotinine levels of ≥ 3 ng/mL were classified as smokers [31]. Demographic and other lifestyle characteristic information was used as covariates in analyses of some parameters in the current study.
Nutrient intakes
The dietary interview component of NHANES is known as “What We Eat in America” (WWEIA) [32, 33]. In this component of the survey, trained dietary interviewers collected detailed information on all foods and beverages consumed by respondents in the previous 24 hour time period (midnight to midnight). A second dietary recall was administered by telephone 3 to 10 days after the first dietary interview, but not on the same day of the week as the first interview. In the current study, nutrient concentration data in the United States Department of Agriculture (USDA) Food and Nutrient Database for Dietary Studies (FNDDS) version 3.0 [34] were used to calculate nutrient intakes by respondents in the combined NHANES 2003–2006. The MyPyramid Equivalents Database (MPED) 2.0 was the source of added sugars concentration data and total fruit and vegetable servings [35]; values for foods consumed in NHANES 2005–2006 but not included in MPED were imputed based on values for similar foods or recipe calculations. Total fruit and vegetable intakes, total intakes of select nutrients, and non-candy intakes of select nutrients were used as covariates in some analyses. Non-candy nutrient intakes were calculated as the difference between total 2-day average nutrient intakes and 2-day average nutrient intakes from foods within the candy category (excluding gum) of the USDA hierarchical food coding scheme as presented in FNDDS [34].
Statistical analysis
Descriptive characteristics including age, sex, race/ethnicity, education, PIR, percent participation in physical activity in the previous 30 days, average daily time sitting and watching TV or videos, and smoking status were summarized by frequency of candy consumption category. Ordered logistic regression models with category of frequency of candy consumption as outcome and each of the descriptive characteristics as predictors were used to compare subjects in the three candy consumption categories.
Estimates of usual nutrient intake (from the 24-hour dietary recalls) were derived using the approach developed by Nusser et al. [36] and Carriquiry [37]. Software for Intake Distribution Estimation (C-Side; version 1.02, 1997, Iowa State University Statistical Laboratory, Ames, IA) which implements this method was used to estimate the usual nutrient intakes. Estimates of usual intakes of energy, protein, fat, saturated fat, total sugars, added sugars, and fiber were generated for subpopulations of adults by frequency of typical candy intake. It was not possible to obtain usual intakes of alcohol using the C-Side model as the model is designed to halt if it fails to result in an acceptable semi-parametric normality transformation. Hence, estimates of 2-day average intakes of alcohol were developed and used in the analyses. Energy adjusted nutrient intakes were determined using the residual method [38].
Obesity was defined as a BMI of 30.0 or higher, and overweight or obese was defined as a BMI of 25.0 or higher. Waist circumference of 102 cm or higher in men and 88 cm or higher in women was defined as elevated based on the defined risk factors for metabolic syndrome [39]. Summed skinfold thickness values at or above the 95th percentile of the age- and sex-specific reference values compiled from the first and second National Health and Nutrition Surveys were considered to represent excessive skinfold thickness [40].
Weight and adiposity status were analyzed using logistic regression with the log of the odds of being obese, overweight/obese, or having an elevated waist circumference or skinfold thickness as outcome variables, with indicator variables for belonging to categories of candy consumption frequency as predictors. Two models were run. The first included non-modifiable confounding factors, namely sex, age and race/ethnicity (model 1), while the second model included these factors as well as categorical factors for education, income, smoking status, physical activity and time watching TV/videos (model 2). Ordered logistic regression models with category of candy consumption frequency as the outcome were used to conduct tests for trend in both models.
Blood pressure, lipids, and insulin sensitivity were analyzed using linear regression models with the physical measures as outcome variables, with indicators for category of candy consumption frequency as predictors. Two models were run. The first included non-modifiable confounding factors, namely sex, age and race/ethnicity (model 1) while the other model also included other variables likely to be associated with the physical measures, including categorical factors for education, income, smoking status, physical activity, time watching TV/videos and whether taking blood pressure, cholesterol or diabetes medication (model 2); and continuous variables for BMI and energy-adjusted dietary factors associated with cardiovascular disease risk [41, 42]. Because blood pressure, cholesterol or diabetes medications have an impact on blood pressure, lipid profile, and insulin sensitivity, they were introduced into the models as dummy variables. This allowed the analysis to take into account the variability of cardiovascular risk factors in all subjects, including those who were on medication and to investigate the potential association of candy consumption with blood pressure, lipid, and insulin sensitivity control for these subjects. Ordered logistic regression models with category of candy consumption frequency as the outcome were used to conduct tests for trend in both models.
The tests were conducted using STATA (StataCorp. 2007, Stata Statistical Software: Release 10, College Station, TX: StataCorp LP) and used the statistical weights developed by NCHS to adjust for the differential probability of selection and non-response and to adjust for the complex statistical design of NHANES. The sampling weight produced by NCHS for FFQ respondents (WTS_FFQ) was used in analyses of demographic characteristics, nutrient intakes, anthropometrics, blood pressure, and HDL cholesterol. Because only a subset of NHANES participants was examined in the morning session and fasted, special sampling weights (referred to as the “fasting” weight) were also generated by NCHS to appropriately analyze outcomes assessed in this subsample. The fasting sampling weight was used in analyses of LDL cholesterol, triglycerides, plasma glucose and insulin.