Exposure Assessment of Tobacco Smoke Biomarkers

Because NHANES participants are recruited through a multistage probability sample, it is necessary to account for this complex design to estimate variances properly and to produce unbiased, nationally representative statistics. Robust estimation may be accomplished by applying survey sample weights to each participant’s data and using Taylor series linearization to produce variance estimates. We used this estimation approach as implemented in the DESCRIPT subroutine of SUDAAN version 11.0.0 (Research Triangle Institute, Research Triangle Park, NC) called from the SAS statistical software application version 9.4 (SAS Institute, Cary, NC), as well as the SURVEYREG and SURVEYMEANS subroutines of SAS 9.4. Sample-weighted linear regression models stratified by tobacco use were fit to data from NHANES survey cycles 2007 – 2008, 2009 – 2010, and 2011 – 2012, where the dependent variable was blood nitromethane concentration (ng/L). Because the distribution of measurements was strongly right-skewed, which could have adversely affected hypothesis testing, the blood nitromethane concentration data was transformed with the natural log for regression analysis. We report slopes from these models along with their 95 percent confidence intervals and p-values. In addition, to facilitate interpretability, we report the slopes transformed to represent the absolute change in biomarker concentration $\Delta Y$ associated with a unit-increase in the predictor $\Delta X$ , as adapted from Rodríguez-Barranco, et al., 2017^{16 (link)}: $\Delta Y=\left[\text{exp}\left(\Delta X\cdot \beta \right)-1\right]\cdot \text{GM}\left[Y\right]$ , where $\text{GM}\left[Y\right]$ is the sample-weighted geometric mean of biomarker concentration. The tabulated regression results in Table 3 assume $\Delta X=1$ , so that $\Delta Y$ represents the absolute change associated with a unit-increase in the predictor. The 95 percent confidence interval is: $95\%\text{CI}\hspace{0.17em}\Delta Y=\left[\text{exp}\left[\Delta X\cdot \left(\beta \pm 1.96\cdot \text{se}\left[\beta \right]\right)\right]-1\right]\cdot \text{GM}\left[Y\right]$ , where $\text{se}\left[\beta \right]$ is the sample-weighted standard error of the slope. Both $\Delta Y$ and its 95%CI are calculated at $\text{GM[}Y]$ , which is reported in the caption accompanying the tabulated regression results. Since this geometric mean is treated as a fixed quantity, the width of the 95%CI may be slightly underestimated. In addition, at values different from the geometric mean, the value of $\Delta Y$ and the width of its 95%CI will vary, owing to the log transformation of the dependent variable. Statistical significance was set to α ≤ 0.05.
Sample-weighted regression models were stratified by combusted tobacco use, and the following variables were included as predictors: sex, age, race/ethnicity, poverty income ratio (PIR; ratio of self-reported family income to the U.S. Census poverty threshold), and fasting time (time elapsed since participant last ate or drank anything other than water and the time of specimen collection). Information for these potential confounders was self-reported. Age [year] was categorized into the following ranges: 12 – 19, 20 – 39, 40 – 59, and ≥60. Poverty level was determined by whether the ratio of a family’s income to poverty (INDFMPIR) was greater or less than the poverty threshold, which is represented by the ratio equaling unity.¹³ In addition, body mass index (BMI) from measurements at the physical examination was included as a predictor. Since standard definitions for underweight (BMI < 18.5 kg/m²), healthy weight (18.5 ≤ BMI < 25), and overweight/obese (BMI ≥ 25) apply to adults ≥20 y, participants younger than 20 y were classified based on their BMI percentile for their sex and age: below the 5th percentile (underweight), between the 5th and 85th percentile (healthy weight), and above the 85th percentile (overweight/obese).¹⁷ NHANES cycle was also included as a predictor.
Food consumption information was collected from participants using structured questionnaires administered by trained interviewers who used intensive elicitation techniques to translate a participant’s recollection of the type and amount of food consumed to a standardized numerical encoding and food mass. Dietary exposure was explored by assessing the mass NHANES participants consumed within each USDA (US Department of Agriculture) food group for the 24-hour period (midnight to midnight) preceding the in-person dietary recall interview conducted as part of the physical examination. Data for the 24-hour recall period are contained in the publicly available NHANES Individual Foods – First Day file (NHANES dataset: DR1IFF), which provides a record describing each food, water, or beverage consumed by the participant, including the mass reported consumed and eight-digit USDA food code. Standardized hierarchical food groups can be identified from the USDA code, where the first digit represents one of nine major food groups, and each subsequent digit represents subgroups of increasing specificity.^{18 (link)} The mass consumed in each food group was summed so that each participant was represented by a single record describing their dietary intake for the previous 24 hours. Each participant’s dietary intake was first apportioned over nine food groups: milk products; meat/poultry; eggs; legumes/nuts/seeds; grain products; fruits; vegetables; fats/oils/salad dressings; and sugars/sweets/beverages. In addition, we distinguished three subgroups: cured meats, luncheon meats and hot dogs, and tap water. The cured meats food group was constructed using the search term “cured” in the USDA What We Eat In America search tool and selecting all food codes referring to meats.¹⁹ The luncheon meat and hot dog food group was constructed by searching the term “luncheon” and selecting all meat food codes and searching “hot dog.” The tap water food group was constructed by searching “tap water.” To avoid double counting, the mass consumed in each subgroup was subtracted from the mass consumed in their respective food group. The USDA food codes and logic for apportioning dietary intake are detailed in Supplemental Table S1.
Serum cotinine was used as a continuous predictor to represent tobacco smoke exposure for both exclusive combusted tobacco users and non-users of tobacco products. Cotinine is a highly specific metabolite of nicotine, the primary addictive agent in tobacco and tobacco smoke, and is thereby present in the blood serum of tobacco smokers. Likewise, since tobacco smoke exposure among non-users of tobacco products is attributable to inhalation of secondhand tobacco smoke (SHS), this exposure can be quantified with serum cotinine. In addition, to provide an alternative representation of tobacco smoke exposure, we ran a regression model where exposure among exclusive smokers was represented by the self-reported average number of cigarettes smoked per day (CPD) over the five days preceding the NHANES physical exam. This CPD regression model was sample-weighted, unstratified, and comprised the same predictors as in the stratified models, except that exposure was classified as ≤0.05 ng/mL serum cotinine (non-exposed to tobacco smoke); >0.05 – ≤10 ng/mL (presumptively exposed to second-hand tobacco smoke); 1 – 10 cigarettes per day (CPD; 0.5 pack), 11 – 20 (1 pack), and >20 (>1 pack), where the reference category was non-exposed participants. The non-exposed category was defined at ≤0.05 ng/mL serum cotinine, which was its LOD in the 1999 – 2000 NHANES cycle, and although this improved in 2001 to 0.015 ng/mL, we use 0.05 ng/mL to permit historical comparison of serum cotinine results.^{20 (link)} The analytic dataset for the CPD model comprised the same participants as in the stratified models, but excluded participants who could not be assigned to a CPD category (N=266), leaving 6,464 participants.