Participants
This study utilized data from the National Health and Nutrition Examination Survey (NHANES) collected from 2015 to 2016 because this cycle included both fluoride biomonitoring data and a breadth of self-reported sleep outcome measures. Plasma fluoride concentrations were measured among 2145 participants aged 6–19 years and household tap water fluoride concentrations were measured among 3987 participants aged 0–19 years. Sleep outcomes were assessed among participants aged 16 and over, and thus we limited our analyses to participants ages 16–19. We included participants who had either plasma or water fluoride measurements, complete covariate data, and data on at least one sleep outcome measure. We excluded 3 participants who reported taking medications for sleep disorders. There were 512 participants who met inclusion criteria. Of those, 473 had plasma fluoride measurements and were included in analyses. For analyses including water fluoride as a predictor, there were 503 participants who met inclusion criteria and we excluded an additional 84 who reported that they did not drink tap water, resulting in a final analytic sample of 419. Participant selection is depicted in Additional file 1: Figure S1. A comparison of demographic characteristics between the current study sample and all adolescents aged 16–19 in NHANES is presented in Additional file 2: Table S1. Sampling weights were applied to account for the complex NHANES survey design as recommended by the National Center for Health Statistics (NCHS). The weighted samples for plasma and water fluoride analyses represented 12,531,822 and 11,577,700 U.S. adolescents, respectively.
Fluoride measures
Fluoride concentrations were measured in blood plasma and household tap water. Collection times of blood and tap water were not standardized. Plasma fluoride concentrations were measured using an ion-specific electrode and hexamethyldisiloxane method [14], while household tap water samples were measured electrometrically with an ion-specific electrode [15]. Plasma and water samples were measured for fluoride in duplicate (using the same sample) and the average of the two values was publicly released. Fluoride measurements were conducted at the College of Dental Medicine, Georgia Regents University, Augusta, GA. The lower limit of detection (LLOD) for plasma fluoride was 0.25 nmol, and the LLOD for water fluoride was 0.10 mg/L. Approximately 86 and 100% (all) of participants had values above the LLOD for water fluoride and plasma fluoride respectively. Laboratory generated values for water fluoride concentrations below the LLOD were utilized in analyses; however, values below the LLOD were imputed as \( LLOD/\sqrt{2} \) for descriptive statistics.
Sleep measures
Sleep habits and sleep disorders were ascertained through questionnaires in participants’ homes by trained staff using the Computer-Assisted Personal Interview (CAPI) system. The CAPI system is equipped with built-in consistency verification to reduce data entry errors [16]. The questions included in the sleep questionnaire were not validated.
Bedtime/wake time and sleep duration
Participants were asked to estimate the time that they usually get in and out of bed on weekdays or workdays, excluding naps (i.e. What time do you/does study participant (SP) usually go to sleep on weekdays or workdays? What time do you/does SP usually wake up on weekdays or workdays?). Using bedtime and wake time data, the number of hours of sleep per night (i.e. nighttime sleep duration) was calculated by the NCHS.
Sleep disturbances
Participants were asked how often they snore (i.e. In the past 12 months, how often did you/SP snore while you were/s/he was sleep?), and how often they experience symptoms suggestive of sleep apnea (i.e. “In the past 12 months, how often did (you/SP) snort, gasp, or stop breathing while (you were/s/he was) asleep?). Responses to these questions were categorized as “never”, “rarely – 1-2 nights a week”, “occasionally - 3-4 nights a week”, or “frequently - 5 or more nights a week”. When participants implied that they did not know whether they exhibit these behaviors, the interviewer asked whether anyone told them that they do. Participants were also asked directly whether they experience sleep disturbances (i.e. “Have you/Has SP ever told a doctor or other health professional that you have/s/he has trouble sleeping?”). Responses were dichotomized as “yes” or “no”.
Daytime sleepiness
To assess daytime sleepiness, participants were asked “In the past month, how often did (you/SP) feel excessively or overly sleepy during the day?”. Responses to this variable were categorized as “never”, “rarely - 1 time a month”, “sometimes - 2-4 times a month”, “Often- 5-15 times a month”, or “Almost always - 16-30 times a month”.
Covariates
We selected covariates a priori that are empirically associated with fluoride exposure and sleep in existing literature [17,18,19,20,21,22,23], including age, sex, body mass index (BMI), race/ethnicity, and the ratio of family income to poverty. The NCHS calculated the ratio of family income to poverty by dividing annual family income by the poverty guidelines for a given survey year.
Statistical analyses
We applied survey weights from the mobile exam center visit (i.e. MEC weights) for all analyses to account for the clustered sample design, over-sampling, post-stratification, survey non-response, and sampling error. Survey weights also permit generalization to the entire US population [24]. Given that we utilized a dietary variable as exclusion criteria (i.e. tap water consumption) for analyses including water fluoride, we applied reweighted MEC weights to our dietary sample prior to regression analyses with water fluoride, according to NCHS guidance (as describe elsewhere [25]). Results of regression analyses did not change regardless of whether MEC weights or reweighted MEC weights were applied. Means and proportions were calculated for descriptive analyses of demographic variables as well as fluoride exposure and sleep outcome measures. A Pearson correlation examined the relationship between logarithm (base 2)-transformed plasma and water fluoride concentrations. To examine the relationship between fluoride exposure and sleep duration, sleep duration was transformed into a 3-category variable in which short duration = 1, normal duration = 2 and long duration = 3. Normal duration was determined based on sleep duration recommendations by the National Sleep Foundation. 8–10 h is considered normal duration for 16–17-year-olds, and 7–9 h is considered normal duration for 18–19-year-olds [26]. Values below or above this duration were categorized as short or long duration respectively. Survey-weighted multinomial logistic regression were utilized to model sleep duration or daytime sleepiness as a function of plasma or water fluoride concentrations while adjusting for covariates. For regression analyses of fluoride exposure and symptoms suggestive of sleep apnea or snoring, we created dichotomous variables of 0 = never occurs or 1 = occurs 1 or more times per week (i.e. “never” versus “ever”). We utilized this classification to account for potential underestimation of the occurrence of these behaviors and because occurrence of symptoms suggestive of sleep apnea even once per week may indicate underlying sleep dysfunction. We conducted survey-weighted binomial logistic regression to model snoring, symptoms suggestive of sleep apnea or self-reported trouble sleeping as a function of plasma or water fluoride concentrations while adjusting for covariates. Bedtime and wake time variables were converted to numeric format, and the bedtime variable was rescaled prior to analysis such that 7 pm (i.e. the earliest bedtime in our sample) was set at 0 and 7 am (i.e. the latest bedtime in our sample) was set at 12. This allowed us to consider times occurring after midnight as later than times occurring prior to midnight. Associations between fluoride exposure and sleep and wake time were explored using survey-weighted linear regression adjusted for covariates. We explored potentially influential values using a Cook’s Distance estimate and did not identify any. Assumptions of linear and logistic regression were satisfied for all models except that we detected heteroscedasticity in models of fluoride exposure and bedtime/wake time. As such, we conducted unweighted quasi-likelihood estimation models (see Additional file 2: Table S2) to account for this, but results did not appreciably differ from weighted linear regression. Therefore, we present weighted linear regression models herein. No issues with multicollinearity were detected. We included a fluoride*sex interaction term in our models to test for sex-specific associations; however, this term was removed if non-significant. Additionally, we conducted a sensitivity analysis to examine whether adjusting for serum cotinine - a biomarker of nicotine exposure, influenced associations between plasma fluoride concentrations and sleep outcomes. A two-tailed alpha of 0.05 was the criteria for statistical significance for all main effects in regression analyses, while a two-tailed alpha of 0.1 was the criteria for statistical significance for interactions. We applied a Holm-Bonferroni correction to account for multiple comparisons for each fluoride variable whereby each class of sleep outcomes (e.g. sleepiness; sleep duration) was considered a separate test. SAS (V.9.4) software was used for all analyses.