In this study, we evaluated the association between arsenic in toenail clippings, household water arsenic, and average daily consumption of 120 different diet items for which data were available. Study participants provided toenail clippings and a household tap water sample for analysis of total arsenic concentration using previously established protocols [19, 37]; most of the arsenic in both of these matrices is likely to be in the inorganic form (see e.g.,  and , respectively). Toenail clippings (mass 0.04 ± 0.01 g, mean ± 1 standard deviation [SD]) were analyzed at the University of Missouri Research Reactor Center using standard-comparator instrumental neutron activation analysis (NAA). Nails were washed carefully to remove external contamination, freeze-dried, and then stored in sealed vials until testing [19, 37]. Samples, certified reference material, and a keratin quality control sample were irradiated for 60 min at a thermal neutron flux of 8 × 1013 neutrons cm-2 s-1, then live-counted for 2 hours at a sample-to-detector distance of ~10 cm using a high resolution gamma-ray spectrometer after a decay period of ~24 hours. Arsenic was calculated from the 559-keV gamma ray from the decay of As-76, relative to known standards and after correction for physical decay. Quality control samples were within 1 SD of the expected value .
Drinking water samples were analyzed in the Dartmouth Trace Element Analysis Core using a Finnigan MAT GmbH ELEMENT high resolution inductively coupled mass spectrometer equipped with an MY hydride generator (Finnigan MAT GmbH, Bremen, Germany) . Samples were acidified to pH 1 with ultrapure nitric acid upon arrival at the laboratory, then spiked with Suprapur H2O2 (Merck KGaA, Darmstadt, Germany) to 0.01% at least 24 hours prior to analysis. During analysis, hydride generation was used to separate arsenic from ArCl + species, increasing the ability to detect arsenic at concentrations <5 μg/L .
Participants also completed a written, validated, semi-quantitative food frequency questionnaire (FFQ) [39, 40] to quantify diet over the previous year. An annual FFQ should provide a good match to the time scale over which the toenail clippings provide an integrated measure of exposure [6, 37]. The FFQ asked about the consumption of specific portion sizes of 120 different items from seven broad categories (dairy, fruits, vegetables, eggs and meat, breads, beverages, and baked goods) over the previous 12-month period; we analyzed associations between toenail arsenic and each of these diet items. We converted all responses to servings per day using the midpoint of each interval and assuming that a month has 30 days (Additional file 1: Table S1). When participants skipped a question, we set the frequency of consumption to missing. As in MacIntosh et al. , we focused on whole foods; the associations between toenail arsenic concentrations and micronutrient and vitamin consumption, both with and without supplements, were analyzed separately .
In addition, participants were interviewed, usually in their home, to obtain information on sociodemographic and lifestyle factors (e.g., smoking history, drinking water source, ) that may have affected the association between toenail arsenic concentration, drinking water arsenic concentration, and individual diet items.
Prior to analysis, we normalized data on toenail arsenic concentrations using natural-log (ln) transformation. Analyses reported here exclude the 70 subjects who did not report using their household water for drinking and cooking as well as subjects who did not meet the caloric thresholds suggested by Willett : 18 men below 800 calories and 13 above 4000 calories, and three women below 500 calories and four above 3500 calories. We also excluded from the analysis one individual with an extremely high toenail arsenic concentration (7.6 μg/g), which was 420% higher than the next highest concentration . This left us with a sample size of 852 subjects, down from 934.
In all analyses, we accounted for the previously observed non-linear association between toenail arsenic concentrations and household water arsenic concentration  by including natural-log transformed household water arsenic, an indicator variable  for whether subjects had concentrations of arsenic in household water <1 μg/L or ≥1 μg/L, and their interaction in a general linear model (GLM, SAS version 9.2).
We then evaluated associations with the self-reported, estimated daily rate of consumption of each of the 120 diet items in the FFQ. In the first stage of the analysis, we determined whether the association between water-corrected toenail arsenic concentration and each diet item differed between the two household water arsenic groups (<1 μg/L vs. ≥1 μg/L) using an interaction between the consumption rate of the diet item and the indicator variable for household water arsenic. If there was no statistically significant (α = 0.05) interaction between the diet item and the water exposure group, we concluded that the association between water-corrected toenail arsenic and the diet item was not affected by water exposure, and fit a GLM to the full dataset in the second stage of analysis (Model 1). However, if there was a statistically significant interaction between the diet item and the water exposure group, we concluded that the association between water-corrected toenail arsenic and the diet item differed between the two water exposure groups and so conducted analyses separately for the two groups (Models 2a and 2b). The slope coefficients (
) for each dietary item have the units natural-log transformed (toenail arsenic concentrations, μg/g) · (servings/d)-1.
For those diet items for which the slope coefficient was statistically significant for the appropriate model (1 or 2a/2b), we evaluated robustness to extreme values in the predictors by looking at unadjusted scatterplots, then recalculating regression coefficients after systematically deleting visually apparent outliers. Seven diet items were no longer significant after removal of such values and were not considered further.
Although associations between toenail arsenic and demographic characteristics such as age have previously been described , the mechanisms behind these associations have not been elucidated. For example, we do not know whether age directly affects toenail arsenic, or whether age influences diet, which in turn influences exposure as indicated by toenail concentrations. We therefore reported “crude” unadjusted associations between water-corrected toenail arsenic and each diet item, as well as analyses after adjustment for covariates that were deemed important from previous literature [19, 44, 45], biological plausibility, and univariate associations . We adjusted for four categorical variables (sex, smoking status [never/ever], season of toenail collection, case–control status [control, bladder cancer, basal cell carcinoma, squamous cell carcinoma]) and three continuous variables (age, daily intake of water from the household water source [ounces · d-1], and total energy intake [kcal · d-1]).
To help interpret regression coefficients from these adjusted models, we determined the percent change in predicted (back-transformed) toenail arsenic concentrations between 5th percentile and 95th percentile consumers for each food, using an approach similar to that described in Gruber et al. . Predictions were made for non-smoking, control subjects whose toenails were collected during the most common season (fall), separately for males and females at the mean age, caloric consumption, and water consumption for their sex. For Model 1 foods (those for which associations were consistent across household water arsenic concentrations), we used the overall median household water concentration. For Model 2 foods (those for which associations differed between water arsenic exposure groups), we used the median household water concentration for the appropriate exposure group.
We accounted for multiple testing across the individual foods using the false discovery rate (FDR) procedure implemented in the R package qvalue . Specifically, we calculated the Q-value, the minimum FDR at which a test may be called statistically significant , from the combined list of P-values for the association with each of the 120 foods, as generated by the appropriate model (1, 2a or 2b). We considered variables with a Q-value > 0.1 as less likely due to multiple testing.