Study population. NHANES uses a complex multistage sampling design to obtain representative samples of the noninstitutionalized U.S. population [National Center for Health Statistics (NCHS) 2008]. For this analysis, we used data from 14,213 adults ≥ 20 years of age who participated in the NHANES 1999–2004 interviews and examinations. The overall participation rate was 70%. We excluded 772 pregnant women, 661 participants who were missing blood cadmium measures, 4 participants with urine cadmium corrected for molybdenum interference equal to zero, 1,934 participants who were missing information on serum cotinine or self-reported smoking variables, and 1,853 participants who were missing other variables of interest such as sociodemographic and cardiovascular risk factors. The final sample size was 8,989 participants. These participants had similar sociodemographic characteristics compared with the overall NHANES 1999–2004 population. The NCHS Research Ethics Review Board approved the NHANES protocols. All the participants provided oral and written informed consent.
Blood and urine cadmium. Blood and urine cadmium were measured at the Environmental Health Laboratory (Atlanta, GA, USA), Centers for Disease Control and Prevention, National Center for Environmental Health (NCEH). Extensive quality control procedures were followed including confirmation that collection and storage materials were not contaminated with background cadmium or other metals (NCHS 2008).
Blood cadmium and lead were measured simultaneously in whole blood using a multielement atomic absorption spectrometer with Zeeman background correction (SIMAA 6000 model; PerkinElmer, Norwalk, CT, USA) in 1999–2002 and an inductively coupled plasma-mass spectrometer (PerkinElmer/SCIEX model 500; PerkinElmer, Shelton, CT, USA) in 2003–2004. National Institute of Standards and Technology whole blood standard reference materials were used for external calibration (NCHS 2008). The interassay coefficients of variation for blood cadmium ranged from 3.2% to 9.4%. The limit of detection (LOD) was 0.3 μg/L in NHANES 1999–2002 and 0.2 μg/L in NHANES 2003–2004, resulting in 15% and 5% of observations below the LOD, respectively.
Urine cadmium was measured only in a randomized subsample of one-third of the 1999–2004 NHANES population (n = 2,867 in our study population). Thus, measurements of urine cadmium were missing completely at random in the remaining two-thirds of the study population. Urine cadmium was measured using inductively coupled plasma-mass spectrometry (PerkinElmer/SCIEX model 500). The interassay coefficients of variation for urine cadmium ranged from 1.2% to 4.7%. The LOD was 0.06 μg/L, which resulted in 3% of the observations falling below the LOD. In NHANES 1999–2002, cadmium levels in urine were corrected for interference from molybdenum oxide. NIST urine reference materials were used for external calibration (NCHS 2008). Urine cadmium data was reported in micrograms cadmium per gram creatinine. Urine creatinine was measured by the modified kinetic Jaffé method (NCHS 2008).
For participants with blood cadmium concentrations below the LOD, and for participants with urine cadmium concentrations missing completely at random, cadmium concentrations were imputed as the median of the subject-specific posterior distribution of predicted levels obtained from a Markov Chain Monte Carlo with Gibbs sampling nested linear model (Tellez-Plaza et al. 2010
). The variables used to predict cadmium levels were selected by a backward stepwise process using linear regression models that included age, sex, smoking status, serum cotinine, and blood or urine cadmium. The imputation methods have been described in detail elsewhere (Tellez-Plaza et al. 2010
). Urine cadmium measurements below the LOD were imputed as the LOD divided by the square root of two (NCHS 2008).
Baseline data collection.
Information on age; sex; race/ethnicity; education; menopause status; smoking; and medication for treating hypertension, diabetes, and hypercholesterolemia was based on self-reported information. Body mass index and blood pressure were measured during the examination (NCHS 2008). Hypertension was defined as a mean systolic blood pressure ≥ 140 mmHg, a mean diastolic blood pressure ≥ 90 mmHg, a self-reported physician diagnosis, or medication use. Diabetes was defined as a fasting glucose ≥ 126 mg/dL, a nonfasting glucose ≥ 200 mg/dL, a self-reported physician diagnosis, or medication use. Serum C-reactive protein was analyzed using high-sensitivity latex-enhanced nephelometry using the Behring Nephelometer Analyzer II (Dade-Behring Diagnostics Inc., Somerville, NJ, USA). Serum total cholesterol was measured enzymatically using the Cholesterol High Performance reagent (no. 704036; Roche Diagnostics, Indianapolis, IN, USA). High density-lipoprotein (HDL) cholesterol was measured using a direct HDL reagent (no. 1661442; Roche Diagnostics). High cholesterol was defined as a serum total cholesterol > 200 mg/dL or medication use. Serum cotinine was measured by an isotope-dilution high-performance liquid chromatography/atmospheric pressure chemical ionization tandem mass spectrometric method (NCHS 2008). Serum creatinine was measured using a kinetic rate Jaffé method in a Hitachi 704 multichannel analyzer (Boehringer Mannheim Diagnostics, Indianapolis, IN, USA). Estimated glomerular filtration rate (eGFR) was calculated from calibrated creatinine, age, sex, and race/ethnicity using the Modification of Diet in Renal Disease Study formula (Selvin et al. 2007
NHANES 1999–2004 participants were followed for mortality through 31 December 2006. Vital status and cause of death were determined by probabilistic matching between NHANES records and death certificates from the National Death Index (NDI) (NCHS 2010). The NCHS submitted NHANES records to the NDI for the implementation of a probabilistic algorithm that used identifying data elements such as social security number, first name, middle name, last name, date of birth, race, sex, state of birth, and state of residence to match records according to established criteria (NCHS 2010). Before the matching algorithm was applied, each record was screened to determine if it contained sufficient combinations of identifying data elements, with < 0.2% of NHANES population being considered ineligible for the matching. A calibration study on NHANES I Epidemiological Follow-up Study, which used similar methodology, found that 96.1% of deceased participants and 99.4% of living participants were correctly classified (NCHS 2010). The cause of death was determined using the underlying cause listed on death certificates, and was coded using the International Classification of Diseases
, 10th Revision
(ICD-10; World Health Organization 2007
). Cause-specific mortality was ascertained for CVD (ICD-10 codes I00-I78), heart disease (ICD-10 codes I00-I09, I11, I13, I20-I51) and ischemic heart disease (ICD-10 codes I20-I25). NCHS ensures that the identity of the participants is not disclosed (NCHS 2010). All direct identifiers, as well as any characteristics that might lead to identification, were omitted from the linked dataset used in the present study.
Follow-up time for each individual was calculated as the difference between the age at the date of the NHANES examination and the age at the date of death, age on 31 December 2006, or age 90 years, whichever occurred first. Follow-up was censored at age 90 years because of the high mortality after this age and the low number of participants who were contributing person-time experience.
To perform all the statistical analyses we used the survey package (Lumley 2004
) in R software (version 2.12.1; R Development Core Team 2011
) to account for the complex sampling design and weights of the NHANES and to obtain appropriate standard errors. Blood and urine cadmium levels were right skewed and log-transformed for the analyses. Cut-offs for blood cadmium quintiles were based on weighted distributions in the study sample. Cut-offs for creatinine-corrected urine cadmium quantiles were based on weighted distributions in the originally measured one-third random subsample.
We estimated hazard ratios [95% confidence intervals (CIs)] for mortality end points using Cox proportional hazards regression with age as the time scale and individual starting follow-up times (age at examination) treated as the staggered entries. Cadmium was modeled as a log-linear term, which was used to estimate hazard ratios for mortality comparing the 80th with the 20th percentiles of the blood and creatinine-corrected urine cadmium distributions. To test for nonlinear relationships in all-cause and CVD mortality models, we used restricted quadratic splines with knots at the 20th, 50th, and 80th percentiles of each cadmium distribution, and we applied the Wald test adjusted for the survey design to the nonlinear spline terms.We observed no statistically significant departures from linearity (p-values for nonlinearity in all-cause and CVD mortality models were 0.40 and 0.20, respectively, for blood cadmium and 0.56 and 0.89, respectively, for urine cadmium). The level of statistical significance used for hypothesis testing was 0.05.
Statistical models were progressively adjusted to evaluate the potential confounding effect of different groups of variables. Model 1 accounted for sociodemographic variables including race/ethnicity (non-Hispanic white, non-Hispanic black, Mexican-American, other), sex (men, women), education (≥ high school, < high school) and low income (≥ $20,000, < $20,000). Model 2 further adjusted for established CVD risk factors including postmenopausal status for women (no, yes), body mass index (continuous, kilograms per meter squared), hypertension (no, yes), diabetes (no, yes), blood lead (continuous, log micrograms per deciliter), total cholesterol (continuous, milligrams per deciliter), HDL cholesterol (continuous, milligrams per deciliter), cholesterol lowering medication (no, yes), C-reactive protein (continuous, log milligrams per liter), and estimated glomerular filtration rate (eGFR; continuous, milliliter per minute per 1.73 meters squared). Smoking could be a major confounder. Thus, in model 3, we further adjusted for current smoking status (never, former, current), intensity of current exposure to tobacco smoke using serum cotinine levels (continuous, log nanograms per milliliter), and cumulative smoking using number of pack-years (modeled as restricted cubic splines with knots at 10, 20, and 30 pack-years). The assumption of hazards proportionality was assessed visually based on the smoothed relationship between time and scaled Schoenfeld residuals (Grambsch and Therneau 1994
). We observed no major departures from proportionality. The primary statistical analyses were conducted in the dataset including imputed cadmium values. As sensitivity analyses, we repeated the analyses restricting the sample to the one-third random subsample with the originally measured urine cadmium levels.
We had a priori interest on evaluating differential associations by sex and smoking subgroups. Subgroup analyses were conducted by including interaction terms for log-transformed blood or creatinine-corrected urine cadmium with the corresponding indicator variables for subgroups defined by sex (men, women) and smoking status (never, former, current) in separate models. In the Cox models, the nonparametric baseline hazards were allowed to differ by subgroup categories. We used the Wald test that was adjusted for the survey design to obtain the p-values for the interaction.
Adjusted population attributable risks (PARs) for cadmium exposure were calculated by adapting the standard formula PAR = 1 – Σj
; Bruzzi et al. 1985
) to the survey data. In this formula, the subscript i
denotes one of two categories of cadmium exposure (with each participant classified as exposed or unexposed if above or below the percentile being used to calculate the PAR, respectively), the subscript j
is an index for all strata obtained after cross-classifying the study sample for all adjusted covariates, pij
is the survey-weighted proportion of cases over all cases in the study population in each stratum after cross-classifying the dichotomous cadmium exposure and all adjusted covariates, and RRi|j
is the adjusted hazard ratio for mortality comparing participants exposed to cadmium with those unexposed in stratum j
of covariates. We calculated adjusted PARs for each percentile of cadmium exposure starting at the 10th percentile of each cadmium distribution and displayed the results after smoothing using lowess with 0.90 bandwidth. For each cadmium concentration cut-off value, adjusted PARs thus represent the estimated fraction of deaths that would be avoided in the population had cadmium exposure in participants with levels above that concentration been similar to cadmium exposure in participants with levels below that concentration, assuming that the effects of cadmium are causal and that other risk factors remained unchanged.