The present study utilized samples from the 4th and 5th Korean National Health and Nutrition Examination Survey (KNHANES). The original data were obtained from July 2008 to May 2011. Heavy metal levels, bone mass density (BMD), general health, and nutrition were assessed during this period. The KNHANES is a nationwide cross-sectional survey conducted each year by the Korea Centers for Disease Control and Prevention. The goal of the KNHANES is to assess and monitor health and nutritional status in the Republic of Korea. The target population consists of all non-institutionalized citizens residing in Korea. The KNHANES utilizes a complex survey design and is constructed via two-stage stratified cluster sampling. The KNHANES represents the target population using sampling weights. Additional details have been described in a previous study .
The population of the present study consisted of post-menopausal women ≥50 years of age. Among 37,753 participants in the 2008-2011 KNHANES, a total of 5432 participants were randomly sampled for measurements of heavy metal concentrations in the blood, BMD, and nutrient intake. Male participants (n = 2569), those under age 50 (n = 1766), and non-post-menopausal women (n = 66) were excluded from the analysis. Multiple imputations were then used for missing covariates in the remaining 1031 participants (Fig. 1). Other exclusion criteria (e.g., early menopause, cancer) were not adopted, as these variables are effect modifiers or mediators  rather than confounders in the present model [15, 16] (Online Resource 1).
Blood cadmium levels
Blood samples were obtained from participants for the measurement of blood cadmium levels. Samples were collected in vacuum tubes (vacutainer), mixed with anticoagulants, and stored in a refrigerator. They were analyzed within 24 h in the diagnostic laboratory of NeoDin Medical Institute in Seoul, South Korea. Blood levels of cadmium were measured via graphite furnace atomic absorption spectrometry (GFAAS) using a Perkin Elmer AAnalyst 600 system (PerkinElmer, Turku, Finland). In the internal validation step, inter-assay coefficients of variation (CVs) continued to be within an acceptable range (≤10%). External validation was performed by the Korea Occupational Safety and Health Agency and German External Quality Assessment Scheme (G-EQUAS). The benchmark dose of blood cadmium has yet to be determined for the Korean post-menopausal population. Therefore, blood cadmium levels were categorized into quartiles following previous study .
BMD, osteopenia, and osteoporosis
BMD was measured for the total hip, femoral neck, and lumbar spine (L1–L4) using dual-energy X-ray absorptiometry (DXA, DISCOVERY-W fan-beam densitometer; Hologic Inc., Bedford, MA, USA). For precision assessment, radiologic technologists were trained to make the CVs of randomly double-checked participants ≤1.8% for the total hip, ≤2.5% for the femoral neck, and ≤1.9% for the lumbar spine. Osteoporosis and osteopenia were defined according to T-scores. The T-score is the standard deviation from the average peak BMD for sex-matched and race-matched (Japanese) populations . Osteoporosis was defined as a T-score of −2.5 or less at any site, while osteopenia was defined as a T-score less than −1.0 and greater than −2.5.
To minimize confounding bias, we selected confounders based on the disjunctive cause criterion. In this criterion, variables that are cause of exposure, outcome, or both are controlled. If exposure is the cause of a variable, it is regarded as a mediator or collider, rather than a confounder. Adjusting these variables would lead to bias . A detailed explanation of confounder selection is presented in the form of a directed acyclic graph (DAG) (Online Resource 1).
Age, education level, household income (quartile), occupation, and residence were included as confounders reflecting socioeconomic status [1, 19]. To minimize residual confounding, age was adjusted as both a continuous and categorical variable. The year in which participants were surveyed was also adjusted .
Body mass index (BMI) was calculated based on measured height and weight . Lifestyle factors such as current alcohol consumption, smoking exposure, and physical activity were also included [1, 19, 22]. When defining smoking exposure, we considered that women smokers tend to under-report their smoking status by more than 50%, as identified by a previous KNHANES . Therefore, we defined smoking exposure as 50 ng/mL or more urinary cotinine, based on the results of previous studies [24, 25]. Gas chromatography-mass spectrometry (GC-MS) with a Clarus 600/600 T system (Perkin Elmer, Waltham, MA, USA) was used to measure urinary cotinine. Other information was obtained using questionnaires. Alcohol consumption was categorized into “less than once per month,” “greater than once per month but less than twice per week,” and “greater than twice per week” . Physical activity was categorized as vigorous (3 days/week, ≥20 min/session), moderate (5 days/week, ≥30 min/session), or daily walking (5 days/week, ≥30 min/session).
Vitamin D levels were measured based on serum 25-hydroxyvitamin D(25(OH)D3) concentration using a radioimmunoassay (1470 WIZARD gamma-Counter, PerkinElmer, Turku, Finland) . Food consumptions were recorded based on the 24-h dietary recall method, following which nutrient intake was calculated. Fish and seaweed consumption  as well as energy, protein, calcium, phosphate, potassium, and retinol intake  were included in the model. Intake of certain foods was classified as less than once a week, once a week, or more than once a week . Nutrient intake was adjusted as a continuous variable. The previous use of hormone therapy or oral contraceptive (>1 month of use) was obtained using a questionnaire. The use of hormone therapy was categorized into “never,” “less than 2 years,” and “2 years or more” [31, 32].
Among the study population (n = 1031), 122 participants (11.8%) had at least one missing covariate or outcome value. Multiple imputations were used to impute item non-response. Assuming missing at random (MAR), all variables in the model were included in the multiple imputations. For continuous variables, we used predictive mean matching due to their non-normal distribution. For categorical variables, we used different types of logistic regression for categorical variables based on the variable type (i.e., binary, multinomial, or ordinal). Imputation was performed using the Markov chain Monte Carlo (MCMC) method, and 20 imputed sets were created.
The design and sampling weights of all participants in the KNHANES were included in the analysis to account for the complex survey design. Multinomial logistic regression was used to analyze the associations between cadmium and the risk of osteopenia and osteoporosis. Blood cadmium levels were divided into quartiles, as previously described . In the multinomial model, we used replicate weights with the bootstrap method to estimate 95% confidence intervals (CIs). Bootstrapping was performed 250 times with five imputed sets each. Regression methods were used to examine linear P trends in blood cadmium levels.
Exploratory and sensitivity analyses
For exploratory purposes, we investigated the association of cadmium with osteopenia and osteoporosis by subregion. Osteopenia and osteoporosis in the total hip, femoral neck, and lumbar spine were defined according to T-scores. The multinomial model was respectively used in each subregion except for osteoporosis in total hip due to the low number of cases.
Sensitivity analyses were performed based on three different outcome definitions. (1) Some patients may undergo treatment following the development of osteoporosis, leading to an increase in T-score. If the proportion of treated patients varies based on cadmium levels, using only current T-scores to define osteoporosis can lead to biased estimates. Therefore, patients with T-scores greater than −2.5 answering “yes” to “Have you ever been diagnosed with osteoporosis by a doctor?” and “Are you currently under treatment for osteoporosis?” were classified into the osteoporosis group. (2) The reference for calculating T-scores was derived from a Japanese population ; however, reference values may differ for Korean and Japanese populations . Therefore, osteopenia and osteoporosis were defined according to T-scores that had been re-calculated using Korean reference values. (3) T-scores based on Korean reference values and current treatment information were both used to define osteoporosis and osteopenia.
To examine the robustness of the results, additional sensitivity analysis was performed. (1) Blood cadmium level was categorized into quintile (five sections). (2) Outcome was defined as T-score −1.8 or less and −2.0 or less. For each outcome, logistic regression was used. (3) BMD and T-score at each site were analyzed. Moreover, the lowest BMD and T-score among sites were analyzed.
The study protocol was approved by the Institutional Review Board of Jaseng Hospital of Korean Medicine (JASENG 2020-08-019), who waived the requirement for informed consent due to the nature of the study.