The NHANES study employed a comprehensive multistage survey method that incorporated stratification to enhance precision and accuracy in its outcomes. The research received ethical clearance from the National Center for Health Statistics Research Ethics Review Board, and prior to participation, all individuals provided written informed consent [19]. We acquired data from adult participants (20–59 years) in NHANES from 2011–2016. We excluded participants from the study if they had incomplete data on serum ApoB, TB-BMD, LS-BMD, or covariates (Fig. 1).

The BMD measurements were performed utilizing fan-beam densitometers (Hologic QDR-4500A, Bedford, MA, USA) through DXA scans [20]. ApoB within human serum can engage in an immunochemical reaction, resulting in the creation of immune complexes. These complexes induce alterations in the intensity of scattered light as it traverses through them, and this modified intensity is quantified using a Siemens Prospec chemistry analyzer, enabling precise determination of the ApoB concentration [7]. For further detailed information, please consult the official NHANES website.

We selected these confounders on the basis of judgment, previous scientific literature, all significant covariates in the univariate analysis, or their associations with the outcomes of interest or a change in effect estimate of more than 10% [7, 21]. Standardized questionnaires devised by NHANES researchers collected demographic data, including gender, age, race, education level, poverty-to-income ratio (PIR), marital status, drinking status (determined by whether participants consumed at least 12 alcoholic drinks annually), and smoking status (assessed based on whether participants had smoked a minimum of 100 cigarettes in their lifetime), and physical activity (min/week). Body mass index (BMI) was calculated from weight and height (kg/m²). Co-morbidities diagnosed by physicians, such as liver disease and cancer or malignancies, were determined based on self-reporting by participants. The levels of serum urea nitrogen, serum phosphorus, serum total protein, and serum uric acid were determined using standardized protocols. The intake of magnesium, calcium, sodium, and potassium was obtained through 24-hour dietary recalls conducted by trained interviewers at the mobile examination center.

Categorical variables were presented using frequencies, whereas continuous variables that followed a normal or skewed distribution were expressed as the mean or median, respectively. We employed One-Way ANOVA for normally distributed data, the Kruskal-Wallis H test for skewed distributions, and the chi-square test for categorical variables to compare the clinical characteristics across ApoB quartiles. We assessed the relationship between serum ApoB levels and TB-BMD as well as LS-BMD using linear regression models, presenting regression coefficients ($\beta$) along with their respective 95% confidence intervals [CI]. We utilized both unadjusted and multivariable-adjusted models. In our study, Model 1 was adjusted for gender, age, race, education level, marital status, PIR, and BMI. Model 2 was adjusted for Model 1 + drinking status, smoking status, physical activity, liver condition, cancer or malignancy. Model 3 was completely adjusted, including Model 1 + Model 2 + serum urea nitrogen, serum phosphorus, serum total protein, serum uric acid, calcium, magnesium, sodium, and potassium. To demonstrate the stability of the results and assess potential non-linearity, ApoB was divided into quartiles, and trend p-values were analyzed. In Model 3, we further evaluated the dose-response relationship using a restricted cubic spline (with four knots positioned at the 5th, 35th, 65th, and 95th percentiles of ApoB levels). A multicollinearity check was performed to evaluate the collinearity between the covariates and ApoB. In addition, we conducted multiple sensitivity analyses to assess the robustness of our findings. Subgroup analyses were performed to evaluate potential variations in the relationship between ApoB and BMD, stratified by age (using 45 years as the cutoff), gender, drinking status, and smoking status. Interactions between subgroups were tested using likelihood ratio tests. To further evaluate the robustness of the results, participants with extreme ApoB levels (outside the range of mean $\pm$ 3 standard deviations) were excluded. For handling missing data in the study, we employed a multiple imputation approach based on 5 repetitions and the chained equations method using the “mice” package in R, aiming to maximize statistical power and minimize potential bias arising from missing data.

The present MR study was composed of two steps (Fig. 1). In the first step, a two-sample univariable mendelian randomization (UVMR) approach was employed to assess the causal relationships between serum ApoB levels and BMD (TB-BMD, LS-BMD, and FN-BMD). The results from the UVMR analysis suggested a causal relationship between ApoB and TB-BMD as well as LS-BMD, while no causal relationship was observed between ApoB and FN-BMD. In the second step, both UVMR and multivariable mendelian randomization (MVMR) were employed to screen for mediators of the association between ApoB and the correlation of TB-BMD as well as LS-BMD among six cardiovascular diseases (stroke, coronary artery disease, atrial fibrillation, heart failure, venous thromboembolism, and peripheral atherosclerosis). Subsequently, a two-step MR analysis was performed to compute the mediating effects.

We utilized summary statistics obtained from extensive genome-wide association studies (GWAS) or meta-analyses of GWAS data to perform estimation of genetic correlations and MR analysis. The summary statistics for the serum ApoB dataset were derived from the UK Biobank GWAS summary statistics, which included 435,744 participants of European ancestry. Basic quality control measures were applied to the serum ApoB trait, including the exclusion of extreme outliers, stratification by gender and menopausal status, inverse normal transformation, removal of covariates, and subsequent reapplication of inverse normal transformation [22].

The summary statistics for TB-BMD were obtained from a meta-analysis of GWAS involving 56,284 participants. TB-BMD was assessed using DXA scans following standard manufacturer protocols [23]. The data is publicly available from the Genetic Factors for Osteoporosis (GeFOS) consortium [24]. The most extensive GWAS dataset for BMD assessed using DXA was acquired from GEFOS, encompassing FN-BMD data from 32,735 European participants and LS-BMD data from 28,498 European participants [25]. Additionally, we selected six cardiovascular diseases including stroke, coronary artery disease, atrial fibrillation, heart failure, venous thromboembolism, and peripheral atherosclerosis as candidate mediators. The GWAS datasets for these conditions were obtained from various organizations or consortia, namely MEGASTROKE, the UK Biobank study, Center for Statistical Genetics, Heart Failure Molecular Epidemiology for Therapeutic Targets, and the FinnGen consortium. All participants included in these studies shared European ancestry. Detailed information can be found in (Supplementary Table 1).

The MR analysis is based on three fundamental assumptions [26]. The initial assumption necessitates that instrumental variables (IVs) exhibit an association with the risk factor. Second, these genetic variants should not be connected to confounding factors. Third, IVs should influence the outcome risk solely through the risk factor, without involving other pathways. To meet these three assumptions in the statistical analysis, we conducted quality control procedures to select appropriate IVs. (1) We opted for single-nucleotide polymorphisms (SNPs) meeting the genome-wide significance threshold (p $\le$ 5 $\times$ 10^-8), ensuring their strong association with the exposure of interest. (2) We excluded palindromic SNPs due to their uncertain direction effect. Subsequently, we performed a linkage disequilibrium (LD) clumping procedure to preserve independent SNPs, setting an r² threshold of 0.001 and a clumping window size of 10,000 kb. Any SNPs exhibiting an r² value exceeding 0.01 with the lead SNP within 1000 kb were eliminated from consideration. (3) We assessed the robustness of IVs using the F statistic (detailed formulae are shown in Supplementary Table 2). SNPs having an F-statistic 10 are classified as weak IVs and were omitted from the analysis due to their potential to introduce bias into the results. (4) We corrected for the influence of ambiguous SNPs with non-concordant alleles and palindromic SNPs with ambiguous strands when possible. When correction was not feasible, we excluded these ambiguous and palindromic SNPs from the initially selected instrumental SNPs during the harmonization process [27]. Following a stringent screening process, the selected SNPs were employed as IVs in subsequent MR analyses.

In the two-sample UVMR analysis, we separately evaluated the causal relationships of ApoB with TB-BMD and LS-BMD, denoting the estimated coefficients as $\beta$0. We then conducted a two-step MR analysis to explore whether the six cardiovascular diseases could mediate the causal relationship between ApoB with TB-BMD and LS-BMD. The first step involved employing UVMR to assess the causal effects of ApoB on each cardiovascular disease, denoted as $\beta$1. To ensure the validity of the mediating effects, UVMR was employed to evaluate the absence of reverse causation between ApoB and each cardiovascular disease. The second step entailed the concurrent use of UVMR and MVMR to assess the causal relationship of each cardiovascular disease with BMD, both before and after adjustment for ApoB. The estimated values from MVMR were noted as $\beta$2. The proportion of mediation was calculated using the formula $\beta$1 $\times$ $\beta$2/$\beta$0, and the 95% CI for this mediation effect was computed using the delta method [28].

For UVMR, we assessed the potential causal associations between the exposure variable and the outcome variable using the fixed effects inverse variance weighted (IVW-FE) method. Additionally, we employed three other methods for additional analysis, including Weighted Median, MR Egger, and MR pleiotropy residual sum and outlier (MR-PRESSO). Sensitivity analyses were conducted to validate and ensure the reliability and stability of the results. These analyses encompassed various tests, including the assessment of heterogeneity using Cochrane’s Q test, as well as pleiotropy testing through the MR Egger intercept test [29]. In cases where heterogeneity was detected (p 0.05), the multiplicative random effects IVW (IVW-MRE) method was employed to evaluate the causal effect [30]. The IVW method served as the primary analytical approach in this study. Additionally, we employed the Weighted Median method, renowned for its resilience to potentially flawed instrumental variables [31]. Moreover, we utilized the MR Egger method, proficient at detecting and mitigating bias stemming from directional pleiotropy, operating under the assumption of the instrument strength independent of direct effect (InSIDE) principle [29]. To identify and address the influence of outlier SNPs exhibiting pleiotropic effects, we utilized the MR-PRESSO method, assessing their impact on causal relationships [32]. Finally, we utilized UVMR to assess the presence of reverse causation between ApoB with TB-BMD and LS-BMD. The MVMR analysis was conducted using the IVW method. The odds ratio (OR), $\beta$ coefficient, or proportions are presented along with their respective 95% CI to indicate the magnitude of the effects. We applied the Bonferroni correction to address multiple comparisons, determining the critical p value according to the number of exposures and outcomes [33].

This study involved 5695 prospective participants (20–59 years) from NHANES (2011–2016) who completed interviews. We excluded participants with missing data for ApoB, TB-BMD, and LS-BMD (n = 1687). Finally, 2930 participants with complete datasets were included (Fig. 1). The mean participant age (years) was 38.9 $\pm$ 11.5, and 1530 (52.2%) were men. The mean serum ApoB, TB-BMD, and LS-BMD were 91.0 $\pm$ 25.7 mg/dL, 1111.7 $\pm$ 108.0 mg/cm², and 1031.0 $\pm$ 149.6 mg/cm², respectively. The comprehensive population characteristics categorized by serum ApoB quartiles can be found in Supplementary Table 3.

The results of the univariate analysis for TB-BMD and LS-BMD are presented in Supplementary Table 4. Following multivariable adjustment, a significant association was observed between ApoB and TB-BMD. Model 3 revealed an adjusted $\beta$ of –0.26 (95% CI: –0.41 to –0.12, p 0.001) for TB-BMD when ApoB was considered a continuous variable. When compared to participants within the lowest ApoB quartile (20–72 mg/dL), the adjusted $\beta$ for the serum ApoB and TB-BMD in the 4th quartile (107–260 mg/dL) was –17.41 (95% CI: –28.12 to –6.70, p 0.001) in Model 3. Similarly, ApoB was also significantly correlated with LS-BMD. In Model 3, when ApoB was handled as a continuous variable, the adjusted $\beta$ associated with LS-BMD was –0.53 (95% CI: –0.75 to –0.31, p 0.001). Compared to participants within the lowest ApoB quartile, those in the 4th quartile exhibited an adjusted $\beta$ for serum ApoB and LS-BMD of –37.12 (95% CI: –53.21 to –21.03, p 0.001) in Model 3 (Table 1). There was a linear relationship between serum ApoB with TB-BMD (p for non-linearity = 0.771) and LS-BMD (p for non-linearity = 0.164) (Fig. 2). All VIFs were 2.0, indicating no significant multicollinearity. The subgroup analysis revealed no significant interactions in any of the stratified groups (Supplementary Figs. 1,2). After excluding extreme ApoB values, the results remained stable upon adjustment for all covariates (Supplementary Table 5). Missing data, as detailed in Supplementary Table 6, were addressed using multiple imputation, and the results remained robust following this procedure (Supplementary Table 7).

Fig. 2.

Restricted cubic spline model of ApoB with TB-BMD and LS-BMD. (A) Restricted cubic spline model of ApoB and TB-BMD. (B) Restricted cubic spline model of ApoB and LS-BMD. Solid line represents $\beta$; lightly shaded area represents 95% CI.

In the UVMR analysis, a causal relationship was observed between a 1-standard deviation (SD) increase in serum ApoB levels and a subsequent decrease in TB-BMD ($\beta$_{I⁢V⁢W-F⁢E} = –0.0424, 95% CI: –0.0658 to –0.0190, p = 0.0004) and LS-BMD ($\beta$_{I⁢V⁢W-F⁢E} = –0.0806, 95% CI: –0.1317 to –0.0296, p = 0.0020). The MR-PRESSO method confirmed the existence of the causal relationship, while the Weighted Median method and the MR Egger method demonstrated trends consistent with the study results. None of the methods indicated a causal relationship between ApoB and FN-BMD (Fig. 3). The detailed information of the SNPs is presented in Supplementary Table 8.

Fig. 3.

UVMR analysis of the effect of serum ApoB on TB-BMD, LS-BMD and FN-BMD. *The number of outliers; IVW-FE, fixed effects inverse variance weighted; IVW-MRE, multiplicative random effects inverse variance weighted; MR PRESSO, mendelian randomization pleiotropy residual sum and outlier; UVMR, univariable mendelian randomization; FN-BMD, femoral neck bone mineral density.

The calculations determined that the F-statistic of every SNP exceeded the threshold of 10 (Supplementary Table 2). Therefore, the likelihood of a weak IV affecting the results was relatively low. The sensitivity analysis revealed heterogeneity in the results (Cochrane’s Q test p 0.05), but no evidence of pleiotropy (MR-Egger intercept test p $>$ 0.05). The detailed results can be found in Supplementary Table 9. The subsequent application of the IVW-MRE method further confirmed the causal effect of ApoB on TB-BMD ($\beta$_{I⁢V⁢W-M⁢R⁢E} = –0.0424, 95% CI: –0.0746 to –0.0103, p = 0.0096) and LS-BMD ($\beta$_{I⁢V⁢W-M⁢R⁢E} = –0.0806, 95% CI: –0.1384 to –0.0229, p = 0.0062). In addition, we performed reverse association analyses of serum ApoB on TB-BMD, LS-BMD, and FN-BMD, and these analyses showed no evidence of reverse causality (Supplementary Table 10).

In the UVMR, employing the IVW-FE method revealed a causal relationship between genetically determined serum ApoB levels (per 1-SD increase) and an elevated risk of all five cardiovascular diseases (except venous thromboembolism) (Table 2). All SNPs for each variable exhibit sufficient instrument strength, with F-statistics exceeding 10 (Supplementary Table 2). The MR-Egger intercept test indicated the absence of horizontal pleiotropy across all analyses (all p $>$ 0.05). The Cochrane’s Q test indicated heterogeneity across all analyses (Supplementary Table 11). Therefore, employing the IVW-MRE method to assess causality revealed that serum ApoB levels were associated with stroke (OR_{I⁢V⁢W-M⁢R⁢E} = 1.1068, 95% CI: 1.0505 to 1.1662, p = 0.0001), coronary artery disease (OR_{I⁢V⁢W-M⁢R⁢E} = 1.3693, 95% CI: 1.2649 to 1.4822, p = 7.80 $\times$ 10^-15), heart failure (OR_{I⁢V⁢W-M⁢R⁢E} = 1.0729, 95% CI: 1.0202 to 1.1284, p = 0.0062), and peripheral Atherosclerosis (OR_{I⁢V⁢W-M⁢R⁢E} = 1.2790 , 95% CI: 1.2620 to 1.4199, p = 4.88 $\times$ 10^-15). In the reverse MR analysis, only a causal relationship between coronary artery disease and ApoB was observed (Supplementary Table 10).

The causal relationship between heart failure and TB-BMD remained regardless of ApoB adjustment ($\beta$_{M⁢V⁢M⁢R} = –0.1126, 95% CI: –0.1955 to –0.0296, p = 0.0078) (Fig. 4). Sensitivity analysis indicated that there was no heterogeneity or horizontal pleiotropy in the causal relationship between heart failure and TB-BMD (Supplementary Table 12). Based on the following screening process, we ultimately confirmed that heart failure can mediate the causal relationship between ApoB and TB-BMD. (1) There should be a unidirectional causal relationship between ApoB and TB-BMD. (2) A causal relationship between heart failure and TB-BMD should still exist regardless of whether ApoB is adjusted. (3) The relationship between ApoB and heart failure should be in the same direction as the relationship between heart failure and TB-BMD. We applied a two-step MR analysis to assess the mediating effect of heart failure on the causal relationship between ApoB and TB-BMD. Heart failure explained 18.69% (2.18%, 35.21%) of the total effect of serum ApoB levels on TB-BMD (p 0.05). No causal relationships were identified between other cardiovascular diseases and either TB-BMD or LS-BMD.

Fig. 4.

UVMR and multivariable mendelian randomization (MVMR) analysis of the effect of potential mediator on TB-BMD and LS-BMD.