The NHANES study is a multi-stage, stratified, and nationally representative study of the US population conducted by the National Center for Health Statistics of the Centers for Disease Control and Prevention. It aims to assess the nutrition and health status of Americans. The survey includes demographic, dietary, examination, laboratory, and questionnaire data. Prior to data collection, all study procedures were authorized by the ethics review committee of the National Center for Health Statistics, and informed consent forms were signed by all participants. For our study, we specifically analyzed data collected between 2007 and 2018. Subjects were excluded from our study for the following reasons: (1) Missing T2DM data; (2) Missing HDL-C data; (3) Aged under 20 years old; (4) Missing covariate (such as smoking, alcohol consumption, hypertension, hyperlipidemia, etc.) data ( Figure 1 ).

We define T2DM as any of the following: (1) HbA1c levels equal to or greater than 6.5%; (2) serum glucose levels exceeding 200 mg/dL at 2 hours after a 75 g glucose load (OGTT); (3) fasting glucose levels equal to or greater than 126 mg/dL; (4) self-reported diagnosis of diabetes; (5) self-reported use of insulin or other diabetes medication. The duration of diabetes was determined by subtracting the participant’s current age from the self-reported age at diagnosis, or zero for individuals diagnosed during the NHANES examination.

The HDL-C data (Code: LBDHDDSI) included in our study, spanning from 2007 to 2018, were stored and analyzed by the University of Minnesota in Minneapolis, Minnesota, USA. Throughout this period, there were no changes in the laboratory methods or sites, but there were updates in the laboratory equipment. Specifically, from 2007 to 2012, the Roche Modular P Chemistry Analyzer was used for measuring the analyte, while from 2013 to 2018, the Roche Cobas 6000 and Roche Modular P Chemistry Analyzer were employed. The contracted laboratory adhered to the Westgard rules during the analysis of NHANES specimens (16), and the NHANES Quality Assurance and Quality Control (QA/QC) protocols complied with the requirements of the Clinical Laboratory Improvement Amendments of 1988, ensuring the reliability of the data source. Furthermore, several other common lipid traits, such as Low-density lipoprotein cholesterol (LDL-C) (code: LBDLDLSI), TG (code: LBDTRSI), and Total cholesterol (TC) (code: LBDTCSI), have been included in this study for further sensitivity analysis.

Our covariates included age, sex, race/ethnicity (Mexican American, other Hispanic, non-Hispanic white, non-Hispanic black, non-Hispanic Asian, other race), BMI, smoking status, alcohol consumption status, marital status, education level (less than 9th grade, 9-11 grade, high school graduate, some college/AA degree, and college graduate), poverty income ratio (PIR), cardiovascular disease, hypertension, and diabetes. BMI was divided into three categories: Normal (<25), Overweight (≥25, <30), and Obese (≥30). PIR was divided into three categories: Low (≤1.39), Medium (>1.39, <=3.49), and High (>3.49). Sedentary status is classified as severe (Defined as being sedentary for 480 minutes or more per day) and mild (Defined as being sedentary for less than 480 minutes per day) according to World Health Organization guidelines on physical activity levels (17). Smoking status was classified as Current Smoker (Defined as having smoked more than 100 cigarettes in a lifetime and still smoking), Former Smoker (Defined as having smoked more than 100 cigarettes in a lifetime but no longer smoking), Never smoke (defined as smoking less than 100 cigarettes in their lifetime). Alcohol status was classified as Current drinker (Defined as having consumed more than 12 alcoholic beverages of any type in a lifetime and still drinking), Former drinker (Defined as having consumed more than 12 alcoholic beverages of any type in a lifetime but not in the past year), Never drinker (Defined as not having had more than 12 alcoholic beverages of any type in a lifetime). Hypertension was defined according to the American Heart Association/American College of Cardiology (AHA/ACC) 2017 guidelines as systolic blood pressure≥130 mmHg or diastolic blood pressure ≥80 mmHg and self-reported diagnosis or use of antihypertensive medication. As per the guidelines set by the Adult Treatment Panel III of the National Cholesterol Education Program (NCEP-ATP III), hyperlipidemia is defined by the following criteria: total cholesterol levels equal to or exceeding 200mg/dL, triglyceride levels equal to or exceeding 150mg/dL, HDL cholesterol levels below 40mg/dL for men and below 50mg/dL for women, or LDL cholesterol levels equal to or exceeding 130mg/dL (18). Additionally, individuals who self-reported the use of cholesterol-lowering medications were also classified as having hyperlipidemia. For CVD, a positive response to any of the following questions was defined as CVD: “Has a doctor or other health professional ever told you that you have congestive heart failure (CHF)/coronary heart disease (CHD)/angina/heart attack/stroke?”.

The NHANES database release files provide weights for each sample, such as the Mobile Examination Center (MEC) examination weights (19). The calculation of weights enables NHANES’ complex sampling design to achieve national representativeness. In this study, due to the inclusion of a large amount of laboratory data obtained in the MEC, we utilized the combined wtmec2yr for analysis. Additionally, since NHANES 2007-2018 combines data from 5 consecutive survey cycles (NHANES operates on a biennial cycle), the new weight (2007-2018) is calculated as 1/2 × wtmec2yr. The “survey” package’s svydesign function is employed for weighted computations. By utilizing quartiles, we can divide HDL-C levels into four categories (Q1, Q2, Q3, and Q4), with Q1 used as the reference. Continuous variables were reported as mean ± standard deviation (SD), while categorical variables were presented as individual counts (N) and percentages (%). Weighted t-tests (for continuous variables) or weighted chi-square tests (for categorical variables) were used to assess differences between T2DM and non-T2DM subjects. Kruskal-Wallis tests (for continuous variables) or weighted chi-square tests (for categorical variables) were used to evaluate differences among the four groups based on HDL-C exposure levels. Multivariable linear regression and restricted cubic splines (RCS) were used to analyze the correlation between HDL-C and T2DM. Initially, an unadjusted model was fitted, followed by stepwise adjustment for covariates. Model 1 adjusted for age, sex, and race; Model 2 further adjusted for PIR, BMI, marital, education level, cotinine, sedentary, alcohol, and smoking based on Model 1; Model 3 additionally adjusted for CVD, hyperlipidemia and hypertension based on Model 2. Results were presented as odds ratios (OR) with their corresponding 95% confidence intervals (95% CI). Subgroup analyses were conducted for significant results. Finally, a logistic regression model was used to assess the significance of the interaction between HDL-C and covariates on T2DM. In addition, considering the potential links of LDL-C, TC, and TG with DM, we categorized LDL-C, TC, and TG levels into three groups—Normal, Moderately elevated, and High - based on the recommendations for the general population outlined in the NCEP-ATP III. We conducted a sensitivity analysis to further adjust these lipids in the association between HDL-C and T2DM. Furthermore, we investigated the relationships between LDL-C, TC, TG, and T2DM. Finally, based on the above analyses, we conducted a reevaluation of subgroup analyses and interaction analyses.

Under the framework of two-sample Mendelian randomization studies, this study utilizes SNPs as instrumental variables (IVs) for MR analysis, and adheres to the three fundamental assumptions of MR analysis (20). Assumption I posits that genetic variations are robustly associated with exposure (HDL-C). To minimize bias, it is important to use IVs with strong correlations (P< 5 × 10^-8), and the F-value of the genetic tool should be above 10. Assumption II states that any confounding factors that influence the relationship between genetic variation and the outcome of exposure are not relevant. To address this issue, we restricted the sample population to Europeans. Additionally, we obtained phenotype information for each SNP from Phenoscanner (http://www.phenoscanner.medschl.cam.ac.uk/) and manually excluded any SNPs that were found to have an impact on outcome-related phenotypes. Assumption III asserts that the impact of genetic variation on the outcome can only be achieved through its association with the exposure. To address this issue, we utilized MR-Egger regression and the MR-PRESSO method to minimize the potential for horizontal pleiotropy and conducted sensitivity analyze.

To construct IVs for proxying HDL-C, we obtained summary statistics data from the Global Lipids Genetics Consortium (GLGC) GWAS, which included 94,595 individuals of European ancestry (21). The IVs for T2DM were obtained from a GWAS meta-analysis, which included a total of 12,931 T2DM patients and 57,196 healthy controls of European ancestry (22).

We assessed the strength of the SNPs used as instruments using the F-statistic and considered only SNPs with an F-statistic greater than 10 to minimize weak instrumental bias (20). The instrument strength (F-statistic) was calculated using the following equation: F= [R² (N-K-1)/K (1- R²)]. [N denotes the sample size of the exposure factor, k denotes the number of SNPs in each instrument, and R² denotes the variance explained by the instrument. R² = 2×EAF× (1-EAF) × Beta², which was calculated from the equation proposed by Shim et al. (23)].

We used a threshold of P < 5 × 10^-8 and LD r² ≤0.001 as the extraction criteria for exposure’ IVs. We employed the inverse variance weighting (IVW) method as our primary analysis, supplemented by weighted median, weighted mode, MR-Egger and simple mode. Due to the potential association of TC, TG, and LDL-C with T2DM, we constructed a multivariable Mendelian randomization (MVMR) model to eliminate their impact on the causal relationship between HDL-C and T2DM. MVMR is an extension of the standard univariate MR, designed to analyze the causal effects of different exposures on outcomes and estimate the direct causal effects of each exposure in a single analysis (24).

We used MR-Egger regression to assess the possibility of horizontal pleiotropy, and the average pleiotropic effect of the IVs was represented by the intercept term of the MR-Egger regression (25). Additionally, the MR Pleiotropy REsidual Sum and Outlier (MR-PRESSO) were used as a supplement to evaluate horizontal pleiotropy. Its functionalities include detecting horizontal pleiotropy, correcting for horizontal pleiotropy by removing outliers, and determining if there are substantial changes in the causal effects before and after removing outliers (26). Heterogeneity was quantified using Cochran’s Q statistic. Moreover, a leave-one-out analysis was performed to assess the influence of each peripheral SNP on the results.

The statistical analyses for this study were conducted using R (version 4.2.2). Additionally, MR analysis utilized the R package “TwoSampleMR”. A two-tailed P-value < 0.05 was considered statistically significant.

Finally, a total of 9,420 individuals were included in this study. Based on the exclusion criteria, a total of 2,224 participants (18%) were classified as having T2DM. Compared to non- T2DM participants (7,196 individuals), T2DM patients generally tend to be older, have lower household income, engage in sedentary behavior more frequently, and exhibit a higher prevalence of obesity. Additionally, T2DM participants are more likely to have comorbidities such as hypertension, hyperlipidemia, and CVD. Finally, it is important to emphasize that compared to individuals without T2DM, those with T2DM exhibit a significant reduction in HDL-C levels. ( Table 1 ) These findings highlight the distinct characteristics and health challenges faced by individuals with T2DM. According to the quartile method, we divided HDL-C levels into four intervals: Q1 (0.28-1.09], Q2 (1.09-1.32], Q3 (1.32-1.60], and Q4 (>1.60), with units in mmol/L. We observed that individuals with higher HDL-C levels had lower BMI, fewer smokers, and a lower prevalence of hypertension, hyperlipidemia, cardiovascular disease, and T2DM. On the other hand, individuals with lower HDL-C levels had a higher prevalence of higher PIR levels and alcohol consumption ( Supplementary Table S1 ).

By constructing a multiple linear regression model, we found that when using Q1 (low HDL-C) as the reference, in all models, the Q2, Q3, and Q4 groups were significantly negatively associated with the risk of T2DM (P < 0.001). The risk of T2DM decreased as HDL-C levels increased, and this result remained robust after adjusting for covariates. Particularly, after adjusting for strict variables such as age, gender, race, education level, marital status, simvastatin, sedentary behavior, PIR, alcohol consumption, smoking, hypertension, hyperlipidemia and CVD (model 3), compared to the Q1 group, the average risk ratios for T2DM decreased by 23%, 49%, and a remarkable 71% in the Q2, Q3, and Q4 groups, respectively. These results did not differ significantly across different models, indicating the reliability of HDL-C in predicting T2DM risk ( Table 2 ). We further utilized the restricted cubic spline (RCS) model to capture the non-linear relationship between HDL-C and T2DM. After adjusting for covariates, this relationship remained significant (P = 0.028, Figure 2A ).

We conducted interaction analyses for the covariates that showed significant differences between T2DM and non-T2DM in Table 1 . The results showed that the effect of HDL-C levels on T2DM varied with age (P for interaction = 0.006, Table 2 ), while there were no significant interactions for the other covariates (P > 0.05). Using the Q1 range of HDL-C levels as the reference, we subdivided age into young (20-30], middle-aged (31-45], middle-old (46-60], and elderly (>60) groups (while adjusting for other covariates). We found that there was no association between HDL-C and T2DM in the young population, but as age increased, this association strengthened. Particularly for the elderly group, individuals in the Q2 range already had a 40% lower risk of T2DM [OR=0.60, 95%CI (0.46, 0.79), P < 0.001], and the Q4 range had a 73% lower risk [OR=0.27, 95%CI (0.20, 0.36), P < 0.001]. HDL-C levels showed varying degrees of association with reducing T2DM risk in other subgroups.

In the further sensitivity analysis, we conducted a reanalysis after including the additional variables LDL-C, TC, and TG (baseline characteristics of patients are provided in Supplementary Table S2 ). The results showed that despite changes in the included population, the association between HDL-C in the Q4 group (>1.60) and T2DM remained robust [OR=0.43, 95%CI (0.20, 0.92), P = 0.03]. In subgroup analysis results, there was no interaction between other lipid levels and the reduction of T2DM risk by HDL-C, demonstrating the reliability of HDL-C in predicting T2DM. ( Supplementary Table S3 ) Furthermore, after adjusting for age, sex, race, education, marital status, PIR, BMI, sedentary behavior, cotinine, alcohol, smoking, hypertension, hyperlipidemia, CVD, LDL-C, HDL-C, TC, and TG separately (with lipid excluded as independent variable), LDL-C, TC, and TG were no longer statistically associated with T2DM. ( Supplementary Table S4 ).

When using HDL-C as an exposure instrument, the assessment of genetic instrument strength, as indicated by a Total-F-statistic of 106.94 (R² = 4%, with individual SNP F-values ranging from 29.95 to 962.62, Supplementary Table S5 ), suggests that there is no evidence of statistical bias resulting from weak instrument bias. In the IVW analysis, we observed a significant inverse association between HDL-C levels and the risk of T2DM (OR = 0.69, 95% CI = 0.52-0.82; P = 1.41× 10^-13). This finding indicates that for per 1 mmol/L increase in HDL-C. All alternative analytical approaches yielded statistically significant results that were consistent with the IVW analysis. Moreover, no evidence of heterogeneity or horizontal pleiotropy was observed, as indicated by non-significant Cochran’s Q test and MR-Egger intercept/MR-Presso p-values >0.05 ( Table 3 ). Additionally, a sensitivity analysis (leave-one-out analysis) demonstrated the stability of the results, showing consistent findings even when individual SNPs were excluded ( Figure 2B ). These findings provide robust statistical evidence supporting the negative association between HDL-C levels and the risk of T2DM. The MVMR analysis, adjusted for LDL-C, TC and TG, showed that HDL-C results for T2DM remained robust. (P =8.02× 10^-3, Supplementary Table S6 ). The consistency across different analytical methods strengthens the validity and reliability of our study’s conclusions. The results of the reverse MR analysis elucidated that having T2DM is not causally associated with HDL-C levels [IVW, OR: 1.00, 95% CI (0.97, 1.03), P = 0.089]. Additionally, the heterogeneity, horizontal pleiotropy and sensitivity analysis of the reverse MR analysis was also deemed reliable ( Supplementary Table S7 ).