To ensure the integrity of the MR analysis, this study adhered to the STROBE-MR guidelines framework (13). The exposure factors were selected from the largest available public GWAS database, from which SNPs representing LSB were screened as IVs for genetic prediction. The outcome factors were chosen from the largest non-overlapping GWAS data on T2D and impaired glycemic homeostasis. In the first stage, two-sample MR was employed to derive significant causal relationships. In the second stage, utilizing multivariate MR, we investigated the interrelationships among three LSB phenotypes, the occurrence of T2D, and impaired glycemic homeostasis. In the third stage, the potential mediating pathways between LSB and T2D and glycemic homeostasis impairment were examined, along with their genetically predicted effects. The flowchart of the overall study design is shown in Figure 1 . The STROBE-MR guidelines2 checklist is presented in Supplementary Table 1 .

In our two-sample MR analysis, we used the Inverse Variance Weighted (IVW) method as the key metric to evaluate the predicted effect sizes of the causal association between genetically determined LSB and both T2D and impaired glycemic homeostasis. The IVW method has been widely used in TSMR studies, and its validity and accuracy exceed those of other MR analysis methods (22). However, if the study includes invalid IVs in the SNPs, they may affect the outcome through pathways other than exposure factors, leading to significant errors in the IVW method results (23). Compared to the IVW (fixed effects), the IVW (random effects) produce the same beta, but with slightly larger standard errors. IVW (random effects) reduces the bias of results due to the presence of unavoidable heterogeneity, and given the potential heterogeneity between the estimates of different IVs, we chose IVW (random effects) as the main analytical outcome (24). To ensure the reliability of the MR analysis results, we employed a combination of MR-Egger (24), weighted median (25), simple median, and MR-RAPS methods to assess the results jointly. When more than 50% of the included IVs are valid, the weighted median and simple median method results can be considered reliable (24). When the Instrument Strength Independent of Direct Effects (InSIDE) assumption holds, which means that horizontal pleiotropy arises not only due to a single confounder but through multiple confounders and vice versa, the MR-Egger regression method can correct for the effects of genetic predisposition confounding and provide more accurate estimates of causality (24). MR-RAPS (Mendelian Randomization Robust Adjusted Profile Score), which adjusts for genetic predisposition confounding when estimating the result of a cause between the exposure variable and outcome variable, provides stronger robustness to heterogeneity and weak IVs with genetic predisposition confounding.

A reverse MR analysis was performed to assess potential reverse causal effects. In the second stage of Multivariate Mendelian randomization (MVMR) analysis, the independent effect of each LSB phenotype on the causal relationship between T2D and impaired glycemic homeostasis was assessed. The IVW method was similarly employed as the primary outcome for genetic predictive effects in MVMR analysis.

Two-sample MR analysis revealed causal associations of genetically predicted LSB with the risk of developing T2D and impaired Glycemic Homeostasis. To investigate potential mediating pathways in the TSMR results, two-step MR and mediation analysis were employed to examine potential factors that play a mediating role in the development of T2D and impaired Glycemic homeostasis. From previous studies, obesity and body size-related factors (BMI (26), body fat percentage (27), waist circumference, hip circumference, waist-hip ratio (28), triglyceride (29)), systemic inflammation level (CRP (30)), and educational attainment (31) were found to be associated with sedentary behavior and T2D onset. Consequently, these factors were incorporated as potential mediators in this stage of the study. The first step investigated the causal association between genetically predicted LSB and potential mediators, using IVW as the primary method (23) to estimate the predicted effect values of LSB for each mediator. The outcomes were presented as correlation coefficients (β) along with their corresponding 95% confidence intervals. For mediators extracted from the same GWAS or database, we used Bonferroni-corrected P-values to assess whether the results were statistically significant, with a corrected p-value of 0.017 for waist circumference, hip circumference, and waist-to-hip ratio. The second step employed MVMR analysis to evaluate the influence of risk factors for T2D (i.e., results with significance in the first step) on the association with the development of T2D after correcting for the effect of LSB. Considering that including too many variables in the study may introduce a more serious problem of covariance, multivariate LASSO regression was used to help screen out unnecessary exposure factors (32, 33).In the third step, based on the above analysis, the predicted values of the respective mediating effects of each risk factor on T2D were obtained using the coefficient product method (34) to assess the proportion of mediating effects of each mediator on T2D.

MR studies rely on three assumptions (10): 1) genetic IVs selected to represent exposure factors must be strongly correlated with exposure factors; 2) these genetic IVs cannot be correlated with any other confounding factors; 3) these genetic IVs can only affect the outcome event by acting on the exposure factor and cannot be directly correlated with the outcome event. Violation of any of these three assumptions would result in significant errors in the MR study.

For each SNP used as an IV in this study, the strength of the effect was assessed using the F statistic (35), calculated as F   s t a t i s t i c = N − k − 1 k ( R 2 1 − R 2 ) (when there were less than 5 SNPs: R i 2 = 2 * β i 2 * E A F * 1 − E A F , when there were more than 10 SNPs: R i 2 = 2 * E A F i * 1 − E A F i * β i 2 2 * E A F i * 1 − E A F i * β i 2 + 2 * N * E A F i * 1 − E A F i * s e i 2 ) (36), where EAF is the effect allele frequency, β_i is the estimated genetic effect on LSB, se_i is the standard error of the estimated genetic effect of the exposure, N denotes the sample size of the exposure group, and k signifies the total number of IVs incorporated. An F statistic ≥ 10 is generally considered to indicate that the results are less influenced by weak IVs (37).

In addition, various sensitivity analyses were conducted in this study to examine potential violations of the three main MR assumptions (10) and potential errors in the IVW method results. Cochran’s Q test and the I² statistic were employed to assess heterogeneity and gauge the consistency of the IVs in estimating causal effects. A P-value of 0.05 or lower indicated the existence of pleiotropy (38). When heterogeneity was present, the IVW method served as the primary outcome of the MR analysis. Sensitivity was evaluated using MRPRESSO and leave-one-out test analyses, examining the effect of each SNP on the association with the outcome to ensure the results were free from statistical horizontal pleiotropy (39). Pleiotropy presence was assessed using the intercept obtained from MR-Egger regression (40), with P< 0.05 indicating statistically significant pleiotropy. In this study, we integrated and formatted the GWAS summary data for exposure and outcome variables. By sampling one million unique variants in the formatted data, we estimated a range of interference parameters. Ultimately, we employed the formatted data, estimated interference parameters, and a set of variants adjusted for LD to fit the CAUSE model, which enabled us to evaluate the likelihood of causal and sharing effects (41).

All data analyses in this study utilized R 4.2.2 software and several associated R packages, including TwoSampleMR (ver.0.56) (42), MRPRESSO (ver.1.0) (39), Mendelian Randomization (ver.0.7), and CAUSE (version 1.2.0) (41).

We utilized the GWAS associated with LSB from the UK Biobank. Following the exclusion of SNPs that failed to meet the genome-wide correlation threshold (5*10^-8) and the removal of SNPs in LD with higher P-values, we subsequently screened and eliminated SNPs that might potentially cause confounding bias. Ultimately, three phenotypic SNPs were obtained as IVs. The GWAS features and related literature used in the study are shown in Table 1 . The F statistics for all SNPs included and used in this study are presented in Supplementary Tables 2 , 3 .

In the first stage, the two-sample MR results indicated a significant positive causal relationship between genetically predicted leisure television watching behavior and T2D prevalence, as well as glycemic homeostatic impairment (including HbA1c, fasting glucose, fasting insulin, 2-hour glucose, HOMA-IR, and HOMA-B). After log-transformation to odds ratio (OR) values, each one log odds increase in leisure television watching time was associated with a 64% increase in the prevalence of T2D (OR 1.64, 95% CI 1.39-1.93, P< 0.001), a 3% increase in HbA1c (OR 1.03, 95% CI 1.01-1.05, P< 0.05), a 6% increase in fasting glucose(OR 1.06, 95% CI 1.02-1.10, P< 0.001), an 8% increase in fasting insulin (OR 1.08, 95% CI 1.04-1.12, P< 0.05), a 13% increase in 2h glycemic (OR 1.13, 95% CI 1.00-1.28, P< 0.05), a 19% increase in HOMA-IR (OR 1.19, 95% CI 1.10-1.28, P< 0.001), and a 10% increase in HOMA-B (OR 1.10, 95% CI 1.03-1.17, P< 0.05). The results of the four complementary methods—MR-Egger, simple median, weighted median, and MR-RAPS—remained consistent (as shown in Figure 2 ). The wider confidence intervals for MR-Egger compared to the other methods may be due to its weaker statistical power relative to the IVW method (43). Cochran’s Q statistic in the heterogeneity test ranged from 87.0 to 209.4 (P = 7.76*10^-8 - 0.53). The MR-Egger intercept analysis in the test of pleiotropy did not exhibit statistically significant directional pleiotropy (P > 0.05). MR-PRESSO showed no impact on the significance of the results after removing potential outlier loci. The leave-one-out sensitivity analysis demonstrated that the overall results’ statistical significance was not substantially impacted by the MR outcomes of other IVs after sequentially removing SNPs. No significant causal association was found between leisure TV watching behavior and fasting insulinogenic values at the genetic level. Figure 2 depicts a forest plot of the risk association between leisure TV watching behavior and T2D prevalence and impaired Glycemic homeostasis, with different colors representing the results of various study methods.

A significant negative causal relationship was observed between leisure computer use and the prevalence of T2D, with a 35% reduction in T2D prevalence per 1 log odds increase in leisure computer use time after log-transformation to OR (OR 0.65, 95% CI 0.50-0.84, P< 0.001). No notable gene-level associations were identified between leisure computer use and other forms of glycemic homeostasis impairment. The results of the four complementary methods remained consistent. Cochran’s Q statistic equaled 60.0 (P< 0.05) in the test of heterogeneity. MR-PRESSO in the test for pleiotropy did not identify potentially pleiotropic outlier loci. Additionally, no evidence of targeted pleiotropy was found by MR-Egger intercept analysis (P > 0.05).

Moreover, CAUSE analyses indicated that the causal model exhibited a superior fit compared to the sharing model (P< 0.001) when examining the relationship between leisure TV watching and T2D prevalence, as illustrated in Figure 3 and Tables 2 , 3 . This finding suggests a causal association between the two phenotypes. Nevertheless, no significant causal impact of leisure computer use on T2D (P > 0.05) was detected.

At the genetic level, no significant causal associations were found between driving behavior and T2D and glycemic homeostasis impairment. The results of the TSMR analysis between LSB exposure factors and T2D and Glycemic Homeostasis impairment are presented in Supplementary Table 4 .

Additionally, we conducted a reverse MR analysis to determine the genetic basis for the association between diabetes, glycemic homeostasis impairment, and sedentary leisure behaviors, but did not obtain statistically significant results. The findings are displayed in Supplementary Table 5 .

In the second stage, considering the potential interaction between LSB, we performed MVMR to assess the direct effects of different LSB on T2D and glycemic homeostatic impairment after accounting for gene-level interactions. After adjusting for the effect of leisure computer use on leisure TV watching behavior, the results for T2D and impairment of glycemic homeostasis remained consistent with univariate TSMR, except that the results for 2-hour postprandial glucose became nonsignificant (P=0.77). After controlling for the effect of driving behavior on leisure TV watching behavior, the results aligned with those of TSMR. Moreover, after correcting for the effect of Leisure watching TV behavior on driving behavior, a significant causal association between driving behavior and 2h postprandial glucose emerged (P< 0.05). Detailed results can be found in Supplementary Table 7 .

In the third stage of the study, the predicted effect values from the first step of genetic prediction of LSB and potential mediation analysis revealed that for each 1 standard deviation (SD) increase in the IVs for genetic prediction of leisure television viewing, the corresponding increase in BMI (effect prediction coefficient β=0.28, 95% CI 0.18-0.39, P<0.001), body fat percentage (effect prediction coefficient β=0.28, 95% CI 0.17-0.40, P<0.001), CRP (effect prediction coefficient β= 0.13, 95% CI 0.04-0.21, P< 0.05), waist-to-hip ratio (effect prediction coefficient β= 0.16, 95% CI 0.08-0.25, P< 0.001), and triglycerides (effect prediction coefficient β= 0.26, 95% CI 0.20-0.32, P< 0.001) was observed. Conversely, the years of education (effect prediction coefficient β= -0.47, 95% CI -0.52- -0.42, P< 0.001) levels decreased. The correlation between LSB and waist and hip circumference was not statistically significant.

For each 1 SD increase in the IV for genetic prediction of leisure computer use behavior, there was a corresponding increase in educational attainment (effect prediction coefficient β=0.54, 95% CI 0.39-0.62, P<0.001). However, no statistically significant causal associations were observed for obesity and inflammation-related indicators. All these results met the threshold range of significant P-values after Bonferroni correction. The detailed results are presented in Supplementary Table 8 .

In the second step of the MVMR analysis for leisure TV watching behavior and its potential mediators, we found that the LASSO regression screening removed the CRP factor. The results of the remaining mediator analysis demonstrated that, after correcting for the effect of TV watching behavior, the prevalence of T2D increased by 126% (OR 2.26, 95% CI 1.95-2.61, P< 0.001), 33% (OR 1.33, 95% CI 1.18-1.49, P< 0.001), and 64% (OR 1.64, 95% CI 1.31-2.06, P< 0.001) for each 1-unit SD increase in the levels of genetic IVs representing BMI, triglycerides, and waist-to-hip ratio, respectively. The results for body fat percentage (P = 0.25) were not significant.

In a MVMR analysis examining leisure computer use and its potential mediators, we found that the LASSO regression screening removed the years of school factor. Upon completion of the screening and removal process in the initial two stages of the study, BMI, triglycerides, waist-to-hip ratio, and years of schooling emerged as potential factors influencing the prevalence of T2D associated with leisure TV watching behavior.

In the third step, mediation analysis assessed the effect values of genetic prediction for the potential mediators. The mediation analysis results indicated that the proportion of mediating effects of BMI, triglycerides, and educational attainment on the causal relationship between leisure television watching and the occurrence of T2D was 46.57% (95% CI 33.98%-59.16%, P< 0.001), 14.94% (95% CI 11.10%-18.78%, P< 0.001), and 32.30% (95% CI 24.40%-40.19%, P< 0.001), respectively. However, the waist-to-hip ratio did not yield statistically significant results (P = 0.16). Additionally, no significant mediators were identified after screening for leisure computer usage behavior.

Consequently, we concluded that obesity (BMI), abnormal lipid metabolism (triglycerides), and reduced educational attainment serve as significant mediators of the increased prevalence of T2D due to leisure television watching behavior.

As a result, we concluded that obesity (BMI), abnormal lipid metabolism (triglycerides), and reduced educational attainment act as significant mediators in the increased prevalence of T2D associated with leisure television watching behavior. The mediation analysis model diagrams and mediation effect values for the genetic prediction can be found in Figure 4 and Table 4 .