Longitudinal multicohort evidence for bidirectional diabetes arthritis links in older adults

Summary

Diabetes mellitus and arthritis frequently co-occur in older adults, but whether they causally reinforce each other and through which mechanisms remain unclear. We used harmonized longitudinal data from three nationally representative cohorts across China (CHARLS), the United States (HRS), and Europe (SHARE) to test bidirectional relationships between diabetes and arthritis in late life. Cross-lagged panel models over multiple waves quantified reciprocal effects, prespecified mediation analyses evaluated 21 biologically and clinically motivated mediators, and the NOTEARS algorithm probed data-driven causal network structure. Across all cohorts, diabetes consistently increased subsequent arthritis risk and to a lesser extent, arthritis increased subsequent diabetes risk, with effects robust across demographic and clinical subgroups. C-reactive protein, grip strength, and HbA1c emerged as the dominant mediators, linking metabolic, inflammatory, and functional pathways. These findings support viewing diabetes and arthritis as mutually reinforcing conditions that share modifiable upstream mechanisms, motivating integrated prevention and management strategies in aging populations.

Graphical abstract

Highlights

• •Harmonized CHARLS, HRS, and SHARE data test diabetes-arthritis bidirectionality
• •Cross-lagged models show diabetes consistently increases later arthritis risk
• •Arthritis modestly increases later diabetes risk across cohorts and subgroups
• •CRP, grip strength, and HbA1c mediate links via inflammatory/metabolic pathways

Health sciences; medicine; medical specialty; immunology; internal medicine; endocrinology

Introduction

Diabetes mellitus (DM) and arthritis frequently co-occur in aging populations and jointly contribute to pain, disability, and reduced health-related quality of life, with greater burden than either condition alone.^1,2 Shared biologic pathways, particularly involving interleukin-1 and interleukin-6, plausibly link the two diseases.³ Chronic low-grade systemic inflammation is a hallmark of both DM and osteoarthritis (OA) and is associated with higher circulating inflammatory markers, accelerated cartilage degeneration, and worse OA symptoms.⁴ In DM, long-standing hyperglycemia can promote the accumulation of advanced glycation end products (AGEs) and other downstream pathways that stiffen cartilage and impair joint homeostasis, further bridging DM to OA.⁴ Conversely, arthritis-related pain and mobility limitation can reduce physical activity, promote weight gain, and sarcopenia, and worsen insulin resistance and glycemic control, mechanisms supported by Mendelian-randomization evidence that lower habitual activity and greater sedentary behavior causally increase type 2 DM risk.⁵ Clinical studies also show that combined knee OA and DM are associated with reduced muscle strength, physical inactivity, and poorer quality of life compared with OA alone, underscoring the public-health importance of this comorbidity.^2,6

Despite these convergent pathways, the temporal ordering and directionality of DM-arthritis links remain incompletely resolved. Most observational investigations have focused on a single direction of effect—typically DM as a putative risk factor for OA—or have been cross-sectional within one population, limiting temporal inference, raising concerns about reverse causation, and restricting external validity.^1,7 Recent umbrella and scoping reviews confirm that although associations between type 2 DM and OA are frequently reported, much of the evidence still comes from single-region observational designs, and longitudinal, bidirectional analyses remain relatively scarce.^1,7,8 Genetic studies add further complexity: Mendelian randomization has suggested that OA may increase type 2 DM risk via activity limitation, whereas glycemic traits show inconsistent effects on OA.^9,10,11 Recent evidence found genetic associations between type 1 diabetes and rheumatoid arthritis (RA), but not type 2 diabetes,¹² highlighting unresolved directionality and population-specific effects.

In addition, key mechanistic pathways linking DM to arthritis have seldom been quantified in longitudinal settings. Inflammatory biomarkers, physical function, and metabolic traits are repeatedly implicated as correlates of both conditions and as potential intermediaries in the DM-OA nexus, yet few studies have formally decomposed total effects into direct and indirect components through these factors.^2,6,13,14,15 Recent reviews of OA biomarkers emphasize the need for mediation frameworks and harmonized, multi-region analyses to test whether putative pathways generalize across settings.^8,13,14 Such work is particularly relevant because contemporary data indicate that individuals with both DM and knee OA experience worse physical performance and quality of life than those with OA alone, but most available evidence comes from single-population, often cross-sectional cohorts that offer limited mechanistic insight.^2,6

Methodologically, advancing understanding of the DM-arthritis nexus requires longitudinal designs that explicitly model reciprocity and separate within-person change from stable between-person differences, alongside harmonized phenotyping across cohorts and formal mediation frameworks.^16,17,18 Cross-lagged panel models—especially random-intercept variants—can jointly estimate autoregressive stability and bidirectional cross-paths over time while accounting for time-invariant confounding at the person level, thereby strengthening temporal inference under measured-confounding assumptions.^17,18 Complementary causal mediation analysis decomposes total effects into natural direct and indirect effects, providing a principled way to quantify mechanistic pathways when multiple, potentially interrelated mediators are present.^19,20 Modern structure-learning algorithms, such as differentiable directed acyclic graph (DAG) approaches in the NOTEARS family, offer a data-driven way to interrogate network topology among mediators and outcomes and to explore whether causal architectures differ by region.^21,22 Finally, rigorous cross-cohort harmonization improves comparability and external validity when synthesizing evidence across regions and health-care systems.^16,23,24,25

Against this backdrop, we undertook a multicohort longitudinal investigation using nationally representative aging surveys from three continents—the China Health and Retirement Longitudinal Study (CHARLS), the US Health and Retirement Study (HRS), and the Survey of Health, Ageing and Retirement in Europe (SHARE)—to characterize the bidirectional longitudinal associations between DM and arthritis in older adults.^23,24,25 We harmonized core exposures, outcomes, and covariates across cohorts, applied cross-lagged panel models to estimate reciprocal associations over time, and used formal mediation analysis to quantify indirect effects through 21 candidate mediators spanning systemic inflammation, physical function, and glycemic control, informed by contemporary biomarker and functional literature.^{8,13,14,15,26,27} To complement these hypothesis-driven analyses, we further employed structure-learning algorithms to explore data-driven mediator networks and potential region-specific mechanistic architectures.^21,22 We hypothesized that (1) DM and arthritis exhibit robust bidirectional longitudinal associations, with the DM→arthritis pathway exceeding the reverse; and (2) systemic inflammation, functional decline, and dysglycemia constitute primary mediating mechanisms that may vary across regions.^{1,2,6,13,14,15,26,27}

Results

Study population characteristics and selection

Following systematic screening procedures, the final analytical samples comprised 6,628 participants from CHARLS, 2,569 from HRS, and 3,005 from SHARE, totaling 12,202 individuals across three continents (Figure 1). In CHARLS, sequential exclusions eliminated participants with missing diabetes/arthritis data (n = 340), age <45 years or missing age (n = 628), >30% missing mediator variables (n = 9,015), and lost to follow-up at wave 3 (n = 164) or wave 5 (n = 933). HRS excluded none for missing baseline disease data, 531 for age criteria violations, 14,040 for excessive mediator missingness, and 1,329 lost at wave 14. SHARE excluded 165 for missing disease data, 1,254 for age requirements, 23,592 for mediator missingness, and 2,400 lost at wave 8. Baseline characteristics demonstrated substantial heterogeneity across cohorts (Table 1): median age was lowest in CHARLS (57 years, IQR 51–63) compared to HRS (66 years, IQR 61–70) and SHARE (60 years, IQR 55–66, p < 0.001). Female predominance was observed in HRS (60.3%) and SHARE (56.1%), while CHARLS showed more balanced gender distribution (53.3% female). Educational attainment varied markedly, with CHARLS exhibiting the highest proportion of low education (52.2%) versus HRS (16.5%) and SHARE (44.8%, p < 0.001). Mean BMI differed substantially (HRS: 27.4 kg/m², SHARE: 26.0 kg/m², CHARLS: 23.3 kg/m², p < 0.001), reflecting population-specific obesity patterns. Baseline diabetes prevalence ranged from 12.8% in CHARLS to 20.5% in HRS and 15.3% in SHARE; while arthritis prevalence spanned 35.2% (CHARLS) to 48.3% (HRS) and 42.7% (SHARE). Comorbidity profiles revealed higher hypertension rates in HRS (52.4%) compared to CHARLS (24.7%) and SHARE (29.0%), and elevated cancer prevalence in HRS (11.4%) versus CHARLS (0.8%) and SHARE (3.9%). Missing data patterns for mediator variables are detailed in Tables S1, S2, and S3. Selection bias analysis comparing included versus excluded participants revealed that the primary difference was age (CHARLS: SMD = 0.444; HRS: SMD = 0.530; SHARE: SMD = 0.481), which was expected given the age eligibility criteria (≥45 years for CHARLS; ≥50 years for HRS and SHARE) and the exclusion of younger participants with incomplete follow-up (Table S4). Other baseline characteristics, including gender, education, BMI, smoking, drinking, and prevalence of diabetes, arthritis, and other comorbidities, showed negligible differences between included and excluded samples across all three cohorts (all SMD <0.05), suggesting minimal selection bias beyond age-related attrition.

Figure 1.

Flowchart of study participant selection

Participant selection process across CHARLS, HRS, and SHARE cohorts showing sequential exclusion criteria from baseline to final analytical samples. Boxes display sample sizes at each stage with exclusion reasons and numbers.

Table 1.

Baseline characteristics of study populations

Characteristics	Level	Overall (n = 12,202)	CHARLS (n = 6,628)	HRS (n = 2,569)	SHARE (n = 3,005)	p value	SMD
Age, years	–	60.00 [54.00, 66.00]	57.00 (51.00, 63.00)	66.00 (61.00, 70.00)	60.00 (55.00, 66.00)	<0.001	0.519
Gender	Male	5,435 (44.5)	3,096 (46.7)	1,020 (39.7)	1,319 (43.9)	<0.001	0.073
	Female	6,767 (55.5)	3,532 (53.3)	1,549 (60.3)	1,686 (56.1)	–	–
Education level	Low	5,228 (42.8)	3,457 (52.2)	425 (16.5)	1,346 (44.8)	<0.001	0.440
	Medium	4,042 (33.1)	1,630 (24.6)	1,498 (58.3)	914 (30.4)	–	–
	High	2,932 (24.0)	1,541 (23.2)	646 (25.1)	745 (24.8)	–	–
Marital status	Married	9,649 (79.1)	5,640 (85.1)	1,669 (65.0)	2,340 (77.9)	<0.001	0.300
	Widowed	1,171 (9.6)	521 (7.9)	389 (15.1)	261 (8.7)	–	–
	Single	273 (2.2)	34 (0.5)	88 (3.4)	151 (5.0)	–	–
	Divorced	1,109 (9.1)	433 (6.5)	423 (16.5)	253 (8.4)	–	–
Residence	Urban	5,857 (48.0)	4,311 (65.0)	781 (30.4)	765 (25.5)	<0.001	0.485
	Rural	6,345 (52.0)	2,317 (35.0)	1,788 (69.6)	2,240 (74.5)	–	–
Body mass index, kg/m2	–	24.69 (22.21, 27.57)	23.30 (21.23, 25.73)	27.40 (24.60, 30.90)	25.95 (23.67, 28.60)	<0.001	0.555
Current smoking	No	9,362 (76.7)	4,562 (68.8)	2,291 (89.2)	2,509 (83.5)	<0.001	0.286
	Yes	2,840 (23.3)	2,066 (31.2)	278 (10.8)	496 (16.5)	–	–
Current drinking	No	6,193 (50.8)	4,398 (66.4)	1,128 (43.9)	667 (22.2)	<0.001	0.502
	Yes	6,009 (49.2)	2,230 (33.6)	1,441 (56.1)	2,338 (77.8)	–	–
Diabetes mellitus	No	10,367 (85.0)	5,780 (87.2)	2,042 (79.5)	2,545 (84.7)	<0.001	0.105
	Yes	1,835 (15.0)	848 (12.8)	527 (20.5)	460 (15.3)	–	–
Arthritis	No	7,345 (60.2)	4,295 (64.8)	1,328 (51.7)	1,722 (57.3)	<0.001	0.144
	Yes	4,857 (39.8)	2,333 (35.2)	1,241 (48.3)	1,283 (42.7)	–	–
Hypertension	No	8,352 (68.4)	4,994 (75.3)	1,223 (47.6)	2,135 (71.0)	<0.001	0.304
	Yes	3,850 (31.6)	1,634 (24.7)	1,346 (52.4)	870 (29.0)	–	–
Coronary heart disease	No	10,832 (88.8)	5,923 (89.4)	2,138 (83.2)	2,771 (92.2)	<0.001	0.142
	Yes	1,370 (11.2)	705 (10.6)	431 (16.8)	234 (7.8)	–	–
History of stroke	No	11,924 (97.7)	6,506 (98.2)	2,468 (96.1)	2,950 (98.2)	<0.001	0.068
	Yes	278 (2.3)	122 (1.8)	101 (3.9)	55 (1.8)	–	–
Cancer	No	11,739 (96.2)	6,576 (99.2)	2,276 (88.6)	2,887 (96.1)	<0.001	0.241
	Yes	463 (3.8)	52 (0.8)	293 (11.4)	118 (3.9)	–	–
Chronic lung disease	No	11,416 (93.6)	6,067 (91.5)	2,433 (94.7)	2,916 (97.0)	<0.001	0.129
	Yes	786 (6.4)	561 (8.5)	136 (5.3)	89 (3.0)	–	–

Demographic, socioeconomic, and clinical characteristics at baseline (T1) across three cohorts. Continuous variables: mean ± SD; categorical variables: n (%). p values from ANOVA (continuous) and chi-square tests (categorical).

Cross-lagged panel analysis of bidirectional diabetes-arthritis associations

Building upon the longitudinal framework established across three cohorts, cross-lagged panel models revealed consistent bidirectional relationships between diabetes and arthritis in all databases (Figures 2A–2C). In HRS, diabetes at T1 significantly predicted incident arthritis at T2 (β = 0.105, SE = 0.021, p < 0.001; standardized β = 0.086), while arthritis at T1 predicted incident diabetes at T2 (β = 0.044, SE = 0.011, p < 0.001; standardized β = 0.054). These reciprocal associations persisted from T2 to T3, with diabetes predicting arthritis (β = 0.142, SE = 0.022, p < 0.001; standardized β = 0.121) and arthritis predicting diabetes (β = 0.055, SE = 0.013, p < 0.001; standardized β = 0.063). The CHARLS cohort demonstrated similar bidirectional patterns during T1→T2 (diabetes→arthritis: β = 0.161, SE = 0.017, p < 0.001, standardized β = 0.116; arthritis→diabetes: β = 0.048, SE = 0.007, p < 0.001, standardized β = 0.061) and T2→T3 periods (diabetes→arthritis: β = 0.158, SE = 0.016, p < 0.001, standardized β = 0.127; arthritis→diabetes: β = 0.057, SE = 0.009, p < 0.001, standardized β = 0.066). The SHARE cohort confirmed these findings with comparable effect sizes across both time intervals (T1→T2 diabetes→arthritis: β = 0.118, SE = 0.022, p < 0.001, standardized β = 0.088; arthritis→diabetes: β = 0.040, SE = 0.011, p < 0.001, standardized β = 0.050; T2→T3 diabetes→arthritis: β = 0.214, SE = 0.021, p < 0.001, standardized β = 0.178; arthritis→diabetes: β = 0.061, SE = 0.012, p < 0.001, standardized β = 0.069). Notably, the effect of diabetes on arthritis (standardized β range: 0.086–0.178) consistently exceeded the reverse effect of arthritis on diabetes (standardized β range: 0.050–0.069) across all cohorts and time periods. Strong autoregressive coefficients indicated high temporal stability for both conditions (diabetes: standardized β = 0.680–0.719; arthritis: standardized β = 0.569–0.616). Cross-cohort heterogeneity analysis revealed low to moderate heterogeneity for most cross-lagged paths (I² = 0–56.9%), with pooled effects confirming the diabetes→arthritis pathway (pooled β = 0.100–0.141) consistently exceeded the reverse pathway (pooled β = 0.057–0.066) across cohorts (Table S5). Random-intercept CLPM (RI-CLPM) analyses, which decompose between-person and within-person effects, yielded attenuated and non-significant within-person cross-lagged coefficients across all cohorts, suggesting that the observed associations primarily reflect stable between-person differences rather than within-person temporal dynamics (Table S6).

Figure 2.

Cross-lagged panel models of bidirectional diabetes-arthritis associations

Standardized path coefficients from structural equation models examining reciprocal relationships between diabetes and arthritis across T1→T2 and T2→T3 in (A) CHARLS, (B) HRS, and (C) SHARE. Solid arrows: p < 0.05; dashed arrows: non-significant. Models adjusted for age, gender, education, BMI, and comorbidities. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. p values for path coefficients were obtained from two-sided Wald tests in lavaan; sample sizes were CHARLS n = 6,628, HRS n = 2,569, SHARE n = 3,005.

Mediation pathways linking diabetes to arthritis

To elucidate the underlying mechanisms through which diabetes influences arthritis development, formal mediation analyses examined 21 potential intermediary factors across all three cohorts (Table S7; Figures 3A–3C). In CHARLS, the strongest mediator was maximum grip strength (indirect effect = 0.0817, 95% CI [0.068, 0.097], p < 0.001), accounting for 26.5% of the total diabetes-arthritis association, followed by C-reactive protein (CRP; indirect effect = 0.0499, 95% CI [0.039, 0.062], p < 0.001, proportion mediated = 16.2%) and activities of daily living limitations (ADL; indirect effect = 0.0289, 95% CI [0.021, 0.038], p < 0.001, proportion mediated = 9.3%). Path coefficients revealed that diabetes significantly reduced grip strength (path a: β = −6.25, p < 0.001), which in turn predicted higher arthritis risk (path b: β = −0.074, p < 0.001). Similarly, diabetes elevated CRP levels (path a: β = 1.39, p < 0.001), with elevated CRP independently associated with arthritis incidence (path b: β = 0.196, p < 0.001). Additional significant mediators included mobility limitations (6.2%), instrumental ADL impairments (4.6%), cognitive decline (4.2%), depression (4.1%), and HbA1c (3.8%), collectively explaining substantial indirect pathways. All significant mediation effects remained robust after false discovery rate (FDR) correction for multiple comparisons, and variance inflation factors for all mediators ranged from 1.00 to 1.03, indicating no substantive collinearity concerns (Table S7).

Figure 3.

Top three mediators of diabetes-arthritis association

Mediation pathways showing indirect effects through the three strongest mediators in (A) CHARLS, (B) HRS, and (C) SHARE. Values represent standardized coefficients with 95% CI. Adjusted for age, gender, education, BMI, and comorbidities. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. Indirect effects and 95% CIs were estimated with 1,000 bootstrap replications; N as above.

Parallel mediation patterns emerged in HRS and SHARE cohorts with notable consistency. In HRS, CRP demonstrated the largest mediating effect (indirect effect = 0.0788, 95% CI [0.057, 0.103], p < 0.001, proportion mediated = 26.4%), followed by grip strength (indirect effect = 0.0375, 95% CI [0.022, 0.056], p < 0.001, proportion mediated = 12.5%) and mobility limitations (indirect effect = 0.0276, 95% CI [0.015, 0.043], p < 0.001, proportion mediated = 9.2%). Depression (7.1%), functional dependency (6.7%), cholesterol (6.5%), and ADL limitations (5.2%) also mediated significant portions of the diabetes-arthritis relationship. In SHARE, the top three mediators were CRP (indirect effect = 0.0861, 95% CI [0.063, 0.109], p < 0.001, proportion mediated = 25.7%), HbA1c (indirect effect = 0.0545, 95% CI [0.037, 0.073], p < 0.001, proportion mediated = 16.2%), and grip strength (indirect effect = 0.0276, 95% CI [0.015, 0.043], p < 0.001, proportion mediated = 8.2%). Depression, mobility impairment, and IADL limitations contributed 6.5%, 5.7%, and 4.8% of the total effect, respectively. Across all databases, systemic inflammation (CRP), physical function decline (grip strength), and glycemic control (HbA1c) consistently emerged as primary mechanistic pathways, while functional limitations, depression, and lipid dysregulation provided additional explanatory routes. The direct effect of diabetes on arthritis remained statistically significant after accounting for all mediators (CHARLS: β = 0.226, p < 0.001; HRS: β = 0.220, p < 0.001; SHARE: β = 0.249, p < 0.001), indicating that unmeasured mechanisms also contribute to this association. Cross-cohort heterogeneity analysis revealed that most mediators showed low to moderate heterogeneity (I²<50%), including mobility limitations (I² = 0%), depression (I² = 23.2%), and IADL impairments (I² = 16.7%), supporting consistent mediation mechanisms across populations. However, substantial heterogeneity was observed for grip strength (I² = 93.4%), CRP (I² = 80.9%), and HbA1c (I² = 89.9%), suggesting that the relative contribution of these mediators varied across cohorts despite consistent directional effects (Table S8).

Data-driven directed network structure discovery

Beyond confirmatory mediation analysis, the NOTEARS algorithm uncovered data-driven DAG structures linking diabetes, arthritis, and 21 potential mediators at T2, revealing both shared and region-specific mechanistic architectures (Figures 4A–4C). The discovered networks exhibited substantial structural heterogeneity across cohorts: HRS yielded the most complex network with 255 edges generating 2,046 distinct directed paths (total effect = 0.054, indirect effect = 0.054), SHARE demonstrated intermediate complexity with 250 edges and 1,954 paths (total effect = 0.077, indirect effect = 0.044), while CHARLS revealed a sparser structure with 171 edges and 759 paths (total effect = 0.087, indirect effect = 0.043). Notably, CHARLS retained a direct diabetes→arthritis edge with high stability (effect = 0.045, stability = 0.99), whereas this direct path appeared with lower stability in SHARE (effect = 0.033, stability = 0.73) and was absent in HRS, suggesting that the measured mediators more completely accounted for the diabetes-arthritis association in Western populations compared to Chinese populations.

Figure 4.

NOTEARS-discovered causal networks

Directed acyclic graphs showing causal relationships among diabetes, arthritis, and 21 mediators at T2 in (A) CHARLS, (B) HRS, and (C) SHARE. Node size: PageRank centrality; edge width: effect size. Only edges with bootstrap stability >50% shown.

Examination of the top-ranked directed pathways revealed consistent primary routes alongside notable regional variations. In HRS, the strongest indirect pathway was diabetes→CRP→arthritis (path effect = 0.012), followed by diabetes→mobility limitations→arthritis (path effect = 0.008), diabetes→grip strength→arthritis (path effect = 0.007), and diabetes→depression→arthritis (path effect = 0.006), with all intermediate edges achieving perfect bootstrap stability (100%). CHARLS prioritized the direct diabetes→arthritis connection as the dominant pathway (effect = 0.045), with secondary indirect routes through grip strength (path effect = 0.010), CRP (path effect = 0.007), and ADL limitations (path effect = 0.005). SHARE exhibited a pattern intermediate between the two, retaining a moderately stable direct link while featuring prominent indirect pathways through CRP (path effect = 0.010), HbA1c (path effect = 0.008), and depression (path effect = 0.005). Edge weight patterns corroborated earlier mediation findings: diabetes consistently elevated CRP (HRS: 0.117, CHARLS: 0.094, SHARE: 0.122) and reduced grip strength (HRS: −0.081, CHARLS: −0.102), with these mediators subsequently increasing arthritis risk through inflammation (CRP→arthritis: HRS 0.106, CHARLS 0.077, SHARE 0.078) and functional decline (grip→arthritis: HRS –0.088, CHARLS –0.098) mechanisms. The divergent network topologies across populations underscore the importance of considering region-specific mechanistic pathways when designing intervention strategies targeting the diabetes-arthritis nexus.

Subgroup-specific effect heterogeneity

To evaluate potential effect modification, stratified analyses examined bidirectional associations across demographic and lifestyle subgroups during both T1→T2 and T2→T3 periods (Figures 5A–5F). In CHARLS, the diabetes→arthritis effect remained consistently robust across most subgroups, with odds ratios ranging from 6.17 (95% CI: 3.19–11.93, p < 0.001) in obese individuals to 9.05 (95% CI: 5.25–15.60, p < 0.001) in those aged ≥65 years during T1→T2, and comparable magnitudes during T2→T3 (OR range: 4.52–7.54). The reverse arthritis→diabetes pathway demonstrated attenuated effects (T1→T2 OR range: 1.22–2.70; T2→T3 OR range: 1.84–2.91), with loss of statistical significance among obese participants (T1→T2: OR = 1.22, p = 0.546; T2→T3: OR = 1.93, p = 0.044). In HRS, younger individuals (<65 years) exhibited stronger diabetes→arthritis associations during T1→T2 (OR = 7.10, 95% CI: 4.01–12.57) compared to older adults (OR = 4.04, 95% CI: 2.53–6.46), although this pattern reversed during T2→T3. Gender differences were consistently observed, with males demonstrating stronger effects in both directions across time periods (T1→T2 diabetes→arthritis: male OR = 5.67 vs. female OR = 4.25; arthritis→diabetes: male OR = 2.95 vs. female OR = 1.38, p = 0.258). SHARE exhibited similar heterogeneity patterns, with particularly pronounced diabetes→arthritis effects among older adults during T2→T3 (OR = 10.08, 95% CI: 5.79–17.55) and overweight individuals (OR = 12.16, 95% CI: 7.72–19.16). Smoking status revealed divergent patterns across cohorts: current smokers in SHARE demonstrated amplified arthritis→diabetes effects (T1→T2: OR = 3.23; T2→T3: OR = 3.31) compared to non-smokers, whereas this differential was less pronounced in CHARLS and HRS. Overall, bidirectional associations persisted across nearly all examined subgroups, with effect magnitudes consistently larger for the diabetes→arthritis pathway than the reverse direction.

Figure 5.

Subgroup analyses by time period

Forest plots showing odds ratios (95% CI) for diabetes→arthritis and arthritis→diabetes pathways across subgroups in CHARLS (A: T1→T2, B: T2→T3), HRS (C: T1→T2, D: T2→T3), and SHARE (E: T1→T2, F: T2→T3). Subgroups: age, gender, education, BMI, residence, smoking, and drinking. Models adjusted for covariates.

Sensitivity analyses confirming robustness

To assess the robustness of observed bidirectional associations, multiple sensitivity analyses were conducted across varying analytical specifications (Tables S9, S10, S11, and S12). Covariate adjustment strategies ranging from unadjusted models to fully adjusted models yielded highly consistent results: in HRS, the diabetes→arthritis effect remained stable across model 0 (standardized β = 0.085, p < 0.001), model 1 (β = 0.086, p < 0.001), model 2 (β = 0.085, p < 0.001), and model 3 (β = 0.086, p < 0.001) during T1→T2, with parallel consistency observed in CHARLS and SHARE. Exclusion of participants with both conditions at baseline strengthened effect estimates, particularly for diabetes→arthritis pathways (HRS T1→T2: β = 0.136, p < 0.001; CHARLS: β = 0.162, p < 0.001; SHARE: β = 0.113, p < 0.001), suggesting that prevalent comorbidity may attenuate observed associations. Long time-lag analyses examining direct T1→T3 effects while controlling for T2 status demonstrated persistent long-term relationships (HRS: diabetes→arthritis β = 0.103, p < 0.001; CHARLS: β = 0.085, p < 0.001; SHARE: β = 0.140, p < 0.001), though model fit indices indicated reduced adequacy compared to sequential models (RMSEA range: 0.285–0.291). Complete case analyses restricted to participants without missing data produced comparable estimates to multiple imputation approaches across all cohorts and time periods, validating the imputation strategy. All models demonstrated excellent fit with CFI>0.99, TLI>0.99, RMSEA<0.02, and SRMR<0.015 across databases and analytical approaches, confirming that bidirectional diabetes-arthritis associations are robust to analytical choices.

Discussion

Principal findings

In three nationally representative aging cohorts from Asia, Europe, and North America, we found robust bidirectional longitudinal associations between DM and arthritis. However, the pathway from DM to subsequent arthritis was consistently stronger than the reverse pathway from arthritis to later DM, even after extensive adjustment and sensitivity analyses. These findings support the growing view that the co-occurrence of DM and OA is not simply a cross-sectional clustering of common age-related disorders, but reflects partially shared pathophysiological processes that unfold over time in older adults.^1,6,11 By integrating harmonized data across three continents, the present study extends prior single-cohort investigations and recent Mendelian-randomization work on the DM-OA nexus, suggesting that the temporal pattern of a stronger DM→arthritis direction may generalize across diverse populations, although genetic evidence indicates potential differences between type 1 and type 2 diabetes in their associations with inflammatory arthritis.^9,10,11,12

Mechanistically, our observation that systemic inflammation, physical function, and glycemic control jointly mediate a substantial proportion of the DM→arthritis association is consistent with contemporary models of OA as a systemic-metabolic as well as local joint disease.^4,8,28,29 In contrast, mediating pathways from arthritis to DM were weaker and less consistent, although disability and functional limitation still contributed meaningfully. Taken together, these results support a conceptualization in which DM acts as a “driver” that accelerates joint pathology and symptom burden, while arthritis feeds back more modestly through activity restriction, weight gain and metabolic deterioration.^5,6,18

Interpretation and plausible mechanisms

Several biological and behavioral mechanisms may explain the observed asymmetry. Chronic hyperglycemia activates pro-inflammatory pathways, oxidative stress and AGE formation in cartilage, bone and synovium.^4,28,30,31 Experimental and translational studies show that hyperglycemia-induced accumulation of AGEs stiffens the cartilage matrix, alters chondrocyte metabolism and increases susceptibility to mechanical damage, thereby predisposing to OA onset and progression.^30,31 At the systemic level, DM is characterized by low-grade chronic inflammation and dysregulated lipid metabolism, with elevated cytokines, C-reactive protein (CRP) and atherogenic lipids that have all been linked to OA risk and symptom severity.^{4,28,29,32,33} Our mediation findings, in which CRP and glycemic markers were among the strongest indirect pathways, are biologically plausible. CRP not only reflects systemic inflammation but independently predicts adverse cardiometabolic outcomes and partially mediates the protective association between lifestyle factors and type 2 DM risk.^34,35 Meta-analytic evidence further demonstrates that resistance exercise effectively reduces CRP levels in patients with type 2 DM,³⁶ suggesting potential intervention targets along the inflammation-mediated pathway.^8,29,32,33

Physical function and disability emerged as additional mediators, with grip strength showing particularly strong mediating effects across cohorts. The prominence of grip strength as a mediator likely reflects multiple converging pathways: DM-related sarcopenia characterized by loss of muscle mass and quality, peripheral neuropathy impairing neuromuscular coordination, and microvascular complications reducing muscle perfusion.³⁷ Recent evidence confirms that prediabetes is already an independent risk factor for sarcopenia in older men, suggesting that muscle deterioration begins early in the glycemic continuum.³⁷ Furthermore, higher relative grip strength and regular muscle-strengthening activity are independently associated with lower type 2 DM risk, indicating a bidirectional muscle-metabolic relationship.³⁸ Grip strength may thus serve as both a proxy for overall musculoskeletal health and a direct consequence of DM-induced neuromuscular dysfunction, explaining its robust mediating role across populations.^2,6,39 Furthermore, fatigue, depressive symptoms and fear of pain can reduce physical activity, promoting weight gain and further metabolic deterioration.^5,18,39 Our results suggest that these functional and behavioral sequelae of DM contribute to arthritis risk beyond purely biochemical pathways, reinforcing the importance of integrated management of mobility, muscle strength and mental health in people with DM.^2,6,18

The reverse pathway—from arthritis to incident DM—was weaker but still present. Chronic pain and stiffness can limit physical activity, reduce cardiorespiratory fitness and increase sedentary time, all of which are established risk factors for type 2 DM.^5,18 Observational and genetic studies have suggested that arthritis-related disability and activity restriction may promote DM through weight gain, reduced muscle mass and worsening insulin resistance.^9,27 In our analyses, however, these pathways accounted for a relatively small share of the total arthritis→DM effect, and point estimates were more heterogeneous across cohorts. One possibility is that contemporary DM screening and treatment in older adults mitigate the incremental metabolic impact of arthritis-related inactivity. Another is that residual confounding by unmeasured lifestyle or clinical factors (e.g., diet, analgesic use, etc.) differentially influences this direction across regions.

Mediation and directed network structure learning

By applying formal mediation analysis, we were able to quantify the proportion of the DM→arthritis association explained by predefined sets of intermediates. Systemic inflammation, physical function and glycemic control together accounted for a meaningful but incomplete share of the total effect, implying that additional pathways—such as local joint microvascular disease, altered pain processing or genetic susceptibility—remain to be elucidated.^{28,29,30,31,32,33} The persistence of sizable natural direct effects after adjusting for these mediators suggests that treating only inflammation or glycemia may not fully prevent DM-associated arthritis risk, echoing recent work on multimodal OA pathogenesis.^8,29

Our use of differentiable structure-learning algorithms provided a complementary, data-driven view of mediator networks.^21,22 NOTEARS-based graphs highlighted partially overlapping but distinct mediator topologies across cohorts. Notably, CHARLS retained a direct diabetes→arthritis edge with high stability, whereas this direct path was absent in HRS, suggesting that the measured mediators more completely accounted for the association in Western populations. Several factors may explain this regional difference: first, the CHARLS cohort was younger (median age 57 vs. 66 years) and may represent earlier disease stages where direct metabolic effects of hyperglycemia on joint tissues predominate over inflammation-mediated pathways; second, population-specific differences in obesity patterns (mean BMI 23.3 vs. 27.4 kg/m²) may alter the relative contribution of mechanical versus metabolic pathways; third, differences in healthcare access and DM treatment intensity could modify how hyperglycemia translates into joint pathology; and fourth, genetic or dietary factors specific to east Asian populations may confer distinct susceptibility to AGE-mediated cartilage damage.^11,29,32,40 While the learned graphs remain observational and hypothesis-generating, they help prioritize mechanisms and mediator combinations for future intervention studies.

Heterogeneity and generalizability

Subgroup analyses showed that the bidirectional DM-arthritis association was broadly robust across sex, age, body mass index and educational strata, with only modest variation in effect sizes.^{11,26,40,41,42,43} This suggests that the core phenomenon of mutual reinforcement between DM and arthritis is fairly general in older adults. Nonetheless, we observed some heterogeneity across regions and socioeconomic strata. Participants with lower education or unhealthy lifestyle profiles tended to exhibit stronger associations and larger mediated effects through inflammation and function, in line with reports that social disadvantage amplifies the burden of multimorbidity, including arthritis and DM.^40,42,43,44 Cross-cohort differences in baseline risk, obesity prevalence and access to care may also partly explain why certain mediators (e.g., dyslipidemia or CRP) appeared more salient in some regions than others.^28,32,40,41

Importantly, our harmonized analyses of CHARLS, HRS and SHARE suggest that the broad pattern of a stronger DM→arthritis pathway, partially mediated by inflammation, function and glycemia, is reproducible across diverse health-care systems and cultural contexts.^{23,24,25,26,40} This bolsters the external validity of our findings and supports the notion that integrated management of DM and arthritis is likely to yield benefits in many settings, not only where OA or DM prevalence is highest.

Clinical and public health implications

Our results carry several implications for clinical practice and prevention. First, in older adults with DM, clinicians should maintain a high index of suspicion for early joint symptoms and functional decline, recognizing that DM may actively drive arthritis development rather than being a mere comorbidity.^1,2,4,6,8 Routine assessment of pain, stiffness, mobility and muscle strength in diabetes care could facilitate earlier identification of patients at high risk of arthritis and enable timely non-pharmacologic interventions such as exercise therapy, weight management and physical rehabilitation.^2,6,29,40 Second, the prominent mediating role of systemic inflammation and dysglycemia underscores the potential value of treatment strategies that simultaneously target metabolic control and inflammatory pathways—for example, optimizing glucose-lowering therapies with favorable inflammatory profiles, addressing dyslipidemia and hypertension, and promoting anti-inflammatory dietary patterns.^4,8,29,32,33

Third, the modest but non-negligible pathway from arthritis to DM suggests that arthritis services should incorporate basic metabolic screening and lifestyle counseling, particularly for older adults with marked disability or obesity.^5,18,40 Multidisciplinary models that integrate rheumatology, endocrinology, rehabilitation and primary care may be especially suited for managing this bidirectional comorbidity. At the population level, our findings support policies that promote physical activity, reduce obesity and improve access to early arthritis and DM management, especially in socioeconomically disadvantaged groups where the combined burden is highest.^40,42,43,44

Future directions

Future research should extend this work in several directions. Integrating richer biomarker panels, imaging-based phenotypes and detailed treatment data into longitudinal mediation frameworks would help refine the relative importance of systemic versus local joint mechanisms along the DM→arthritis pathway.^{8,29,31,32,33,45} Combining cohort data with genetic instruments for key mediators and outcomes could strengthen inference and clarify the roles of glycemic control, obesity and inflammation using Mendelian-randomization and related approaches.^9,10,27 Pragmatic trials and implementation studies are needed to test whether multifaceted interventions that jointly target DM control, inflammation, and physical function can prevent or delay arthritis in people with DM, and to evaluate how best to integrate such strategies into real-world health systems.^29,32,33,40 Finally, future network discovery work could explore how mediator structures evolve over time and differ across subgroups, thereby informing precision-prevention strategies for DM-arthritis multimorbidity in aging populations.

Key strengths of this study include the use of multicontinental, nationally representative cohorts with harmonized phenotypes; replication of findings across three independent datasets; and the integration of cross-lagged panel modeling, formal mediation and modern causal structure learning.^{16,18,21,22,23,24,25} Notably, our RI-CLPM analyses yielded attenuated and non-significant within-person cross-lagged coefficients, suggesting that the observed associations primarily reflect stable between-person differences rather than within-person temporal dynamics. This finding implies that individuals who develop DM tend to be those already predisposed to arthritis through shared risk factors, and highlights the importance of early, integrated prevention targeting common upstream determinants.^17,18 Extensive subgroup and sensitivity analyses further enhance confidence in the robustness of our conclusions.^11,26,40

In three nationally representative cohorts across Asia, Europe, and North America, DM and arthritis exhibit reciprocal longitudinal links, with a consistently stronger DM→arthritis pathway largely transmitted through inflammation, functional decline, and dysglycemia. These convergent results argue for an integrated prevention and management paradigm that targets shared pathophysiology and preserves function in aging populations—while allowing for regional adaptation informed by cohort-specific network structure.

Limitations of the study

Several limitations merit consideration. First, both DM and arthritis were ascertained primarily via self-report, which may introduce misclassification; however, prior validation studies in similar aging cohorts suggest reasonable specificity for self-reported diagnoses, and any non-differential misclassification would likely bias associations toward the null.^23,24,25 Second, mediators such as inflammatory markers, functional measures and metabolic traits were assessed at discrete survey waves, and we could not capture short-term fluctuations or cumulative exposure, potentially leading to underestimation of indirect effects.^8,13,14,45 Third, although we adjusted for a wide range of covariates and used advanced longitudinal models, residual and unmeasured confounding cannot be excluded, particularly for lifestyle factors and medication use; as with all observational studies, causal interpretations should therefore remain cautious.^11,22,40

Fourth, our measures of arthritis relied on self-reported doctor-diagnosed arthritis and did not distinguish between OA and inflammatory arthritides such as RA in all cohorts. This is an important limitation because OA and RA have distinct pathophysiological relationships with DM: OA is primarily linked to metabolic and mechanical factors, whereas RA involves autoimmune inflammation with different associations with glucose metabolism, as evidenced by studies showing that DM-related autoantibodies are more frequent in RA patients with comorbid DM,⁴⁶ and that specific immune cell alterations (Th2 and Treg reductions) characterize RA patients who develop type 2 DM.⁴⁷ Furthermore, we lacked detailed information on joint site, radiographic severity and treatment, which may differentially relate to DM and its complications.^1,8,40 Fifth, mediator sets were necessarily limited by the variables harmonized across cohorts; emerging biomarkers (e.g., novel inflammatory, lipidomic or imaging markers) and psychosocial factors could not be included.^8,29,33 Finally, our causal discovery analyses, while informative, remain exploratory and sensitive to modeling assumptions; their primary role is to generate hypotheses that require confirmation in interventional or quasi-experimental studies.^21,22

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Chuan He (20240910@kmmu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The datasets analyzed in this study are publicly available from their respective sources: CHARLS data can be accessed at http://charls.pku.edu.cn, HRS data at https://hrs.isr.umich.edu, and SHARE data at https://share-eric.eu. These datasets are fully publicly accessible, and no reviewer-only access, temporary links, or embargoed access restrictions apply.
• •The analysis code and scripts used in this study are publicly available in the repository listed in the key resources table (deposited data), at https://github.com/Paperaceepted/Iscience (accessed 2026-03-08).
• •Any additional information required to reanalyze the data reported in this study is available from the lead contact upon request.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (81560372); the “Xingdian Talents” Support Project of Yunnan Province; the “535” Talent Project of the First Affiliated Hospital of Kunming Medical University (2023535D11); and the Project of Yunnan Orthopedics and Sports Rehabilitation Clinical Medical Research Center (202102AA310068). We thank all participants and investigators of the CHARLS, HRS, and SHARE studies for their invaluable contributions to data collection and management.

Author contributions

Conceptualization, M.Z. and C.H.; data curation and validation, Y.C. and T.Z.; methodology and formal analysis, M.Z. and K.L.; visualization, M.Z. and F.G.; writing – original draft, M.Z., Y.C., and C.H.; writing – review and editing, all authors; supervision and project administration, C.H. C.H. is the corresponding author and had full access to all the data and the final responsibility for the decision to submit.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Publicly available longitudinal cohort datasets: China Health and Retirement Longitudinal Study (CHARLS); Health and Retirement Study (HRS); Survey of Health, Aging and Retirement in Europe (SHARE).	National School of Development, Peking University (CHARLS); University of Michigan, Institute for Social Research (HRS); SHARE-ERIC (SHARE).	CHARLS: Waves 1–5 (2011–2020), https://charls.pku.edu.cn/ (accessed 2026-02-26); HRS: Waves 8–14 (2006–2018), https://hrs.isr.umich.edu/ (accessed 2026-02-26); SHARE: Waves 1–8 (2004–2020), https://share-eric.eu/ (accessed 2026-02-26).
Analysis code (R) for this study	GitHub	https://github.com/Paperaceepted/Iscience
Software and algorithms
Statistical software and algorithms: R (version 4.3.0); mice; lavaan; mediation; (additional R packages used for data management/visualization and causal graph analysis, as applicable).	R Foundation for Statistical Computing; Comprehensive R Archive Network (CRAN).	R (v4.3.0): https://www.r-project.org/ (accessed 2026-02-26); mice (Multivariate Imputation by Chained Equations): https://cran.r-project.org/package=mice (accessed 2026-02-26); lavaan (structural equation modeling/CLPM): https://cran.r-project.org/package=lavaan (accessed 2026-02-26); mediation (causal mediation analysis): https://cran.r-project.org/package=mediation (accessed 2026-02-26); igraph (network analysis/centrality metrics): https://cran.r-project.org/package=igraph (accessed 2026-02-26).

Experimental model and study participant details

Ethics

This study was conducted in accordance with the Declaration of Helsinki and adhered to ethical guidelines for secondary analysis of de-identified public datasets. The original data collections were approved by the relevant ethics committees and obtained informed consent from all participants. Specifically, CHARLS was approved by the Biomedical Ethics Review Committee of Peking University (IRB00001052-11015); HRS was approved by the University of Michigan Institutional Review Board (HUM00061128); and SHARE was approved by the Ethics Committee of the University of Munich (142/16 S). As this study involved retrospective analysis of publicly available de-identified data, no additional ethical approval was required.

Study participants and inclusion criteria

This study is a secondary analysis of de-identified data from three longitudinal aging cohorts across three continents: CHARLS (waves 1–5; 2011–2020), HRS (waves 8–14; 2006–2018), and SHARE (waves 1–8; 2004–2020). Participants were eligible if they had complete information on diabetes mellitus and arthritis at three harmonized time points: T1 (CHARLS wave 1; HRS wave 8; SHARE wave 1), T2 (CHARLS wave 3; HRS wave 11; SHARE wave 6), and T3 (CHARLS wave 5; HRS wave 14; SHARE wave 8). Participants were required to be ≥ 45 years at baseline in CHARLS and ≥50 years in HRS and SHARE. Individuals with >30% missing mediator data at T2 were excluded, and complete diabetes/arthritis information at T2 and T3 was required.

Method details

Study design and harmonization

We conducted a multicohort longitudinal analysis to investigate bidirectional temporal relationships between diabetes mellitus and arthritis in older adults using harmonized variables across CHARLS, HRS, and SHARE. Core disease variables and covariates were standardized to enable cross-cohort comparisons using consistent coding schemes.

Definitions of diabetes mellitus and arthritis

Diabetes mellitus and arthritis were defined using self-reported physician diagnoses and coded as binary indicators (0 = absent, 1 = present) within each cohort and then harmonized across cohorts.

Covariates

Demographic and socioeconomic covariates were harmonized as follows: gender (0 = male, 1 = female), education (low/medium/high), marital status (married/widowed/single/divorced), residence (0 = rural, 1 = urban), smoking (0 = never/former, 1 = current), and drinking (0 = non-drinker, 1 = current drinker). Chronic comorbidities (hypertension, heart disease, stroke, cancer, and chronic lung disease) were coded as binary variables. Depression was assessed using cohort-specific validated instruments (CES-D-10 in CHARLS, CES-D-8 in HRS, EURO-D in SHARE) and harmonized into both a standardized depression score and a dichotomous indicator using prespecified cutoffs.

Missing data management

Missing covariate data were handled using multiple imputation by chained equations (MICE) implemented in R. Five imputed datasets were generated using logistic regression for binary variables, predictive mean matching for continuous variables, and proportional odds models for ordinal variables, and pooled using Rubin’s rules.

Cross-lagged panel modeling (CLPM)

Bidirectional longitudinal associations between diabetes and arthritis were estimated using cross-lagged panel models (CLPM) within a structural equation modeling framework over two intervals (T1→T2 and T2→T3). Models simultaneously estimated (i) autoregressive paths for each condition, (ii) cross-lagged paths capturing reciprocal effects, and (iii) contemporaneous correlations, adjusting for prespecified covariates (age, gender, education, BMI, and comorbidities including hypertension, heart disease, stroke, cancer, and chronic lung disease). Model fit was evaluated using CFI, TLI, RMSEA, and SRMR with standard thresholds (CFI>0.95, TLI>0.95, RMSEA<0.06, SRMR<0.08). Random-intercept CLPM (RI-CLPM) was additionally used to decompose between-person stable differences and within-person temporal dynamics.

Mediation analysis

Prespecified mediation analyses quantified indirect effects of diabetes on arthritis through 21 biologically and clinically motivated mediators. For each mediator, path a (diabetes→mediator), path b (mediator→arthritis), the direct effect (c′), total effect (c), and the average causal mediation effect were estimated, adjusting for the same covariate set as above. Confidence intervals were obtained using 1,000 bootstrap replications.

Causal structure learning (NOTEARS)

To complement hypothesis-driven mediation, causal network structure among diabetes, arthritis, and 21 mediators at T2 was explored using the NOTEARS algorithm, which formulates directed acyclic graph learning as continuous optimization with a differentiable acyclicity constraint. Hyperparameters were set as follows: L1 regularization λ = 0.015, augmented Lagrangian growth rate ρ = 0.1, maximum iterations = 1000, acyclicity tolerance = 1e-8, and edge threshold = 0.015. Bootstrap stability analysis (100 resamples) retained edges appearing in >50% of samples. Network centrality metrics (degree, betweenness, PageRank) were computed to identify influential nodes.

Quantification and statistical analysis

All analyses were conducted in R v4.3.0 with two-sided p < 0.05 considered statistically significant. Missing-data imputation used MICE (m = 5) with pooling via Rubin’s rules. CLPMs were fitted using lavaan with robust maximum likelihood estimation, and model fit was assessed using CFI, TLI, RMSEA, and SRMR. Formal mediation analyses were performed using the mediation package with 1,000 bootstrap replications to obtain 95% confidence intervals for indirect effects. NOTEARS structure learning was implemented with the prespecified optimization and sparsity parameters described above, and network measures were computed using igraph. Subgroup analyses evaluated effect modification by age group (<65 vs. ≥ 65), gender, education, BMI categories (Asian cutoffs for CHARLS; WHO cutoffs for others), residence, smoking, and drinking. Sensitivity analyses included multiple covariate adjustment strategies, exclusion of baseline comorbidity, long time-lag models (T1→T3 controlling for T2), and complete-case analyses.

Additional resources

CHARLS data access: http://charls.pku.edu.cn.

HRS data access: https://hrs.isr.umich.edu.

SHARE data access: https://share-eric.eu.

Clinical trial registration: Not applicable.

Published: March 24, 2026

Supplemental information

Table S4. Selection bias analysis comparison of baseline characteristics between included and excluded participants across three cohorts

Standardized mean differences reported to assess potential selection bias.

Table S5. Cross-cohort heterogeneity analysis for CLPM cross-lagged effects cochran’s Q statistic, I2 index, and pooled effect estimates for cross-lagged paths (diabetes→arthritis and arthritis→diabetes) across three cohorts

Random-effects meta-analysis used for pooling. I² values < 25%, 25%–50%, 50%–75%, and >75% indicate low, moderate, substantial, and high heterogeneity, respectively.

Table S6. Comparison of standard CLPM and random-intercept CLPM cross-lagged path coefficients and model fit indices comparing standard CLPM with RI-CLPM across three cohorts

RI-CLPM separates between-person and within-person effects.

Table S7. Mediation analysis results with FDR correction and collinearity diagnostics path coefficients, indirect effects, proportion mediated, FDR-adjusted p values, and variance inflation factors for all tested mediators

Results include path a (diabetes→mediator), path b (mediator→arthritis), direct effect c’, and indirect effect with 95% CI from 1,000 bootstrap replications.

Table S8. Cross-cohort heterogeneity analysis for mediation effects Cochran’s Q statistic, I2 index, and pooled indirect effects for each mediator across three cohorts

Random-effects meta-analysis with REML estimation used for pooling. I² values < 25%, 25%–50%, 50%–75%, and >75% indicate low, moderate, substantial, and high heterogeneity, respectively.

Table S9. Model comparison sensitivity analysis

CLPM results across four covariate adjustment strategies: model 0 (unadjusted), model 1 (age + gender), model 2 (age + gender + education + BMI), model 3 (full adjustment). Standardized coefficients, standard errors, p values, and model fit indices reported.

Table S10. Baseline dual-disease exclusion analysis

Comparison of main analysis versus analysis excluding participants with both diabetes and arthritis at baseline. Standardized coefficients and fit indices for both samples.

Table S11. Long time-lag analysis (T1→T3)

Direct effects from T1 to T3 controlling for T2 status, examining long-term causal relationships. Standardized coefficients, standard errors, p values, and model fit indices.

Table S12. Complete case analysis

CLPM results restricted to participants with no missing data on key analysis variables. Sample sizes and results compared with multiple imputation analysis.