Abstract
Background
Composite metabolic-oxidative scores have shown inconsistent associations with cognitive outcomes, yet whether their components move in the same direction remains unexamined.
Objective
To test whether Metabolic Oxidative Risk Score (MORS) tracks cognitive change, and to decompose MORS to assess whether component effects on cognitive slope are concordant or opposing.
Methods
We constructed a four-component MORS (uric acid, triglycerides, fasting glucose, reverse-scored HDL; range 4–20) and examined its association with cognitive change across four waves (2011–2018) in CHARLS (N = 7,300, age ≥ 45 years) using linear mixed models. Equivalence testing distinguished null findings from insufficient power. Components were decomposed individually and simultaneously. Cross-sectional triangulation used NHANES 2011–2014 (N = 1,340, age ≥ 60 years).
Results
The adjusted MORS × time interaction was β = −0.00046 SD/year (95% CI −0.00185 to 0.00093; p = 0.516); equivalence testing rejected any clinically meaningful effect (p = 0.0002). Component decomposition revealed opposing effects: higher uric acid predicted steeper decline (β = −0.0055; p = 0.003) while higher triglycerides predicted attenuated decline (β = + 0.0042; p = 0.045), with near-perfect cancellation at the composite level. In NHANES, the MORS–cognition association reversed sign after socioeconomic adjustment and showed a significant non-linear association (p = 6 × 10−4); leave-one-out analysis identified uric acid as the dominant signal carrier.
Conclusions
An equal-weight metabolic-oxidative composite does not track cognitive decline because its components exert opposing effects that cancel. Uric acid emerges as an independent, potentially modifiable risk factor with implications for Alzheimer's disease prevention research.
Keywords: Alzheimer's disease, CHARLS, cognitive aging, dementia risk, equivalence testing, linear mixed models, NHANES, triglycerides, uric acid
Introduction
Population aging has made cognitive decline one of the defining public health challenges of this century. The 2024 report of the Lancet standing Commission on dementia prevention estimated that approximately 45% of dementia cases are attributable to 14 modifiable risk factors across the life course. 1 Several of these factors cluster within the metabolic syndrome: hypertension, diabetes, dyslipidemia, and obesity. A recent meta-analysis of 21 prospective studies comprising 411,810 participants reported that metabolic syndrome was associated with a 33% increased risk of all-cause dementia (RR 1.33, 95% CI 1.03 to 1.71) and more than a doubling of vascular dementia risk (RR 2.07, 95% CI 1.32 to 3.24). 2 A separate meta-analysis of 28 prospective cohorts encompassing 6.75 million participants reached similar conclusions. 3 The mechanistic link is biologically coherent. Chronic metabolic stress amplifies systemic reactive oxygen species, and mitochondrial dysfunction together with impaired antioxidant defense accelerates neuronal damage, synaptic loss, and cognitive deterioration in the aging brain.4–8
Given these links, a natural strategy for epidemiology is to construct composite scores that integrate several metabolic and oxidative markers into a single exposure variable. Such aggregation follows the tradition of cardiovascular risk scoring and has the appeal of statistical parsimony and clinical translatability. However, the empirical record on composite metabolic scores and cognition has been mixed. Longitudinal studies of metabolic syndrome status changes in the Lifelines cohort found no differences in cognitive performance among remitting, de novo, and persistent phenotypes. 9 Several Chinese community studies have likewise reported null or weakly positive associations between cardiometabolic burden and cognitive decline among older adults.10–12 These inconsistent findings raise a question that composite-score studies rarely examine directly: do the components of the score move in the same direction with respect to the cognitive outcome?
There are strong a priori reasons to expect that they may not. Uric acid has a dual biological identity. It is the most abundant aqueous antioxidant in human plasma, yet chronic hyperuricemia promotes endothelial dysfunction, cerebral small vessel disease, and neuroinflammation. 13 A systematic review of 19 observational studies (49,901 participants) reported that the pooled analysis of prospective cohorts indicated higher uric acid exerted a deleterious effect specifically on attention and executive function (β = −0.02, 95% CI −0.04 to 0.00), even while the cross-sectional relationship with global cognition appeared marginal. 14 In a China Health and Retirement Longitudinal Study (CHARLS)-based prospective analysis of 7828 non-hyperuricemic adults, participants with low or low-normal plasma uric acid had poorer cognitive performance and steeper decline in executive function over four years compared with those with high-normal levels, with a significant interaction by age. 15 These domain-specific and bidirectional signals describe a complex, potentially non-linear relationship between uric acid and cognition.16,17
Triglycerides present an equally heterogeneous picture. Measured in late life, they may reflect nutritional adequacy, caloric intake, and metabolic resilience rather than atherogenic burden. In a prospective analysis of the ASPREE trial, Zhou and colleagues examined 18,294 community-dwelling adults aged 65 years and older and reported that higher triglyceride levels were associated with a lower risk of all-cause dementia, with a hazard ratio of 0.82 (95% CI 0.72 to 0.94) per doubling of triglycerides, and with slower decline in composite cognition and memory. 18 External validation in 68,200 UK Biobank participants produced an essentially identical estimate. The authors explicitly noted that late-life triglyceride concentrations are influenced by contemporaneous lifestyle behaviors, nutrition, and general health, providing a plausible pathway through which triglycerides could act as a marker of vitality in older adults. High-density lipoprotein cholesterol, typically viewed as cardioprotective, shows a U-shaped association with cognitive impairment in Chinese and American older adults, with both low and very high levels associated with poorer cognition.19–22 Glucose regulation, unambiguously harmful at the level of diabetes, shows weaker and less consistent associations at fasting plasma glucose levels below the diabetic threshold among older adults. 23
These heterogeneous component-level findings are pre-analytic facts in the published literature. They imply that an equal-weight additive composite such as the Metabolic Oxidative Risk Score (MORS) carries an implicit assumption of directional concordance across its components that may not hold for cognitive outcomes. If the components cancel, the composite will appear null even when individual biomarkers carry real and opposing biological signal. The consequence is a masked heterogeneity that standard composite analyses cannot detect.
The present study had three objectives. First, we estimated the longitudinal association of MORS, calculated from uric acid, triglycerides, fasting plasma glucose, and reverse-scored HDL cholesterol, with four-wave trajectories of cognition in the CHARLS cohort. Second, motivated by the pre-specified heterogeneity hypothesis grounded in the component literature cited above, we decomposed MORS into its four constituents and assessed their individual and simultaneous contributions to cognitive slope. Third, we triangulated our CHARLS findings in an independent older adult sample from the National Health and Nutrition Examination Survey (NHANES) 2011 to 2014 using the same score and a validated cognitive battery. Throughout, we distinguished carefully between rejecting a null hypothesis and providing evidence for equivalence, using two one-sided tests with a pre-specified smallest effect size of interest.24,25 Our intent was neither to affirm nor reject MORS per se, but to understand what an aggregated metabolic-oxidative exposure variable reveals, and conceals, about cognitive aging in an older Chinese population. Given that metabolic dysregulation is among the established modifiable risk factors for Alzheimer's disease and related dementias,1–3 understanding which components of such composites carry genuine signal has direct relevance to prevention research targeting the preclinical trajectory of dementia.
Methods
Study designs and populations
We used two complementary data sources. The longitudinal analysis drew on the CHARLS, a nationally representative cohort of community-dwelling adults aged 45 years and older, initiated in 2011 and followed biennially. 26 We used waves 1 (2011), 2 (2013), 3 (2015), and 4 (2018), which together span seven years of follow-up and four cognitive assessments. Venous blood samples collected in the 2011 baseline wave provided all biomarker exposures. 27 After exclusions for age below 45, missing biomarker data, fewer than two cognitive assessments, baseline brain damage or memory-related disease, and extreme biomarker values, the analytic sample comprised 7300 participants. The cross-sectional triangulation used the NHANES 2011 to 2012 and 2013 to 2014 cycles. Eligibility required age 60 years or older, complete cognitive test data, and non-missing biomarker values, yielding an analytic sample of 1340 participants. Both studies obtained informed consent from participants, and the original data collection was approved by their respective institutional review boards. The present secondary analysis used de-identified public data.
Construction of the metabolic oxidative risk score
MORS integrates four fasting blood biomarkers whose dysregulation has been mechanistically implicated in oxidative stress, systemic inflammation, and endothelial dysfunction in aging. The four components were serum uric acid, triglycerides, fasting plasma glucose, and HDL cholesterol. Within each cohort, the first three components were coded by sex-specific quartiles on the directly harmful scale, with quartile 1 receiving 1 point and quartile 4 receiving 4 points. HDL cholesterol was reverse-coded, so that the lowest quartile (most harmful) received 4 points and the highest quartile received 1 point. The total score therefore ranged from 4 to 20, with higher values representing greater cumulative metabolic-oxidative burden. Quartile cut-points were derived within each cohort to preserve cohort-specific distributional characteristics and to avoid spurious population-level effects.
Cognitive outcomes
In CHARLS, cognitive function was measured at each wave using a harmonized battery adapted from the Health and Retirement Study. 28 Episodic memory was captured through immediate and delayed recall of 10 Chinese words, yielding a score ranging from 0 to 20. Mental status was assessed by a modified Telephone Interview for Cognitive Status, incorporating serial-7 subtraction, date orientation, and figure drawing, with scores from 0 to 11. The two domains were summed to a total cognitive score ranging from 0 to 31. To align with NHANES, we also computed a standardized global cognition Z score by centering and scaling each wave-specific total score to the baseline wave. In NHANES, cognitive function was evaluated using the Consortium to Establish a Registry for Alzheimer's Disease Word Learning subtest (CERAD-WL) for immediate and delayed memory, the Animal Fluency test (AFT) for executive function and verbal fluency, and the Digit Symbol Substitution Test (DSST) for processing speed.29–31 A composite global cognition Z score was calculated by averaging the four domain-specific Z scores, computed relative to the full analytic sample.
Covariates
Three hierarchical adjustment models were prespecified. Model 1 adjusted for age and sex. Model 2 additionally adjusted for educational attainment (five categories) and, in NHANES, self-reported race or ethnicity. Model 3 further adjusted for marital status, residence (urban or rural, CHARLS only), smoking status (never, former, current), alcohol consumption, body mass index, self-reported hypertension, diabetes, and cardiovascular disease. In NHANES, family income-to-poverty ratio replaced residence in Model 3. All covariates were measured at baseline. To mitigate bias from missing data, multivariate imputation by chained equations with 20 imputations was performed for variables with more than 1% missingness, and pooled estimates were reported. 32 Complete-case analysis was performed as a sensitivity check.
Statistical analysis
For CHARLS, the primary analysis was a linear mixed model for the standardized global cognition Z score with random intercepts and random slopes by participant, and fixed effects for baseline MORS, time in years since baseline, and their interaction. The MORS × time interaction coefficient is the quantity of interest because it describes how MORS modifies the rate of cognitive change. Restricted maximum likelihood estimation was used. We prespecified three hierarchical adjustment models as described above. As incident cognitive impairment was a dichotomous complement, we fitted logistic regression for new-onset cognitive impairment at any follow-up wave, defined as a global cognition Z score below the 10th percentile of baseline. For NHANES, we fitted linear regression of global cognition Z score on MORS with the same three-model adjustment strategy.
Because our primary CHARLS analysis anticipated a null or small effect, we performed an equivalence test using two one-sided tests (TOST) against bounds of ±0.003 SD/year.24,25,33 This bound was chosen conservatively and transparently. Typical minimal clinically important differences for annual cognitive change in older adults lie between 0.05 and 0.5 SD/year. A bound of 0.003 SD/year is therefore two to three orders of magnitude below any clinical threshold and corresponds roughly to the smallest slope difference our sample size could detect with 80% power. Rejection of the TOST null at this bound constitutes formal evidence that no slope difference larger than a trivially small effect is consistent with the data.
Restricted cubic splines with three knots at the 10th, 50th, and 90th percentiles of MORS were fitted for both the baseline cross-sectional association and the longitudinal slope, with a likelihood ratio test comparing linear and non-linear specifications. 34 Pre-specified sensitivity analyses comprised: (1) complete-case analysis, (2) exclusion of participants with baseline cognitive impairment, (3) exclusion of participants with diabetes, (4) lagged exposure using wave 2 biomarker values where available, (5) unweighted analysis, (6) exclusion of participants with extreme MORS values, and (7) restriction to participants with three or more cognitive assessments. Pre-specified subgroup analyses were conducted by age (< 60 and ≥60), sex, education (lower and higher), residence (urban and rural), and baseline cognitive tertile, with p values for interaction from product terms in the fully adjusted model.
Component decomposition followed a pre-specified two-step strategy motivated by published heterogeneity in the component literature. First, we fitted four separate single-component LMMs replacing MORS with each component, specified both as continuous (per SD) and as quartiles. Second, we fitted a single simultaneous LMM with all four components entered as quartiles together, each with its own time interaction. This design allowed direct assessment of each component's independent contribution to cognitive slope while accounting for correlations among components. The sum of the four component × time coefficients in the simultaneous model approximates the MORS × time coefficient and provides a decomposition of the composite null into its directional parts.
Analyses were performed in R version 4.3.2, using lme4 for mixed models, mice for imputation, TOSTER for equivalence testing, rms for restricted cubic splines, and survey for NHANES complex survey weights. 35 Two-sided p < 0.05 was considered statistically significant, with no multiplicity correction applied to the pre-specified primary analyses. Sensitivity and subgroup analyses are interpreted descriptively.
Results
Baseline characteristics
Among the 7300 CHARLS participants, the mean baseline age was 58.0 years (SD 8.5), 48.9% were female, 77.8% resided in rural areas, and the mean MORS was 12.5 (SD 3.2). Baseline characteristics stratified by MORS tertile are summarized in Table 1. As expected from the biological construction of the score, higher MORS tertiles were associated with higher body mass index (p < 0.001), higher prevalences of hypertension (30.0% in T3 versus 16.9% in T1, p < 0.001) and diabetes (14.8% in T3 versus 2.9% in T1, p < 0.001), and marginally higher cardiovascular disease prevalence. These gradients confirm that MORS behaves as a valid indicator of cardiometabolic burden.
Table 1.
Baseline characteristics of the CHARLS analytic sample, stratified by MORS tertile (N = 7300).
| Characteristic | T1 (low) | T2 | T3 (high) | p (trend) |
|---|---|---|---|---|
| N | 2434 | 2431 | 2435 | NA |
| Age, years (mean ± SD) | 57.8 ± 8.8 | 58.0 ± 8.4 | 58.3 ± 8.3 | 0.062 |
| Female, n (%) | 1082 (44.5) | 1198 (49.3) | 1292 (53.1) | <0.001 |
| Rural residence, n (%) | 1960 (80.5) | 1905 (78.4) | 1815 (74.5) | <0.001 |
| BMI, kg/m2 (mean ± SD) | 22.1 ± 3.4 | 23.6 ± 3.5 | 25.1 ± 3.7 | <0.001 |
| Hypertension, n (%) | 412 (16.9) | 565 (23.2) | 730 (30.0) | <0.001 |
| Diabetes, n (%) | 70 (2.9) | 176 (7.2) | 360 (14.8) | <0.001 |
| Current smoker, n (%) | 770 (31.6) | 728 (30.0) | 702 (28.8) | 0.041 |
| MORS (mean ± SD) | 8.9 ± 1.3 | 12.3 ± 0.9 | 16.0 ± 1.4 | <0.001 |
| Baseline global cognition | 11.28 ± 4.9 | 11.60 ± 4.8 | 11.79 ± 4.6 | <0.001 |
Tertiles were defined within the analytic sample. p values for trend were calculated by including the tertile rank as a continuous variable in a linear or logistic regression model as appropriate. BMI: body mass index; MORS: Metabolic Oxidative Risk Score.
A paradoxical observation emerged for the baseline cognitive outcome. Participants in the highest MORS tertile had a higher baseline global cognition score than those in the lowest tertile (11.79 versus 11.28, p < 0.001). This cross-sectional positivity is inconsistent with a simple harmful effect of metabolic-oxidative burden on cognition. It is, however, entirely consistent with two epidemiological phenomena commonly observed in late-life cohorts: depletion-of-susceptibles-type selective survival and the fact that urban participants, with generally higher educational attainment and cognitive scores, also carry greater metabolic burden.36,37 We return to this observation in the Discussion.
Primary longitudinal analysis in CHARLS: null and equivalent
The primary mixed model showed no association between baseline MORS and the rate of cognitive change. In Model 3, the MORS × time interaction was β = −0.00046 SD/year (95% CI −0.00185 to 0.00093, p = 0.516; Table 2). Estimates in Model 1 (β = −0.00052, p = 0.437) and Model 2 (β = −0.00056, p = 0.401) were essentially identical, indicating insensitivity to covariate adjustment. Figure 1 displays the estimated cognitive trajectories by MORS tertile, which are visibly parallel over the 7-year follow-up.
Table 2.
Primary longitudinal analysis in CHARLS. Linear mixed model estimates for the MORS × time interaction on global cognition Z score.
| Model | β (MORS × time), SD/year | 95% CI | p | N (obs.) |
|---|---|---|---|---|
| Model 1 (age, sex) | −0.00052 | −0.00183 to 0.00079 | 0.437 | 27,462 |
| Model 2 (+ education) | −0.00056 | −0.00186 to 0.00074 | 0.401 | 27,462 |
| Model 3 (fully adjusted) | −0.00046 | −0.00185 to 0.00093 | 0.516 | 27,462 |
| TOST equivalence test (bounds ±0.003 SD/year) | NA | NA | 0.0002 | NA |
| Incident cognitive impairment (logistic) | OR = 0.987 | 0.959 to 1.016 | 0.372 | 7300 |
Models include random intercept and slope per participant. TOST: two one-sided tests. The equivalence bound of ±0.003 SD/year corresponds to approximately 2–3 orders of magnitude below the minimal clinically important difference for annual cognitive change.
Figure 1.

Estimated cognitive trajectories over 7 years of follow-up in CHARLS by MORS tertile, from the fully adjusted linear mixed model (model 3). The three tertiles show essentially parallel slopes, reflecting the null MORS × time interaction (β = −0.00046 SD/year, p = 0.516). The baseline offset between tertiles is cross-sectional and does not represent a longitudinal effect of MORS.
To distinguish this null result from mere insufficient power, we performed TOST equivalence testing against bounds of ±0.003 SD/year. The test rejected the null of a meaningful effect with p = 0.0002, providing positive evidence that the true MORS × time effect lies within the trivially small interval. Incident cognitive impairment, analyzed as a binary outcome, showed the same pattern: OR = 0.987 per 1-point MORS increase (95% CI 0.959 to 1.016, p = 0.372). Both the continuous slope and the dichotomous incidence analyses converged on the same conclusion.
Sensitivity and subgroup analyses
Seven prespecified sensitivity analyses, summarized in Table 3, produced uniformly null estimates. Complete-case analysis (β = −0.00051), exclusion of baseline cognitive impairment (β = −0.00040), exclusion of diabetes (β = −0.00027), lagged exposure using wave 2 biomarkers (β = −0.00034), unweighted analysis (β = −0.00048), exclusion of extreme MORS values (β = −0.00042), and restriction to participants with three or more assessments (β = −0.00058) all yielded confidence intervals that straddled zero and fell well within the equivalence bounds.
Table 3.
Sensitivity analyses of the MORS × time interaction in CHARLS (model 3).
| Sensitivity specification | β (SD/year) | 95% CI | p |
|---|---|---|---|
| S1 Complete-case analysis (no MI) | −0.00051 | −0.00193 to 0.00091 | 0.481 |
| S2 Exclude baseline cognitive impairment | −0.00040 | −0.00180 to 0.00100 | 0.574 |
| S3 Exclude diabetic participants | −0.00027 | −0.00175 to 0.00121 | 0.723 |
| S4 Lagged exposure (wave 2 biomarkers) | −0.00034 | −0.00188 to 0.00120 | 0.669 |
| S5 Unweighted (no sampling weights) | −0.00048 | −0.00187 to 0.00091 | 0.499 |
| S6 Exclude extreme MORS (< 6 or > 18) | −0.00042 | −0.00184 to 0.00100 | 0.562 |
| S7 Participants with ≥3 cognitive assessments | −0.00058 | −0.00197 to 0.00081 | 0.414 |
All specifications yield confidence intervals within the pre-specified equivalence bounds. MI: multiple imputation.
Subgroup analyses by age, sex, education, residence, and baseline cognition (Figure 2) showed no heterogeneity in the MORS × time interaction, with interaction p values ranging from 0.095 to 0.813. A formal test for interaction between baseline cognitive tertile and MORS × time (p = 0.278) yielded no evidence that the null result was driven by a selection-bias interaction at the cognitive distribution's extremes.
Figure 2.

Forest plot of MORS × time slope estimates across pre-specified subgroups in CHARLS (model 3). Subgroups are defined by age, sex, diabetes status, and residence. p values for interaction from product terms ranged from 0.095 to 0.813, providing no evidence of subgroup heterogeneity in the MORS-cognition slope association. The dashed vertical line at zero represents no effect.
Component decomposition: opposing directions within the composite
Given the robust null at the composite level, we proceeded to the prespecified component decomposition. The single-component LMMs are summarized in Table 4A. Uric acid, specified as a continuous variable in SD units, was the only component with a statistically significant time interaction: β = −0.00512 SD/year per SD (p = 0.002). The coefficients for triglycerides (β = + 0.00002, p = 0.428), fasting plasma glucose (β = 0.00000, p = 0.972), and HDL cholesterol (β = −0.00008, p = 0.524) were all indistinguishable from zero in isolation.
Table 4.
Component decomposition of MORS in CHARLS. (A) Four single-component linear mixed models. (B) A single simultaneous four-component model with components entered as quartiles (reference quartile 1).
| A. Single-component LMMs (continuous, per SD) | |||
|---|---|---|---|
| Component × time | β (per SD/year) | 95% CI | p |
| Uric acid × time | −0.00512 | −0.00841 to −0.00183 | 0.002 |
| Triglycerides × time | +0.00002 | −0.00003 to 0.00007 | 0.428 |
| Fasting glucose × time | 0.00000 | −0.00018 to 0.00018 | 0.972 |
| HDL × time | −0.00008 | −0.00032 to 0.00016 | 0.524 |
The simultaneous four-component model, with all four components entered as quartiles together, revealed a more informative structure (Table 4B). Higher uric acid quartile independently predicted steeper cognitive decline (β = −0.00552 SD/year per quartile, p = 0.003). Higher triglyceride quartile independently predicted attenuated decline, that is, a protective slope effect (β = + 0.00424 SD/year per quartile, p = 0.045). Fasting plasma glucose (β = −0.00034, p = 0.856) and HDL cholesterol (β = −0.00168, p = 0.422) contributed no slope effects. The uric acid and triglyceride coefficients were nearly equal in magnitude and opposite in sign. Their near-perfect mutual cancellation accounts mechanically for the MORS × time null: the four component × time coefficients sum to −0.00130 SD/year, of the same order as the observed MORS × time estimate.
B. Simultaneous four-component LMM (quartiles, per quartile increment).
| Quartile × time | β (per quartile/year) | 95% CI | p |
|---|---|---|---|
| Uric acid quartile × time | −0.00552 | −0.00914 to −0.00190 | 0.003 |
| Triglyceride quartile × time | +0.00424 | +0.00010 to +0.00838 | 0.045 |
| FPG quartile × time | −0.00034 | −0.00406 to 0.00338 | 0.856 |
| HDL quartile × time | −0.00168 | −0.00580 to 0.00244 | 0.422 |
All estimates from Model 3 specifications. In B, HDL quartiles are on the reverse-scored scale (higher quartile = lower HDL, harmful direction), matching the MORS coding convention. The uric acid and triglyceride coefficients are nearly equal in magnitude and opposite in sign, accounting for the composite-level null. FPG: fasting plasma glucose; HDL: high-density lipoprotein cholesterol.
These component-level findings are not post hoc. They were pre-specified on the basis of the heterogeneous component literature cited in the Introduction. The ASPREE prospective analysis of triglycerides and dementia 18 and the Tang et al. systematic review of uric acid and cognitive decline 14 both antedated the present analysis and provided biologically grounded directional predictions that our CHARLS data now confirm quantitatively in an older Chinese cohort.
Figure 3 depicts the cross-sectional baseline relationships in CHARLS for MORS (Figure 3A) and uric acid (Figure 3B). MORS shows a monotonically positive association with baseline cognition, the survivor-bias signature noted above. Uric acid shows a log-shaped plateau, consistent with the dual antioxidant-prooxidant biology of urate. The contrast between the cross-sectional positive pattern and the longitudinal harmful slope for uric acid illustrates that baseline level and rate of change are distinct dimensions that can coexist without contradiction.
Figure 3.

Restricted cubic spline curves for the cross-sectional associations of MORS (A) and serum uric acid (B) with baseline global cognition in CHARLS. MORS shows a monotonically positive relationship, consistent with a depletion-of-susceptibles survivor-bias pattern and with urban/education confounding. Uric acid shows a log-shaped plateau beginning at approximately 7 mg/dL. The longitudinal MORS × time interaction is null (see Table 2 and Figure 1), while the longitudinal uric acid × time interaction is negative (see Table 4), illustrating that baseline level and rate of change carry distinct information.
Cross-sectional triangulation in NHANES: the role of adjustment
In the NHANES 2011 to 2014 sample of 1340 adults aged 60 years and older, the crude and minimally adjusted association between MORS and global cognition was negative: Model 1 β = −0.023 (95% CI −0.038 to −0.008, p = 0.004). With adjustment for race and ethnicity, educational attainment, and family income-to-poverty ratio in Model 2, the coefficient attenuated toward the null: β = −0.007 (p = 0.149). In Model 3, the fully adjusted estimate reversed sign: β = + 0.015 (95% CI −0.013 to 0.043, p = 0.294). A sign reversal on full adjustment is a recognized feature of either confounding by socioeconomic position acting through multiple metabolic components, or overadjustment when mediators on the causal pathway are included in the model. The direction of attenuation from Model 1 to Model 3 indicates that a substantial fraction of the crude negative association between MORS and cognition was attributable to SES-related confounding rather than to a direct metabolic effect.
Restricted cubic spline analysis in NHANES revealed a significant non-linear association between MORS and global cognition (p non-linear = 6 × 10−4; Figure 4), with cognitive scores declining across the MORS range and steepening at higher values.
Figure 4.

Restricted cubic spline curve for the cross-sectional association between MORS and global cognition Z score in NHANES 2011 to 2014 (model 3). A significant non-linear association was detected (p for non-linearity = 6 × 10−4), with predicted cognitive scores declining across the MORS distribution and the decline steepening at higher MORS values. The spline has three knots at the 10th, 50th, and 90th percentiles of MORS.
The non-linearity, manifesting as a steepening decline at higher MORS values, is compatible with two non-exclusive explanations: a dose-dependent accumulation of metabolic-oxidative burden that accelerates cognitive deterioration at higher composite scores, and reverse causation from undiagnosed incipient dementia producing metabolic deterioration. Leave-one-out analysis, in which the association was re-estimated after removing each component in turn, identified uric acid as the component whose removal most substantially attenuated the composite signal (β dropped from −0.023 to −0.013 and the p value rose from 0.004 to 0.637; Figure 5). This pattern identifies uric acid as the dominant signal carrier within the NHANES composite.
Figure 5.

Leave-one-out (LOO) analysis of MORS components in NHANES. Each bar shows the re-estimated main-model coefficient (β) and 95% CI after removing the indicated component from MORS. Removal of uric acid produces the largest attenuation (β from −0.023 to −0.013; p from 0.004 to 0.637), identifying uric acid as the dominant signal carrier within the composite. Removal of other components (HDL-C, triglycerides, fasting glucose) produced negligible attenuation (all Model 3 p > 0.05).
All seven sensitivity analyses in NHANES, including restriction by age, exclusion of diabetes, and re-scoring with sex-combined quartiles, produced estimates consistent with the main analysis (all Model 3 p > 0.05).
Discussion
Principal findings
In a nationally representative longitudinal cohort of 7300 Chinese adults followed for 7 years, an equal-weight additive MORS showed no association with the rate of cognitive decline. This null was not the result of insufficient power. A formal equivalence test rejected the presence of any slope effect larger than 0.003 SD/year, an effect size two orders of magnitude below any clinical threshold of meaningful cognitive change. Component decomposition revealed why the composite was null: two of its four components moved the cognitive slope in nearly symmetric opposite directions. Higher uric acid accelerated decline, while higher triglycerides slowed it. Fasting glucose and HDL cholesterol contributed negligible slope effects in this age range. Cross-sectional triangulation in NHANES 2011 to 2014 reproduced the pattern at the component level, identified uric acid as the dominant signal carrier through leave-one-out analysis, and demonstrated that the crude MORS-cognition association is largely a function of socioeconomic confounding.
The aggregation problem: when composites mask the biology
The central methodological insight of this analysis is that composite scores make an implicit assumption. When several risk factors are summed with equal weights and a single directional coding (higher = worse), the investigator implicitly assumes that all components move the outcome in the same direction. This assumption is approximately valid for cardiovascular outcomes, where the components of metabolic syndrome indeed accumulate risk in a mostly additive fashion. For neurocognitive outcomes, the assumption is more fragile. Uric acid and triglycerides in late life show not just weak discordance, but active opposition in their association with cognitive slope. When the two components are weighted equally and summed, their effects cancel arithmetically. The composite appears null, and the investigator is left with the impression that metabolic-oxidative burden does not matter for cognition. The component decomposition shows that the opposite is true: metabolic-oxidative markers matter a great deal, but they matter in opposing directions.
This problem is not new in epidemiology, but it is underappreciated in the composite-score era. It follows from a simple algebraic identity: the coefficient of a sum is the sum of the coefficients, only when covariances are ignored. Whenever components are correlated and their individual slope effects differ in sign, the composite slope is a weighted average that tends toward zero. The Lifelines cohort null for metabolic syndrome transitions and cognition, 9 the Qatar Biobank finding that fasting glucose was the sole component associated with cognition in an otherwise null metabolic syndrome model, 38 and the Chinese MetS severity score findings that emerge only with sex- and age-specific weights 11 all point to the same phenomenon in different guises.
Uric acid: a domain-specific, potentially modifiable risk factor
The uric acid slope effect warrants particular attention because it was the only individual component that survived the simultaneous four-variable decomposition, and because it is the one component with plausible routes to clinical intervention. The magnitude we observed (β = −0.0055 SD/year per quartile, or approximately 4% of a standard deviation of cognitive decline over 7 years per quartile) is small at the individual level, but it is potentially actionable at the population level given the high prevalence of hyperuricemia in older Chinese populations.14,16 Our finding is biologically coherent with prior work. Tang et al. 14 reported a similar coefficient (β = −0.02 per category) for the attention and executive function domain, and Huang et al. 15 showed stronger adverse effects of uric acid extremes on cognitive performance in participants aged 60 and older. The specificity of the effect, in magnitude and in direction, is striking given the methodological independence of our analysis.
A natural question is why the harmful direction of uric acid should dominate in this cohort, given its dual identity as both an antioxidant and a vasculotoxic agent. We propose that the balance tips toward harm in the context of the CHARLS population for two reasons. First, hyperuricemia in older Chinese adults is predominantly driven by purine-rich dietary patterns, impaired renal excretion, and metabolic syndrome rather than by conditions associated with antioxidant depletion, and therefore tracks endothelial injury, cerebral small vessel disease, and chronic low-grade inflammation more faithfully than it tracks oxidative stress protection.13,16 Second, the time scale of our analysis is seven years, which is long enough for the cumulative vascular neurotoxicity of persistently elevated urate to manifest as measurable cognitive slope differences, whereas acute antioxidant benefits would operate on a much shorter timescale. In this context, the protective antioxidant capacity of uric acid may be more relevant to acute neurological events than to the slow, vascular-mediated trajectory of cognitive aging in older adults living with chronic metabolic burden.
The distinction between the cross-sectional positive association observed at baseline in CHARLS (higher MORS tertile associated with higher baseline cognition) and the longitudinal harmful slope effect of uric acid is not a contradiction. Baseline level and rate of change are conceptually and statistically distinct quantities. A participant can begin with a high cognitive score because of early-life education and other reserve factors, and still experience a faster rate of decline because of late-life metabolic-oxidative stress. The cross-sectional positivity in our data is compatible with a survivor-bias pattern described by Mayeda et al. 36 and Chung et al.: 37 participants with high metabolic burden who reach the study age without clinical dementia are preferentially drawn from the subset whose cognitive reserve is relatively high. The longitudinal analysis, which conditions on baseline and tracks change, is therefore less vulnerable to this bias than cross-sectional estimation.
Triglycerides in late life: nutritional adequacy, not atherogenic signal
The protective slope effect of higher triglyceride quartile, although modest in magnitude, replicates in direction the large ASPREE + UK Biobank prospective analysis that reported a dementia hazard ratio of 0.82 per doubling of triglycerides and slower cognitive decline at higher triglyceride levels in community-dwelling older adults. 18 The biological interpretation proposed in that study, namely that late-life triglycerides reflect nutritional adequacy, caloric intake, and contemporaneous lifestyle behaviors rather than lifetime atherogenic burden, maps directly onto our Chinese cohort. In CHARLS, the majority of participants were rural residents with modest nutritional status, among whom low triglyceride concentrations plausibly signal undernutrition, cachexia, or chronic illness burden rather than a healthy metabolic phenotype. This population-specific dynamic is the opposite of what applies in midlife populations, where high triglycerides are a robust marker of cardiovascular risk. The same biomarker carries opposite information in different life stages and nutritional contexts, and this is precisely the kind of heterogeneity that equal-weight composites cannot accommodate.
This life-stage-specific interpretation is not a Chinese peculiarity. The ASPREE participants were predominantly non-Hispanic white Australian and American adults with adequate nutrition, yet the triglyceride-dementia relationship was identical in direction and nearly identical in magnitude to ours. 18 That convergence across two vastly different populations suggests that the reversal of the triglyceride signal from midlife atherogenesis to late-life nutritional vitality is a general biological phenomenon driven by the metabolic physiology of aging, not a regional dietary artifact. The implication for composite score design extends well beyond Chinese cohorts: any equal-weight metabolic risk score applied to community-dwelling older adults will likely carry this internally contradictory component structure, rendering the composite uninformative or actively misleading for cognitive outcomes regardless of the population studied.
Interpreting the NHANES non-linear and sign-reversing findings
In NHANES, the non-linear declining association and the sign reversal from Model 1 to Model 3 both have principled explanations. The steepening decline at higher MORS values is consistent with a dose-dependent metabolic-oxidative burden on cognitive function, and with the biology of uric acid as a vasculopathic agent at chronically elevated concentrations.13,16 The sign reversal under SES adjustment is most parsimoniously explained by a directed acyclic graph structure in which educational attainment and income act as common upstream determinants of both metabolic health and cognitive performance. Adjusting for these variables blocks the primary confounding path, revealing the residual association between MORS and cognition after removing the socioeconomic route. At that point, the remaining variation in MORS is concentrated among participants with relatively low SES who nonetheless carry high metabolic burden and yet have survived cognitively intact, a pattern that reflects selective survival: having high metabolic burden and reaching old age with preserved cognition may require precisely the kind of cognitive reserve that would produce a positive association under SES conditioning. Ferguson et al. 19 documented a similar adjustment-sensitivity for HDL in a cohort of 184,367 health plan members, underscoring that cross-sectional MORS-cognition associations reflect a complex superposition of confounding, survival, and biological effects that no single model can fully disentangle. Our result should therefore be read as a demonstration that cross-sectional MORS-cognition associations are adjustment-sensitive, and that the longitudinal CHARLS findings carry more inferential weight for causal inference.
Methodological considerations
The primary strength of our design lies in the combination of a longitudinal cohort with four cognitive assessments over 7 years, a pre-specified analysis plan with component decomposition grounded in prior literature, formal equivalence testing that distinguishes null findings from indeterminate ones, and cross-sectional triangulation in an independent cohort with a different population, different cognitive battery, and different assessment protocol. Few published MORS-like composite analyses have combined these features. Several limitations nonetheless apply.
First, biomarker exposures in CHARLS were measured only at baseline, so we cannot estimate associations between longitudinal biomarker trajectories and cognitive change. Repeated-measures biomarker data would allow separation of mean and variability effects. Year-to-year lipid variability has been independently associated with dementia and cognitive decline in ASPREE-XT 39 and is not captured by our single-measure exposure. Second, the cognitive batteries in CHARLS and NHANES differ, which prevents direct pooling and makes the triangulation qualitative rather than quantitative. Third, we did not adjust for urate-lowering therapy, statin use, or fibrate use. Pharmacological modification of the biomarker components may induce differential misclassification, which would typically bias component-specific estimates toward the null. However, nationwide pharmacy data suggest that statin prescription rates among rural Chinese adults in the 2011 to 2018 period remained low, and formal urate-lowering therapy was prescribed in a small minority of participants without clinical gout. Under conditions of low treatment prevalence, the magnitude of pharmacological misclassification is unlikely to fully account for the directional uric acid and triglyceride effects we observed, and would be expected to attenuate rather than generate opposing associations. Fourth, our conclusions apply to a Chinese middle-aged and older adult population with predominantly rural backgrounds; generalization to other populations requires local validation. Fifth, the cross-sectional component of the study cannot exclude reverse causation in which incipient dementia induces metabolic changes. The longitudinal CHARLS component mitigates but does not eliminate this concern. Finally, selection bias at recruitment remains an irreducible challenge in geriatric cohort research.40–42 Only participants who survived to age 45 and consented to biomarker sampling entered the analytic sample; the cross-sectional positive gradient at baseline illustrates this dynamic directly.
Implications for research and practice
The practical consequence of our findings is that equal-weight additive composite scores are not a reliable tool for studying metabolic-oxidative risk for cognitive decline in older adults. Three alternative strategies deserve consideration. First, component-specific analyses, including pre-specified decompositions and leave-one-out strategies, should accompany any composite-score analysis. Second, weighted composite construction, with weights derived from independent populations to avoid overfitting, may accommodate directional heterogeneity. Third, biomarker-specific intervention research, particularly targeting uric acid given its domain-specific and modifiable nature, may deliver more informative evidence than aggregate score studies. The broader lesson is that cognitive aging does not resemble cardiovascular disease in its response to metabolic burden. The biomarker biology in late life is life-stage-specific, nutrition-sensitive, and subject to survival and reverse-causation dynamics that simple additive constructs cannot capture.
Interpreting biomarker endpoints in therapeutic trials
These findings also bear on the interpretation of biomarker endpoints in therapeutic trials targeting cognitive outcomes in older adults. A central distinction is between a biomarker that lies on the causal pathway to the outcome and a biomarker that serves as an indirect marker of an upstream physiological state. 43 Our triglyceride finding belongs to the latter category: in late life, circulating triglycerides plausibly index nutritional adequacy and general vitality rather than acting directly on neuronal pathways. The same numerical change in such a marker can therefore carry opposite prognostic meaning depending on what produced it. Reductions in total cholesterol and triglycerides observed with advancing age in population studies are frequently artifacts of selective survival, whereby individuals with low values due to underlying frailty or illness are progressively removed from the observed population, while those who remain healthy retain higher values for longer. 44 A formally analogous phenomenon has been described for unintentional weight loss in older cohorts, where a composite metabolite score dominated by lower triglycerides and altered lipid fractions tracked subsequent mortality and mobility decline, indicating that a triglyceride decline driven by deteriorating nutritional status is a marker of harm rather than of metabolic improvement. 45 By contrast, when triglyceride reductions are achieved through an intentional lifestyle intervention that simultaneously improves fitness, glycemic control, and other markers of metabolic health, as in the Look AHEAD trial, the same directional change in triglycerides accompanies an overall improvement in health status rather than its erosion. 46 More generally, a triglyceride decline following a metabolic or lifestyle intervention should not be read in isolation as evidence against benefit, nor automatically as the loss of a protective compensation. Its meaning depends on the accompanying trajectory of weight, nutritional indices, and markers of frailty: a triglyceride fall accompanied by preserved or improving nutritional status and physical function is consistent with genuine metabolic improvement, whereas the same fall accompanied by weight loss, declining albumin, or emerging frailty would instead suggest that the intervention has eroded a compensatory reserve that, in this population, was associated with cognitive stability. This reasoning extends the broader argument of the present study: because the prognostic meaning of metabolic-oxidative biomarkers in late life is contingent on the physiological context that generates their values, endpoint selection and interpretation in cognitive aging trials cannot rely on the assumed directionality that would apply in midlife cardiometabolic disease, and should instead be anchored to a profile of co-occurring indicators of nutritional and functional status.
Conclusion
An additive metabolic oxidative risk score, constructed from uric acid, triglycerides, fasting glucose, and reverse-scored HDL cholesterol, does not track cognitive decline in a representative cohort of Chinese middle-aged and older adults. Equivalence testing provides formal evidence that the null is not a reflection of inadequate power. Decomposition of the composite reveals the biological reason: higher uric acid accelerates cognitive decline and higher triglycerides slow it, with near-symmetric cancellation at the composite level. Cross-sectional triangulation in NHANES confirms the component-level structure and highlights uric acid as the dominant signal carrier. Our findings argue against the uncritical use of equal-weight composite metabolic scores in cognitive aging research and in favor of component-specific longitudinal analyses that respect the directional heterogeneity of metabolic-oxidative biomarkers in late life.
Acknowledgements
We thank the CHARLS team at Peking University and the NCHS for making these data publicly available, and the study participants for their invaluable contributions.
During the preparation of this manuscript, ChatGPT (OpenAI, GPT-4) was used to assist with language editing and formatting of the text. All scientific content, analyses, and interpretations were performed exclusively by the authors. The authors reviewed and edited all content and take full responsibility for the integrity and accuracy of the published work.
Footnotes
ORCID iDs: Tian-Yu Li https://orcid.org/0009-0004-2263-1927
Zhi-Ming Yu https://orcid.org/0000-0003-3980-2134
Ethical considerations: CHARLS was approved by the Ethics Review Committee of Peking University (IRB00001052-11015). NHANES protocols were approved by the NCHS Research Ethics Review Board. The present secondary analysis of de-identified public data required no additional ethics approval.
Consent to participate: All CHARLS and NHANES participants provided informed consent to the original study teams at the time of data collection. The present study is a secondary analysis of de-identified public data, for which no additional consent was required.
Consent for publication: Not applicable.
Author contribution(s): Tian-Yu Li: Conceptualization; Data curation; Formal analysis; Methodology; Software; Visualization; Writing – original draft.
Zhi-Ming Yu: Funding acquisition; Project administration; Writing – review & editing.
Funding: The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors declare that financial support was received for the research and publication of this article from the YanZhen Talent Program of Wuxi People's Hospital (Grant No.: 2024-YZ-XKDTR-YZM-2024).
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data availability statement: CHARLS data are publicly available at https://charls.pku.edu.cn/; NHANES data at https://www.cdc.gov/nchs/nhanes. Analytic code is available from the corresponding author upon reasonable request.
References
- 1.Livingston G, Huntley J, Liu KY, et al. Dementia prevention, intervention, and care: 2024 report of the Lancet standing commission. Lancet 2024; 404: 572–628. [DOI] [PubMed] [Google Scholar]
- 2.Qiu SD, Zhang DD, Ma LY, et al. Associations of metabolic syndrome with risks of dementia and cognitive impairment: a systematic review and meta-analysis. J Alzheimers Dis 2025; 105: 15–27. [DOI] [PubMed] [Google Scholar]
- 3.Wang Q, Zhang L, Xu R, et al. Metabolic syndrome and risk of dementia and cognitive decline: a systematic review and meta-analysis of prospective cohort studies from 6,753,197 participants. GeroScience 2025. Epub ahead of print 18 Nov 2025. DOI: 10.1007/s11357-025-02014-9. [DOI] [PubMed] [Google Scholar]
- 4.Singh P, Barman B, Thakur MK. Oxidative stress-mediated memory impairment during aging and its therapeutic intervention by natural bioactive compounds. Front Aging Neurosci 2022; 14: 944697. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Teleanu DM, Niculescu AG, Lungu II, et al. An overview of oxidative stress, neuroinflammation, and neurodegenerative diseases. Int J Mol Sci 2022; 23: 5938. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Ionescu-Tucker A and Cotman CW. Emerging roles of oxidative stress in brain aging and Alzheimer’s disease. Neurobiol Aging 2021; 107: 86–95. [DOI] [PubMed] [Google Scholar]
- 7.Korovesis D, Rubio-Tomás T, Tavernarakis N. Oxidative stress in age-related neurodegenerative diseases: an overview of recent tools and findings. Antioxidants (Basel) 2023; 12: 131. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Maldonado E, Morales-Pison S, Urbina F, et al. Aging hallmarks and the role of oxidative stress. Antioxidants (Basel) 2023; 12: 651. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Marcolini S, Schneider CCI, Liemburg EJ, et al. Metabolic syndrome status changes and cognitive functioning: insights from the lifelines cohort study. J Prev Alzheimers Dis 2024; 11: 1283–1290. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Chen X, Li W, Huang Y, et al. Association of type 2 diabetes mellitus with cognitive function in adults: a prospective cohort study. J Alzheimers Dis 2023; 93: 1509–1520. [DOI] [PubMed] [Google Scholar]
- 11.Chen M, Liu L, Liu N, et al. Association of metabolic syndrome severity with cognitive decline among Chinese older adults: evidence from two prospective cohort studies. Front Neurosci 2026; 20: 1780804. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Ma X, Yan N, Song J. Association between metabolic score for insulin resistance and cognitive impairment in Chinese middle-aged and older adults: findings from the CHARLS study. J Alzheimers Dis 2026; 109: 152–163. [DOI] [PubMed] [Google Scholar]
- 13.Mijailovic NR, Vesic K, Borovcanin MM. The influence of serum uric acid on the brain and cognitive dysfunction. Front Psychiatry 2022; 13: 828476. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Tang X, Song ZH, Cardoso MA, et al. The relationship between uric acid and brain health from observational studies. Metab Brain Dis 2022; 37: 1989–2003. [DOI] [PubMed] [Google Scholar]
- 15.Huang Y, Zhang S, Shen J, et al. Association of plasma uric acid levels with cognitive function among non-hyperuricemia adults: a prospective study. Clin Nutr 2022; 41: 645–652. [DOI] [PubMed] [Google Scholar]
- 16.Liu Q, Peng M, Yang T, et al. Uric acid levels and risk of cognitive impairment: dose-response meta-analysis of prospective cohort studies. PLoS One 2023; 18: e0293832. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Huang W, Zhou Q, Sun Y, et al. Association between higher serum uric acid levels and cognitive function: a systematic review and meta-analysis. J Gerontol A Biol Sci Med Sci 2025; 80: glaf174. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zhou Z, Ryan J, Tonkin AM, et al. Association between triglycerides and risk of dementia in community-dwelling older adults: a prospective cohort study. Neurology 2023; 101: e2288–e2299. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Ferguson EL, Zimmerman SC, Jiang C, et al. Low- and high-density lipoprotein cholesterol and dementia risk over 17 years of follow-up among members of a large health care plan. Neurology 2023; 101: e2172–e2184. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Hussain SM, Robb C, Tonkin AM, et al. Association of plasma high-density lipoprotein cholesterol level with risk of incident dementia: a cohort study of healthy older adults. Lancet Reg Health West Pac 2023; 43: 100963. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Wang F, Shang X, Yang Q, et al. High-density lipoprotein cholesterol and cognitive impairment: a U-shaped relationship in China’s aging population. PLoS One 2026; 21: e0343768. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Kjeldsen EW Thomassen JQ Juul Rasmussen I, . et al. Plasma high-density lipoprotein cholesterol and risk of dementia: observational and genetic studies. Cardiovasc Res 2022; 118: 1330–1343. [DOI] [PubMed] [Google Scholar]
- 23.Yuan Z, Liu H, Zhou R, et al. Association of serum uric acid and fasting plasma glucose with cognitive function: a cross-sectional study. BMC Geriatr 2023; 23: 271. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Lakens D. Equivalence tests: a practical primer for t tests, correlations, and meta-analyses. Soc Psychol Personal Sci 2017; 8: 355–362. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Lakens D, Scheel AM, Isager PM. Equivalence testing for psychological research: a tutorial. Adv Methods Pract Psychol Sci 2018; 1: 259–269. [Google Scholar]
- 26.Zhao Y, Hu Y, Smith JP, et al. Cohort profile: the China health and retirement longitudinal study (CHARLS). Int J Epidemiol 2014; 43: 61–68. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Chen X, Crimmins E, Hu P, et al. Venous blood-based biomarkers in the China health and retirement longitudinal study: rationale, design, and results of the 2015 wave. Am J Epidemiol 2019; 188: 1871–1877. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Crimmins EM, Kim JK, Langa KM, et al. Assessment of cognition using surveys and neuropsychological assessment: the health and retirement study and the aging, demographics, and memory study. J Gerontol B Psychol Sci Soc Sci 2011; 66: i162–i171. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Brody DJ, Kramarow EA, Taylor CA, et al. Cognitive performance in adults aged 60 and over: national health and nutrition examination survey, 2011–2014. Natl Health Stat Report 2019; 126: 1–23. [PubMed] [Google Scholar]
- 30.Geng R, Huang X, Li Y. Elevated serum uric acid is associated with cognitive improvement in older American adults: a large, population-based analysis of the NHANES database. Front Aging Neurosci 2022; 14: 1024415. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Fillenbaum GG, van Belle G, Morris JC, et al. Consortium to establish a registry for Alzheimer’s disease (CERAD): the first twenty years. Alzheimers Dement 2008; 4: 96–109. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.van Buuren S, Groothuis-Oudshoorn K. Mice: multivariate imputation by chained equations in R. J Stat Softw 2011; 45: 1–67. [Google Scholar]
- 33.Caldwell AR. Exploring equivalence testing with the updated TOSTER R package. PsyArXiv, 2022.
- 34.Harrell FE. Regression modeling strategies. 2nd ed. Cham, Switzerland: Springer International Publishing, 2015. [Google Scholar]
- 35.Bates D, Mächler M, Bolker B, et al. Fitting linear mixed-effects models using lme4. J Stat Softw 2015; 67: 1–48. [Google Scholar]
- 36.Banack HR, Bea JW, Kaufman JS, et al. Investigating and remediating selection bias in geriatrics research: the selection bias toolkit. J Am Geriatr Soc 2019; 67: 1970–1976. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Chung Y, Glymour MM, Power MC, et al. Common methodological issues in observational epidemiological studies of older adults. Nat Ment Health 2025; 3: 865–874. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Al Hayek AA, Robert AA, Al Dawish MA, et al. Association between metabolic syndrome and decline in cognitive function: a cross-sectional study. Diabetes Metab Syndr Obes 2023; 16: 851–859. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Zhou Z, Moran C, Murray AM, et al. Association of year-to-year lipid variability with risk of cognitive decline and dementia in community-dwelling older adults. Neurology 2025; 104: e210247. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Mayeda ER, Tchetgen Tchetgen EJ, Power MC, et al. A simulation platform for quantifying survival bias: an application to research on determinants of cognitive decline. Am J Epidemiol 2016; 184: 378–387. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Weuve J, Tchetgen Tchetgen EJ, Glymour MM, et al. Accounting for bias due to selective attrition: the example of smoking and cognitive decline. Epidemiology 2012; 23: 119–128. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology 2004; 15: 615–625. [DOI] [PubMed] [Google Scholar]
- 43.Ciani O, Manyara AM, Davies P, et al. A framework for the definition and interpretation of the use of surrogate endpoints in interventional trials. eClinicalMed 2023; 65: 102283. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Roche-Lima A. Modeling multivariate distributions of lipid panel biomarkers for reference interval estimation and comorbidity analysis. Healthcare (Basel) 2025; 13: 2499. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Yao L, Simonsick EM, Newman AB, et al. A metabolite score of unintentional weight loss explained a substantial proportion of associated mortality and mobility limitation risk in a biracial older cohort. Aging Cell 2025; 24: e70181. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Wing RR; Look AHEAD research group. Cardiovascular effects of intensive lifestyle intervention in type 2 diabetes. N Engl J Med 2013; 369: 145–154. [DOI] [PMC free article] [PubMed] [Google Scholar]
