From October 2014 to March 2015, individuals from a village in the suburbs of Xi’an, northwestern China, were selected as the study population for cross-sectional analysis. The research procedure has been described in detail elsewhere (Shang et al., 2016). The screening process for participants is shown in Figure 1. Written informed consent was obtained from all participants, and the study was approved by the Medical Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University.

The inclusion criteria were (1) permanent residents of Qubao village and Bitou village in Huyi District, Xi’an; (2) age ≥40 years at the time of the baseline survey; and (3) participants who provided written informed consent and completed the baseline survey. Exclusion criteria: (1) individuals with conditions that may affect cognitive function, such as stroke, chronic alcoholism, brain trauma, previous craniocerebral surgery, central nervous system tumor, intracranial infection, epilepsy, organic psychosis, schizophrenia, affective psychosis, congenital mental retardation or untreated hypothyroidism; (2) individuals with severe visual or hearing impairment that prevents completion of the cognitive assessment; (3) participants with severe cardiac, hepatic, or renal diseases, acute or terminal stages of a variety of chronic illnesses; and (4) participants with missing or abnormal WHR and BMI data (more than the mean ± 3SD).

Anthropometric indicators, including height, weight, waist circumference, and hip circumference, were measured by trained nurses according to standardized protocols (Kerwin et al., 2010). Participants were measured while standing, wearing light clothing, and without shoes where applicable. Waist circumference was measured 1 cm above the upper edge of the navel, and hip circumference was measured at the widest point between the waist and thigh. WHR was calculated as waist circumference divided by hip circumference. For 6 participants with missing anthropometric measurements, available self-reported data were used to calculate BMI and WHR, and 7 participants with outlying BMI or WHR values were excluded.

Chinese Mini-Mental Status (CMMS) was used to assess global cognitive function. Cognitive impairment was defined as an CMMS score below the education-specific cutoff, specifically, ≤17 for illiterate participants, ≤20 for participants with secondary school education, and ≤24 for participants with junior high school education or above (Katzman et al., 1988).

Clinical characteristics, including demographic information, health-related behaviors, comorbidities, blood pressure measurements, and biochemical indicators, were collected. These variables included sex, age, years of education, education level, smoking status, alcohol consumption, physical activity, hypertension, diabetes, systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C). In the multivariable analyses, covariates included sex, age, education level, smoking status, alcohol consumption, physical inactivity, hypertension, FBG, TG, HDL-C, LDL-C, and BMI. Hypertension status was primarily determined from measured blood pressure, and self-reported medical history was used for the 12 participants without blood pressure measurements. Health-related behavioral information was collected through self-report. Regular exercise was defined as engaging in at least moderate-intensity physical activity, equivalent to brisk walking, at least three times per week for at least 30 min per session, or routinely performing moderate- or vigorous-intensity physical labor. Participants who did not meet this criterion were classified as physically inactive. Missing questionnaire-based covariate information was supplemented through telephone interviews within 3 months after the survey, and missing biochemical measurements were retested when possible. After follow-up interviews and retesting of biochemical measurements, missingness in the variables included in the final models was minimal. Complete-case analysis was used for multivariable modeling, and all primary fully adjusted models included 1,792 participants.

Data were analyzed by applying SPSS 26.0 software and R software (Version 4.4.1; R Core Team, 2024). Categorical variables were summarized as counts and percentages. Continuous variables were assessed for normality. Approximately normally distributed variables were summarized as mean ± standard deviation (SD), whereas skewed variables were summarized as median and interquartile range. Univariate analyses were performed using t-tests, analysis of variance, χ² tests, and the Kruskal–Wallis rank sum test, depending on the purpose of the analysis and the type of data. All statistical tests were two-tailed with statistical significance set at 5% (α = 0.05).

Multivariate logistic regressions were used to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for the association between WHR and cognitive impairment, adjusting for covariates. Cognitive impairment (yes/no) was the dependent variable, and WHR was the independent variable. The fully adjusted model included covariates selected based on previous evidence and between-group differences in univariate analyses, including sex, age, education level, smoking status, alcohol consumption, physical inactivity, hypertension, FBG, TG, LDL–C, HDL–C, and BMI. We additionally performed sensitivity analyses using sequential adjustment models. Model 1 adjusted for age, sex, education level, smoking status, alcohol consumption, and physical inactivity. Model 2 further adjusted for hypertension, FBG, TG, HDL-C, and LDL-C. Model 3 additionally adjusted for BMI. Multicollinearity among covariates was assessed using variance inflation factors. Multicollinearity was considered acceptable because all variance inflation factors were below 1.87. For stratified analyses by age and sex, stratification variables were not included as covariates in the corresponding subgroup models.

The key analytical steps were as follows: First, the association between WHR and cognitive impairment was examined in the total population using logistic regression. WHR was modeled using both restricted cubic splines (RCS) with three knots at the 25th, 50th, and 75th percentiles to assess potential nonlinear associations, and as a standardized continuous variable to estimate the linear association. In the RCS analyses, WHR was modeled on its original scale, with the median value (0.88) used as the reference point (OR = 1). For the linear analyses, WHR was standardized as a z score by subtracting the sample mean and dividing by the sample SD, and ORs were reported per 1-SD increase in WHR (SD = 0.06) to facilitate interpretation. Second, stratified analyses were performed by sex (females and males) and age group (40–59 vs. ≥60 years) to explore potential sex- and age-dependent patterns. Then, participants were divided into four subgroups: middle-aged females (females aged 40–59 years), older females (females aged ≥60 years), middle-aged males (males aged 40–59 years), and older males (males aged ≥60 years), and the same analyses were repeated within each stratum. Third, age-related effect modification was assessed separately in females and males by including WHR, age group (40–59 vs. ≥60 years), and their interaction term in logistic regression models. WHR was modeled both linearly per 1-SD increase, and using RCS. All models were adjusted for the same set of covariates as previously specified, and ORs with 95% confidence intervals were estimated across the WHR distribution.

A total of 1792 participants were included in the analysis, of whom 726 (40.5%) were male. Participants were aged between 40 and 85 years (mean 55.53 ± 9.92 years). The median WHR was 0.88 (0.84, 0.92). The median CMMS score was 27 (24, 29). In total, 230 participants (12.8%) were identified as having cognitive impairment. Participant characteristics according to cognitive impairment status are shown in Table 1.

WHR was positively associated with the odds of cognitive impairment in the total population (Table 2, OR = 1.245 per SD, 95% CI: 1.022–1.517, P = 0.029), while the test for nonlinearity showed borderline evidence (Figure 2A, P_overall = 0.019, P_nonlinear = 0.090).

Participants were stratified into female and male subgroups, and the association between WHR and cognitive impairment was examined separately in females and males. Among females, WHR was positively associated with the odds of cognitive impairment (Table 2, OR = 1.308 per SD, 95% CI: 1.016–1.684, P = 0.037). The spline curve was broadly similar to that observed in the total population, but evidence of nonlinearity was stronger among females (Figure 2B, P_overall = 0.014, P_nonlinear = 0.049). Specifically, using 0.88 as the reference value (OR = 1), the spline curve suggested relatively stable estimated odds below this value and increasing estimated odds at higher WHR levels. However, among males, WHR was not significantly associated with cognitive impairment (Table 2, OR = 1.149 per SD, 95% CI: 0.830–1.591, P = 0.402; Figure 2C, P_overall = 0.326, P_nonlinear = 0.227).

To examine the age-dependent association between WHR and cognitive impairment, the total population was divided into middle-aged (40–59 years) and older (≥60 years) groups. In the 40–59-year-old subgroup, WHR was positively associated with the odds of cognitive impairment (Table 2, OR = 1.406 per SD, 95% CI: 1.061–1.863, P = 0.018), with no statistically significant evidence of nonlinearity (Figure 2D, P_overall = 0.016, P_nonlinear = 0.143); however, in the ≥60-year-old subgroup, no significant association between WHR and cognitive impairment was observed (Table 2, OR = 1.090 per SD, 95% CI: 0.822–1.446, P = 0.550; Figure 2E, P_overall = 0.452, P_nonlinear = 0.272).

These findings suggest that the association between WHR and cognitive impairment may vary by age and sex. To further explore this interaction, the total population was stratified into four subgroups based on sex and age: females aged 40–59 years, females aged ≥60 years, males aged 40–59 years, and males aged ≥60 years.

Among females, the RCS curves for the two age groups ran nearly parallel, with no evidence of effect modification by age (RCS WHR × age group P_interaction = 0.927; Figure 3A). Consistent with this visual pattern, the linear interaction term was also non-significant (continuous WHR × age group P_interaction = 0.790; OR = 0.939 per SD, 95% CI: 0.590–1.493; Table 3), indicating that the association between WHR and cognitive impairment did not differ meaningfully between the middle-aged and older females.

Among males, the RCS curves for the two age groups diverged in opposite directions, suggesting potential age-dependent heterogeneity in the association between WHR and cognitive impairment. The RCS interaction showed borderline evidence of interaction (RCS WHR × age group P_interaction = 0.062; Figure 3B). Given the absence of strong evidence for nonlinearity in the overall WHR–cognitive impairment association among males, the pre-specified linear interaction model was regarded as the summary of age-related heterogeneity and demonstrated a significant WHR × age group interaction (continuous WHR × age group P_interaction = 0.010; OR = 0.462 per SD, 95% CI: 0.258–0.828; Table 3). Specifically, WHR was positively and significantly associated with the odds of cognitive impairment in males aged 40–59 years (OR = 1.652 per SD, 95% CI: 1.052–2.595, P = 0.029, Table 2), whereas no significant association was observed in males aged ≥60 years (OR = 0.804 per SD, 95% CI: 0.492–1.314, P = 0.384, Table 2). These findings suggest age-related heterogeneity in males, with the positive association observed between WHR and cognitive impairment in middle-aged but not older males.

Sensitivity analyses using sequential adjustment models are presented in Supplementary Table 1. In the total population, the OR per 1-SD increase in WHR increased modestly from 1.152 (95% CI: 0.963–1.378) in Model 1 to 1.245 (95% CI: 1.022–1.517) in Model 3. Among females and participants aged 40–59 years, the direction and magnitude of associations remained broadly similar across the sequential models, whereas no significant associations were observed among males or participants aged ≥60 years. In the sex-by-age subgroup analyses, the ORs among middle-aged males increased from 1.266 (95% CI: 0.861–1.862) in Model 1 to 1.652 (95% CI: 1.052–2.595) in Model 3, whereas the inverse association among older males was progressively attenuated (OR = 0.599, 95% CI: 0.389–0.922 in Model 1 vs. OR = 0.804, 95% CI: 0.492–1.314 in Model 3). The corresponding RCS analyses yielded similar results. Across Models 1–3, evidence of nonlinearity was limited to females (P_nonlinear = 0.047, 0.041, and 0.049), whereas no significant nonlinear associations were observed in the other subgroups.

In sequential interaction analyses (Supplementary Table 2), no significant WHR × age group interaction was observed among females across Models 1–3. In contrast, the linear WHR × age group interaction remained statistically significant among males after sequential adjustment (continuous WHR × age group P_interaction = 0.017, 0.014, and 0.010), with borderline RCS interaction results in Models 2–3 (RCS WHR × age group P_interaction = 0.082 and 0.062).