We collaborated with university teachers in several Chinese cities to recruit community-dwelling adults aged ≥40 years. Participants were initially invited through local networks and then asked to share the survey link within their social circles (snowball sampling). Convenience sampling was chosen to enable rapid data collection under limited resources.

The online questionnaire was administered between July and September 2023. Before entering the survey, participants read an information sheet and provided electronic informed consent. To ensure data quality, each device/IP address could submit only one completed questionnaire. Completion times were monitored, and records showing automated response patterns were removed. All data were processed confidentially and in line with relevant ethical and data protection standards.

Inclusion criteria were: (1) age ≥40 years (based on commonly used thresholds for middle-aged and older adults in China (19–21)); (2) voluntary participation and online consent; and (3) permanent residence in communities within official administrative boundaries. Exclusion criteria were: (1) completion time <270 s (determined via pilot testing as the minimum reasonable duration); and (2) duplicate or obviously invalid responses. After applying these criteria, 508 participants were retained for analysis.

Guided by socio-ecological frameworks and previous work (22–24), we grouped independent variables into four domains: individual characteristics, behaviors, interpersonal relationships, and community/socioeconomic factors.

Because we had many independent variables relative to the sample size, we first used LASSO regression to identify variables showing robust associations with each outcome (32–34). LASSO adds an L1 penalty term λ to the loss function, shrinking some coefficients to zero and yielding a parsimonious set of independent variables, which is generally more robust than stepwise selection (35). Continuous outcome data were modeled using LASSO-regularized linear regression under a Gaussian family (family = “gaussian”). Unordered categorical variables (e.g., marital status, primary economic source, and employment status) were coded as factors. Before applying LASSO, we calculated variance inflation factors (VIFs) to assess multicollinearity, using VIF > 5 as a threshold (36). According to this criterion, no evidence of multicollinearity was observed (Supplementary Table S1).

Among the 508 participants aged ≥40 years, 60.8% were men and 39.2% women (Table 1). Most were between 40 and 54 years old (70.6%), lived in urban areas (94.7%), and had at least a bachelor’s degree (85.3%). Regarding employment, 74.2% were full-time workers and 19.5% were retired; only 1.6% were farmers. The majority were married (92.3%) and not living alone (94.9%). Salaries were the main income source for 80.1% of participants, and 59.1% had monthly incomes between 5,001 and 12,000 RMB. With respect to behaviors, 27.0% smoked, 46.5% drank alcohol, 63.2% reported regular exercise, and 78.3% underwent regular health check-ups. Sedentary time was most commonly 4–<6 h/day (40.4%). Sleep duration typically ranged from 6–<7 h/day (45.3%) or ≥7 h/day (36.4%). A total of 29.5% reported at least one chronic disease. The mean (SD) BMI was 23.94 (2.81). Mean (SD) physical activity and health literacy scores were 18.45 (14.50) and 35.86 (7.93), respectively. Please note that because only one participant in the sample reported “student” as their employment status, this category was combined with the similarly small “farmer” group (n = 8). The merged category was subsequently labeled as “Other” (n = 9) in all subsequent analyses.

Using LASSO regression with 10-fold cross-validation, correlates of physical activity (PA) and health literacy (HL) were identified separately (Table 2; Figure 1). For PA, the optimal penalization parameters were λ_min = 0.3140 and λ_1se = 1.3913, whereas for HL they were λ_min = 0.2890 and λ_1se = 1.1669. Results below focus on the more conservative λ_1se solutions. For physical activity, six correlates were observed, including regular exercise and sleep duration, sex and daily sedentary time, as well as two employment-status contrasts with opposite directions. For health literacy, only three correlates were observed: monthly income, regular exercise, and other employment status (full-time employment as the reference). Please note that “other employment status” does not represent a clearly defined category and therefore does not allow for meaningful interpretation. Therefore, this result is not reported or discussed further.

For physical activity, the λ_1se model achieved CV-MSE = 129.92 and CV-RMSE = 11.40, corresponding to a cross-validated R² of 0.383. For health literacy, the corresponding performance was more modest (CV-MSE = 59.64; CV-RMSE = 7.72; CV-R² = 0.053). Across both outcomes, λ_min solutions yielded slightly improved predictive performance but at the cost of reduced parsimony (Supplementary Table S2).

For physical activity, four correlates consistently met predefined stability criteria (selection frequency ≥ 0.70 and sign consistency ≥ 0.90): regular exercise, sleep duration, sex, and daily sedentary time. For health literacy, three correlates met the same criteria: monthly income, regular exercise, and other employment status (with full-time employment as the reference) (Table 3). Repeated CV results closely mirrored bootstrap findings for both outcomes (Supplementary Table S3).

For physical activity, five correlates were retained across all selection procedures: sleep duration (positive), routine health check-ups (positive), retired (with full-time employment as the reference) (positive), sex (negative), and daily sedentary time (negative). These variables were retained at λ_1se and showed high selection frequencies with consistent directions in both bootstrap resampling and repeated cross-validation (Supplementary Tables S4–S7). For health literacy, monthly income (positive) and other employment status (full-time employment as the reference) (negative) were the only correlates that consistently passed all procedures (Supplementary Tables S8–S11).

Across both outcomes, LASSO with selection tests identified a small set of associated factors. For physical activity (PA), regular exercise and sleep duration were positively associated with PA, whereas sex (female vs. male) and daily sedentary time were negatively associated. For health literacy (HL), monthly income and regular exercise showed positive associations.

While the LASSO-selected correlates of physical activity showed moderate cross-validated explanatory power, the corresponding cross-validated R² values for health literacy were small. The results suggest that, while several sociodemographic and behavioral correlates were robustly identified, a large proportion of variance in health literacy remains unexplained by the current set of correlates. This may be because the conceptualization of health literacy is a multifaceted construct shaped by long-term educational, cognitive, and contextual factors that are not fully captured by cross-sectional sociodemographic measures. In this context, the present findings are best interpreted as identifying correlates rather than providing a highly predictive model of health literacy.

Our finding that men have higher physical activity levels than women aligns with studies showing sex differences across multiple domains of physical activity (45) and lower prevalence of physical inactivity in retired Chinese men compared with women (17). These patterns likely reflect both biological factors and gendered divisions of labor and caregiving.

Regular exercise emerged as a correlate of overall physical activity in this study. In theory, exercise is only a subset of physical activity (46). Importantly, regular exercise was operationalized as a behavioral pattern characterized by temporal regularity—namely, the habitual engagement in planned or structured exercise over time—typically regarded as the opposite of irregular exercise (47). In contrast, overall physical activity captured a broader construct reflecting cumulative activity across daily life, including occupational, transport-related, and leisure-time domains (48). In our questionnaire, physical activity was assessed across three dimensions—frequency, typical duration, and intensity of activities of various types. In contrast, regular exercise specifically referred to whether individuals engaged in deliberate exercise according to a fixed schedule or at regular time points. The observed phenomenon may be attributable to that individuals who maintain regular exercise habits may develop stronger self-regulatory capacities, such as routine formation, planning, and behavioral consistency, which facilitate higher levels of physical activity beyond formal exercise sessions and across multiple contexts.

Our results also support prior evidence linking sleep and physical activity. Longer sleep duration has been associated with sufficient physical activity among older adults in institutional settings (49), although causal pathways remain unclear. The associations between higher physical activity, and less sedentary time in our study are consistent with the idea that these behaviors cluster and compete for limited daily time.

With respect to health literacy, our findings are broadly consistent with studies showing strong “social gradient” effects: higher socioeconomic status, including education status and income, is associated with better health literacy or health information literacy (50–52). We also observed robust associations between personal income and health literacy. Differences in conceptualization (health literacy vs. health information literacy), measurement (continuous vs. categorical outcomes), and sample characteristics (hospitalized vs. community-dwelling, older vs. middle-aged and older adults) may explain why education did not emerge as a major associated factor in our models.

Finally, it should be noted that in the sensitivity analyses excluding regular exercise, being retired (with full-time employment as the reference category) and regular health check-ups were identified as associated factors. Retirement is often accompanied by greater discretionary time and fewer occupational constraints, which may facilitate engagement in various forms of non-exercise physical activity, such as walking, household tasks, and leisure-time movement. In contrast, full-time employment may involve prolonged sedentary periods and limit opportunities for spontaneous or accumulated physical activity throughout the day. In addition, regular health check-ups may be associated with higher physical activity levels, as individuals who engage in preventive health services are generally more health-conscious and proactive in managing their health, which may extend to maintaining more active lifestyles.

Several limitations should be considered. First, our convenience online snowball sampling likely selected individuals who are urban, digitally connected, and relatively well-educated, which limits generalizability to the broader population of middle-aged and older Chinese adults. Estimates for these subgroups are therefore likely unstable and should be interpreted with caution. Future research should use probability-based or stratified sampling, or deliberately oversample rural residents and farmers, to validate whether the observed socioeconomic and occupational patterns hold in more representative samples. Second, the cross-sectional design and reliance on self-reported measures for health literacy, physical activity, and behaviors may introduce recall and social desirability biases and do not permit causal inference. Third, we used short-form instruments, which cannot differentiate specific domains of physical activity or sub-dimensions of health literacy in detail. Finally, although we used LASSO to handle many variables with a modest sample size, results may still be sensitive to model specification and unmeasured confounding. Our findings should therefore be interpreted as preliminary and hypothesis-generating.