This study conducted a secondary analysis of deidentified, publicly available China Health and Retirement Longitudinal Study (CHARLS) data, exempt from additional ethics review. The original CHARLS was approved by Peking University’s Biomedical Ethics Review Committee (IRB00001052-11015), obtained written informed consent from all participants, and complied with the Declaration of Helsinki and China’s Personal Information Protection Law. All procedures adhere to CHARLS protocols.

Participants were enrolled in 3 follow-ups of the CHARLS from 2015 to 2020. It is a nationally representative sample of the middle-aged and older adult population (aged ≥45 years) in China, a community-based longitudinal study conducted by the China Center for Economic Research at Peking University [35]. The CHARLS used a multistage, probability proportional to size sampling technique to select participants. The primary sampling units were administrative villages in rural areas and urban neighborhoods. The CHARLS data can be accessed through its official website [36]. It is designed to better understand the socioeconomic determinants and consequences of aging.

Data were drawn from 3 waves (2015 [T1], 2018 [T2], and 2020 [T3]) of the CHARLS. The baseline cohort included 23,459 participants. Of these 23,459 participants, we excluded individuals aged <45 years or >100 years and those with household consumption per capita equal to 0 at baseline (n=10,026, 42.7%), yielding 13,433 (57.3%) eligible respondents. From this group, we further excluded those with missing data on broadband access, internet use, or cognitive function (362/13,433, 2.7%) and those with a total cognitive score of 0 (143/13,433, 1.1%). An additional 1.1% (153/13,433) of the participants were excluded due to missing values for key time-invariant covariates (employment status, self-rated health, or household consumption per capita) at baseline. Of the remaining 12,775 participants with at least one valid cognitive measure, 6712 (52.5%) were lost to follow-up across the 3 waves, resulting in a final analytical sample of 6063 (47.5%) participants with sufficient longitudinal data for modeling.

R (version 4.1.0; R Foundation for Statistical Computing) was used for the purposes of descriptive and correlational analyses. Descriptive statistics were used to summarize the characteristics of the participants, specifically frequencies with percentages for categorical variables and medians with IQRs for continuous variables. The lavaan package in R was used to perform structural equation modeling. PP-LGCM was implemented to investigate the cascading effects of digital divides on trajectories of cognitive function. For each domain, the baseline level (intercept) and the rate of change over time (slope) were modeled [39]. The cascade effect framework suggests that broadband access influences internet use, which subsequently affects the intercepts and slopes of cognitive function. Full information maximum likelihood was applied to address missing data by leveraging all available information from each participant. The following fit indexes were evaluated: the comparative fit index (CFI), Tucker-Lewis index (TLI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR). Conventional thresholds stipulate that CFI and TLI values above 0.90 are acceptable, with values exceeding 0.95 indicating a good fit. For RMSEA and SRMR, values below 0.05 indicate a good fit, whereas values between 0.05 and 0.08 are acceptable [40]. In the conditional latent growth curve modeling and PP-LGCM analyses, demographic variables were incorporated as covariates.

A total of 6063 Chinese middle-aged and older adults were included in this study (Figure 2). Table 1 shows significant sociodemographic disparities in digital access and use. Individuals with broadband access and internet use were significantly younger, more educated, and more likely to reside in urban areas; reported better health; engaged in social activities; and had a higher household consumption (P<.001 in all cases). Men were more likely to be internet users (P<.001) but not among broadband owners (P=.79). Marital status and employment status showed minimal or no association with internet use (P=.10 and P=.29, respectively), although both differed slightly by broadband subscription (P<.001). Those who had broadband access and internet use also exhibited significantly higher cognitive function scores (median 15.5, IQR 14.5-16.5 vs median 12.5, IQR 10-14.5; P<.001), although the difference for broadband access was less pronounced (median 14.0, IQR 12-15.5 vs median 12.5, IQR 9.5-14.5; P<.001).

The correlation matrix presented in Table 2 delineates the relationships among broadband access, internet use, and general cognition at 3 distinct time points. The findings demonstrate moderate stability within each construct over time, with autocorrelation coefficients ranging from 0.30 to 0.62. Conversely, cross-construct correlations were consistently weaker, typically below 0.40, thereby providing evidence of adequate discriminant validity among the constructs. This pattern of associations indicates that the 3 constructs may be validly treated as separate latent variables in longitudinal modeling, such as latent growth curve analyses.

The fit indexes presented in Table 3 strongly support the measurement models for each domain. Overall, these results indicate that the latent construct representations were adequate over time, confirming the validity of the developmental trajectories studied.

Table 4 shows that sociodemographic factors influenced initial levels and changes over time in broadband access, internet use, and cognitive function. Older age correlated with lower baseline levels (β=–0.327 to –0.144) and slower growth in internet use (β=–0.356) and general cognition (β=–0.151). Male individuals reported slightly less broadband access (β=–0.032) but higher baseline internet use and general cognition (β=0.107), with steeper declines over time (β=–0.152 and –0.160). Higher educational level boosted initial general cognition (β=0.518) but predicted slower growth in internet use (β=–0.332) and slightly faster cognitive gains (β=0.123). Rural residents started lower (eg, broadband; β=–0.220) but showed accelerated growth in broadband access (β=0.170) and internet use (β=0.202). Health status and social activity predicted higher levels of digital inclusion early on. For example, at T2, better health status was associated with greater broadband access (β=0.066), increased internet use (β=0.081), and higher general cognition scores (β=0.039). Social activity demonstrated a stronger link specifically with internet use (β=0.122). By T3, these effects were weakened. Employment status reduced broadband access (β=–0.062 to –0.041 across waves) but had minimal impact on internet use and a small positive association with general cognition (β=0.034 at T3).

The PP-LGCM model demonstrated an excellent fit (χ²₉₇=335.1; P<.001; CFI=0.976; TLI=0.991; RMSEA=0.020; SRMR=0.026). This finding revealed a significant mediating pathway connecting broadband access to cognitive function through internet use, as illustrated in Table 5 and Figure 3. Consistent with our hypothesized cascade model, baseline broadband access strongly predicted initial internet use (β=0.453; P<.001), and increases in broadband access over time forecasted steeper growth in internet use (β=0.625; P<.001). In turn, higher baseline internet use was associated with better initial cognition (β=0.470; P<.001), and greater increases in use were associated with slower cognitive decline (β=0.444; P=.006). Although older age, lower educational level, and rural residence were associated with lower initial digital engagement, rural participants exhibited significantly faster adoption rates (broadband slope: β=0.163; internet slope: β=0.102; P=.10), suggesting a partial narrowing of the use gap. Educational level boosted broadband access (β=0.238), internet use (β=0.337), and general cognition (β=0.299; P<.001 in all cases); and older age negatively affected broadband access (β=–0.324) and internet use (β=–0.167; P<.001 in both cases) and slowed internet use growth (β=–0.338; P<.001), with no significant impact on cognitive slope. Men reported higher baseline internet use (β=0.119; P<.001) and general cognition (β=0.071; P<.001), although they experienced steeper declines in internet use over time (β=–0.149; P=.02). Household consumption predicted higher baseline broadband access (β=0.203) and internet use (β=0.116; P<.001 in both cases) but was negatively related to their slopes. Social activity had a consistent association with internet use, with β coefficients of 0.117 at T1, 0.122 at T2, and 0.109 at T3. Furthermore, it positively predicted cognitive function (β=0.037 at T1, β=0.040 at T2, and β=0.035 at T3). Health status demonstrated a positive relationship with both internet use and cognition at T1 and T2; however, this association did not persist at T3. Employment displayed a negative correlation with broadband access across all waves (β=–0.062 at T1, β=–0.073 at T2, and β=–0.041 at T3), indicating lower adoption rates among employed individuals. Its influence on cognition remained negligible.

The findings of this study demonstrated that the cascading potential influence of household broadband access predicts internet use and subsequent codevelopment of this prediction, both of which are associated with a possible slower decline in cognitive trajectories among middle-aged and older Chinese adults. The PP-LGCM model confirmed the 3-level digital divide framework and extends the perspective from static views to a process-based approach, emphasizing how disparities in infrastructure could influence actual internet use and cognitive aging through codevelopmental pathways. Our findings substantially enhance the discussion of digital inclusion as a vital component of healthy aging, emphasizing the importance of not only physical access but also creating supportive environments that enable older adults to thrive.

This aligns with the first progress report of the United Nations Decade of Healthy Ageing, which underscores the vital role of digital engagement in sustaining intrinsic capacity in later life [26]. The report cautions that, despite investments in connectivity, the resources for engagement remain insufficient, especially among older adults in low- and middle-income countries. Van Dijk [41] similarly argues that access alone is insufficient; individuals require motivation, digital skills, and supportive opportunities to translate connectivity into benefit. Our results support this view: China’s infrastructure expansion from 2015 to 2020 narrowed the rural-urban gap in broadband access, yet this did not translate into comparable gains in internet use or cognitive outcomes among rural respondents. Early advantages among urban and highly educated individuals likely reflect the Matthew effect [42,43], in which initial resources accumulate over time. However, what appears to be a cumulative advantage may stem from persistent conversion gaps [44,45]: the failure to translate access into meaningful use, especially in mid to late life, when cognitive resilience is more vulnerable [46,47].

Moreover, the association between internet use and cognitive resilience is plausibly mediated by neurocognitive mechanisms aligned with the cognitive reserve hypothesis [48,49]. Internet use is positively associated with cognitive stimulation and cognitive reserve [50,51]. Activities such as seeking health information, comparing services, or navigating digital interfaces engage executive functions, working memory, and attention [52,53]. Social relationship factors are correlated with cognitive function and its decline, and engaging in social activities via calls or messaging may reduce social isolation, which is identified as a risk factor for cognitive decline [54,55]. The internet offers problem-solving opportunities through online tasks such as navigating websites, managing banking, or learning, which demand cognitive effort and decision-making, helping maintain mental agility [56] and thereby stimulating cognitive abilities [57]. Although our binary measure of internet use limits granularity, the findings are consistent with those of research linking online behaviors to delayed recall and processing speed [58], indicating that continuous engagement activates these neural pathways [59].

Structural inequalities shape both digital and cognitive trajectories [60]. Age is a predictor of disadvantage [23,43]: older adults tend to have less broadband access, lower internet use, and slower progress in these domains. Educational level functions as a lifelong enabler, increasing broadband access, internet use, and cognitive capacity [60]. Men in this study reported higher initial levels of internet use and general cognition. However, there is a significant decline in internet use over time, rather than in cognitive function, aligning with previous research [18,61], some of which is potentially attributable to cohort attrition in men. Household consumption predicted higher initial digital engagement but also correlated with slower growth, implying saturation among more advantaged individuals [22]. Self-rated health and social activity were associated with better outcomes in the initial wave but lost significance in later waves, potentially due to survivor bias or the stabilization of digital habits [32]. Employment status consistently showed a negative relationship with broadband access across waves, likely reflecting work-related constraints or measurement limitations.

While our PP-LGCM model supports a directional cascade from access to use to general cognition, we acknowledge the plausibility of reverse causation: individuals with higher baseline cognitive capacity may be more likely to adopt and sustain internet use [32]. Future studies using cross-lagged panel models or instrumental variables are recommended to further elucidate the causal direction.

This study has several limitations. First, attrition (6712/12,775, 52.5%) may introduce bias if those lost to follow-up were disproportionately cognitively impaired or digitally excluded. Second, our binary measures of internet use could not distinguish between passive (eg, scrolling) and cognitively stimulating activities (eg, learning and problem-solving)—a key nuance for future research. Third, although the Mini-Mental State Examination cognitive composite is widely used in surveys among large aging populations, it has limited ceiling effects in high-functioning individuals, sensitivity to early or specific decline, and possible cultural or educational biases in literacy or numeracy items [62,63].

This study provides longitudinal evidence that internet use mediates the association between broadband access and general cognition among middle-aged and older Chinese adults. The growth trajectory in broadband access predicted increased internet use, which in turn predicted slower cognitive decline. These results argue for integrating digital literacy and socially engaging online opportunities into public health strategies for healthy aging.