Data source and variable selection

This study utilizes data from the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative survey that collects high-quality microdata on middle-aged and older adults in China. CHARLS employs a rigorous probability proportional to size (PPS) sampling method, with waves conducted in 2013, 2015 and 2018, covering 19,000 respondents across 150 counties and 450 communities. This dataset is widely recognized for its reliability and comprehensive coverage, making it ideal for studying ageing and cognitive health. We include both middle-aged and older adults in our sample to capture early cognitive changes starting from midlife and to study how internet use across this age range may impact cognitive function. This broader age coverage allows for more robust analysis of heterogeneous effects and better statistical power, while reflecting the gradual nature of cognitive decline.

The dependent variable, cognitive ability, includes five components: immediate memory, delayed memory, temporal orientation, figure memory and calculation ability. Each component is assessed through standardized tasks, with a total possible score of 33, capturing a broad spectrum of cognitive function.Footnote ¹

The core explanatory variable is internet use, measured as a binary variable (1 for users, 0 for non-users). Control variables include demographic and socio-economic factors such as gender, age, urban/rural status, education, income, health status and family structure (Cai Reference Cai2022). Table 1 presents descriptive statistics.

Table 1.

Descriptive statistics of the main variables

Note: Due to varying degrees of missing data among different variables, there are discrepancies in the sample sizes for each variable.

Identification strategy: generalized random forests

Causal inference is one of the core issues in econometrics. Traditional econometric models, based on limited data and parametric settings, primarily focus on average treatment effects and do not adequately consider individual heterogeneity in treatment effects. In recent years, with the advent of big data, machine learning methods, which boast powerful generalization abilities and high explanatory power for real-world phenomena, have inspired an increasing number of social science researchers to optimize and upgrade algorithms for solving causal inference problems. Among these, the method of GRFs, further developed by Wager and Athey (Reference Wager and Athey2018) from foundational causal trees (Athey and Imbens Reference Athey and Imbens2016), allows for effective estimation of heterogeneous treatment effects through machine learning techniques, thereby providing more credible causal inferences. This represents a significant breakthrough in the explanatory power of supervised machine learning (such as random forests) in economics (Athey and Imbens Reference Athey and Imbens2016; Knittel and Stolper Reference Knittel and Stolper2019).

Using GRFs enhances causal inference by leveraging a data-driven, non-parametric approach that mitigates the ‘curse of dimensionality’ and provides robust counterfactual simulations. Unlike traditional random forests, GRFs estimate treatment effects at an individual level, ensuring more granular insights. The key advantage lies in their ability to account for complex interactions between covariates, enabling a nuanced understanding of how treatment effects vary across different subpopulations. This flexibility makes GRFs particularly well-suited for studies with high-dimensional datasets, where traditional econometric techniques might struggle to deliver accurate causal estimates.

In this study, GRFs facilitate a detailed exploration of the impact of internet use on cognitive ability, allowing us to uncover nuanced heterogeneity across demographic groups. Moreover, they provide a robust comparison between machine learning–based methods and traditional econometric approaches, reinforcing the validity of our findings. For technical details, including model construction and estimation algorithms, readers are referred to Appendix 1, which offers a comprehensive discussion of the GRF methodology and its implementation.

Regression analysis results

The fixed effects model estimated in this article is a two-way fixed effects model that incorporates individual and year dummy variables:

(1)\begin{equation}{\text{Cognitio}}{{\text{n}}_{it}} = {\beta _0} + {\beta _1}{\text{Interne}}{{\text{t}}_{it}} + {\beta _2}{X_{it}} + {u_i} + {\lambda _t} + {\varepsilon _{it}}\end{equation}

where subscript $i$ is for individual $i$, $t$ for year $t,{u_i}$ is individual fixed-effect, ${\lambda _t}$ is for year fixed-effect, ${\varepsilon _{it}}$ is random perturbation. ${\text{Cognitio}}{{\text{n}}_{it}}$ is for cognitive ability of individual $i$ at year $t$, ${\text{Interne}}{{\text{t}}_{it}}$ is whether individual $i$ uses the internet at year $t$, ${X_{it}}$ are control variables in Table 1.

The fixed effects model includes individual and year dummy variables, forming a two-way fixed effects framework. Column (1) of Table 2 presents OLS baseline results, while Column (2) adds fixed effects, showing that internet use positively impacts cognitive abilities, significant at the 0.01 level. However, potential endogeneity arises from self-selection (e.g. higher-educated individuals are more likely to use the internet), reverse causality (cognitively stronger individuals may adopt internet use more readily) and omitted variables. To address these, the study combines propensity score matching (PSM) with fixed effects. Using a Logit model, individuals are matched within common support, and unmatched samples are excluded. Column (3) of Table 2 shows that the positive effect of internet use remains robust, though balance tests suggest residual endogeneity related to education. Instrumental variable estimation further mitigates endogeneity. Community internet usage and broadband installation rates serve as instruments, capturing infrastructure and network effects while maintaining exogeneity. Column (4) of Table 2 confirms the robustness of results, with a strong first-stage F-value (123.19) and no over-identification issues (Sargan p-value = 0.972). Internet use consistently enhances cognitive abilities, even after addressing potential biases.

Table 2.

Regression analysis results

Notes: All standard errors used in this article are robust.

*** Further, ** and *, respectively, indicate significance at the 0.01, 0.05 and 0.10 levels, with t-statistics in parentheses, unless otherwise specified.

Causal identification based on machine learning

This study combines both traditional econometric methods (OLS, fixed effects and IV) with modern machine learning techniques (GRF) to strengthen the credibility of causal inference. Instead of picking one baseline model and using the others as robustness checks, we take a comparative approach to show the strengths and weaknesses of each method. The IV method identifies local average treatment effects (LATE) under the assumption of valid instruments, while the GRF method provides individual-level treatment effect estimates and captures effect heterogeneity. We give equal analytical attention to both methods: the IV method offers strong internal validity and interpretability, while the GRF method adds more flexibility and external validity to the analysis.

Traditional regression models, including OLS, fixed effects and IV approaches, show that internet use significantly enhances cognitive abilities in older adults. However, OLS and fixed effects fail to address endogeneity, particularly reverse causality, while IV methods rely on strong, unverifiable assumptions of exogeneity. To overcome these limitations, this study employs GRFs, a machine learning method that provides robust causal inference by estimating heterogeneous treatment effects (Athey and Imbens Reference Athey and Imbens2016; Knittel and Stolper Reference Knittel and Stolper2019). Unlike traditional models, GRFs diminish the emphasis on estimating equation coefficients, focusing more on accurate predictions of the dependent variable. This allows for more precise counterfactual predictions using existing data, and, thus, causal effects can be computed by comparing actual observations to counterfactual predictions (Ejlskov et al. Reference Ejlskov, Wulff, Bøggild, Kuh and Stafford2018; Li et al. Reference Li, Luo, Reangsing and Schneider2022; Lin et al. Reference Lin, Wu, He and Fang2023).

Counterfactual simulation

This study employs random forests to model counterfactual scenarios based on respondent characteristics, providing a nuanced comparison of cognitive abilities under two conditions: internet use and non-use. Figure 1 illustrates these simulations, highlighting a marked cognitive decline in older adults, particularly beyond age 60. The left side of the figure demonstrates that non-users would experience significant cognitive gains if they adopted internet use, while the right side shows a substantial decline in cognitive abilities for current users if they ceased internet use. These patterns underscore the positive impact of internet use on mitigating age-related cognitive decline. The comparison between users and non-users also reveals that internet users consistently exhibit higher cognitive abilities and slower decline rates. This suggests inherent differences in cognitive capacity, further affirming the role of reverse causality and other endogeneity issues.

Figure 1.

Counterfactual simulation.

Note: The accuracy of the counterfactual predictions for older adults who use the internet may be affected due to the smaller sample size.

Table 3 quantifies these findings, indicating that internet use boosts cognitive abilities by approximately 16 per cent compared to non-use. Conversely, non-users experience a 24 per cent deficit relative to their potential cognitive capacity if they were to adopt internet use. These results underscore the significant role of internet use as a protective factor against cognitive decline. To ensure robust model performance, the study employs tenfold cross-validation, a rigorous method that divides the sample into subsamples for iterative validation. This process minimizes prediction error and enhances the accuracy of counterfactual estimates, with consistently reliable results across multiple iterations.

Table 3.

Average treatment effect estimates based on counterfactual predictions

Generalized random forest results

In constructing the GRF model, this study begins by incorporating respondent characteristics to identify key predictors, reporting their variable importance. Variables with importance below the average are excluded, and the model is recalibrated. This process enhances the accuracy of the estimates by removing predictors with weak explanatory power. Additionally, as random forests are ensemble methods, the number of decision trees plays a crucial role in model performance. Too few trees can compromise accuracy and introduce bias (Athey et al. Reference Athey, Tibshirani and Wager2019). To address this, the study tests different tree counts and employs the clustering method proposed by Athey and Wager (Reference Athey and Wager2019) to account for heterogeneity across groups.

Table 4 shows that with the increase in the number of decision trees, the average treatment effect remains around 3.7 with little change in t-values, indicating that the selected number of decision trees meets accuracy requirements. Column (4) presents results after applying the clustering method (Athey and Wager Reference Athey and Wager2019), showing that internet use significantly improves cognitive scores by 3.726 points. These GRF estimates exceed those from the IV model due to the LATE nature of IV, which offers strong internal validity but limited external validity, potentially underestimating the broader impact of internet use. In contrast, GRF captures a more comprehensive effect, aligning better with the study’s context. Furthermore, GRF parameter estimates yield significantly higher t-values compared to fixed effects and IV models, indicating greater precision. Overall, GRF provides more robust and reliable estimates than traditional regression approaches.

Table 4.

Generalized random forest results

* P < 0.1, **P < 0.05, ***P < 0.01.

Heterogeneity analysis

Heterogeneity test

Traditional linear regression, based on limited data and parametric settings, typically assumes that the average treatment effect is constant, which does not sufficiently consider the heterogeneity of individual treatment effects. The GRF model, however, allows for estimating the average treatment effect for individual subjects and uses clustering to obtain the average effect estimate for the entire sample. Figure 2 displays the distribution of the estimated average treatment effects of internet use on cognitive abilities. It can be observed that most sample average treatment effects are concentrated around 2.7, but the overall range of estimates varies from 0.6 to 5.0, which visually suggests that the impact of internet use on cognitive abilities exhibits significant heterogeneity across different individuals.

Figure 2.

Distribution of average treatment effect estimates for internet use.

Further, this article employs the ‘optimal linear prediction method’ proposed by Chernozhukov et al. (Reference Chernozhukov, Demirer, Duflo and Fernandez-Val2018) to test for heterogeneity. The results in Table 5 show that, regardless of whether clustering is applied, the coefficient of ${D_i}$ is significantly positive, indicating significant differences in the impact of internet use on the cognitive abilities of older adults.

Table 5.

Heterogeneity test

* P < 0.1, **P < 0.05, ***P < 0.01.

Heterogeneity analysis

In contrast to traditional linear regression, which analyses heterogeneity through grouping or interaction terms, the GRF model can directly estimate the average treatment effects for different sample groups, facilitating more precise heterogeneity analysis. This article specifically selects variables of interest such as urban/rural residence, education level, gender and age for heterogeneity estimation, with the results reported in Figure 3.

Figure 3.

Heterogeneity test.

Note: The grey dotted line is the baseline result value of 3.726 for the conditional average treatment effect estimate of the generalized random forest model.

The analysis reveals that internet use provides greater cognitive benefits to rural older adults compared to their urban counterparts (Panel A). This disparity likely stems from the richer social environments and better infrastructure in urban areas, which already support cognitive engagement. In contrast, the internet offers rural older adults a vital channel for mental stimulation, significantly enhancing their cognitive outcomes. Panel B indicates that the cognitive gains from internet use are more pronounced among individuals with lower education levels. Those with higher education already possess greater cognitive reserves, limiting the incremental benefits. Conversely, less-educated individuals experience substantial cognitive improvements, as the internet serves as a key source of cognitive stimulation. Panel C highlights minimal gender differences, with slightly stronger cognitive gains observed in women, suggesting broadly similar benefits across genders. Finally, Panel D shows that cognitive improvements are most significant when internet use begins in middle age (45–60). This may be due to a greater capacity for learning and higher internet usage frequency in this group, as older individuals (61–70) generally exhibit reduced adaptability to new technologies, resulting in smaller cognitive gains.

Mechanism analysis

The empirical research confirms the positive impact of internet use on cognitive abilities among older Chinese adults. Based on estimates from the GRF model, internet use has been shown to improve older adults’ cognitive abilities by an average of 28 per cent. How is this improvement achieved? The article next discusses the potential underlying mechanisms suggested in the theoretical hypotheses section – social interaction, stress levels and health behaviours – using data. For social interaction, the study considers involvement in social activities (total involvement in activities such as playing mahjong or cards, dancing or exercising, participating in club activities, volunteering and attending classes or training over the past month, with each activity scored 1 point, total score = 5) and social engagement (whether they visited friends or socialized in the past month, scored as 1). For stress levels, the article uses depression (based on related depression scale questions, total score = 24) and life satisfaction (total score = 5) as characteristic variables. For health behaviours, current smoking (yes = 1) and drinking habits (frequency of drinking in the past year, 1 = does not drink, 2 = less than once a month, 3 = more than once a month) are used as proxy variables. The article first verifies the impact of mechanism variables on cognitive abilities using linear regression, then tests whether internet use affects these mechanism variables using the GRF model. This approach effectively combines the advantages of traditional linear econometric methods and machine learning, offering better interpretative power for mechanism analysis.

Panel A of Table 6 shows that all six variables significantly affect cognitive abilities, highlighting the roles of increased social interaction, reduced stress and healthier behaviours. Panel B further demonstrates that internet use significantly boosts social activities and engagement, aligning with prior findings (Cotten et al. Reference Cotten, Anderson and McCullough2013; Ejlskov et al. Reference Ejlskov, Wulff, Bøggild, Kuh and Stafford2018). By connecting older adults to broader social networks, the internet mitigates isolation and fosters cognitive engagement (Cai Reference Cai2022). Additionally, internet use reduces depression and enhances life satisfaction, supporting its role in stress reduction (Nilsen et al. Reference Nilsen, Andel, Fors, Meinow, Darin Mattsson and Kåreholt2014; Turner et al. Reference Turner, James, Capuano, Aggarwal and Barnes2017; Lu and Kandilov Reference Lu and Kandilov2021). This stress relief complements social interaction, collectively preserving cognitive health as suggested by the ‘use it or lose it’ hypothesis. However, internet use shows no significant impact on smoking and drinking habits, suggesting that health behaviours play a limited role in cognitive improvements for older adults through digital engagement. This finding aligns with mixed evidence in the literature (Maher et al. Reference Maher, Lewis, Ferrar, Marshall, De Bourdeaudhuij and Vandelanotte2014), that is, the health behaviour mechanism may not play a role or may play a small role in the cognitive improvement of middle-aged and older adults caused by internet use.

Table 6.

Mechanism test

* P < 0.1, **P < 0.05, ***P < 0.01.

Further analysis

To address the limitations of using a binary indicator for internet usage, we further explore the effects on cognitive ability of internet usage frequency (internet frequency) and duration (internet time). Table 7 shows a significant positive relationship between internet usage frequency and cognitive ability (Column 1), as well as between the amount of time spent online and improved cognitive ability (Column 2). These findings suggest that both frequent and prolonged internet usage provide valuable information and stimulation, which can enhance cognitive ability.

Table 7.

Further analysis

* P < 0.1, **P < 0.05, ***P < 0.01.

With the rapid expansion of mobile internet, the penetration of mobile internet tools has continued to rise. Building on survey items related to mobile internet usage, we extend our analysis to examine its potential impact on cognitive ability, particularly considering how smartphones and social media platforms may influence cognitive function among older adults. In Columns (3) to (6), we include indicators for cellphone use, WeChat use, WeChat usage frequency and broadband access. The results show that cellphone use is significantly and positively associated with cognitive ability (Column 3), as is WeChat use (Column 4). Even more, frequent WeChat use (Column 5) is associated with better cognitive performance. These findings suggest that mobile internet tools – especially social platforms like WeChat – may enhance cognitive ability by facilitating information exchange, social interaction and cognitive stimulation. In contrast, broadband access alone does not show a significant effect (Column 6), indicating that it is not merely the availability of digital infrastructure but the active engagement with mobile internet tools that matters for cognitive outcomes.