This paper uses the data from the China Health and Retirement Longitudinal Survey (CHARLS) 2018 for analysis. The CHARLS project is led by the Peking University, which aims to collect a large and representative sample of the population aged 45 years and above in China. Specifically, the CHARLS 2018 covers ~150 counties, 450 communities or villages in China, and consists of ~23,000 respondents in 28 provinces. The sample was selected following the probability proportional to the size sampling strategy. Due to their high-quality nature, these data have been widely used in scientific research and have generated many publications.

In the questionnaire, 10 questions from the CES-D (Center for Epidemiological Studies-Depression) are designed and drawn up for the depression test (Supplementary Table A1). For example, the respondent was asked if in the past week he/she was bothered, depressed, fearful, lonely, or always had trouble in mind, felt hard to do anything, could not sleep well or get going. The answer options for each question were “little or no,” “not much,” “sometimes or half the time,” and “most of the time,” which were assigned with values of 0, 1, 2, and 3, respectively. Out of the 10, there were two opposite questions, i.e., “I am full of hope for the future” and “I am very happy,” with the same answer options. We used the reverse scoring method for the answers to the two questions. The scores of the 10 questions were then summed to generate the CES-D score as the outcome variable. The respondent with a CES-D score ≥10 is often defined as having symptoms of depression (28). Our results show that the overall incidence of depression among the Chinese elderly was ~33.84% in 2018.

Besides the module of health status, the questionnaire also collected information about individual and family characteristics, e.g., age, gender, education, marriage status, family size, living habits, and daily activities. In particular, a question on learning about the respondent's behavior of internet use is designed. In that question, the respondent was asked if he/she had used the internet in the past month, with two answer options of “yes” or “no.” We used the information from this question to generate our explanatory variable. During the data cleaning process, we excluded those observations of persons whose age was below 45 years. We also removed those observations that had missing values in our selected variables. The final sample for analysis in this paper has 17,365 observations.

The major challenge in identifying the effect of internet use on the elderly's depression status pertains to the fact that the elderly's behavior of internet use could be endogenous, i.e., internet use and depression might be affected by many common factors. This endogenous behavior on the part of the elderly may result in the concern of selection bias, if internet users are compared to non-users directly (29). To address such concern of selection bias, we employed the propensity score matching (PSM) technique. The PSM technique eliminates the impact of covariates by resembling the randomized assignment to treatment, to create conditions of a random experiment (30). The PSM has been widely employed in empirical work (31–33). Specifically, the following probit model is estimated in the first stage:

where INTi* is a latent variable that indicates the utility of respondent i's choice of internet use. If the utility exceeds 0, we observe that the respondent chooses to use the internet (INT_i = 1); otherwise, the respondent does not use the internet (INT_i = 0). X_i is a set of exogenous variables, which affects the respondent's choice of internet use. Estimation from Equation (1) reports the determinants of the elderly's choice of internet use and allows us to predict the propensity score of each observation to use the internet.

The predicted propensity scores are then used to find one or more matching partners for internet users from non-users. We use the popular matching algorithms, i.e., nearest neighbor, radius, and kernel matching, to find matching partners (34). Taking advantage of the large sample, we select one, five, and ten matching partners in the nearest neighbor matching for robustness tests. A matching caliper of 0.001 is set to reduce potential matching bias. After matching, we compute the average treatment effect on the treated (ATT) of internet use on depression status following Equation (2):

where Y₁ and Y₀ are the elderly's depression statuses of the matched internet users and non-users, respectively.

The validity of PSM depends on three assumptions. First, before matching there must be a sufficient overlap of propensity scores between internet users and non-users. Second, after matching, the covariates must statistically have no difference between internet users and non-users. Third, in the selection function there is no omitted variable that is correlated with internet use and the elderly's depression status (35). In the Results section, we present supportive evidence for the satisfaction of these assumptions.

A major problem of the PSM is that it can only mitigate selection bias due to observables but not due to unobservables. We therefore employ the endogenous switching regression (ESR) model, which accounts for selection bias from both observables and unobservables, to complement the PSM. The ESR model has also been widely employed in empirical work (36–39). The typical ESR model has three equations. The first equation, as shown in Equation (1), determines two regimes that a respondent may fall into, namely, using the internet or not. The rest of the two equations explain the outcome variable, i.e., the depression status, under different regimes:

where Y_1i and Y_0i are the measures of a respondent's depression status, which are observed only under regimes one and two, respectively. Z_i contains all the variables in X_i and at least one instrumental variable. The error terms ε_i, η_1i, and η_0i presumptively follow a joint normal distribution. In the estimation process, two inverse Mills ratios λ_1i and λ_0i are predicted using Equation (1) for internet users and non-users, respectively. The outcome equations are then updated by including the inverse Mills ratios:

where δ₁ and δ₀ are the parameters to be estimated for the inverse Mills ratios. To estimate the selection and outcome equations simultaneously, a full information maximum-likelihood method should be employed (40). The estimated parameters in the outcome equations are then used to predict the expected depression status.

where E(Y_1i|INT_i = 1) is the expected depression status of the elderly who use internet and E(Y_0i|INT_i = 1) is the expected depression status of the elderly in the counterfactual scenario. The average treatment effect on the treated is then computed following Equation (6):

Heterogeneous effects of internet use on depression status are estimated by restricting Equation (6) to sub-group samples.

Table 1 presents the variable definitions. The outcome variable is the CES-D score, which is a reliable and widely used measure of the depression status in clinical practices and epidemiological studies (41, 42). The explanatory variable is internet use, defined as 1 if the respondent used the internet in the last month, and 0 otherwise.

The variables included in the selection function are important. Some scholars propose that the variables should be theoretically important for outcomes (43). Whereas some other scholars argue that they should be simultaneously associated with selection and outcomes (34). In practice, empirical studies using PSM may choose variables that are important for selection (44), outcomes (45), or both (30, 46). In this paper, we follow the literature and select a set of variables, e.g., age, gender, education, martial status, family size, and living habits, which are associated with internet use and mental health to better predict the choice of internet use and mental health (47, 48).

The selected instrumental variable is the ratio of internet use by other elderly in the same communities. We propose that the ratio of internet use by other elderly in the same communities should be correlated with a respondent's behavior of internet use due to the peer effect (49). Besides, internet use by other people is simply their personal behavior, which is unlikely to have an impact on a specific elderly's depression scores. Such choice of instrumental variable has been proved valid in many empirical studies (30, 50, 51).

Figure 1 shows the differences in depression scores between internet users and non-users. In general, the depression scores of the elderly who use the internet are smaller than those of the non-users. Table 2 further reports differences in the average depression scores between the elderly who are internet users and non-users. We find that the average CES-D score for the full sample is 8.709, while for the sub-samples of the internet users and non-users, the average CES-D scores are 6.374 and 9.061, respectively. That is, compared with the elderly who do not use the internet, those who use the internet score lower in the depression measure. The difference in depression score between the two groups is significant at 1%. These results tend to suggest to us that internet users have a lower risk of depression than non-users.

There are also differences in covariates between internet users and non-users. Generally, the young, male, highly educated, and married respondents are more likely to use the internet than their counterpartners. Respondents with an urban Hukou have a higher probability of using the internet than those respondents with a rural Hukou. Internet users often have a smaller family size and better eyesight than non-users. Besides, internet users tend to get retired, less likely to have agricultural work, and more likely to receive basic subsistence allowance from the government. Internet users also drink and smoke more often than non-users.

The above results give a preliminary indication about internet use to prove that it is highly associated with the elderly's depression status. However, since there are also significant differences in many covariates, we cannot confidently conclude that there is a causal effect of internet use on the incidence of depression in the elderly. It is therefore necessary to further address the concern of potential selection bias in the rest of the analysis.

Table 3 reports the determinants of the elderly's choice of internet use from Equation (1). The regression coefficients are reported in column (1). As shown in column (2), we also computed the average marginal effects of the variables. We find that age has a significant and negative impact on the probability of internet use. A 1-year increase in the respondent's age decreases the probability of using the internet by 0.8%. The probability of using the internet for men is 1.5% higher than that for women. The elderly who have an urban Hukou are more likely to use the internet. Compared with the elderly who have an education of primary school or below, those who have a higher level of education are more likely to use the internet, implying that education increases the likelihood of using the internet. Besides, the elderly person who has a larger family is less likely to use the internet, probably because internet use, as a substitution of social communication, is less necessary for the elderly in such large families.

Vision has a significant and positive impact on internet use. It is intuitive that good vision is a prerequisite for using the internet, especially for the elderly who often suffer from eye diseases in China. The elderly who are engaged in agricultural work are less likely to use the internet. Apart from the reason that access to the internet remains limited in a rural area, another reason could be attributed to the fact that farmers may have limited knowledge of internet and may evince less interest in internet use in China.

The retired respondents have a higher probability of using the internet than the respondents who have not retired from active service, probably because the retired have more time available at their disposal for leisure. The elderly who receive subsistence allowance from the government, arguably the poor people, are unlikely to use the internet. Indeed, internet use could cost a lot of money for the poor in China, which excludes the poor elderly from internet use. The elderly who drink alcohol are more likely to use the internet, probably because wine is a luxury good and drinking alcohol may also represent the income level.

Table 4 reports the effects of internet use on the elderly's depression status from the propensity score matching. Different matching algorithms generate similar results. Overall, the elderly who use the internet have a lower depression score than those who do not use the internet. The size of the effects ranges from −0.916 to −1.042, which accounts for a reduction effect of ~13% in the depression scores. The statistical significances are all set at the 1% level, regardless of the matching algorithms. These results support our expectation that internet use functions to reduce the incidence rate of depression in the elderly group in China.

Several tests are conducted to support the validity of our results. First, Supplementary Figure A1 shows that the overlap of propensity scores between internet users and non-users is sufficiently large, implying that most internet users can find a matching partner. Thus, our results can be generalized to the whole sample. Second, results from the balancing test (Supplementary Table A2) show that there is no significant difference in covariates between internet users and non-users after matching. Third, the sensitivity tests of the results report that the Rosenbaum bounds are all larger than 1.75, which implies that the significance of our results is not sensitive to omitted variables.

In this paper, we also test the heterogeneous effects across age. Specifically, we split the sample into two groups, namely the elderly aged 60 or below and the elderly aged over 60. We then estimate the impact of internet use on depression status using the PSM. Table 5 shows that, on average, the impact of internet use on depression status for the elderly aged 60 or below ranges from −0.647 to −0.790, while for the elderly aged over 60 it ranges from −1.264 to −1.987. That is, the reduction effect of internet use on the depression status is larger for the even older groups.

To test the heterogeneous effects across gender, we estimate the effects of internet use on the depression status for men and women separately. Table 6 reports the results. We find that the effect of internet use on depression score for the male elderly ranges from −0.684 to −0.851, using different matching algorithms, while the effect of internet use on depression score for the female elderly ranges from −0.977 to −1.028.

We also test the heterogeneous effects across occupation. We split the sample into two groups according to whether the elderly has any agricultural work or not. Table 7 reports the results. We find that the effects of internet use on depression status for the elderly with agricultural work range from −0.970 to −1.080. For the elderly who do not have any agricultural work, the effects range from −0.742 to −0.888.

Table 8 reports the results from the endogenous switching regression (ESR) model. The F-statistic from tests on the strength of the instrumental variable in the selection function is 1,335.285 (P-value = 0.000), which exceeds the critical value of 10. This result implies that the concern of a weak instrumental variable should not be a problem in our work. Table 8 shows that the ESR model reports negative but larger effects of internet use on the elderly's depression scores. All the effects are statistically significant at 1% level. These results highlight the importance of addressing selection bias from unobserved confounds. On average, internet use reduces the elderly's depression scores by 3.370 points, which accounts for an approximate reduction of 37.19% from the average depression scores of the non-users.

Table 8 also shows that the general pattern derived from PSM still holds. That is, the effect of internet use on the depression status is larger for the older, female elderly, or the elderly who have agricultural work. Specifically, the effects of internet use on the depression status for the older and younger elderly are −3.772 and −3.251, respectively. The effects of internet use on the depression status for the male and female elderly are −3.145 and −3.648, respectively. While the effect on the elderly who have agricultural work is −3.627, for the elderly who have no agricultural work, it is −3.252. The T-statistics from tests over differences in the effects between different ages, genders, and employment status are −11.12, −12.85, and 8.74, respectively. The corresponding P-values are all smaller than 1%, which supports the heterogeneous effects of internet use.