The ELSA is a nationally representative sample of the English population aged 50 years and older. Established in 2002, the sample has been followed up every 2 years [29]. Data were collected through computer-assisted personal interviews, self-completion questionnaires, and nurse assessments (every 4 years) [29]. The original sample was selected from the Health Survey for England (1998, 1999, and 2001) [30]. This analysis used data from waves 6 (2012-2013), 7 (2014-2015), and 8 (2016-2017). The analytical sample (n=9169 persons aged 50 years and older at wave 6, with 27,507 person-years of follow-up between waves 6 and 8) comprised sample members first recruited at waves 1 (n=5659), 3 (n=888), 4 (n=1796), and 6 (n=826). ELSA participants provided signed consent, and the London Multicentre Research Ethics Committee (MREC 01/2/91) granted ethical approval for the study.

As an aspect of the affective dimension of hedonic well-being, mental health was measured using the validated 8-item version of the Center for Epidemiologic Studies Depression Scale (CES-D) [31]. Items were answered using a dichotomous response (yes or no), which were summed to result in a total score ranging from 0 to 8, where a score of 0 indicates no depressive symptoms. As an aspect of the cognitive dimension of hedonic well-being, life satisfaction was measured using the Satisfaction with Life Scale [32,33], which has been explicitly tested on older persons and consists of 5 items. The individual item responses ranged from strongly agree to strongly disagree on a 7-point Likert scale, which was summed to result in a total score ranging from 5 (lowest satisfaction) to 35 (highest satisfaction).

Data on the frequency of internet use were collected through self-completion questionnaires. The responses were recoded as daily, weekly, monthly or less, or never.

The purposes of internet use were assessed only for participants who reported using the internet at least once every 3 months. For this study, 12 purposes of internet use at wave 6 were grouped into 6 as follows: (1) communication (sending or receiving emails; use of social networking sites; and creating, uploading, or sharing content), (2) entertainment (news, newspaper or blog websites, streaming or downloading live or on demand television or radio, music, electronic books, and games), (3) information access (finding information about goods and services; searching for information for learning, research, and fact finding; and looking for a job or sending a job application), (4) managing finances, (5) electronic commerce (shopping and buying goods or services over the internet or selling goods or services over the internet), and (6) other. The questions on the purposes of internet use changed in waves 6, 7, and 8; therefore, we could not use it as a time-varying variable in the modeling. Our analyses focused on the 5 categories mentioned earlier.

Participants’ age was measured in a single year (coded to 99 to indicate those aged 91 years or older, whose individual age was not released). We tested nonlinear age and mental health associations using second-order polynomial terms. Couple status was measured according to whether a participant was cohabiting (married or unmarried) or not. Working status (currently working) was derived from economic activity status by dichotomizing those who were employees or self-employed from those who were seeking work, sick and not seeking work, retired, or unoccupied. Health status (presence of a limiting long-term illness) was measured through whether a participant self-reported an illness or disability that limited their activities in any way.

SEP at wave 6 was measured using 3 different domains [34]: occupational class, wealth, and education. On the basis of degree of control and autonomy, occupational class (based on the household reference person’s last occupation) was measured using the 3-category version of the National Statistics Socio-Economic Classification: managerial and professional, intermediate, and routine and manual occupations [35]. Wealth was measured using quintiles of total net nonpension household wealth, a summary measure of the value of financial, physical, and housing wealth minus any debt [35]. It represents the permanent economic status of older adults [36], as most participants in aging cohort studies are retired. The highest educational qualification was grouped into 3 categories: degree level, below degree level, and no qualifications. Education among the ELSA cohort members is generally determined earlier in the life course, whereas wealth and occupational class are more likely to change during mid to later life.

To examine the longitudinal association between internet use and mental health, we used a series of fixed and random effects linear models. A nonlinear model for depression did not satisfy the condition of the fixed effects model, and the results from the random effects logistic regression model were substantively similar (not shown here). Therefore, the CES-D score was treated as a continuous dependent variable in both the fixed and random effects models. The fixed effects models control for all measured and unmeasured time-invariant characteristics by using each participant as their own control over survey waves. We adjusted for time-varying age, age squared, working status, couple status, and health status to address our first hypothesis.

H1: More frequent internet use predicts better mental health longitudinally.

Internet use was entered into the models as a four-category independent variable, with daily use as the reference. Random effects models do not control for the correlation between measured and unmeasured characteristics and, therefore, analyze both within- and between-participant variance. We adjusted for the same time-varying variables as for the fixed effects model and additionally adjusted for time-invariant SEP and sex to address our second hypothesis.

H2: SEP moderates the association between internet use and mental health.

An assessment of moderation was performed by including an interaction term between the frequency of internet usage and each marker of SEP; these were retained if statistically significant (likelihood ratio test, P<.10). Separate random effect models tested the association between the purpose of internet use and mental health to address our final hypothesis.

H3: There is a positive association between communication internet use and mental health and a negative association between information internet use and mental health.

The analysis not presented here showed similar results when limiting the analytical sample to those who used the internet compared with the results for all participants presented here.

To address attrition from wave 6 onward and item nonresponse within a survey wave, we fitted multiple imputation models by chained equations on all variables (outcomes, exposures, and covariates) with missing values. The purpose of internet use was imputed conditionally on whether a participant reported internet usage at wave 6. All statistical analyses were conducted using cross-sectional survey weights to minimize bias arising from sample attrition from the beginning of the ELSA study to participation in wave 6. The use of imputation and survey weights did not substantively alter the main findings.

Table 1 presents the means, SDs, or percentages across person-waves for the key variables by frequency of internet usage. Overall, 56% of older adults used the internet daily, 12% weekly, 6% monthly or less, and 26% had never used the internet. The most common purpose was information access (68%), followed by communication (66%), electronic commerce (52%), entertainment (45%), and managing finances (39%). Those who used the internet more frequently had better life satisfaction scores; lower depression scores; and were more likely to be younger, currently working, be in a couple, have no limiting long-term illness, male, have a degree, be in managerial and professional occupations, and be in the wealthiest quintile compared with those who used the internet less frequently or not at all. For example, the mean age of a daily internet user was 66 (SD 7.64) years, whereas the mean age of a participant who had never used the internet was 75 (SD 9.30) years. Overall, 49% of daily users were male, compared with 38% of those who had never used the internet.

Table 2 presents the fixed effects panel modeling estimates for the frequency of internet usage on depression and life satisfaction. The model estimates from the full models are shown in Multimedia Appendix 1. The coefficients represent estimates of change within participants between waves 6 and 8, and therefore positive estimates represent an increase in the depression score (a poorer health outcome) and an increase in the life satisfaction score (a positive health outcome). The frequency of internet use was not predictive of within-person change in depression when taking into account time-varying age, working, being in a couple, and health status. Daily internet usage did predict better life satisfaction over time compared with using the internet on, at most, a monthly basis when controlling for the same time-varying variables (β for monthly or less=−0.512; P=.02). On the basis of model estimates, Figure 1 shows that those using the internet daily were predicted to have more than a half-point increase in life satisfaction scores between waves 6 and 8. There was no significant within-person change in life satisfaction for weekly or never users (vs daily users).

Figure 2 graphically illustrates the predictions from the random effects modeling, which showed that education and occupational class significantly moderated the longitudinal association between internet use frequency and mental health (likelihood ratio test P<.10 for interaction terms). The model estimates from the full models are presented in Multimedia Appendix 2. The differences in the predicted scores of depression (P=.09) and life satisfaction (P=.02) by internet use were steeper for those with a degree. For example, the predicted life satisfaction score for those who had no formal educational qualifications was almost 25 irrespective of internet use frequency, whereas the predicted score for those with a degree was over 1 unit higher for daily users compared with monthly or never users. Similarly, the models suggested that occupational class moderated the association between internet use frequency and life satisfaction (P=.05) but not depression. There was no evidence to suggest that wealth moderated the associations.

Table 3 presents the estimates from the random effects model for mutually adjusted purposes of internet usage on mental health (full estimates are presented in Multimedia Appendix 3). Compared with those not using the internet for this stated purpose, using the internet for communication purposes was associated with a lower depression score between waves 6 and 8 (β=−0.24; P=.002) and a higher life satisfaction score (β=.97; P<.001) when controlling for time-varying and time-invariant variables. In contrast, using the internet for information access was associated with a worse life satisfaction score (β=−0.86; P=.001).

Analysis of more specific purposes (Multimedia Appendix 4) showed that using the internet for email was the only communication purpose associated with better mental health longitudinally and that using the internet for job searching was the only information access purpose associated with deteriorating mental health. Adjustment for more detailed working status categories did not change our findings, suggesting that using the internet for job searching was associated with lower life satisfaction longitudinally over and above being unemployed or out of work.

This study partially confirmed our first hypothesis (H1) that more frequent internet use improves mental health. After adjustment for time-varying confounding factors, there was a positive longitudinal effect of using the internet daily compared with monthly or less on life satisfaction. Internet use frequency was not associated longitudinally with depression. The second hypothesis (H2) that SEP moderates the association between internet usage frequency and mental health was confirmed. This association was stronger for those with an educational degree. There was also a stronger association in those in managerial occupations and professional occupations in life satisfaction. We confirmed our third hypothesis (H3) that using the internet for communication purposes (specifically use of email) was associated with better mental health and that using the internet for information access (specifically for job searching) was associated with worse mental health.

Existing studies have found mixed results regarding the association between internet use and mental health. Notwithstanding methodological differences, our findings are consistent with studies that found that frequent internet use can have benefits for older adults’ depression and life satisfaction [13,15,21,34]. The discrepancy between the findings could be because of methodological issues, such as measurement of mental health [18], and varying sample characteristics, sizes, and context [37]. Very few previous studies have been able to analyze internet use frequency and its relationship with mental health longitudinally. Our findings are consistent with the limited literature available [15,16].

This is the first study to our knowledge that explores SEP as a potential moderator of longitudinal associations between internet use and mental health among older adults. Several studies draw conclusions from sample sizes of fewer than 300 respondents and are not generalizable to middle- and older-aged adults. In contrast, the results of this study come from a large and nationally representative sample of persons aged 50 years and older. Furthermore, the range of questions on internet use allowed us to extend knowledge on the impact of different purposes of internet use on mental health.

Evidence from this study suggests that markers of SEP moderate the association between internet use and mental health; frequent usage in the highest educational and occupational groups had a larger positive impact on mental health than frequent use among those in the lowest groups. This might suggest that low usage in higher SEP groups through digital exclusion, for example, may be particularly problematic for mental health. Our findings, however, need to be viewed in light of our mental health outcomes. The measure of mental health used in our study is determined solely on the affective (depression) and cognitive (life satisfaction) aspects of hedonic well-being, which may not be fully indicative of overall mental health, which also includes eudemonic aspects such as a sense of purpose or control. We cannot rule out the possibility that internet use fulfills eudemonic needs among older adults in disadvantaged SEP groups. Future studies should use a broader measure of mental health to assess whether the longitudinal association between internet use and mental health varies significantly across SEP groups. Further work could also explore whether the relationship between internet use frequency and mental health is moderated by SEP in younger samples in whom almost all use the internet. One might posit that our finding is a cohort effect because later-born generations are expected to enter older age with a much higher prevalence of internet usage.

Our findings highlight heterogeneity among different purposes of internet use. Using the internet for communication through email was associated with better mental health. This provides empirical support for the importance of communication for positive mental health, as it allows individuals to maintain strong social ties and contacts, thereby reducing the negative impacts of social isolation and loneliness on mental health [7,15,38,39]. In contrast, using the internet for job searching was associated with higher depression and worse life satisfaction over time. This is comparable with previous analyses of ELSA data. This finding may indicate dissatisfaction with one’s current job or economic activity status rather than a direct effect of using the internet for this purpose [21]. Our results emphasize the importance of disaggregating different internet usage purposes, as internet use among older adults is not homogenous [40] and impacts different aspects of mental health in different ways.

Our findings point to clear policy implications, the most pressing of which is the need for interventions to encourage older adults to use the internet, so as to build and maintain strong social ties, thereby avoiding the damaging impacts of social isolation and loneliness on mental health. Strategies to promote and increase internet usage among older adults could include setting up public Wi-Fi in places frequently visited by older people and where usage might be low because of barriers to access and digital exclusion. Design of policy interventions to increase internet usage should consider the fact that nonusers of the internet in older age are not homogenous entities but are heterogeneous with respect to gender, age, SEP, household composition, and attitudes toward the internet [41]. It is also important to recognize that for many older people, intrinsic training in how to use the internet did not take place during their formal education or working life, and therefore, the barriers to access can be much more than about availability. Therefore, modifying any psychological barriers to active engagement with the internet is also important. More specifically, to promote internet use for communication, public policy makers should encourage private providers of social networking tools to create age-friendly interfaces. Hardware also needs to be appropriate for novice users and users with impairments such as poor eyesight. This requires manufacturers to be sensitive to these needs when designing products and making the accessibility settings easy to navigate to change the text size, for example.

Several limitations of this study need to be acknowledged. The 8-item CES-D administered in ELSA does not measure depression, rather symptomology, and its association with internet usage may be confounded by omitted variables not included or imperfectly measured in this study. These might include the diagnosis of mental health problems and access to mental health services. Further research should theorize the possible pathways between internet usage and mental health to minimize potential confounding. The measurement of our main exposure is also imperfect, as we were unable to disaggregate daily use of the internet (almost half of the sample) by the number of hours spent on the internet on a typical day. Evidence from other studies suggests that spending an excessive amount of time on the internet correlates with negative mental health in younger age samples [42]. We are not aware of any evidence of the relationship between very frequent internet usage and mental health in older age. A further limitation is the limited 4 years of follow-up. More data points are necessary to explore in detail how any change in the frequency and purposes for internet usage impact mental health trajectories.

A final limitation of our study is the potential bias in our estimates because of the selective nature of participation in aging cohort studies, such as ELSA. Our study sample included participants from the refreshment samples added at waves 3, 4, and 6 to ensure a nationally representative sample of persons aged 50 years and older living in private households at the time of wave 6 (2012-2013). Although our analyses used the appropriate weights available for use with the ELSA study and used multiple imputed data to reduce (wave and item) nonresponse bias, the healthier nature of surviving participants could have inevitably led to an underestimation of the internet use and mental health associations.

In summary, internet use can serve as a channel for improving the mental health of older adults. With population aging and increases in the prevalence of poor mental health, promotion of (affordable) internet use and minimization of psychological barriers to digital engagement could potentially help to promote healthy living in older age.