We analyzed data from the National Health and Aging Trends Study (NHATS), which has been collecting data since Wave 1 in 2011. For the discovery sample, we used data from Wave 1 (2011) through Wave 8 (2018), encompassing a total of 7 waves of follow-up. In Wave 1, 8245 participants were initially recruited into the study. To ensure analytical accuracy, we first excluded 2230 participants lacking key data (missing data related to digital isolation) and further excluded 26 participants missing baseline sleep disorder data. This resulted in 5989 participants included in the cross-sectional analysis of the discovery sample. To examine the longitudinal association between digital isolation and sleep disorders, we excluded 3711 participants with confirmed sleep disorders at baseline. As a result, 2278 participants were included in the longitudinal analysis of the discovery sample.

To validate our findings, we used an independent validation sample comprising 4182 participants newly recruited in Wave 5 (2015) of NHATS. For these participants, data from Wave 5 to Wave 12 (2015‐2022) were used, encompassing seven waves of follow-up. In the validation sample, 723 participants lacking digital isolation-related information and 16 with missing baseline sleep disorder data were excluded. Thus, 3443 participants remained for the cross-sectional analysis in the validation sample. Similarly, for longitudinal analysis, 2197 participants with baseline sleep disorders were excluded, resulting in 1246 participants in the longitudinal validation cohort.

In the final analysis, the discovery and validation samples were combined, yielding a total of 9432 participants for the cross-sectional study. After excluding baseline cases of sleep disorders, 3524 participants were included in the longitudinal cohort. Combining the discovery and validation samples enhanced the reliability and generalizability of the findings.

In this study, digital isolation was defined based on participants’ engagement across four key digital domains: (1) mobile phone ownership and usage (no ownership classified as digitally isolated in this domain), (2) functioning computer ownership and usage (no ownership or inability to use classified as digitally isolated), (3) electronic communication via email or text message in the past month (no engagement classified as isolated), and (4) engagement in other internet-based activities (excluding email or text message) in the past month (no participation classified as isolated). A score of 1 was assigned for nonengagement in each domain and 0 for engagement. These scores were summed to create a composite digital isolation score, ranging from 0 to 4, with higher scores indicating greater digital isolation. It is important to acknowledge that this scoring method treats all digital domains with equal weight and was not formally validated within this study. The distribution of the composite digital isolation scores in our sample ranged from 0 to 4. A substantial portion of participants scored at the extremes, reflecting either high digital engagement (0‐1) or high isolation (3-4). Based on this distribution, participants scoring ≤2 were categorized into the low isolation group, representing at least some engagement across 1 or 2 digital domains. Conversely, participants scoring ≥3 were categorized into the high isolation group, indicating a lack of engagement across the majority (3 or more) of the assessed domains. This cutoff aimed to distinguish between individuals with some level of digital connection and those experiencing a more substantial degree of digital disconnection. This classification formed the basis for subsequent statistical analyses [27].

This study aimed to assess sleep disorders based on participants’ sleep patterns during the past month. A total of 3 key indicators were used to identify potential sleep disorders. First, participants were asked whether they frequently took more than 30 minutes to fall asleep. Those who reported this occurring every night, most nights, or some nights were classified as having difficulty initiating sleep. Second, participants were asked whether they had difficulty returning to sleep after nocturnal awakenings. Those who experienced this every night, most nights, or sometimes were considered to have difficulty maintaining sleep. Finally, the study assessed whether respondents needed to use medication to help them sleep. Those who reported taking medication every night, almost every night, or on some nights were also classified as having a sleep disorder.

Based on these 3 questions, respondents were considered to have a sleep disorder if they experienced issues in any of the three areas and were assigned a value of 1 in the composite sleep disorder variable. If no issues were reported in these areas, they were assigned a value of 0. For missing or refused responses, the data were marked as missing values. This sleep disorder variable provided crucial data for subsequent statistical analyses, helping to reveal potential associations between sleep issues and other health variables [27].

We controlled for several covariates to minimize potential confounding. Respondents’ age was divided into two categories: 75 years and younger and 75 years and older, to analyze the effect of different age groups on the study results. Sex was based on respondents’ self-reports and categorized as male or female [28-30]. Marital status was classified as married or other statuses (such as cohabitating, separated, divorced, widowed, or never married) to evaluate the potential impact of marital status on the study variables [31-33]. Race and ethnicity included categories for non-Hispanic White, non-Hispanic Black, Hispanic, and other races (such as Native American, Asian, and Native Hawaiian) [34,35]. Smoking status was determined based on whether respondents were regular smokers, aiming to assess the impact of smoking on health outcomes [36].

Hearing impairment was defined as the use of hearing aids or reported difficulty hearing in quiet or noisy environments [37]. Vision impairment was determined based on respondents’ reports of difficulty seeing, even with corrective lenses, such as glasses [38]. For mental health, the Patient Health Questionnaire-2 scale was used to measure depressive symptoms. The total score ranged from 0 to 6, and participants with a score of 3 or higher were classified as having potential depression. Similarly, the Generalized Anxiety Disorder-2 scale assessed anxiety symptoms, with total scores ranging from 0 to 6, and a score of 3 or higher indicating potential anxiety [39-41]. In addition, we assessed respondents’ chronic conditions based on the self-reported number of chronic diseases and categorized them into groups with 3 or fewer conditions and more than 3 conditions [42]. These covariates were included in subsequent multivariable analyses to control for potential confounding effects of health and demographic characteristics on study outcomes.

All statistical analyses were sequentially conducted on the discovery, validation, and combined samples to ensure the robustness and generalizability of the findings. Descriptive analyses of baseline characteristics were first performed. χ² tests were applied to examine differences among groups in baseline characteristics across levels of digital isolation (discovery vs validation samples). These descriptive analyses provided an initial overview of population characteristics and informed subsequent regression analyses.

In the cross-sectional analysis, multivariable logistic regression models were used to evaluate the association between digital isolation and sleep disorders. Model 1 was unadjusted, reflecting the crude association, whereas Model 2 adjusted for potential confounders to minimize bias. Covariates included in the adjusted model were age group, sex, race and ethnicity, smoking status, marital status, chronic disease status, hearing impairment, vision impairment, depression, and anxiety. Results were reported as odds ratios (ORs) with 95% CIs; P values <.05 were considered statistically significant. Analyses were initially conducted in the discovery sample to identify preliminary associations, then replicated in the validation sample to confirm robustness and consistency. Finally, analyses using the combined sample were performed to enhance statistical power and generate more generalizable conclusions.

For longitudinal analysis, Cox proportional hazards regression models were used to examine the link between digital isolation and the likelihood of incident sleep disorders among participants without baseline sleep disorders. Model 1 and Model 2 would also be calculated. Kaplan–Meier survival curves were used to estimate the incidence rate of sleep disorders across digital isolation groups, with differences assessed using the log-rank test. Survival analyses were initially performed in the discovery sample, validated in the validation sample, and subsequently applied to the combined sample to ensure generalizability.

In addition, subgroup analyses were conducted to explore potential effect modifications of digital isolation on specific populations (eg, different age groups, sex, and races and ethnicities). Interaction terms were added to the Cox regression model to detect potential interactions between subgroups, with corresponding P values used to assess significance. To ensure robustness, sensitivity analyses were also performed to verify the stability and consistency.

To assess the robustness of our longitudinal findings, particularly the nonsignificant result in the multivariable model of the validation sample, we planned a sensitivity analysis. This analysis involved incrementally introducing potential confounding variables (age group, sex, race and ethnicity, smoking status, marital status, hearing impairment, visual impairment, depression, anxiety, and chronic disease) into the Cox regression models to observe their impact on the hazard ratio between digital isolation and sleep disorders.

All statistics were performed using R (version 4.4.1; R Foundation for Statistical Computing), using 2-tailed tests with statistical significance set at P<.05.

The study protocol was approved by the Ethics Committee of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology. The data for this study were obtained from the NHATS study, which is a publicly available dataset. The original NHATS study was approved by the Institutional Review Board of Johns Hopkins University, and all participants provided written informed consent before data collection. The dataset provided for public use is fully deidentified to ensure participant privacy and confidentiality.

Figure 1 shows a flowchart illustrating participant inclusion. Table 1 presents the baseline characteristics of the study population. In the discovery sample (n=5989), 54.73% (3279/5989) of participants were aged 75 years and older, whereas in the validation sample (n=3443), this proportion was slightly lower at 48.57% (1673/3443). In the combined sample (n=9432), 52.5% (4952/9432) of participants were aged 75 years and older. Regarding sex distribution, females comprised 55.82% (3343/5989) of the discovery sample and 55.27% (1903/3443) of the validation sample, resulting in an overall female proportion of 55.62% (5246/9432) in the combined sample. Analysis of marital status revealed that approximately 52.81% (4981/9432) of participants were married across the combined sample. In terms of racial and ethnic composition, non-Hispanic Whites were the dominant group (6503/9432, 68.95%), while non-Hispanic Blacks and Hispanics accounted for 20.48% (1932/9432) and 5.59% (527/9432), respectively. In addition, the distribution of digital isolation differed significantly between samples; the proportion of the high-isolation group was lower in the validation sample compared to the discovery sample (960/3443, 27.88% vs 2166/5989, 36.17%).

Table 2 shows the cross-sectional association between digital isolation and sleep disorders. In the discovery sample, the prevalence of sleep disorders was significantly higher in the high-isolation group compared to the low-isolation group (67.03% vs 59.06%). In the unadjusted model (Model 1), the odds of sleep disorders were 41% higher in the high-isolation group compared to the low-isolation group (OR 1.41, 95% CI 1.26-1.57; P<.001). After adjusting for potential confounders, the association was still significant (OR 1.23, 95% CI 1.09-1.39; P<.001). Similar results were observed in the validation sample, with significantly increased odds of sleep disorders in the high-isolation group (OR 1.22, 95% CI 1.02‐1.47; P=.03), further supporting this association. In the combined sample, the adjusted odds of sleep disorders were 21% higher in the high-isolation group compared to the low-isolation group (OR 1.21, 95% CI 1.10‐1.34; P<.001), demonstrating a consistent cross-sectional relationship across the entire study population. The unadjusted models in all samples showed even stronger associations, suggesting a partial attenuation of the effect by the included covariates.

Table 3 summarizes the longitudinal association between digital isolation and incident sleep disorders among participants without baseline sleep disorders. In the discovery sample, the unadjusted Cox regression analysis showed a significantly increased hazard of sleep disorders in the high-isolation group (hazard ratio [HR] 1.30, 95% CI 1.15‐1.47; P<.001). After adjusting for age, sex, race and ethnicity, and other confounders, the association remained significant (HR 1.21, 95% CI 1.05‐1.38; P=.006). In the validation sample, the unadjusted model also showed a higher hazard in the high-isolation group (HR 1.29, 95% CI 1.07‐1.55; P=.008), but this association was no longer statistically significant after adjustment (HR 1.11, 95% CI 0.91‐1.36; P=.30), indicating a less clear longitudinal relationship in this specific sample. In the combined sample, the adjusted hazard of sleep disorders in the high-isolation group was 17% higher (HR 1.17, 95% CI 1.05‐1.31; P=.005), providing overall support for a longitudinal association, although the nonsignificant finding in the validation sample warrants further investigation. The unadjusted HRs were higher across all samples, suggesting that the included covariates partially explain the association. To ensure the validity of our Cox proportional hazards models, we assessed the proportional hazards assumption for all covariates using the Schoenfeld residuals test in R. The global test results indicated that the proportional hazards assumption was not significantly violated in any of the samples: discovery sample (χ²₁₂=5.71; P=.93), validation sample (χ²₁₁=4.49; P=.95), and pooled sample (χ²₁₂=4.94; P=.96). Furthermore, the individual covariate tests also showed no significant violations of the proportional hazards assumption (all P>.05). The detailed results of the test are presented in the supplementary material (see Table S1 in Multimedia Appendix 1).

Table 4 displays the link between different items of digital isolation and sleep disorders. In the discovery sample, phone isolation was not significantly associated with sleep disorders (HR 1.01; P=.88), but in the validation sample, phone isolation significantly increased the hazard of sleep disorders (HR 1.46; P=.03). Computer isolation, email isolation, and internet-based activity isolation were all significantly associated with sleep disorders in both the discovery and combined samples, with adjusted HRs ranging from 1.16 to 1.20 (P<.05). However, in the validation sample, the associations for computer isolation and internet-based activity isolation lost significance after adjustment. These findings suggest that different aspects of digital isolation may have varying impacts on sleep disorder risk, and these associations may differ between the discovery and validation samples.

Kaplan–Meier survival curves further illustrated the incidence of sleep disorders during follow-up across different digital isolation groups (see Figure 2). The incidence of sleep disorders was significantly higher in the high-isolation group compared to the low-isolation group, with the log-rank test showing a statistically significant difference between the two groups (P<.001). In addition, the forest plot in Figure 3 demonstrated the link between digital isolation and sleep disorder among subgroups. The analysis showed consistent effects of digital isolation across different subgroups (eg, age, sex, and race and ethnicity), but stronger effects were observed in certain subgroups, such as the older age group and female participants, where interaction term P values indicated significant heterogeneity (P<.05).

Stepwise sensitivity analyses were conducted in the validation sample to examine how the association between digital isolation and incident sleep disorders was attenuated (see Table S2 in Multimedia Appendix 1). In the unadjusted model (Model 1), digital isolation was significantly associated with an increased risk of sleep disorders (HR 1.29; P=.008). After sequential adjustment for demographic characteristics (sex, marital status, and race), smoking status, hearing impairment, and visual impairment (Models 2‐5), the HR for digital isolation was attenuated but remained marginally significant (eg, Model 5: HR 1.23; P=.04). A key change emerged after adjusting for mental health factors. With the inclusion of depression (Model 6), the association between digital isolation and sleep disorders became non-significant (HR 1.20; P=.08), while depression emerged as a strong predictor (HR 1.52; P=.005). Adding anxiety (Model 7) revealed it as another significant predictor (HR 1.65; P=.01), while digital isolation remained nonsignificant (HR 1.18; P=.11). In the fully adjusted model (Model 9), which also included chronic disease burden and age group, digital isolation was no longer significantly associated with sleep disorders (HR 1.11; P=.30). In this model, depression (HR 1.52; P=.01), anxiety (HR 1.67; P=.009), chronic disease burden (HR 1.34; P=.006), and older age (HR 1.32; P=.001) were all independent and significant predictors. These analyses suggest that the initial association between digital isolation and incident sleep disorders was largely explained by confounding or mediation from mental health status, chronic disease burden, and age.