Associations of face-to-face and non-face-to-face social isolation with all-cause and cause-specific mortality: 13-year follow-up of the Guangzhou Biobank Cohort study

The Guangzhou Biobank Cohort Study (GBCS) in China is a 3-way collaborative project among the Guangzhou Twelfth People’s Hospital and the Universities of Hong Kong and Birmingham. Participants were recruited from a community social and welfare association, the “Guangzhou Health and Happiness Association for the Respectable Elders” (GHHARE). From 2003 to 2008, 32,850 members of the GHHARE were invited, and of them, 30,430 agreed to participate and signed informed consents, with a response rate of 92.6%. GHHARE is a large unofficial organization with ten branches throughout all districts of Guangzhou. Membership of this association is open to Guangzhou residents aged 50 years or older for a nominal, monthly fee of 4 Chinese yuan renminbi (CNY) (about 50 US cents). Seven percent of local residents aged 50+ years enrolled in the GHHARE. All surviving participants were invited to return for follow-up examination from March 2008 to December 2012. Details of the GBCS, GHHARE, and some prospective study results have been reported previously [45‐47]. Briefly, baseline information was collected using a computer-based questionnaire in face-to-face interviews by trained nurses in the Guangzhou Twelfth People’s Hospital. Information of demographic characteristics, lifestyle factors, family and personal medical history, and anthropometrics, blood pressure, fasting plasma glucose, lipids, and inflammatory markers was collected following standard protocols. The reliability and validity of the questionnaire were tested 6 months into recruitment by recalling 200 randomly selected participants for re-interview, and the results were satisfactory [45].

Since no study has reported face-to-face and non-face-to-face contact separately, we first proposed the 4 types of social contact using the validated questions in the Berkman-Syme Social Network Index (SNI) [48] with appropriate revise. The four types of social contact were illustrated in Fig. 1 and described in Table 1. We included mail as another way of non-face-to-face contact besides telephone, since the telephone and mail are the commonest non-face-to-face ways to contact others in 2003, before smartphone, Internet, and social media had become popular. A composite social isolation score was derived from the sum of four types of social isolation, with a score from 0 to 7. A higher score indicates greater social isolation (Table 1).

As described in our previous papers [49, 50], information on causes of death up to December 31, 2019, was obtained via record linkage with the Death Registry of the Center for Disease Control and Prevention in Guangzhou. Causes of death were coded according to the 10th Revisions of the International Classification of Diseases (ICD-10) by trained clinical coding officers in each hospital. The ICD-10 codes of the cause-specific mortality were as follows: cardiovascular disease (I00-I99, excluding I26, I27), cancer (C00-C96), and other diseases (all remaining ICD-10 codes). A physician panel including 5 chief physicians from various disciplines reviewed all available medical records of the same individuals and assigned in a standard manner a cause of death, with the assistance of an epidemiologist in the last meeting for unsettled cases.

To examine the extent to which baseline socioeconomic position (SEP), biological, behavioral, and psychological factors explained the associations, we included them in two models. Sex, age, and self-rated health (good/very good, poor/very poor) were included in minimal model. The fully adjusted model was adjusted for thirteen factors (sex, age, self-rated health, SEP, biological and behavioral factors): (1) SEP: education (primary and below, middle school, and college or above), occupation (manual, non-manual, and others), family income (≤ 30,000 CNY/year, > 30,000 CNY/year, and not known; US$1 = 7 CNY); (2) biological factors: height and weight were measured, and body mass index (BMI) was calculated by weight in kilogram divided by height in meters squared (kg/m²). After an initial 5-min rest, seated blood pressure was measured 3 times at 1–5 min’s intervals using an Omron 705CP sphygmomanometer (Omron Corp, Kyoto, Japan). Systolic and diastolic blood pressure (SBP and DBP) were calculated as the average of the last 2 measurements. Fasting glucose was measured after an overnight fast; (3) behavioral factors: smoking status (never, former and current), alcohol use (never, former, and current drinkers), physical activity assessed by International Physical Activity Questionnaire (inactive, moderate, and active) [46]; (4) psychological factors: stress level assessed by Perceived Stress Scale-14 items (PSS-14) [51], and cognitive function assessed by Delayed Word Recall Test (DWRT) [52]. The DWRT was assessed in the full dataset, but PSS-14 was only assessed in a subsample of GBCS in phrases 1 and 2 (n = 19,947). All potential confounders were categorized as in Table 2. We calculated the percentage of excess risk mediated (PERM), which represented the percentage of risks that explained by SEP, and biological and behavioral factors as in the two previous UK Biobank papers [5, 32]:

Chi-square test or analysis of variance was used to compare baseline characteristics by four types of social isolation. Associations of four types of social isolation with all-cause and cause-specific mortality were estimated by Cox proportional hazards model giving hazard ratios (HRs) and 95% confidence intervals (CI). Schoenfeld’s residuals were used to assess the proportional hazard assumption, and no major violations were observed. In sensitivity analysis, competing risk analysis (Fine-Gray’s model) [53] were used to assess the association of social isolation with cause-specific mortality, where each cause was simultaneously modeled as a different event. In sensitivity analysis, fully adjusted model was additionally adjusted for psychological factors using the subgroup dataset because PSS-14 data were available in the subsample only. To explore potential bias due to missing data, we assessed the proportion of missing data for all variables. The proportion of missing data in all variables was very low (i.e., less than 3%) (Additional file 1: Table S2). Therefore, we used complete case analysis in the current study. Subgroup analyses were done by sex, age groups (< 60 years, ≥ 60 years), education (≤ primary, ≥ middle school), and self-rated health (poor/very poor, good/very good) to investigate whether the associations of social isolation with mortality varied by these factors [32, 54, 55]. For sensitivity analysis, in order to assess the independent associations of face-to-face and non-face-to-face isolation with mortality, we analyzed the associations of 3 types of contact (excluding club/organization isolation which showed no association with mortality) with all-cause mortality after mutual adjustment in a sensitivity analysis. To partly address reverse causation, we excluded participants who died within the first 2 years in the main analysis and provided the original data (without excluding death within first 2 years) analysis for comparability. Also, a subsample of 18,104 participants who returned for repeated examination during 2008–2012 was analyzed to examine the associations in survivors after baseline examination. All statistical analyses were done using the Stata version 15.0 (STATA Corp LP). All tests were two-sided, with P < 0.05 as statistically significant.

We have first, in a long-term prospective study (with about 13 years follow-up), show the positive associations of face-to-face and non-face-to-face isolation and a composite isolation score, with all-cause, CVD- and other-cause (but not cancer) mortality. Only 18–36% of the excess risk was attributable to the known risk factors such as SEP, biological, and behavioral factors. No association of club/organization contact (< 1 time/month versus ≥ 1 time/month) with all-cause and cause-specific mortality was found.

Our findings are in line with previous studies showing social isolation was associated with a higher risk of all-cause and CVD mortality [8, 13, 24, 32]. However, no previous studies reported the associations with face-to-face isolation and non-face-to-face isolation separately. For example, the UK Biobank study [32] assessed social isolation using a social isolation index, which only included questions on face-to-face contact. The Million Women Study [3] and English Longitudinal Study of Ageing [54] included results of a composite score with e-mail, phone and face-to-face contact but did not report results on different isolation types. The Whitehall II Study [55], the Nurses’ Health Study [56], and the Heinz Nixdorf Recall Study [24] also showed a higher mortality risk related to social isolation but did not specify the types of contact. Therefore, our study has provided additional evidence that both face-to-face and non-face-to-face isolation are associated with increased mortality risk. Our results on non-face-to-face contact are also consistent with a prospective study showing that participants with online social media experience (Facebook use) had a lower risk of mortality during two years follow-up than those without [57]. However, this study did not report the association of face-to-face contact with mortality, and the association of non-face-to-face contact was not adjusted for SEP and behavioral factors.

Furthermore, whether the associations of social isolation with mortality varied by sex, age, education, and health status is inconclusive [6, 18, 32, 54]. We found consistent associations of social isolation with all-cause, CVD-, and other-cause mortality after adjustment for sex, age, education, and health status. Our sex-stratified results are not consistent with one previous study showing the stronger associations of social isolation with CVD in women than men [32], neither consistent with other two reported stronger associations in men [24, 58]. We found the education did not modify the associations of social isolation with mortality, which was not consistent with previous studies showing the mortality risk of social isolation change depend upon SEP but without clarifying the types of social contact [27, 59, 60]. Moreover, living alone was particularly increased the CVD mortality in unhealthy participants suggesting the importance of immediate help among people who live alone, as a recent study reported [3].

Socially integrated people usually have more access to resources for health-promoting behavioral and chronic disease management [61]. Our results suggest that 18–36% of the excess all-cause, CVD-, and other-cause mortality risks could be attributable to SEP, biological, and behavioral factors, which is lower than those reported in two previous studies [5, 32], which showed that 64% and 84% mortality risk were explained by confounders or mediators. The fraction of excess risk remaining unexplained indicates some biological mechanisms beyond traditional pathways. For example, individuals with greater social support might have better immune function [62, 63]. Other pathways such as neuroendocrine mechanisms might also play a role [64, 65]. A recent genome-wide association study (GWAS) identified 38 significant genetic variants for social interaction, highlighting the possible genetic basis for social isolation [66]. Furthermore, another study found that gene expression differed between socially isolated and non-isolated individuals [67].

A major strength of the current study was that we, for the first time, analyzed the risk of mortality related to four important social isolation types: face-to-face isolation from co-inhabitants, face-to-face isolation from non-co-inhabitants, non-face-to-face isolation, and club/organization isolation, and provided the independent associations of face-to-face isolation and non-face-to-face isolation after mutual adjustment, which have never been reported in previous studies. The comprehensive analyses of four aspects of social isolation are essential because social contact patterns have been changing given the social-economic development. Our results demonstrated the beneficial role of non-face-to-face contact, which, if causal, would have important public health implications, although virtual contact has been the dominant form for non-face-to-face contact nowadays. Other strengths of our study included the long-term follow-up and the comprehensive adjustment for potential confounders. However, some limitations need to be considered. First, as about 98% of participants were retired at the time of baseline examination, workplace contact was not included in our study. Although the absence of workplace contact was unlikely to affect the internal validity of the present study, generalizability of our results to younger populations, in whom workplace contact may represent a major mode of contact, might be limited. Also, the other non-face-to-face contact such as instant message applications (video contact) and e-mail were not included. However, since this cohort study was initiated in 2003, such non-face-to-face contact was not prevalent, especially in older people. Specifically, the new technologies have been found to be effective in social contact for older people, but the findings were often mixed or inconclusive [68, 69]. Further studies are warranted to clarify. Second, we did not assess the effect of loneliness, which is a negative emotional state resulting from isolation, on mortality. However, isolation, which can be measured more objectively, rather than loneliness, was found to be an independent predictor of mortality [5]. Third, as evaluation of social isolation was based on self-report, reporting error was possible, although such error was more likely to be non-differential. Fourth, although residual confounding cannot be completely excluded in observational studies, the adjustment for a wide range of potential confounders in the current study should have minimized this bias. Fifth, reverse causality was possible, but we excluded deaths within 2 years in the main analysis and further conducted subgroup analyses in participants only with self-rated good/very good health status with no substantial changes in the original results. Sixth, we assessed social isolation at baseline and did not account for the changes in social isolation during the follow-up [70]. However, our sensitivity analysis of a subgroup of 18,104 participants who returned for repeated examination during 2008–2012 showed similar results to the main analyses (Additional file 1: Fig S1-2). Finally, as all GBCS participants were middle-aged and older Chinese, generalizability of our results to other populations or younger age groups may be limited.