Methods

The survey for this study was based in Okayama prefecture, which is west of Osaka. A two-stage stratified random sampling procedure was employed in which we targeted 60 companies headquartered in Okayama prefecture. Companies were stratified into three categories according to the number of their employees; 50–99, 100–299 and 300 or more employees. The numbers of companies in each category were 92, 104 and 57, respectively. We randomly selected 20 companies in each category. In the second stage, we randomly sampled 30 employees from each company. To standardise the sampling procedure, we instructed the human resources department of each company on how to randomly sample employees from their payroll. Part-time workers were included but board members were excluded to limit our investigation to the general employees. Questionnaires were distributed to each selected participant and returned in anonymous, sealed envelopes.

Social network at work was assessed by asking each respondent to indicate the number of co-workers in the same company whom they consult with ease on personal issues. Responses were obtained from seven predetermined categories: from 0 to 6 or more people. They were categorised as low social network (0 people), medium (1–3 people) and high social network (4 or more people). The cut-off was determined a priori to examine the effect of social isolation (0 people) compared with evenly split size of the social networks. We also inquired about social ties outside the company in the same manner. The quality of social networks was not assessed.

Our health outcome variable was self-assessed health, measured by responses to a single item: “Would you say that in general your health is excellent, very good, good, fair or poor?” From this item, we created a dichotomous outcome measure (1 = fair or poor; 0 = excellent, very good or good). An extensive literature has documented the reliability and validity of this single-item self-rated health measure, including the finding that it predicts subsequent mortality and hospitalisation risk in prospective studies.1112

We constructed a directed acyclic graph (DAG) of the possible confounders of the association between social networks and health status based on prior theory (fig 1),1314 from which we selected the following variables as relevant confounders for statistical control: sex, age (10-year categories), occupational status (non-manual vs manual), education (less than senior high school education vs more advanced educational attainment), smoking status (never/former vs current), frequency of alcohol consumption and leisure-time physical activity (any vs none). The basic education was defined to be less than senior high school education according to the conventional education system in Japan. Alcohol consumption was categorised into two groups as follows: low (none/rarely/1–3 days a month) vs high (1–2 days a week/3–4 days a week/5–6 days a week/everyday). In addition, height and weight were reported in the questionnaires, from which the body mass index (BMI) was calculated and dichotomised at 25 kg/m² according to the guidelines of the Japan Society for the Study of Obesity. We also created indicator variables for whether the respondent had been told by their physician that they had one of the following chronic conditions during recent health check-ups: high blood sugar, high blood pressure and low high-density lipoprotein (HDL) and/or high triglyceride.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

A directed acyclic graph (DAG) showing the association between social networks and self-rated health, Japan, 2007. U, unknown or unmeasured variables. *Social networks at work and outside companies are defined as the number of people who are easy to consult about some (personal) issues within and outside companies, respectively. †Individual characteristics consist of sex, age, occupational status and education. ‡Health-related behaviours consist of smoking status, alcohol use and physical activity. §Obesity is defined as being greater or equal to 25 kg/m² of body mass index. Co-morbid conditions include the following three problems in recent health check-ups; high blood sugar, high blood pressure and lower high-density lipoprotein (HDL) and/or high triglyceride.

Tests for linear trend of the association between social networks and poor health were carried out using social network per person as a continuous variable. The relationship between work-based and home-based social networks was examined by Pearson’s correlation coefficient. We examined the association between work-based social networks and poor health in the crude logistic regression analysis (model 1). In the multiple regression analysis, all the relevant confounders were adjusted (model 2), and we further examined the independent effects of social networks at work and outside work by mutually adjusting for them. To block the spurious association caused by the dashed arrow in fig 1, both BMI and the presence of co-morbid conditions were adjusted for in the fully adjusted model (model 3). Finally, we evaluated whether the effect of social networks was influenced by sex and age using a test for interaction, by entering multiplicative interaction terms between each factor into model 3. Age was divided into three groups: younger than 30, 30–49 and 50 or older. In every model, the OR and 95% CI for poor health status were obtained for each variable, and a p value of less than 0.05 (two-sided test) was considered statistically significant. All analyses were conducted using SPSS 15.0J (SPSS Inc., Chicago, Illinois).

Results

Of the eligible total of 1800 subjects, questionnaires were returned from 1218 employees (67.6% response) from 46 companies. We excluded respondents with missing values on the social network questions, self-rated health, sex or age, which resulted in 1105 subjects available for the present study. The demographic characteristics and the proportion of subjects reporting poor health status are shown in table 1. No significant differences were observed in any of the variables in table 1 between the analysed and the excluded subjects. Among the 1105 subjects, 407 reported poor health (36.8%). Poor health was more common among workers who were obese, sedentary and diagnosed with chronic conditions during recent health check-ups. We also found that male workers with lower levels of educational attainment also tended to report worse health status. Social networks at work and outside work were inversely associated with poor health (table 2). The mean and the standard deviation of the size of social networks within companies and outside the workplace were 3.71 (1.96) and 4.56 (1.89), respectively. Pearson’s correlation coefficient between work-based and home-based social networks was 0.435 (p<0.01).

The ORs and 95% CIs for each variable are shown in table 3. In the crude regression model (model 1), we found that poor health is more common in those who have lower social networks at work (OR 1.53, 95% CI 1.03 to 2.27, comparing those with the lowest vs highest level of social network, p for trend = 0.003). When the relevant confounders were adjusted for, the association became attenuated and no longer statistically significant (model 2). By contrast, we found a dose–response relationship such that a lower social network outside companies was associated with poor health (OR 2.05, 95% CI 1.14 to 3.69, comparing lowest vs highest level of network outside the workplace, p for trend = 0.012) (model 2). Adjusting for the BMI and the self-reported co-morbid conditions did not change these results materially (model 3). In further analyses, no substantial difference was observed between men and women in the relationships reported above. On the other hand, we found some suggestion of effect modification by the age groups of the employees (p for interaction between work-based social networks and age groups = 0.049, p for interaction between home-based social networks and age groups = 0.090). In the stratified analyses (table 4), among the oldest group (50 years or older) we found a significant association between medium work-based social networks and poor health, and the test for trend was marginally significant (p for trend = 0.087). No consistent association was found between home-based social networks and health. By contrast, among middle-aged workers (30–49 years), we found a more consistent relationship such that the lower level social network outside the workplace was associated with poor health, whereas there was no such relationship between work-based social networks and health. Among the youngest group (<30 years), no consistent relationships were found with work-based or home-based social networks.

Discussion

To our knowledge, this is the first study to examine the separate and independent contributions of social ties within and outside the workplace on employees’ health status. In the job stress literature, previous studies have focused on a lack of co-worker and supervisor social support as a determinant of poor health (ie, the “iso-strain” model).15 However, no studies have simultaneously considered the influence of social ties within and outside the workplace. In this study of Japanese companies, we found that the association between health status and social networks within the workplace becomes attenuated to statistical non-significance once we control for the presence of social ties outside the workplace. In other words, social ties in the home trump social ties at work. However, we also found a suggestion of an important interaction by age groups.

We hypothesised a posteriori that the differential pattern by the age groups may reflect a generational difference among Japanese workers with regard to their attachment to work (ie, stronger attachment among older workers who grew up in a culture of lifetime employment guarantees), as well as generational differences in sources of social connections (ie, more likely to be outside the workplace for younger workers). Among older workers (50 years or older), we found that social ties within the workplace are more strongly associated with health status than social ties outside work. These are likely to be the workers who experienced long-term job security under the system of lifetime employment guarantees typical of many Japanese companies prior to the decade-long recession that occurred during the 1990s. Workers under this system were often encouraged to view the company as part of their family, and many employees often centred their lives around their workplace, for example working long hours at the office (including weekends), socialising every night with co-workers and supervisors, and even taking annual vacations in company-owned facilities. Within a culture of such high social cohesion, it is perhaps not surprising that work-based social networks have more meaning for these workers than networks outside the workplace. By contrast, middle-aged workers are more likely to have experienced the post-recession period during their early employment, when job security was no longer guaranteed and the threat of “restructuring” loomed large in employees’ lives. For these workers, we found that social networks outside the workplace are strongly associated with health status, whereas there appeared to be no relationship with work-based networks. In other words, middle-aged Japanese workers may feel more alienated from their companies due to the breakdown in the lifetime employment system and increased job insecurity, and they may seek social networks outside workplaces. Among younger workers, however, no consistent associations between social networks and health were observed. They are more likely to have been hired during the post-recession period.

Interestingly, when we checked for effect modification by gender, we found no suggestion of a difference in these relationships between men and women, even though social support at work has been suggested to be more important for the health of women than for men.1516 This inconsistency could be due to the lack of adjustment for the social network outside companies in previous studies. In line with this, the relatively higher risk of poor health was observed among women with low network at work when the lack of social network outside companies was not accounted for in the supplementary analysis (data not shown). Since the cultural differences between Western countries and Japan may also explain this inconsistency, our research needs to be replicated in further studies.

In this study, a DAG was drawn a priori to make the research hypothesis more explicit, and relevant variables were measured accordingly.1314 In our models, we attempted to block potentially spurious associations by adjusting for the relevant confounders based on the DAG. To check the robustness of the associations, we adjusted for the BMI and co-morbid diagnoses in the fully adjusted model (model 3). This additional adjustment did not alter the current findings substantively. The dashed arrow in fig. 1 is therefore less likely.

Possible limitations of the current study include the lack of validated, comprehensive assessment of social networks. We assessed social networks with respect to their size only, and we did not assess additional characteristics such as directionality of relationships or the number and identities of people to whom the respondent provided advice. In this regard, more sophisticated assessment using network analysis is warranted in further studies.2 Second, we sampled companies with 50 or more employees since they are assumed to have enough colleagues for the formulation of social networks within the companies. The size of companies was additionally adjusted for, which resulted in comparative findings (data not shown). Thus, the current findings are less likely to be influenced by the size of companies, as long as there are at least 50 employees. Third, the employment status (permanent or part-time/temporary) of the subjects was not available. Although we believe that most subjects are permanent workers, the possibility of residual confounding cannot be ruled out. The cross-sectional design of our study is a further limitation, in that we could not rule out reverse causation (poor health resulting in network pruning) or the possibility of selection bias caused by the out-migration of the unhealthiest and most isolated workers from the workplace. This possible selection bias may have led to the underestimation of the current finding of the older workers without work-based social networks. A further limitation is that both exposure and outcome were measured subjectively. Further studies using more objective measures are thus warranted.

In conclusion, we have found that the lack of a social network at work has adverse effects on self-rated health among older generation Japanese workers while the lack of a social network outside companies is apparently associated with higher risk of poor health among middle-aged workers. Given the culture of the traditional Japanese workplace (eg, facilitating the development of social networks through company trips or club activities, etc), it is possible to envisage future interventions to improve social cohesion within companies and thereby promote workers’ health.