Study design and participants
This study is a longitudinal study using the two-point panel data obtained from the Japan COVID-19 and Society Internet Survey (JACSIS) in 2021 and the Japan Society and New Tobacco Internet Survey (JASTIS) in 2022 [8]. The JACSIS aimed to evaluate the health conditions and social determinants of the COVID-19 pandemic in Japan. The JASTIS aimed to evaluate the status of new tobacco products and socio-demographic factors among the Japanese general population. Both surveys were administered through Internet questionnaires and longitudinally share the same respondents. JACSIS, the baseline survey in 2021, was distributed to 224,389 candidates registered as panellists at a Japanese Internet research company (Rakuten Insight, Inc., Tokyo, Japan) between 28 July and 30 August 2021. We determined a sample size of 31,000 panellists based on Japan’s population distribution in 2019. The survey was terminated once the target number of respondents was reached for each category (sex, age, and prefecture). Consequently, 31,000 panellists responded to the survey. To validate the quality of the data, we excluded responses with discrepancies and/or artificial/unnatural responses. The following criteria were used for exclusions: (1) An invalid response to “Please choose the second alternative from the bottom” (i.e., panellists who failed to select the second from last alternative from the five options available); (2) positive responses to all questions related to drug use (e.g., marijuana, cocaine, or heroin); and (3) positive responses to all questions regarding 16 alternative underlying chronic diseases. We excluded 2,518 respondents with discrepancies and/or artificial/unnatural responses (remaining respondents, n = 28,175).
In 2022, the JASTIS survey was distributed to the participants from the 8th to the 26th of February. We used the same scheme as the JACSIS 2021 study. Consequently, 33,000 panellists responded to the survey and 30,130 remained after validation.
From the two surveys, 22,756 (81%) people responded and we used this cohort for the analysis. The adjustment for non-responders is described below.
This study was reviewed and approved by the Research Ethics Committee of Osaka International Cancer Institute (approved on 19 June 2020; approval number: 20084). Web-based informed consent was obtained from all participants before they responded to the questionnaire, and the option to opt-out at any point was provided. The Internet survey agency respected the Act on the Protection of Personal Information in Japan. A credit point system known as “E-points”, which could be used for Internet shopping and cash conversion, was offered as an incentive.
Social isolation and loneliness
Social isolation refers to the absence of relationships with other people [9]. To measure social isolation, we used the Lubbern Social Network Scale (LSNS-6) [10]. The validity and reliability of the Japanese version of the LSNS-6 were confirmed [11]. The scale was developed to screen for social isolation among community-dwelling adult populations. The scale’s score ranges between 0 and 30, with a higher score indicating more social engagement. A score of 0–11 was regarded as social isolation [10].
The University of California, Los Angeles (UCLA) Loneliness Scale version 3, Short Form 3-item (UCLA-LS3-SF3) was used to assess loneliness [12, 13]. The scale’s three items are: in the past 30 days, how often have you felt (i) a lack of companionship, (ii) left out, and (iii) isolated. The validity and reliability of the Japanese version of the UCLA-LS3-SF3 were confirmed [14]. Participants responded to each item on a four-point scale (“never”=0, “rarely”=1, “sometimes”=2, or “always”=3). The score range was 0–9, with higher scores indicating severe loneliness. A score from 4–9 was defined as loneliness [15].
Vaccination status
In the JASTIS 2022 study, the participants were asked their vaccine status in February 2022. Participants who had not taken any vaccination at that point were categorised as unvaccinated. Participants who had vaccinated at least once were categorised as vaccinated. Those who wanted to get vaccinated but could not due to allergy were excluded.
Covariates
Based on a previous study [16] and theoretical considerations, the following variables were included as covariates: age, sex (female or male), marital status (married or not), education, annual household income, working conditions, living area (category), self-rated health, histories of hypertension, diabetes, hypercholesterolemia, stroke, coronary heart disease, chronic obstructive pulmonary disease, cancer, and/or depression (yes or no), divorced (yes or no), widowed (yes or no), and living alone (yes or no).
Statistical analysis
Internet studies’ respondents are generally not representative of the general population, so we performed statistical adjustment to account for bias. Harmonisation of the data with a major national and representative cross-sectional study would allow us to pool data, providing the potential capacity to adjust for “being a respondent in an Internet survey”. Since this method cannot completely adjust for the difference in respondents between an Internet survey and a nationwide representative survey, the problem of generalisability remains. However, this method can approximate our estimate to a nationally representative estimate, using inverse probability weighting to account for baseline characteristics, such as socio-demographic and health-related factors [8]. Details regarding this method have been given in previous reports [17].
The response rate in the follow-up survey was problematic as non-responders differed in several ways from responders in the survey. There was evidence suggesting that attrition was higher among the younger and socio-economically disadvantaged populations. Therefore, to account for potential non-random non-responses, an additional adjustment for “non-response in the follow-up survey” was conducted, giving inverse probability weighting to the remaining participants in each survey by modelling the probability of not dropping out [18].
The mean values and prevalence of the selected factors were calculated based on social isolation or loneliness, and the overall difference across groups was tested using a chi-square test.
We examined the association between social isolation and/or loneliness and being unvaccinated in 2022. Univariate odds ratios (ORs) and 95% confidence intervals (CIs) of social isolation and loneliness for the being unvaccinated in 2022 were calculated using logistic regression analysis. In a multivariable-adjusted analysis, the model was adjusted for the following variables: age, sex, marital status, education, annual household income, working conditions, living area, self-rated health, history of chronic conditions, being divorced, being widowed, and living alone. The association between social isolation and loneliness was not strong (correlation coefficient = 0.19), so we added both variables to the multivariable logistic regression analysis. We tested statistical interactions for sex and age by adding a cross-product term for social isolation and/or loneliness to the model. The analysis revealed a significant interaction between loneliness and age, so we performed a subgroup analysis.
Propensity Score Matching
Depending on whether or not participants were socially isolated, lonely, or equally scattered, we conducted propensity score (PS) analyses, as sensitivity analyses, to evaluate associations with being unvaccinated. Participants’ characteristics differed according to social isolation or loneliness (Table S1). PS weightings were used to account for the differences. PS (the probability of being socially isolated or lonely for each participant ranging from 0–1) was calculated through multivariable logistic regression using factoring in potential confounders [19]. To judge the success of each PS weighting in terms of creating groups that looked similar based on the observed covariates, we used standardised differences, those being the difference in proportions between the exposed and counterpart groups divided by the standard deviation in the exposed group [20]. Generally, a standardised difference of 0.1 indicates a potentially meaningful imbalance. When PS weighing creates an acceptable balanced univariate logistic regression it can be used for the result.
Additionally, we compared the answers of socially isolated and not socially isolated unvaccinated participants, and lonely and not lonely unvaccinated participants, to identify the reasons why they choose to remain unvaccinated. A chi-square test was performed to reveal any significance differences between the groups.
For validation, P values were obtained using a two-tailed test, and p < 0.05 was regarded as statistically significant. We used SAS version 9.4 (SAS Institute Inc, Cary, NC) for all statistical analyses.