We analyzed data from the Japan COVID-19 and Society Internet Survey (JACSIS), a comprehensive web-based longitudinal survey designed to approximate a nationally representative distribution of adults in Japan. Specifically, this investigation included two iterations of the JACSIS survey: 2020 and 2023. In the 2020 wave, 224,389 registered panellists aged 15–79 years were initially contacted through random sampling, stratified by sex, age, and prefecture, thereby encompassing all 47 prefectures. Between August and September 30, 2020, a total of 28,000 individuals completed the questionnaire. Subsequently, in the 2023 wave, 59,219 individuals—some of whom had participated in 2020—were invited, and responses were collected from September to November, 2023, resulting in 33,000 participants. Japan experienced eight COVID-19 waves from 2020–2023: initial outbreak (March 2020), summer surge (July–August 2020), winter peak (November 2020–February 2021), Alpha variant (March–June 2021), Delta variant (August 2021), Omicron variant (early 2022) with highest cases but lower hospitalization rates, followed by two additional waves in summer and winter 2022 (9). Japan implemented a series of “states of emergency” rather than mandatory lockdowns, focusing on cluster-based contact tracing and voluntary behavioral changes (10). Japan’s vaccination campaign began in February 2021, reaching over 80% of the population with two doses by late 2022 (11, 12).

As a token of appreciation, each participant received electronic reward points (“E-points”) upon completion, and the option to withdraw at any point was clearly communicated to them. All respondents provided informed consent via an online portal before participating. The JACSIS study protocol received ethical clearance from the Research Ethics Committee of the Osaka International Cancer Institute (approval number 20084). We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

The primary outcome was the incidence of COVID-19 infection. Additionally, for participants who developed COVID-19 infection, the severity was defined the necessity of hospital admission for COVID-19 infection requiring oxygen. We defined severity of COVID-19 infection as admission requiring oxygen because it represents an established clinical indicator of disease severity that has been validated in previous studies (13–15). The incidence of COVID-19 infection was ascertained through self-reported questionnaires that assessed both infection history and the requirement for oxygen during hospitalization. The exact questions were: “Have you experienced an event where you were diagnosed with COVID-19?” and “Have you experienced an event where you were infected with COVID-19 and received oxygen therapy requiring hospitalization?”

The exposure of interest was the Social Network Diversity. We extracted nine items from the JACSIS questionnaire, which were subsequently grouped into seven categories, as previously described in detail elsewhere (5). Specifically, these variables were defined as direct interpersonal contact, including: (1) cohabiting or sharing finances with another person; (2) working in an office with at least one other individual; (3) interacting with neighbors; (4) face-to-face meetings with non-cohabiting family or relatives; (5) face-to-face meetings with friends or acquaintances; (6) participation in volunteer groups; (7) in-person participation in sports-related groups or circles; (8) in-person participation in hobby-, learning-, or culture-oriented groups; and (9) attendance at gatherings organized by municipalities or social welfare councils (e.g., salons). Responses for each item were offered on a seven-point frequency scale includes: almost every day (6–7 times a week), 4–5 times a week, 2–3 times a week, once a week, 2–3 times a month, once a month, and rarely.

To quantify SND, we assigned one point to each item if participants reported engaging in that activity at least once a month and zero points otherwise, following previous research that emphasizes the diversity (i.e., the number of different types) of social connections. Items 1 through 5 and item 9 were scored individually. In contrast, items 6, 7, and 8 (volunteer, sports-related, and hobby/learning/cultural groups) were combined into a single category, given their shared nature as organized group activities; participants were assigned one point if they engaged in any of these at least once per month. The total SND score therefore ranged from 0 to 7, with higher scores indicating greater diversity in social network participation (5).

In addition to basic participant characteristics, including age, sex, and body mass index (BMI), the questionnaire included comorbid conditions—hypertension, diabetes mellitus, asthma, and chronic obstructive pulmonary disease (COPD)— as binary variables, following previous studies (16, 17). Socioeconomic and lifestyle factors, which need to be adjusted because they have been consistently associated with variations in health outcomes and may confound the relationship between social determinants and disease severity (18), were also collected: income was classified into low, intermediate, or high categories, while smoking status (current or non-current), and alcohol consumption (none/light vs. moderate/heavy) were based on self-report. Housing status was categorized as owning one’s home, renting, or others. Educational level was grouped as “high school or lower” or otherwise, and employment status was divided into full-time, part-time, and unemployed/student. Living arrangement was classified as “living alone” or “living with others,” and marital status as married, unmarried, or others. These covariates were selected a priori to control for sociodemographic disparities, health behaviors, and chronic disease histories that may confound the relationship between the SND score and clinical outcomes.

Continuous variables are presented as the mean ± standard deviation (SD), and categorical variables are presented as counts and proportions. Baseline participant characteristics were compared between those with and without Social networks using the Student’s t-test for continuous variables and the Chi-square test for categorical variables. To evaluate potential selection bias, we compared the baseline characteristics of these eligible participants with those excluded or lost to follow-up. Given the large sample size, baseline characteristics were analyzed using standardized differences, where an absolute difference of more than 10 percent was considered important.

First, to provide an unadjusted descriptive view of event rates, we constructed bar charts illustrating the proportion of each outcome (COVID-19 incidence or oxygen-requiring admission) at each level of the social network score. Next, to flexibly capture potential nonlinear relationships between the SND score and each outcome, we visualized the associations using restricted cubic splines analysis. The optimal number of knots for the spline functions was determined by minimizing the Akaike Information Criterion (AIC), balancing model fit and complexity. Three models were constructed: Model 1 (unadjusted), Model 2 (adjusted for age, age-squared and sex), and Model 3 (adjusted for age, age-squared, sex, BMI, BMI-squared, hypertension, diabetes, asthma, COPD, income, smoking status, alcohol use, housing, and education level). To quantify these relationships, we employed Poisson regression models with robust standard errors (using sandwich variance estimators) to estimate risk ratios (RRs) and their 95% confidence intervals, with a social network score of 0 serving as the reference point. Additionally, to statistically test for non-linearity, we included quadratic terms for the SND score in our regression models. The significance of the quadratic term coefficient was assessed using robust standard errors to confirm the presence of non-linear relationships between SND and outcomes. There were no missing values in Models 1 and 2. In Model 3, 2,785 (20.3%) participants had missing values in socioeconomic variables (income and education), and we performed a complete case analysis. Details on the directed acyclic graph (DAG) are provided to illustrate the rationale for variable selection (Supplementary Figure 1). In addition, sensitivity analyses stratified by age and sex were conducted to evaluate possible interactions. Statistical significance was set at α = 0.05, and all analyses were conducted using R version 4.2.3 (R Foundation for Statistical Computing, Vienna, Austria).

Over the fixed 3-year follow-up period, 3,251 participants (23.7%) developed COVID-19 infection. Supplementary Table 2 shows the baseline characteristics of the 3,251 COVID-19 cases in our study population, stratified by their SND score. Of these, 277 individuals (8.5%) required hospital admission for oxygen supplementation. The incidence of COVID-19 infection exhibited a monotonic increase as the SND score rose (Figure 2). Among participants diagnosed with COVID-19, the relationship between SND score and the probability of requiring oxygen-requiring admission showed a U-shaped pattern, with the lowest incidence observed at an SND score of 4 (Figure 2).

Restricted cubic splines analysis in the fully adjusted model revealed that participants with higher SND scores had an increased risk of COVID-19 infection. Specifically, compared to those with an SND score of 0 (reference), participants with an SND score of 7 exhibited a significantly higher risk of infection (RR 2.49, 95% CI: 2.11–2.95) (Figure 3A). Among COVID-19 patients, restricted cubic splines analysis demonstrated a clear non-linear, U-shaped relationship between SND scores and risk of oxygen-requiring admission among infected participants. Figure 3B graphically illustrates this curve relationship, showing that participants with moderate SND scores (4, 5) had a significantly reduced risk of severe disease (RR 0.15, 95% CI: 0.11–0.30), while those with the highest SND score of 7 showed no significant risk difference (RR 0.74, 95% CI: 0.48–1.14) compared to those with an SND score of 0. To statistically confirm these non-linear relationships, we tested quadratic terms for SND score in the fully adjusted models. For COVID-19 infection, the quadratic term was statistically significant (coefficient: -0.016, SE: 0.005, p = 0.003), indicating that while infection risk increased with higher SND scores, the rate of increase decelerated at higher levels. For COVID-19 severity among infected individuals, the quadratic term was highly significant (coefficient: +0.142, SE: 0.015, p < 0.001), providing robust statistical evidence for the U-shaped relationship.

Furthermore, even after stratification by age and sex, this trend remained consistent, with higher SND scores associated with increased COVID-19 infection rates (P for interaction 0.50 and 0.75, respectively) (Supplementary Figures 3A,B), while the relationship between SND scores and severity maintained a U-shaped curve (P for interaction 0.11 and 0.16, respectively) (Supplementary Figures 4A,B).

This study has several important limitations. First, our primary outcome measures and SND scores were based on self-reported surveys rather than clinical or administrative records. Specifically, the self-reported items used to assess COVID-19 infection and oxygen-requiring hospitalization were not validated against medical records or registry data in this study. This reliance on self-reported data may have introduced recall bias or misclassification, as participants might have inaccurately remembered their past infection history, hospitalization details, or frequency of social contacts. Second, while our analyses adjusted for various demographic and clinical variables, the observational nature of the study makes it impossible to rule out unmeasured or inadequately measured confounding factors. Crucially, our dataset lacked precise information on the exact timing of infection relative to the varying waves of the pandemic, as well as detailed vaccination status (e.g., number of doses and timing). The absence of these data means we could not fully account for their potential confounding effects on the observed associations between social contact and COVID-19 outcomes. Third, the SND score used in this study was originally developed for this analysis and has not been extensively validated in other populations or settings. Since social interaction patterns can vary significantly across different cultural and demographic contexts, the generalizability of our findings to populations outside of Japan remains uncertain. Although we designed the score to closely align with established methodologies in social network research (5) to ensure construct validity, future studies are needed to validate this specific scoring system in diverse cohorts. Fourth, we must acknowledge the potential for selection bias due to attrition, as approximately 51% of the original cohort was lost to follow-up. Although we adjusted for these sociodemographic factors in our multivariable models to minimize bias, the final cohort may be more socially stable than the general population. Therefore, it is possible that those who remained in the study differed in terms of health consciousness, social network characteristics, or socioeconomic status compared to those who dropped out. Fifth, our analysis relied on a single baseline measurement of SND in 2020. We did not account for potential changes in social behavior over the three-year follow-up period, which may have fluctuated due to evolving public health policies, vaccination status, or changes in individual risk perception. This limitation may have introduced non-differential exposure misclassification, potentially biasing the effect estimates toward the null.