Under the Hong Kong Jockey Club Smart Family-Link Project, the current study was based on the population-based cross-sectional “Family amid COVID-19 survey 2” (FamCov-2) which was conducted using telephone interviews and online questionnaires from 22 February to 23 March 2021. We conducted the FamCov-1 in May 2020 when the second wave of COVID-19 outbreak was under control to assess individual attitudes, concerns, behaviors, and personal and family well-being, including fear (37), COVID-related information sharing (41), and delay in doctor consultation (36). Based on the FamCov-1, we conducted the second phase survey FamCov-2 in March 2021 when the fourth wave of COVID-19 outbreak was under control to assess the longer-term responses to the pandemic, including anti-epidemic fatigue and confidence (16), and adversity coping (42). The inclusion criteria of participants were those who (1) were Hong Kong residents aged 18 years or above, (2) could read or communicate using Cantonese, and (3) had landline or mobile telephone numbers or email accounts to contact. The Hong Kong Public Opinion Research Institute (HKPORI), a well-known local survey agency, was commissioned to conduct the survey.

The survey design and methods have been published (16, 42, 43). Specifically, this survey used a combination of probability and non-probability sampling methods. For the probability sampling, we adopted random telephone interview and probability-based online questionnaire survey. In both the landline and mobile telephone interviews, all telephone interviews were conducted by trained interviewers using random numbers generated by known prefixes assigned to telecommunication service providers. Invalid and working telephone numbers were identified by computers, and the remaining numbers were mixed in random to produce the number pool. A second-level sampling was only adopted in the landline telephone survey using the “next birthday rule” to select one eligible respondent among all those present in a household (i.e., select the qualified family member whose next birthday is nearest to the interview date). The online questionnaire survey invited HKPORI’s probability-panel, representative of the Hong Kong population, who were randomly selected in previous telephone surveys before this study. Totally 1,522 participants (62.9% out of the 2,420 valid telephone numbers) completed the telephone interview and 641 participants (14.9% out of the 4,311 probability-panel) completed the online survey, respectively. For the non-probability sampling, we adopted online questionnaire survey by inviting a convenience sample of HKPORI’s non-probability-panel (voluntarily registered members). Totally 5,372 participants out of the 44,514 non-probability-panel who opened the email invitations completed the online questionnaires.

The questionnaire comprises three subsets of questions on family communication, COVID-19 information, and COVID-19 influence. Each subset included core questions (which were present in all subsets) and subset-specific questions, and was randomly distributed to a third of all participants. The present analysis is based on the subset of COVID-19 influence, which was answered by 2,251 participants. Ethics approval was granted by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (Reference number: UW 20-651). Oral consent for the telephone survey or written informed consent for the online survey was obtained from each participant before the survey commenced. To ensure confidentiality, we used anonymous questionnaire, stored data using a password, and restricted the use of data.

As assessment tools were not available to measure anti-epidemic fatigue, anti-epidemic confidence, and individual and family members’ fear of COVID-19 during our survey time, our research team designed a single item to efficiently and effectively assess each of these variables for Hong Kong community adults. Anti-epidemic fatigue was measured using a single question, “Do you think you have anti-epidemic fatigue?.” Anti-epidemic confidence was measured using a single question, “How much confidence do you have to cope with the COVID-19 pandemic?.” Individual fear and family members’ fear of COVID-19 were measured using a single question, “Has COVID-19 caused you/your family member(s) fear?,” as used in our previous studies (36, 37). Participants responded to each question using an 11-point scale ranging from 0 to 10, with higher scores indicating higher levels of the assessed variable.

Family well-being was measured using family health, harmony, and happiness (3Hs): three separate questions asked, “How healthy/harmonious/happy do you think your family is?” and were answered using an 11-point scale ranging from 0 to 10, which was developed and validated in our previous studies (44, 45). In addition to the three subscales of family 3Hs, we calculated a composite score of family well-being by dividing the sum of family 3Hs by three as in our previous studies, with higher scores indicating higher levels of family well-being (41, 42, 46). In the present study, the Cronbach’s α coefficient of the family 3Hs was 0.94.

Psychological distress was measured using the four-item Patient Health Questionnaire (PHQ-4), including 2 two-item subscales measuring anxiety and depressive symptoms (47). The Chinese version of PHQ-4 has shown good reliability and validity (48, 49). Participants indicated the frequency of the described symptoms over the previous 2 weeks using a 4-point Likert scale from 0 (not at all) to 3 (nearly every day). The total scores for PHQ-4 ranged from 0 to 12, with higher total scores indicating higher psychological distress. In the present study, Cronbach’s α coefficient was 0.88 for the PHQ-4, and 0.82 and 0.79 for the subscales measuring anxiety and depressive symptoms, respectively.

We collected information on socio-demographic characteristics, including sex, age, education level, monthly household income, and number of cohabitants. We recoded age as “18–44 years” or “≥ 45 years” and education level as “secondary or below” or “tertiary.” Monthly household income per person (income divided by household size) was calculated and dichotomized as “lower” or “higher” with reference to the size-specific median monthly household income in Hong Kong’s census statistics (50). Except for the socio-demographic characteristics, each variable was treated as continuous variable without cut-off value.

Descriptive statistics of socio-demographic characteristics were computed and weighted by sex, age group, and education level based on data from the 2019 Hong Kong census (51). Hierarchical regression analysis was used to examine the associations of individual and family factors with psychological distress (Hypotheses 1a to 1d), using a regression coefficient (beta, β) with 95% confidence interval (CIs) to estimate the strength and direction of the associations. To be specific, in Step 1, socio-demographic characteristics were added to examine their associations with psychological distress. In Step 2, individual factors (i.e., anti-epidemic fatigue and confidence) were added to examine the associations between individual factors and psychological distress beyond the effects of socio-demographic characteristics. In Step 3, family factors (i.e., family members’ fear of COVID-19 and family well-being) were added to examine the associations between family factors and psychological distress beyond the effects of individual factors and socio-demographic characteristics. Finally, in Step 4, individual fear was added to preliminarily examine its mediating effect on the associations aforementioned (52).

Structural equation model (SEM) was used to examine the mediating effect (i.e., indirect effect) of individual fear of COVID-19 by decomposing the aforementioned associations into direct and indirect effects (Hypothesis 2) (53). Missing values were handled using full information maximum likelihood estimation. Standardized coefficients and bias-corrected (BC) 95% CI of the direct and indirect effects were estimated by the maximum likelihood and bootstrap methods with 1,000 replications, respectively. We examined the fit indices of the SEM, including (1) root mean square error of approximation (RMSEA) and standardized root mean square residual (SRMR), with RMSEA < 0.06 and SRMR < 0.08 considered as good fit indices; and (2) comparative Fit Index (CFI) and Tucker-Lewis index (TLI), with CFI > 0.95 and TLI > 0.95 considered as good fit indices (54, 55).

Multi-group analysis was used to investigate whether the SEM results differed according to socio-demographic characteristics (Hypothesis 3). To be specific, we conducted four multi-group analyses to test whether the SEM results were equal across sex (male vs. female), age group (18–44 years vs. ≥ 45 years), education level (secondary or below vs. tertiary), and monthly household income per person (lower vs. higher), including three steps for each multi-group analysis (56, 57). First, an unconstrained model (all path coefficients were freely estimated across different groups) and a fully constrained model (all path coefficients were constrained to equality across different groups) were compared using likelihood-ratio tests. Second, if the result of the model comparison indicated a statistically significant difference, we released the constraint of the specific path coefficient with the largest modification index in the fully constrained model and obtained a partially constrained model. Third, we compared the partially constrained model with the unconstrained model. The last two steps were repeated until there was no statistically significant difference between the partially constrained and unconstrained models.

In hierarchical regression analysis, we used the total scores of PHQ-4 and the composite score of family 3Hs to indicate psychological distress and family well-being, respectively. In SEM and multi-group analysis, we used the two PHQ-4 subscales measuring anxiety and depressive symptoms to indicate the latent variable of psychological distress, and the three family 3Hs subscales to indicate the latent variable of family well-being. We used STATA/SE 17.0 and the “lavaan” package in R 4.2.1 for these statistical analyses (58). Two-sided p < 0.05 was considered statistically significant.

Table 1 shows that, after weighting, about half of the participants were female (52.6%). Most participants were 45 years old or above (58.7%), had secondary education level or below (61.2%) and lower monthly household income per person than the median level in Hong Kong’s 2019 census statistics (56.0%). The mean number of cohabitants was 2.20 ± 1.35. For socio-demographic characteristics, the differences between unweighted and weighted data in education level reached a medium level of effect size (Cramer’s V = 0.37), and in monthly household income per person reached a small level of effect size (Cramer’s V = 0.13). However, for individual and family factors, the weighted data showed little difference with the unweighted data.

Supplementary Table 1 shows that all variables of individual and family factors and psychological distress had significant bivariate correlations, except for the association (1) between individual fear of COVID-19 and family well-being, and (2) between family member’s fear of COVID-19 and family well-being.

Table 2 shows the results of hierarchical regression analyses. When adding socio-demographic characteristics in Step 1, age of 45 years or above (β = −1.34, 95% CI [−1.60, −1.07], p < 0.001), tertiary education (β = 0.37, 95% CI [0.04, 0.69], p = 0.03), and higher monthly household income per person (β = −0.37, 95% CI [−0.64, −0.10], p = 0.007) were associated with higher psychological distress. Socio-demographic characteristics explained 7% of the variance of psychological distress. When adding individual factors in Step 2, anti-epidemic fatigue was positively associated with psychological distress (β = 0.23, 95% CI [0.18, 0.28], p < 0.001), while anti-epidemic confidence was negatively associated with psychological distress (β = −0.29, 95% CI [−0.36, −0.22], p < 0.001), supporting H1a and H1b, respectively.

When adding family factors in Step 3, family members’ fear of COVID-19 was positively associated with psychological distress (β = 0.11, 95% CI [0.05, 0.16], p < 0.001), while family well-being was negatively associated with psychological distress (β = −0.57, 95% CI [−0.63, −0.51], p < 0.001), supporting H1c and H1d, respectively. Individual factors accounted for 8% of the variance of psychological distress, and family factors accounted for an additional 13%. Using the subscales of PHQ-4 measuring anxiety and depressive symptoms as two separate outcomes in regression analyses, we found that individual and family factors remained significant, although with smaller coefficients compared to those found when using the total scores of PHQ-4 scores (see Supplementary Tables 2, 3).

When individual fear of COVID-19 was added in Step 4, the coefficient and p-value of anti-epidemic confidence both appeared to decrease slightly, suggesting that individual fear of COVID-19 might have a partial mediating effect; the association between family members’ fear of COVID-19 and psychological distress became non-significant, suggesting that individual fear of COVID-19 might have a full mediating effect.

The hypothesized model with the mediating effect of individual fear of COVID-19 showed good fit indices: χ² = 157.51, df = 29, RMSEA = 0.04, SRMR = 0.01, CFI = 0.99, TLI = 0.98. Supplementary Figure 1 shows the path diagram with the standardized regression coefficients of the hypothesized model. We adjusted sex, age group, education level, monthly household income per person, and number of cohabitants in this model.

Table 3 shows that, for individual factors, the mediating effect of individual fear of COVID-19 was statistically significant on the association between anti-epidemic fatigue and psychological distress (standardized β = 0.010, BC 95% CI [0.003, 0.020], p = 0.011), and the association between anti-epidemic confidence and psychological distress (standardized β = −0.013, BC 95% CI [−0.026, −0.005], p = 0.009), respectively. Given both direct effects were statistically significant, the mediating effect was partial (proportion mediated: 5.92% and 18.31%, respectively). For family factors, the mediating effect of individual fear of COVID-19 was statistically significant on the association between family members’ fear of COVID-19 and psychological distress (standardized β = 0.046, BC 95% CI [0.016, 0.082], p = 0.005), but not statistically significant on the association between family well-being and psychological distress (standardized β = 0.004, BC 95% CI [−0.001, 0.010], p = 0.089), partly supporting our H2. Given the direct effect of family members’ fear of COVID-19 on psychological distress was not statistically significant, the mediating effect was full.

As Supplementary Table 4 shows, the comparisons of multiple-group models stratified by sex [Δχ²₍₁₇₎ = 23.21, p = 0.142], age group [Δχ²₍₁₇₎ = 24.45, p = 0.108], and education level [Δχ²₍₁₇₎ = 17.73, p = 0.406] indicated that the SEM results did not differ by these characteristics. However, when stratified by monthly household income per person, the model comparison showed that the fully constrained model fit worse than the unconstrained model, indicating significant differences in one or more path coefficients across the groups [Δχ²₍₁₇₎ = 29.72, p = 0.028]. Supplementary Table 5 shows that participants in the lower income group reported lower anti-epidemic fatigue, anti-epidemic confidence, and family well-being than those in the higher income group.

We repeated the iterative process to obtain a final partially constrained model (χ² = 464.04, df = 88, RMSEA = 0.07, SRMR = 0.07, CFI = 0.96, TLI = 0.95), which was not statistically significantly different from the unconstrained model [Δχ²₍₁₅₎ = 24.39, p = 0.059]. In the final model, we released the constraints on two path coefficients in the fully constrained model: (1) the direct effect of anti-epidemic confidence on psychological distress, and (2) the direct effect of family well-being on psychological distress, indicating that these two path coefficients were significantly different across lower and higher income groups. Supplementary Figure 2 shows that the direct effect of anti-epidemic confidence on psychological distress was not statistically significant in the lower income group (standardized β = −0.03, p = 0.359) but was statistically significant in the higher income group (standardized β = −0.10, p = 0.001). Besides, the direct effect of family well-being on psychological distress was slightly stronger in the lower income group (standardized β = −0.49, p < 0.001) than in the higher income group (standardized β = −0.42, p < 0.001).