We included data from an anonymous public risk perception survey conducted in Italy and Sweden during two different periods of the COVID-19 pandemic. Detailed information on the study has been published elsewhere31. Briefly, the survey examines public risk perceptions for nine threats (epidemics, floods, droughts, earthquakes, wildfires, terrorist attacks, domestic violence, economic crises and climate change). Data was collected over a one-week period in August and November 2020. The samples were independent and drawn from two existing survey panels of 100,000 people in each country set up by market research firm Kantar Sifo32, and should be considered representative of the Swedish and Italian populations for gender and age. Around 8000 people in the pool were invited to participate, if they didn’t respond up to two reminders were sent out. The capital city regions were over-represented: with a sampling ratio of 1/9 in Italy and a sampling ratio of 4/6 in Sweden) (Supplementary Fig. 1). The missing data were quite small,
People living in the capital region were overrepresented, so specific weights were used in the analysis to account for this. The present study was approved by the Italian Research Ethics and Bioethics Committee (Dnr 0043071/2019) and the Swedish Ethical Review Authority (Dnr 2019-03242). The study was conducted in accordance with the ethical standards set by the European Union as part of Horizon 2020 (EU General Data Protection Regulation and FAIR Data Management). Participants were informed that participation was voluntary and they are giving their informal consent to participate in this study by completing the survey.
Risk perception of epidemics
The present study looked at public risk perceptions of epidemics considering seven domains: likelihood of epidemics, epidemic impact on individuals and communities, individual and agency preparedness, individual and agency knowledge of epidemics using a Likert scale of 1, minimum to 5 , max.
Predictors of risk perception
Information on direct experience with an epidemic and socio-economic factors such as age, gender, employment (yes vs. no), relative income (from 1 to 5), higher education (yes vs. no) were collected in the survey and included in the present study as possible predictors of risk perception.
The study considered excess mortality at the regional level in Italy and Sweden during the first wave of the COVID-19 pandemic (February 15 to May 15 for Italy and March 1 to May 31 for Sweden). The regional level has been defined according to the nomenclature of territorial units for statistics (NUTS) 2 of the European Union33. We have data on excess mortality in the Italian regions from Scortichini et al.17. To estimate excess mortality in Sweden, we compared the COVID-19 outbreak to the time before the outbreak. A two-stage discontinuous time-series approach based on a Poisson model with a function limiting excess risk to zero in early March 2020 was used to calculate excess mortality at the Swedish regional level34. The model was fitted for time-varying confounders such as (i) seasonality using a 3-node natural spline term, (ii) indicators for the day of the week, (iii) air temperature using a mean daily temperature term. Temperature information was retrieved from the ERA-5 reanalysis dataset of the Copernicus climate database35. We ran mixed-effects Poisson regression models with a random term for NUTS-3 administrative units to calculate excess mortality at the regional (NUTS-2) level, accounting for heterogeneity between NUTS-3 administrative units.
National policy response
The stringency index18 is a national response index and is used to quantify the measures implemented in response to the COVID-19 pandemic. The Stringency Index is a daily, country-level measurement that considers nine areas: school closures; workplace closures; cancellation of public events; restrictions on public gatherings; closures of public transport; stay-at-home requirements; public information campaigns; restrictions on internal movements; and international travel controls. In this paper, the level of national policy response was used as a four-level ecological variable (Sweden to August, Italy to August, Sweden to November, and Italy to November) and defined as the area under the curve of the stringency index for each country, between two consecutive days ending August 5, 2020 (first survey) and November 4, 2020 (second survey). This measure has been standardized to the value of Sweden in August (which is used as a reference).
Possible differences in means and confidence intervals for seven items of risk perception across countries and over time were plotted using forest charts and stratified by country and time period. Effect modification by country and period was examined using ordinal logistic regression models with risk perception (independent variables) and country and period as dependent variables. Results were presented as (i) odds ratios (ORs) for each country and period stratum, (ii) ORs for country within period strata and for periods within country strata, and (iii) interaction measures on additive and multiplicative scales36.
Second, multivariable ordinal logistic regression models were performed to evaluate the association of sex, age, occupation, relative income, higher education, and epidemic experience as possible predictors with the seven domains of risk perception (independent variables). The analysis was stratified by country and period.
Third, we examined whether risk perception differed depending on how badly an area was affected by the first wave of the COVID-19 pandemic. We compared the means and confidence intervals for seven points of risk perception between the most affected region in terms of excess mortality (Stockholm region in Sweden with about 60% excess mortality and Lombardy region in Italy with approximately 100% excess mortality) and compared to the rest the country. Ordinal logistic regression models were then performed to examine whether excess mortality at the regional level (dependent variable) was associated with areas of risk perception (independent variables) stratified by country and adjusted for sex, age, and relative income. Finally, the relationship between the level of implemented measures and risk perception was examined using fitted ordinal logistic regression models.
The use of ordinal logistic regression models was based on the assumption that the effect was linear on the logarithmic scale and that each independent variable had an identical effect for a one unit increase in the ordinal dependent variable (proportional odds). In parallel, the goodness of fit of the ordinal logistic models was tested using the deviance fit test. No multicollinearity between independent variables and no correlation between errors from the models were found.
As suggested in the literature, misinterpreting p-values carries significant risks37. Therefore, we decided to interpret the estimates in terms of the possible direction of the effects, using ORs and 95% confidence intervals (CIs) that provide information about significance. In particular, the width of the confidence interval and the size of the p-value are related: the narrower the interval, the smaller the p-value. In addition, the confidence interval provides additional information related to the size of the studied effect.
Statistical analyzes were performed using Stata version 15.0 (StataCorp, College Station, TX, USA) and R version 4.0.338.