**2. Materials and Methods**

### *2.1. Participants and Data*

This is a secondary data analysis of 2053 Italian adults who responded to an online survey administered in March 2020, coinciding with the first wave of the pandemic [3]. Most participants were female (*n* = 1555), 480 were male and 18 reported "other". The respondents had a mean age (SD) of 35.81 (13.19). Please refer to the paper by Flesia et al. [3] for a complete description of the study. The materials are available on Zenodo (10.5281/zenodo.5523260). The present work did not require ethics approval, however the original study was approved by the University of Padova Ethics Committee for Psychological Research (protocol 3576, unique code 189B46FE116994F1A8D1077B835D83BB).

We calculated the adequacy of the sample size using Kock and Hadaya's inverse square root formula [15]. A minimum of 316 people was necessary to achieve 80 percent power, at an alpha of 0.05.

#### *2.2. Measures*

Self-control was assessed using the 13-item Brief Self-Control Scale [16]. Linder et.al. compared unidimensional and two-factor solutions and recommended that the total score be used [17]. The internal reliability of the BSCS in this sample (Cronbach's alpha = 0.84) was identical to that of previous studies.

Socio-economic status (SES) was assessed using participants' typical income, their highest level of education, and how they continued to earn money during the pandemic (i.e., salary or governmental support). These indicators were based on Green's three-item measure of socio-economic status [18]. This was chosen because of its relevance to healthrelated behavior and its parsimony. Since we did not have the exact job titles of respondents, we added a student status. This distinguished established workers and students from having the same attainments. This was necessary because approximately one-fourth of the respondents were students.

The fear of infection was assessed with the questions: (1) How much do you feel in danger of COVID 19 infection? (2) In the last period, are you paying more attention than usual to your physical symptoms? (3) Are you actively searching for information on the progress of the pandemic? These were Likert-type questions with five levels for the first two questions and six levels for the third. The questions were similar in content to "afraid of losing life", "hands getting clammy", "anxiety when watching COVID-19 news in social media" in the Fear of COVID-19 Scale [19]. The survey contained the question, *Do you currently suffer from any of the following diseases?* The available choices were: *immunosuppression, cardiovascular disease, pulmonary disease, cancer, diabetes*, and *none of the above*.

Our dependent variable was a composite of risky behaviors or intentions to disregard restrictions, which we called infractions. This was assessed with six yes-or-no questions: *(1) I respect loyally the rules imposed by ministerial ordinances, (2) I go out regularly in defiance of the ban, (3) I only go out when necessary, (4) I happened to go out for a walk in defiance of the ban, (5) I happened to go to the grocery store without real necessity, (6) I am looking for tricks to bypass the ordinances.* Questions 1 and 3 were reverse-coded to conform to the rest.

We considered self-control, SES, fear of infection and infractions as latent variables, and their respective items as indicators.

#### *2.3. Analysis*

We chose partial least squares structural equation modeling (PLS-SEM) to examine if infractions could be predicted by self-control, health conditions, SES, or a fear of infection. PLS-SEM was chosen because health conditions and socioeconomic status (SES) are more appropriately treated as formative variables instead of reflective variables. Reflective variables are latent constructs that are manifested by empirically measured indicators (or item responses) [20]. Covariance-based SEM (which is usually called SEM) considers underlying constructs as causes. In contrast, formative variables are defined by indicators that are assumed to be the causes of the latent variable [21]. Furthermore, covariance-based SEM requires that the indicators represent a normally distributed latent variable (or be categorized versions thereof) [22,23]. However, using polychoric correlations for ordinal indicators, for example, may still result in biased estimates and standard errors [24]. In contrast, PLS-SEM is a non-parametric method that handles non-normally distributed data, and both reflective and formative indicators [25].

To test hypotheses one to four, we regressed infractions against the four latent variables as shown in Model 1 (Figure 1). To examine if the presence of health conditions indirectly inhibited infractions by increasing the fear of infection, we added a path from health conditions to fear of infection in Model 2 (Figure 2). Confidence intervals and *p* values were calculated based on 5000 bootstrap replicates.

**Figure 1.** *Model 1*: Direct effects only. Please refer to Appendix A Table A1 for the exact wording of indicators. The outcome (infractions) is predicted by four latent variables indicated by circles (self-control, health conditions, SES, and fear of infection). Rectangles are the observed variables. Arrows terminating in infractions are regression coefficients. Arrows originating from a latent variable (reflective) and terminating in a rectangle represent loading. Arrows originating from a rectangle and ending in a latent variable (formative) represent weights.

**Figure 2.** *Model 2:* Direct Effects + 1 indirect Effect. The same as Model 1 except for an added path (regression coefficient) from health conditions to fear of infection. The indirect effect of health conditions on infractions is not significant.

Appendix B Models 1 and 2 were implemented in the Stata package plssem [24] and the results were visualized, assessed for quality, and checked for consistency with SmartPLS 3 [25] and ADANCO 2.0 [26]. All three programs produced identical results.

#### **3. Results**

The direct effects model (Table 1 and Figure 1) shows that only fear of infection had a significant, inverse association with infractions. The other variables had an inverse association with the outcome but were not statistically significant. The indirect effect of health conditions through a fear of infection (0.04 × −0.14) was not significant (Table 2 and Figure 2). Both models had poor predictive value for infractions (R2 = 3.2%)


The overall fit of our two models were assessed using the standardized root mean squared residual (SRMR) [27]. SRMR quantifies the discrepancy between the correlations implied our models and the observed data [28], therefore lower values are better. The SRMRs for Models 1 and 2 were 0.69 and 0.70, respectively. These were both within the suggested cut-off value of 0.80 [29]. However, the direct-effects-only model (Model 1) was more parsimonious.


**Table 2.** Model 2: Direct Effects + 1 Indirect Effect.

The quality of our measured constructs was assessed by inspecting the composite reliability (CR), the average variance extracted (AVE), and the possible multicollinearity. These indices were applicable only for the reflective latent variables (self-control, fear of infection, and infractions). CR is a measure of internal consistency (similar to Cronbach's alpha) but does not require equal loading of the indicators [25]. CR values above 0.7 are preferable, although 0.60 and above are acceptable for exploratory research [25]. AVE is the mean of indicator reliabilities for a construct and should be above 0.5 [21]. (Table 3) Compared to the Fear of COVID-19 Scale which had values of 0.88 and 0.51 for CR and AVE respectively, *fear of infection* had 0.77 and 0.54. Multicollinearity is indicated by a variance inflation factor (VIF) exceeding 3.0 [21]. None of our indicators (items) were collinear, with a VIF which ranged from 1.00 to 1.76 (Appendix A Table A1).

**Table 3.** Reliability of Reflective Latent Variables.

