2.1 Study design and setting

An anonymous online-based cross-sectional survey was conducted from 20th September to 9th December 2022 in 11 countries of the MENA (Egypt, Iraq, Kuwait, Lebanon, Libya, Morocco, Pakistan, Afghanistan, Sudan, Saudi Arabia, and Yemen) by spreading the questionnaire via social media (Facebook, Twitter, WhatsApp, Telegram, etc.).

2.2 Study population

Participants were eligible for participation in this study if they were older than 18 years, could read Arabic or English, could access the Internet through a computer or smartphone to answer the electronic questionnaire, and were residents in the MENA region during the COVID-19 pandemic.

2.3 Sampling method and sample size

Convenience and snowball sampling methods were used to recruit participants with the required sample sizes. Researchers used a combination of convenience and snowball sampling to recruit participants. First, collaborators recruited people who were easy to reach and from different backgrounds to ensure a diverse sample. Then, these participants helped recruit others from their social networks until we had enough participants from each country. According to Maas and Hox [20], 330 observations per country were a sufficient sample size for multilevel logistic regression models. We included 11 countries from Africa and Asia, with 330 observations per country, resulting in 3630 observations. We utilized G. Power software (statistical test: logistic model) to calculate the necessary sample size for the countries included in the study. The minimum required sample size from each country was calculated with an error rate (α) of 0.05, a power of 80%, and a probability of having any of the 5Cs (p1) at 0.4, considering the different prevalences of the 5Cs detected in the sample. The prevalence of having any of the constraints in our data set ranges between 5 and 33%. To cover the whole range, we used 40% (P1 = 0.4) as an assumption of having any of the constraints. Additionally, we accounted for the variance of one of the explanatory variables explained by the other independent variables (0.3). This moderate level of correlation (0.3) is often used in sample size calculations when precise information about the correlations is not available, but some degree of association is expected [21]. The non-response rates between 5 and 10% are typical for many surveys and do not usually compromise the validity of the results if managed properly [22].

Consequently, the minimum required sample size from each country was 307, which was increased to 330 to accommodate a non-response rate of 7% [21]. Equal sample sizes ensure that each country's data contributes equally to the overall analysis, providing consistent statistical power across countries. This uniformity can help detect meaningful differences or relationships among the 5Cs (e.g., confidence, convenience, complacency, calculation, and collective responsibility) across all countries, assuming the effect sizes are similar.

2.4 The data collection tool

The researchers designed an anonymous electronic questionnaire using Google Forms that was connected directly to an Excel sheet for automatic data transfer and analysis. To examine the practicality and accessibility of the online questionnaire, a pilot study was conducted in which each researcher sent the questionnaire to at least three individuals. In total, 185 people participated in the pilot study, representing about 5.1% of our total sample (3630 participants). The results of the pilot study indicated that the questionnaire takes 5–10 min to be fully answered and that few sentences required rephrasing for clarity. Participants were also restricted from submitting multiple responses to ensure data accuracy. To address the potential impact of language and cultural nuances on participant responses, we utilized a validated questionnaire available in both Arabic and English. The questionnaire was revised based on feedback from the pilot study, and participants were encouraged to contact the research team for clarification. These strategies aimed to ensure that the questionnaire was understood and answered accurately by participants from all cultural backgrounds.

The questionnaire consisted of two parts:

The first part included socio-demographic characteristics (participants’ age, sex, nationality, residence, marital status, occupation, and education), medical history, history of previous COVID-19 infection, COVID-19 vaccination status, and history of death from COVID-19 infection among participants’ relatives or friends. The occupations were classified into four major categories: high-skilled (non-manual) occupations include legislators, senior officials, managers, professionals, technicians, and associate professionals; low-skilled (non-manual) occupations include clerks, service workers, and market sales workers; skilled manual occupations include skilled agriculture and fishery workers, craft and related trades workers, plant and machine operators, and assemblers; others include elementary workers, nonworking participants, and students [23].

The second part involved vaccine hesitancy questions, which were investigated using a 15-item tool developed from a “5C model” of psychological antecedents to vaccination. The current research used the validated Arabic and English versions of the 5C scale, a well-established instrument for assessing the psychological antecedents of vaccination. The 5C scale has been validated in previous studies [16, 24]. Using these validated versions of the 5C scale guaranteed the reliability and validity of the data collected in the current study. The psychological antecedents of vaccination of 5C include five domains: confidence, complacency, constraints, calculation, and collective responsibility [16]. Participants responded to each of the 15 items using a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). The scores of the three questions representing each domain were calculated, and the mean scores of each domain were classified as yes and no. In general, a sense of confidence and collective responsibility are associated with a positive attitude toward vaccination, whereas increased complacency, constraints, and calculation are associated with vaccine hesitancy.

The 5C subscales were classified according to the cutoff values according to No for confidence (< 5.7 vs. ≥ 5.7), complacency (< 4.7 vs. ≥ 4.7), and constraints (< 6.0 vs. 6.0), evaluation of the calculation of COVID-19 (< 6.3 vs. ≥ 6.3), and collective responsibility (< 6.2 vs. ≥ 6.2) [25]. The questionnaire was available in Arabic and English (Supplementary material S1). The questionnaire showed an acceptable level of internal consistency, where its Cronbach’s alpha was 0.764 (0.75–0.78).

2.5 Variables

The outcome variable of this study was the 5C scale: confidence, constraints, complacency, collective responsibility, and calculations. The 5C scale was coded as a binary variable based on a threshold available in the literature [25]. This study examined the determinants of the 5C scale by applying a multilevel logistic regression model including 11 countries in the MENA region.

The independent variables were selected at two levels: country and individual. Age, sex, educational level, social status, occupation of the respondents, suffering from a chronic disease, previous COVID-19 infection, and experiencing death from COVID-19 within the family were at the individual level while the continent was added at the country level.

2.6 Statistical analysis

Statistical analysis was carried out using the R package called lme4 [26]. Counts and percentages for categorical variables were applied. The chi-squared (χ2) test of independence was carried out to check the relationship between the dependent variables (confidence, complacency, constraints, assessing calculations, and collective responsibility) and independent variables such as age, sex, educational level, occupation, suffering from a chronic disease, and previous COVID-19 infection. Statistical significance was set at P < 0.05. A point-biserial correlation was used to measure the strength and direction of the association between one continuous variable and one dichotomous variable. This is a special case of Pearson’s product-moment correlation, which is applied when there are two continuous variables; in this case, one of the variables is measured on a dichotomous scale.

2.6.1 Data analysis and model

A multilevel logistic regression model was employed to account for the clustering of observations within each country. Additionally, it assesses the impact of explanatory variables on each component of the 5C psychological factors across 11 countries in the MENA Region. Nevertheless, neglecting the hierarchical structure of the data and assuming independence among observations at the first level can lead to inaccurate model estimates. If the intraclass correlation exceeds 0.05, failing to apply a multilevel approach can bias parameter estimates and standard errors.

The single-level logistic regression model was improved by adding random slopes and intercepts to the model. The analysis was performed in consecutive steps before estimating the final model [27]. First, the null model with only the intercept was estimated with a fixed parameter, and then we added the random-effects parameter. The log-likelihood ratio test and intraclass correlation (ICC) were used to evaluate the multilevel model with a random intercept. The final model takes the following form:

$$Logit(\frac_}_})=_+\left(_+_\right)_+__+_$$

(1)

Where the proportion (\(P(_=1\)) to (1-\(P(_=1))\) was the odds of each of the 5C scale, \(_\) was the fixed intercept (the average log of odds across the countries), \(_\) was the deviation of the country intercept (the random effect of the intercept). It was assumed that the random effect followed a normal distribution with variance \(_^\). Moreover, we added the random effect part to the slope of the independent variables. \(_\) was the value of the independent variable (first-level variables), \(_\) was the fixed slope, \(_\) was the country-specific slope deviation for the selected variables, \(_\) was the value of country-level variables, \(_\) was the fixed slope of country-level variables.

To assess the goodness of fit of the models, we used the Akaike information criterion (AIC), the Bayesian information criterion (BIC), and the likelihood ratio test; the lower the indices, the better the model [28]. Additionally, the Likelihood Ratio Test (LRT) was used to compare the fit of two nested models—one with a random slope and one without. The results of the LRT reveal whether including a random slope significantly improves the model’s fit compared to a simpler model. If the LRT yields a significant p-value, it indicates that the model with the random slope provides a substantially better fit to the data, justifying the inclusion of the random slope to account for variability in the effect of the independent variable across different levels (e.g., countries).

Multi-level logistic regression model

In the following section, we present the results of the null model first with a fixed part of the intercept and then with the addition of the random effects part. The improvement in the model was checked after adding the random-effects part using the likelihood ratio test.

The odds of being confident, complacent, constrained, performing calculations, and having collective responsibility were 0.32, 0.30, 0.05, 0.36, and 0.25 in different countries. We conducted a null model using only the fixed part. The likelihood ratio test indicated a significant difference between the two models (\(^\) = 398.5, 63.6, 34.5, 343.5, and 250.8, at p < 0.001). Additionally, intraclass correlations differed across the 5C scale. However, all of them exceeded 0.05 except complacency (0.05),Footnote 1 indicating that the single-level model was not appropriate for estimating clustered data in each country. Hence, we added a random effect part to the intercept, assuming a multilevel structure of the model. The intraclass correlation coefficient (ICC) values indicate varying degrees of country-level influence on the 5C components. For “confidence” (ICC = 0.22) and “constraints” (ICC = 0.10), there is moderate variability between countries, suggesting that while country-specific factors do play a role, a substantial amount of variance occurs within countries. In contrast, “calculation” (ICC = 0.38) and “collective responsibility” (ICC = 0.33) show higher ICC values, reflecting significant country-level variability, indicating that differences between countries are notably influential for these components. Conversely, “complacency” (ICC = 0.05) has a low ICC, implying that most of the variability is within countries rather than between them. These findings underscore the varying impact of country-level factors across different psychological components of the 5Cs.

The variance of the random intercept presents the average squared deviation of the intercepts from the overall mean intercept (the fixed intercept). A higher variance indicates greater variability in intercepts among countries. The table also shows that the variances of the random effects part of the intercept (σ200) for confidence, complacency, constraints, calculations, and collective responsibility are 0.91. 0.16, 0.36, 2, and 1.64, respectively, indicating high variability in the calculation and collective responsibility model intercepts. Besides, the standard deviations are the square roots of the variances, indicating how much the intercepts vary across different countries (0.95, 0.4, 0.6, 1.4, and 1.2, respectively) (Table 1). The variations among the odds of the 5Cs in the 11 countries under study were presented in the Supplementary material S2: (Figs. A1–A5)].

Table 1 Null model results after adding random effects to the intercept, the Middle East and North African countries, 2022 (N = 3630)

After accepting the assumption of heterogeneity of the logit of odds across different countries, we conducted five models with random intercepts and added explanatory variables. The significance of the random slope for each significant independent variable was tested using the likelihood ratio test for the nested models. Consequently, the random slope of the COVID-19 infection was added to the confidence model. In addition, we added a random slope of educational level to the complacency and constraint models. Finally, we added the COVID-19 death random slope to the calculation model. Therefore, adding the random slope for the variables mentioned above indicates that the effects of these variables differ between countries (level two).Footnote 2