The COVID-19 pandemic has resulted in unprecedented social and economic disruptions, with a profound worldwide impact on public health, lifestyle, and food security [1
]. Currently, more than 4 million people have died from COVID-19 worldwide, and more than 210 million people have been infected [2
]. In Portugal, COVID-19 has killed more than 17 thousand people, and more than one million have been infected [2
]. At the beginning of the year, Portugal implemented the second COVID-19 national lockdown to stop the spread of COVID-19 and to ease pressure on the National Health Service. Thus, the COVID-19 vaccine is a matter of public health since it is considered our main hope to tackle the coronavirus pandemic and to return to everyday life [3
]. Across several countries, the national vaccination program has started, so finding the population’s intention of being vaccinated and knowing why some people do not intend to be vaccinated is important for improving the COVID-19 vaccination communication plan [4
]. Additionally, we can predict vaccination intention at an individual and group level if we identify vaccine hesitancy determinants.
Herd immunity, estimated to occur when a large part of the population has been vaccinated and, hence, renders the coronavirus spread very unlikely, is dependent on the individuals’ willingness to be vaccinated [5
]. Although the reported willingness to be vaccinated in Portugal (75%) was found to be higher than the average willingness in the European countries (73.9%), little is known about the actual numbers for the released vaccines since these judgments were performed based on a hypothetical vaccine [6
]. With the vaccine available, it is essential to replicate these findings primarily due to the current national lockdown context that may influence people’s intention of being vaccinated. The proportion of people willing to receiving vaccination as soon as possible declined due to misinformation and the emerging concerns from governments across Europe [7
]. The AstraZeneca vaccine has suffered constant setbacks due to issues related to blood clots [8
], which decreased individuals’ trust in the vaccine and consequently reduced vaccination intention [9
Vaccine hesitancy has been widely addressed, but few studies have evaluated the success of vaccination hesitancy reduction interventions [10
]. Thus, there is a need to clarify the motives that influence vaccination intention in order to design effective intervention strategies. In a systematic review, there were several implemented strategies identified and grouped into the following categories: individual/social group influences; vaccine and vaccination-specific issues; and contextual issues [10
]. The authors found that the most effective interventions employed multiple strategies, such as increasing vaccination knowledge and awareness, improving convenience and access to vaccination, and engaging influential leaders to promote vaccination.
Specifically, COVID-19 vaccine hesitancy should be strongly considered since the vaccine’s safety and effectiveness have been highly discussed in social media due to the record time of its development process and its unknown long-term side effects [11
]. The literature suggests that individual characteristics play a significant role in the acceptance of the COVID-19 vaccine: older age [12
], male gender [12
], or religion [14
] were frequently reported as good predictors of vaccination intention.
Additionally, the fear of facing adverse side effects from the vaccine may prevent the achievement of a global immunity as people commonly intend to avoid losses more than they avoid gains. This concept was defined by Kahneman and Tversky [14
] as loss aversion. The concept has already been used to frame vaccination intention as a trade-off between risks [15
]. However, there is still a mix in literature as to whether higher risk-averse people are more afraid of the potential vaccine side effects or, on the other hand, whether they show stronger intentions of being vaccinated to avoid the negative consequences of the actual disease [4
]. Despite the fact that the role of emotion in COVID-19 vaccine communication has already been discussed, with fear being associated with vaccine hesitancy [4
], little is known about a possible loss aversion effect.
Framing the COVID-19 vaccine communication plan regarding the vaccination intention’s psychological determinants may significantly diminish vaccine hesitancy [17
]. For example, the theory of planned behavior (TPB) predicts that social norms may influence vaccination intention by shaping the individuals’ perception about the socially desired behavior—injunctive norms [18
]. Social norms are values and beliefs shared within a population, guiding the group’s behavior without explicit laws [19
]. Thus, it is essential to identify the shared beliefs and attitudes towards the coronavirus and the vaccine.
Furthermore, contextual issues such as culture, politics, media environment, or influential leaders should also be considered when addressing communication strategies as they are systematically reported as determinants of vaccine hesitancy [10
]. The increase in COVID-19 new cases in several countries lead governments to declare national lockdowns. Under current restrictions, people must stay at home, with limited exceptions for permitted reasons. Therefore, the actual conditions are of great importance to observe whether adverse settings in which the coronavirus’s negative impacts are highlighted (e.g., staying at home) influence vaccine hesitancy.
During the lockdown, individuals may recall coronavirus’s negative consequences easier and, consequently, heavily weigh its potential effects. One possible explanation, supported by the availability heuristic [20
], is that people perform likelihood judgments, e.g., calculate the risk of becoming seriously ill, based on the ease of retrieval of certain events. In the case of the COVID-19 pandemic, they may perform judgments based on the number of daily new positive cases or for how long they have been in lockdown.
2. Parental Vaccination Intention
Despite the efforts placed towards the rapid development of a COVID-19 vaccine, the actual clinical trials are not clear about the desired dose or the possible side effects for children (under 18 years old) [21
]. Children can also become infected and transmit, and develop clinical complications from the coronavirus, so there is a need to approve children’s COVID-19 vaccine uptake [21
]. However, the success of vaccinating this age group relies on the parents’ willingness to vaccinate their child/children. Parental vaccine hesitancy has also been frequently explained by psychological factors such as a higher perceived risk of the child experiencing heavier adverse effects from the vaccine than from the actual disease or the perceived ability to control children’s exposure to the disease [22
Relative to the COVID-19 vaccine, parents reported to be more likely to vaccinate themselves than their child/children, and the main reasons for not accepting a children’s vaccine were as follows: not enough evidence; safety concerns; and the belief that children are hardly affected [23
]. Given the delay in COVID-19 vaccine clinical trials for children, they will be the last age group to receive the vaccine [21
]. Thus, it becomes essential to explore the most critical factors for explaining parental vaccine hesitancy to prevent widespread parental refusal.
This study answers a public health issue and the achievement of group immunity. Our main objective is to find people’s willingness to be vaccinated and to vaccinate their children. Then, we aim to clarify the underlying factors of that choice and uncover why people do not intend to be vaccinated. These motives will be framed under personal characteristics, psychological factors, and context. Following previous research [6
], which found that psychological factors were the main predictors of vaccination intention, our first hypothesis is that the beliefs and attitudes towards the COVID-19 vaccine will be the main determinants of vaccination intention. Thus, we expect that individuals with more positive beliefs and attitudes about the COVID-19 vaccine will show higher vaccination intentions.
Following Tversky’s and Kahneman’s [20
] availability heuristic predictions, we expect that as the number of days in lockdown increases, people will be more prone to being vaccinated in order to mitigate the adverse effects of the pandemic. Relative to personal characteristics and similarly to previous studies [12
], we expect that age will positively affect vaccination intention. Other socio-demographic variables will also be included in further exploratory data analysis. Ultimately, we intend to build an artificial intelligence model for predicting vaccination intention. We hope that our findings can provide public policy managers with powerful insights into predicting and avoiding widespread COVID-19 vaccine hesitancy.
5.1. Self and Children Vaccination Intention
Participants’ vaccination intention is expressed in Figure 1
. Vaccination intention of the participants showed a left-skewed distribution (M
= 6.03; SD
= 1.63; ME
= 7.00). Vaccination intention of the participants’ children also showed a left-skewed distribution (M
= 5.91; SD
= 1.74; ME
= 7.00). Participants’ intention of receiving the COVID-19 vaccine was, on average, higher than the intention of vaccinating their children, t
(648) = 3.14 and p
5.2. Vaccination Concerns
Participants indicated that their primary concern of receiving the COVID-19 vaccine was the record time of its development process followed by the possible adverse side effects of the vaccine. Moreover, the participants frequently reported belonging to a risk group and doubting the vaccine’s effectiveness as their primary concerns. These results are presented in Figure 2
5.3. Dimensionality Reduction from the COVID-19 Attitudes and Beliefs Scale
We performed two principal component analysis (PCA), one for the COVID-19 attitudes and beliefs scale and the other for the attitudes and beliefs about the COVID-19 vaccine scale. Relative to the coronavirus scale, three main dimensions were extracted: the perceived threat of COVID-19 (α
= 0.58), trust in the management of the COVID-19 (α
= 0.84), and the impact of COVID-19. Thus, confirmatory factor analysis was performed, which showed a comparative fit index of (CFI) = 0.94, a Tucker–Lewis index of (TLI) = 0.90, and a root mean square error of approximation (RMSEA) = 0.07. Data are presented in Appendix B
5.4. Dimensionality Reduction from the COVID-19 Vaccine Attitudes and Beliefs Scale
From the vaccine scale, four main dimensions were extracted: general COVID-19 vaccination beliefs and attitudes (including, e.g., subjective norms, behavioral control, anticipated regret, and vaccine adverse side effects); the “others” intention of being vaccinated; the perceived knowledge sufficiency; and the return to normal life. Thus, a confirmatory factor analysis was performed which showed CFI = 0.82, TLI = 0.79, and RMSEA = 0.09. Data are shown in Appendix C
5.5. Multiple Regression Models of Self and Children Vaccination Intention
We built two nested multiple linear regression models, one for explaining self-vaccination intention and the other for explaining the intention to vaccinate children consisting of only the predictors that showed a significant effect in the complete original models (with all the measured variables).
The first model, relative to the self-vaccination intention, explained 65% of the variance (F(3, 645) = 397.6 and p < 0.001). It was observed that positive beliefs and attitudes towards the vaccine significantly predicted self-vaccination intention (β = 0.75, t(645) = 31.38, and p < 0.001), as did the perceived risk of the COVID-19 (β = 0.07, t(645) = 3.03, and p = 0.003), while the perceived knowledge sufficiency relative to the coronavirus and the vaccine negatively predicted self-vaccination intention (β = −0.05, t(645) = −2.05, and p = 0.040).
The model with the intention to vaccinate the children entered as the outcome variable explained 56% of the variance (F(3, 645) = 278.9; p < 0.001.) It was observed that positive beliefs and attitudes towards the vaccine significantly predicted the intention to vaccinate children (β = 0.72; t(645) = 23.47; p < 0.001), as did COVID-19 perceived threat (β = 0.09; t(645) = 3.21; p = 0.001). On the other hand, perceived knowledge of the coronavirus and the vaccine, despite the fact that it approached the significance level in the global model, did not show a significant effect in this most parsimonious model (β = −0.04; t(645) = −1.53; p = 0.13).
5.6. The Predictive Power of the General Beliefs and Attitudes towards the Vaccine
To understand the contribution of each item from the general attitudes and beliefs towards the vaccine component for predicting vaccination intention, we performed two multiple regression models: one for explaining self-vaccination intention (F
(14, 634) = 101.9; p
= 0.69); and the other for explaining children vaccination intention (F
(14, 634) = 69.44; p
= 0.61). We would like to highlight that health care professionals’ recommendation on receiving the vaccine had a significant positive effect for both self-vaccination intention (β
= 0.25; t
(634) = 6.97; p
< 0.001) and for children vaccination (β
= 0.22; t
(634) = 5.09; p
< 0.001), while the lack of confidence because the vaccine was too new showed a negative effect both for self-vaccination intention (β
= −0.13; t
(634) = −3.95; p
< 0.001) and for (β
= −0.13; t
(634) = −3.31; p
= 0.001). Then, in order to ensure reproducibility of our results, we performed the same analysis in a subset comprising 80% of participants and repeated it in the remaining 20%. The regression coefficients for each item are presented in Figure 3
(code retrieved from [30
5.7. Machine Learning (ML)
We trained an artificial neural network (ANN) machine learning model to predict individuals’ willingness to be vaccinated with respect to the recently released COVID-19 vaccine. ANNs were inspired by how the human brain works and are vastly used for pattern recognition and classification problems [31
]. ANNs have proven useful, for example, for predicting the COVID-19 outbreak [32
To develop our ANN, the entire dataset (649 participants) was divided into two parts in a ratio of 80/20: a training set consisted of 519 answers where the model learned and adjusted its predictions, and a testing set consisted of 130 entries in which we tested the quality of the model’s predictions. During the training phase, hyperparameter tuning was achieved by using exhaustive search through GridSearch with five parameters to find the model with the highest 3-fold cross-validation accuracy. The best model showed an 80% cross-validation accuracy and was defined by the following: a ReLu (rectified linear unit) activation function; a hidden layer with three nodes; a learning rate of 0.0001; the Adam (adaptative moment estimation) optimizer algorithm; and 1000 maximum iterations.
The ANNs have three layers: an input layer, hidden layers, and an output layer (a global view of our model is presented in Appendix D
and is available as (Figure A1
). The output layer consisted of a multi-classification problem in which the model was trained to predict whether the person would have the following: have the COVID-19 vaccine for sure; have moderate intention; or have low intention. These three categories were created from the 7-point Likert type question “what is your intention to be vaccinated?” where “7” was coded as having the vaccine for sure, “6–4” as having moderate intention, and “3–1” as having a low intention. These three categories were formulated from a theoretical and empirical perspective. Theoretically, we were interested in finding whether the person would have any doubts about having the vaccine, so any answer below “7” was considered showing a certain degree of vaccine hesitancy. Empirically, as the responses followed a left-skewed distribution, with more than half participants answering “7” (62.6%), it made sense to try to identify, in the general population, people who have some sort of vaccine hesitancy and to what degree that hesitancy manifests (high vs. moderate hesitancy).
The input features were chosen based on previously identified critical components through a multiple regression model: general attitudes and beliefs towards the vaccine, COVID-19 perceived risk, and perceived knowledge. Initially, we tested the model with 17 items relative to those three components. Then, we computed the relative importance of features and selected those that contributed the most to the model’s performance. The final model comprised six inputs, all belonging to the previously extracted component of general attitudes and beliefs towards the vaccine. The questionnaire’s items were used as inputs and not the components extracted from the previous PCA analysis since the model showed a higher performance with the items entered as features. The relative importance of each feature is presented in Appendix E
In the testing phase, the model showed an accuracy rate of 85% for predicting participants’ vaccination intention, with a sensitivity of 100% for identifying people with high vaccine hesitancy. When compared with other machine learning types of algorithms, our ANN showed the highest accuracy in a 10-fold cross-validation test. The confusion matrices for the model’s predictions vs. the actual values in the testing set, as well as additional evaluation metrics (precision, recall, and the F1-score) for the adequacy of the model to each class, are presented in Appendix F
and Table A5
The main aim of the present study was to find people’s intention of being vaccinated against COVID-19 in Portugal and to vaccinate their children. Moreover, we expected to find the underlying factors behind that decision and the genuine concerns towards the vaccine. For this reason, the survey analyzes participant’s characteristics, psychological factors, contextual factors, and concerns about the vaccine. Ultimately, our goal was to predict vaccination intention by using a machine learning model with the determinants of vaccine hesitancy as inputs. Therefore, we first hypothesized that psychological factors would be the significant determinants of vaccination intention, especially the general beliefs and attitudes towards the vaccine. Moreover, we expected that the lockdown would positively affect individuals’ willingness to be vaccinated and that age would positively predict vaccination intention.
Most people reported having full intentions of being vaccinated and vaccinating their children. Although nearly half the participants were not sure about having the vaccine, the principal reported concerns were included the record time of its development process and its possible adverse side effects.
In order to understand the determinants of vaccine hesitancy, we constructed two multiple regression models for explaining self and children’s vaccination intention. We found that all significant predictors belonged to the group of psychological factors, with positive beliefs and attitudes towards the vaccine being the significant determinants of vaccination willingness. However, we cannot conclude that neither contextual factors, such as the number of days in lockdown, nor personal characteristics such as age influenced vaccination intention. Regarding self-vaccination intention, we also observed that the perceived risk of contracting the coronavirus was positively linked to a higher vaccination intention. In contrast, the perceived knowledge about the disease and the vaccine seemed to reduce vaccination intention. Relative to the intention of vaccinating the children, the model showed that a higher perceived COVID-19 threat was positively linked with the willingness to vaccinate the children.
Finally, our neural network machine learning model for predicting vaccination intention could be operationalized by providing the model with inputs relative to six questions from the COVID-19 vaccine attitudes and beliefs scale [13
]. The model was able to predict with 85% accuracy whether the individual would have low, moderate, or high intention of being vaccinated against COVID-19.
6.1. Theoretical Implications
This study presents essential findings for understanding COVID-19 vaccination intention, its determinants, and the main concerns about the vaccine. Our results suggest that nearly half the Portuguese population shows a certain degree of vaccine hesitancy. Despite vaccine hesitancy being not as high as other authors previously found for the same population [34
], the results that show 56% of individuals reporting that they preferred delaying vaccine uptake and 9% reporting they refused are still alarming. The most-reported concerns are related to the vaccine’s side effects [6
] and lack of trust in the vaccine’s development process due to its record time [11
]. Moreover, despite the priority in vaccinating particularly vulnerable populations, such as cancer patients [35
], belonging to a risk group was commonly reported as a concern about having the vaccine by the participants in our study. This is explained since cancer patients’ leading concern is the vaccine’s possible side effects, similarly to the general population [36
Vaccine’s general beliefs and attitudes, containing questions regarding the vaccine’s side effects, family and “friends” approval, and the recommendation by government and public health professionals, among others, were found to be the most determinant factors in predicting COVID-19 vaccination intention. These findings support existing literature concerning the effect of social norms on COVID-19 vaccination intention [37
]. Moreover, COVID-19 perceived risk and perceived threat and the perceived knowledge about the coronavirus and the vaccine were also significant predictors for explaining self-vaccination intention and children’s vaccination intention, which is particularly relevant as these predictors all belong to the group of psychological factors, similarly to previous studies [6
]. Surprisingly, individuals with higher perceived knowledge about the coronavirus and the vaccine showed lower vaccination intentions. In an experimental study, it was observed that exposure to misinformation decreased participants’ willingness to be vaccinated, especially for scientific-sounding misinformation [7
]. Linden et al. [39
] also warned of the adverse impact of misinformation on strategies to tackle the pandemic since COVID-19 misinformation had been widely spread on social media. Thus, the negative effect of perceived knowledge on vaccination intention could be explained by participants’ prior exposure to misinformation relative to either the disease or the vaccine.
Since only psychological factors were determinants of vaccination intention, we cannot support previous findings of the importance of personal characteristics. A possible explanation for the absence of significant effects is that the influence of personal characteristics on vaccination intention varies between populations. In France [12
] and UK [13
] age had a positive effect on vaccination intention, while in the US [14
] it had no effect. Moreover, men in the US [14
] and in France [12
] showed higher vaccination intention than women, while this effect was not observed in the UK [13
]. Finally, religiosity decreased vaccination intention in the US [14
] but showed no effect neither in France [12
] nor UK [13
Additionally, contrary to our expectations, the number of days in lockdown did not influence vaccination intention. Fridman et al. [40
] tested, by using a longitudinal study, whether the perceived COVID-19 threat would increase over time and, consequently, increase the positive attitudes towards the vaccine and the vaccination intention but found an overall decrease in vaccination intention. Thus, similarly to previous findings [40
], we cannot conclude that changes in context affect people’s willingness to be vaccinated [10
]. Despite the lockdown’s restrictions on individuals’ lives, with the decreasing number of COVID-19 infections during the lockdown [41
], people could have believed that vaccination would no longer be necessary. Finally, despite our findings of some of the psychological factors behind vaccination intention, others remain to be clarified, as with risk aversion. Risk aversion did not significantly affect our multiple regression models, and it decreased the model’s performance when entered the ML model. In this manenr, we cannot provide further insights into the debate on whether risk aversion has a positive or negative role in vaccine hesitancy [42
With respect to children’s vaccination, our results support previous findings of the parents’ higher intentions of receiving vaccination against COVID-19 themselves compared to the intention of vaccinating their children [23
]. One possible explanation is that the COVID-19 perceived threat for children is so low that the advantages of receiving the vaccine are inferior to the vaccine’s potential adverse side effects [23
]. Our results support this influence of the COVID-19 perceived threat on child vaccination, as higher degrees of COVID-19 perceived threat predicted increased willingness to vaccinate children.
The ANN model supports the results of the multiple regression models, with items regarding the general beliefs and attitudes towards the vaccine being the most relevant features for predicting vaccination intention. This study shows that an algorithm can predict vaccination intention of a given population with good accuracy even with only a few inputs. In sum, ANNs complement linear regression models by capturing non-linear relationships and predicting subsequent behavior [44
]. ML models have proved to be important for understanding and predicting the coronavirus spread [45
], and our results suggest that they can also be applied to the vaccine hesitancy issue.
6.2. Practical Implications
Now that COVID-19 vaccines are available and the vaccination plan is being carried out, addressing vaccination intentions is of great importance in order to prevent widespread vaccine hesitancy. Our findings suggest that vaccine hesitancy should be cautiously considered. Although most participants reported wanting to be vaccinated and wanting to vaccinate their children, a large proportion was doubtful. Interestingly, participants who reported the maximum intention of being vaccinated still shared some of the concerns relative to the vaccine. Thus, communication strategies should consider people’s lack of trust in the vaccine’s development process and safety.
Public policymakers may find our results about the determinants of the vaccination intention essential for designing effective strategies for reducing vaccine hesitancy. The most significant predictor was the general beliefs and attitudes towards the vaccine, which encompasses, among other attributes, the vaccine’s side effects, others’ approval, and expert recommendation. Thus, communication plans may use people’s general beliefs and attitudes about the vaccine to reinforce the driving forces and reduce the restraining forces of the vaccination behavior [46
]. For example, similarly to previous findings about the human papillomavirus vaccination [47
], framing COVID-19 vaccination messages using injunctive norms could increase the interest in seeking additional information about the vaccine and, consequently, enhance confidence in receiving vaccinations. Additionally, our findings suggest using an authority principle of persuasion [48
] may reduce vaccine hesitancy. Therefore, experts (health professionals) and authority figures (government’s members) should communicate the vaccine’s importance and safety.
Additionally, our findings suggest that parental hesitancy towards expected child vaccination against COVID-19 should also be considered, as participants reported a lower intention of vaccinating their children than compared to vaccinating themselves. In addition to the already discussed strategies, we would like to highlight the importance of raising awareness about the actual COVID-19 threat for children, especially with the new variants of the virus [49
], as this was found to be a determinant of parents’ intention to vaccinate their children.
The proposed ANN may contribute to managing the vaccination plan since it allows us to predict people’s intention of receiving vaccinations against COVID-19. The model’s simplicity is another practical advantage since it only has six inputs and performs relatively well. Thus, with the results of these six questions, we can predict vaccination intention without having to ask people about their vaccination intentions directly.
6.3. Limitations and Further Studies
This study presents some limitations that should be considered. The main limitation concerns our sample characterization. The number of female participants vastly exceeded the number of male participants. The participants were, on average, relatively young, and only a few participants reported having primary and basic education as their maximum qualification. Thus, future studies must not neglect the commonly reported effects of gender, age, and education on COVID-19 vaccination intention [12
Our machine model approach was not extended to predict children’s vaccination since only 113 participants reported having minor children. However, future studies with a larger sample size may apply our ANN for predicting parents’ intention of vaccinating their children now that child vaccination has been approved.
Additionally, our ML model was trained and validated with data from the Portuguese population. Generalization of the proposed model to other populations should be taken cautiously since predictors of vaccination intention may differ due to societal and cultural beliefs [12
]. Therefore, we highlight the importance of future studies validating our model for different populations. Finally, our sample size comprised 649 participants with unleveled distributed classes (a low percentage of people showing low intentions of receiving vaccination), which poses a risk of overfitting and, consequently, results in lack of generalization power [50
]. Global datasets could be a solution to this problem since the code is freely available online. However, a global approach to solve COVID-19 related issues poses limitations since data variability may reduce the models’ performance [51
]. Therefore, a trade-off between performance and generalization power should be considered when applying the proposed ANN to other populations.