Positive attitudes towards COVID-19 vaccines: A cross-country analysis.

COVID-19 severely impacted world health and, as a consequence of the measures implemented to stop the spread of the virus, also irreversibly damaged the world economy. Research shows that receiving the COVID-19 vaccine is the most successful measure to combat the virus and could also address its indirect consequences. However, vaccine hesitancy is growing worldwide and the WHO names this hesitancy as one of the top ten threats to global health. This study investigates the trend in positive attitudes towards vaccines across ten countries since a positive attitude is important. Furthermore, we investigate those variables related to having a positive attitude, as these factors could potentially increase the uptake of vaccines. We derive our text corpus from vaccine-related tweets, harvested in real-time from Twitter. Using Natural Language Processing (NLP), we derive the sentiment and emotions contained in the tweets to construct daily time-series data. We analyse a panel dataset spanning both the Northern and Southern hemispheres from 1 February 2021 to 31 July 2021. To determine the relationship between several variables and the positive sentiment (attitude) towards vaccines, we run various models, including POLS, Panel Fixed Effects and Instrumental Variables estimations. Our results show that more information about vaccines' safety and the expected side effects are needed to increase positive attitudes towards vaccines. Additionally, government procurement and the vaccine rollout should improve. Accessibility to the vaccine should be a priority, and a collective effort should be made to increase positive messaging about the vaccine, especially on social media. The results of this study contribute to the understanding of the emotional challenges associated with vaccine uptake and inform policymakers, health workers, and stakeholders who communicate to the public during infectious disease outbreaks. Additionally, the global fight against COVID-19 might be lost if the attitude towards vaccines is not improved.

1. Introduction

In an attempt to curb the spread of COVID-19, minimise the loss of life and take the pressure off the national health systems, governments worldwide started their vaccine rollout campaigns late in December 2020. However, this rapid rollout of the COVID-19 vaccine has created different emotional responses across the globe. This is problematic since receiving the COVID-19 vaccine is the best possible solution to open up economies and prevent further loss of life. Compounding the problem is the mistrust in governments’ abilities to procure and administer the rollout of vaccines and the spread of fake news by anti-vaxxers (for example, see Sharma [1] and Bonnevie et al. [2]). Spreading fear and anxiety is a significant problem because we know from the existing literature that vaccine efficacy depends not only on the vaccine, but also on the characteristics of the vaccinated (Madison et al. [3], Glaser et al. [4]). Unfortunately, the COVID-19 pandemic has already led to increased depression, loneliness, and stress levels, increasing the efficacy problem (Madison et al. [3]). Adding to this, vaccine hesitancy is growing worldwide and is seen as one of the top ten threats to global health. This makes it easy to see that governments today face a significant challenge.

To this end, our primary aim is to conduct a cross-country panel analysis to investigate the trend in positive attitudes towards COVID-19 vaccines over time. This will enable us to determine whether people are becoming more or less positive towards accepting the vaccine and likely reasons for these trends. A secondary aim lies with determining those variables which are significantly related to a positive vaccine attitude and can inform policymakers.

Previous studies (Lyu et al. [5], Xue et al. [6], Chopra et al. [7]) analysed the emotions in vaccine-related tweets. However, their primary aim was to better understand the public perceptions, concerns, and emotions related to COVID-19 vaccine topics and discussions on social media. They determined the sentiments related to topics and discussions and investigated the strength of discussions and sentiments over time. The main limitations of these studies include that they: i) only analysed English tweets, with no attention being paid to specific geographical areas or comparing the sentiment across different countries, ii) did not use sentiment analysis in further analyses, and iii) did not investigate the variables related to positive vaccine attitudes. We overcome these limitations by constructing a daily time-series called the Vaccine Positive Attitude Index (VPAI), a real-time measure of people’s positive attitudes toward the COVID-19 vaccine across ten countries for the period 1 February 2021–31 July 2021. The countries span both the Northern and Southern hemispheres and include Australia, Belgium, Germany, Great Britain, France, Italy, the Netherlands, New Zealand, South Africa and Spain.

We derive the VPAI using Big Data by extracting a live stream of tweets for specific geographical areas which contains a list of vaccine-related keywords. After the data is cleaned, we use Natural Language Processing (NLP) to derive the sentiment and emotions of the tweets. After calculating the mean levels of positive sentiment per day, we investigate the trend over time in the VPAI and compare it across our panel of ten countries under investigation. Additionally, we determine which variables are related to the VPAI and, therefore, when addressed, could create a more positive attitude and increase the uptake of COVID-19 vaccines. To limit the effect of confounding factors, we introduce various estimation techniques to address possible endogeneity. We use Pooled Ordinary Least Squares as a base model and extend the analyses to include Panel Fixed Effects and Instrumental Variables regressions.

Our results indicate that the VPAI trends downward over time for the whole sample. We find the same results considering the Northern and Southern hemisphere subsamples. Considering the trends per country, we find in all countries a downward trend except Belgium and the Netherlands, which shows a slightly positive trend. Therefore, interventions are needed to change the attitude toward vaccines and increase the uptake. Our results show those variables that can improve the positive attitude towards the COVID-19 vaccines are information-related to the safety and expected side effects of the vaccines, improving trust in vaccines, reviewing regulations implemented to limit the spread of the vaccines as it seems that people weigh-up the benefits of being vaccinated against lockdown regulations. Additionally, increasing trust in governments to procure and effectively roll out vaccinations should be a priority. Furthermore, social media platforms, such as Twitter, should launch targeted campaigns focusing on educating people about the safety of vaccines, providing progress on the rollout and encouraging all ages to get vaccinated. We are confident that if the factors found significant in the econometric models (confidence levels of 95 per cent or more) are addressed, the positive attitudes towards vaccines will improve. Policy interventions in line with these recommendations will contribute to the universal plan to restore global health and the world economy.

The rest of the paper is structured as follows. The next section contains a brief background of the countries used in our analyses and studies on COVID-19 vaccine hesitancy. Section 3 describes the data and the selected variables, and outlines the methodology used. The results and discussion follow in section 4, while the paper concludes in section 5.

2. Background and literature review

2.1 Country background

This study focuses on three Southern hemisphere countries; South Africa, New Zealand and Australia and seven Northern hemisphere countries; Belgium, Germany, Great Britain, France, Italy, the Netherlands, and Spain. Primarily the choice of countries is determined by data availability. However, the dataset can be extended to include more countries in future studies. The current selection of countries from both hemispheres provides unique insights into people’s attitudes to the COVID-19 vaccine. Table 1 summarises key facts for each country used in the current study.

Table 1

Key summary facts of countries in this study.

Country	Total population	Average happiness levels** (2020)	Oxford Stringency Index (Average for the period)	First confirmed COVID-19 case (2020)	Date of first lockdown (2020)	Total confirmed COVID-19 cases (28 August 2021)	Total confirmed COVID-19 deaths (28 August 2021)	Date of vaccine rollout	Percentage of the population fully vaccinated (31 July 2021)
Australia	25.5 million	7.09	58.64	25 January	17 March*	51,256	999	22 February 2021	15%
Belgium	11.6 million	6.98	58.30	4 February	13 March	1.18 million	25,360	28 December 2020	60%
France	66.99 million	6.66	63.15	24 January	17 March	6.81 million	114,506	27 December 2020	48%
Germany	83.02 million	7.08	72.71	27 January	22 March	3.93 million	92,136	27 December 2020	52%
Great Britain	66.65 million	7.17	66.70	31 January	23 March	6.73 million	132,699	8 December 2020	56%
Italy	60.36 million	6.39	74.90	30 January	9 March	4.52 million	129,002	27 December 2020	52%
Netherlands	17.28 million	7.73	64.88	27 February	15 March¶	1.97 million	18,339	6 January 2021	54%
New Zealand	5.5 million	7.14	26.80	28 February	26 March	3,465	26	19 February 2021	15%
South Africa	57.7 million	6.32	51.90	6 March	27 March	2.76 million	81,461	17 February 2021	5%
Spain	46.94 million	6.40	64.33	31 January	14 March	4.83 million	84,000	27 December 2020	58%

* Australia never officially went into a complete lockdown such as that seen in the other countries. We used the day when the closure of international borders was announced as a proxy for “lockdown.”

¶ The Netherlands started a so-called ’intelligent lockdown’ on this date.

** The happiness scores cited here reflect the average for the period in 2020 before the first COVID-19 case was announced.

Sources: Hale et al. [8], Greyling et al. [9], Google [10, 11], Roser et al. [12], Mathieu et al. [13].

From Fig 1, we can see that the country performing the worst in terms of the total number of people fully vaccinated is New Zealand (approximately 750,000 people). However, if we consider the vaccinated as a percentage of the total population, South Africa performs the worst with 5 per cent as of 31 July 2021. Of interest is the Northern-Southern hemisphere split. The Northern hemisphere outperforms all three the Southern hemisphere countries (Australia, New Zealand and South Africa). Of the Northern hemisphere countries, France is the worst performer with 48 per cent fully vaccinated, whereas Belgium is the best performer (60 per cent) (Mathieu et al. [13]).

Fig 1

COVID-19 number of people vaccinated and the percentage of fully vaccinated people per country (31 July 2021).

Source: Mathieu et al. [13].

2.2 Literature on COVID-19 vaccine hesitancy

There is an exponential growth of studies in the literature on COVID-19 vaccine hesitancy as researchers from all disciplines addresses one of the biggest global health threats.

Research regarding COVID-19 vaccine hesitancy spans across both online surveys (see, for example, Akarsu et al. [14], Fisher et al. [15], Freeman et al. [16], Ward et al. [17], Seale et al. [18]) and in-person surveys (see, for example, Paul et al. [19], Sallam [20]). Primarily these studies found people’s hesitancy and refusal of the COVID-19 vaccine were mostly attributed to i) fear driven by possible side effects of the vaccine, and ii) the unreliability of what is seen as a new vaccine. Paul et al. [19] conducted a study involving surveying 32,361 participants from 7 September to 5 October 2020. The authors found that distrustful attitudes towards vaccination were higher amongst individuals from ethnic minority backgrounds, with lower levels of education, lower annual income, poor knowledge of COVID-19, and poor compliance with government COVID-19 guidelines. Apart from the Paul et al. [19] study, the other aforementioned studies found willingness to take the COVID-19 vaccine was closely related to one’s sense of collective responsibility and campaigning for the ’greater good’. Furthermore, these studies highlighted a need for better and more transparent information, the role of anti-vaccination campaigns, and a lack of trust in the government. Interestingly, it was found that low rates of COVID-19 vaccine acceptance were reported in the Middle East, Russia, Africa and several European countries.

Our current study uses Big Data to construct a Vaccine Positive Attitude Index (VPAI); therefore, the rest of the literature review will focus on those that also use Big Data, with special attention to three studies closest to ours in spirit. We note there is a burgeoning of literature using Big Data in the form of Twitter to analyse vaccine-related topics. Therefore, we cannot possibly discuss all of them.

For example, Yousefinaghani et al. [21] used vaccine-related tweets to track frequent hashtags, frequent mentions, main keywords, and main themes with positive and negative sentiments in the tweets. Hussain et al. [22] used Facebook and Twitter to study people’s hesitancy, perceptions and sentiment towards the COVID-19 vaccine. Küçükali et al. [23], Nuzhath et al. [24], Bonnevie et al. [2] and Thelwall et al. [25] all identified prominent themes about vaccine hesitancy and refusal on social media during the COVID-19 pandemic. These studies found that the most frequent themes that elicit a negative sentiment are anti-vaccination, poor scientific processes, conspiracy theories, mistrust of scientists and governments, lack of intent to get a COVID-19 vaccine, freedom of choice, and religious beliefs.

Sharma et al. [1] and Bonnevie et al. [26] focused on using Twitter to identify suspicious coordinated accounts in the dataset to find misinformation campaigns that drive the conversation against getting the COVID-19 vaccine. Based on an analysis of the collective behaviours and activities of accounts, they found that they correspond to a ’Great Reset’ conspiracy theory and ten additional themes such as research and clinical trials and vaccine ingredients.

Three studies that come the closest to ours in spirit are Lyu et al. [5], Xue et al. [6] and Chopra et al. [7]. Lyu et al. [5] used 1.5 million English vaccine-related tweets collected between March 2020 and January 2021 and categorised the tweets into 16 topics grouped into five overarching themes. Their results showed that under their first theme called "Opinions and Emotions Around Vaccines and Vaccination", the topic out of all 16 topics that were mostly tweeted was opinions about vaccination. In terms of their sentiment analysis (using the Syuzhet lexicon) they found that, apart from fluctuations throughout the period, the sentiment increased regarding the COVID-19 vaccine. Their emotions analysis (using the NRC lexicon) found trust was the most prevalent emotion, followed by anticipation and fear. They found that before Moderna, one of the first to test their COVID-19 vaccine on humans in April 2020, fear was the most prevalent emotion.

Xue et al. [6] analysed 4 million English vaccine-related tweets using a list of 20 hashtags from 7 March to 21 April 2020. Their main aim was to identify popular unigrams (one word) and bigrams (two words), salient topics and themes, and sentiments in the collected tweets. In terms of unigrams, they found "virus", "lockdown", and "quarantine" to be the most popular. Bigrams "COVID-19", "stay home", "corona virus", "social distancing" and "new cases" was the most popular. Furthermore, they identified 13 discussion topics from the tweets and categorised them into five different themes. For example, theme 1, "Public health measures to slow the spread of COVID-19", included topics such as face masks, quarantine, test kits, lockdown, safety, vaccine and US shelter-in-place. Their emotions analysis (using the NRC lexicon) showed that anticipation followed by fear, trust, and anger were prevalent across 12 of the 13 topics.

Chopra et al. [7] collected 1.8 million English vaccine-related tweets from across India, the United States, Great Britain, Brazil, and Australia from June 2020 to April 2021. They aimed to create ten lexical categories, split between two classes, namely emotions (6 categories) and influencing factors (4 categories) and study the temporal evolution of these categories across time. The lexical emotions category includes hesitation, sorrow, faith, contentment, anticipation and rage, while their influencing factors are misinformation, vaccine rollout, inequities, and health effects. The authors used the word-count approach to measure each category’s strength in a given tweet. They calculated the strength of the categories monthly and split their period under investigation in two; Before and After the date when each country’s government approved the first COVID-19 vaccine. Their results differed across countries with, for example, India experiencing a decrease in the strength of hesitation experienced after vaccine approval, with mentions of health effects contributing the most in tweets with a positive hesitation score. The United States experienced a significant increase in contentment after their vaccine approval. Rage and discussions on misinformation became significantly higher after vaccine approval in India, whereas the opposite was true for the United States.

Given the above literature review, no other study has done what we propose to do. We will be the first study to use Big Data to determine the sentiment and emotions related to COVID-19 vaccines through a vaccine positive attitude index. Additionally, no other study has followed the trends in attitudes over time and derived emotion and sentiment time-series data across countries to determine the variables that significantly influence a positive attitude towards the COVID-19 vaccine.

3. Data and methodology

3.1 Data

In the analyses, we use a cross-country panel dataset with high-frequency daily data (see section 3.2). We analyse the time period from 1 February 2021 to 31 July 2021 (181 days) across ten countries.

3.1.1 Constructing time-series data using sentiment and emotions analysis

To derive our time-series data which capture sentiment and emotions, we construct variables using Big Data by extracting tweets from Twitter. In our analysis, we extracted two sets of tweets based on keywords, the one related to COVID-19 vaccines and the other related to government. The tweets containing these words amounted to 1,047,000 tweets. We extracted all tweets according to specific geographical areas (country).

The first step in our analysis is to determine the tweets’ language (we detected 64 different languages), and all non-English tweets were translated to English. After the translation process, we use NLP to extract the sentiment and the underlying emotions of the tweets. To test the robustness of the coding of the sentiment of the translated tweets, we use lexicons in the original language, if available, and repeat the process. We compare the coded sentiment of the translated and original text and find the results strongly correlated.

We make use of a suite of lexicons. Each of them differs slightly but with the primary aim to determine the sentiment of unstructured text data. The two lexicons mostly used in our analysis are Sentiment140 and NRC (National Research Council of Canada Emotion Lexicon developed by Turney and Mohammad [27]). The other lexicons are used for robustness purposes and are part of the Syuzhet package. The lexicons include Syuzhet, AFINN and Bing. The sentiment is determined by identifying the tweeter’s attitude towards an event using variables such as context, tone, etc. It helps one form an entire opinion of the text. Depending on the lexicon used, the text (tweet) is coded. For example, if a tweet is positive, it is coded as 0, if neutral 2 and if negative 4.

We use the NRC lexicon to code the sentiment (as explained above) and analyse the underlying tweets’ emotions. It distinguishes between eight basic emotions: anger, fear, anticipation, trust, surprise, sadness, joy and disgust (the so-called Plutchik [28] wheel of emotions). NRC codes words with different values, ranging from 0 (low) to 8 (the highest score in our data), to express the intensity of an emotion or sentiment.

To construct the time-series data, we use the coding of the tweets and derive daily averages. In this manner, we derive a positive sentiment, a negative sentiment and eight emotion time-series. We derive the sentiment time-series using different lexicons as a robustness test and compare these results using correlation analyses. We perform various additional robustness tests, for example, to determine whether the sampling frequency significantly influences the results.

To test the robustness of the frequency, we construct the relevant index (time-series) per day (the norm), we repeat the exercise but construct the time-series per hour. We find similar trends in our hourly and daily time-series, indicating that the timescale at which sampling takes place does not significantly influence the observed trend.

To test whether the volume of tweets affects the derived time-series data, we extract random samples of differing sizes from the daily text corpus of tweets. The time-series based on these smaller samples (50 per cent and 80 per cent of the daily extracted tweets) are highly correlated to the original time-series.

3.2 Selection of variables

3.2.1 The outcome variable: Vaccine positive attitude index (VPAI)

To construct the VPAI index, we follow the method explained above and extract COVID-19 related tweets using the keywords: vaccinate, vacc, vaccine, Sputnik V, Sputnik, Sinopharm, Astrazeneca, Pfizer (if NEAR) vaccine, Pfizer-BioNTech, Johnson & Johnson, and Moderna.

To ensure that the extracted vaccine-related tweets discuss attitudes related to receiving the COVID-19 vaccine, we first constructed word clouds per country. For example, Fig 2 illustrates the word cloud generated for Great Britain. After generating word clouds for all countries, we returned to the original tweets and confirmed the context of the words with high frequencies. We determined that these vaccine-related tweets are indeed related to receiving the vaccine and expressed that "it’s good to receive a vaccine" and that people are happy after receiving their second vaccination. For example, tweets that generated the word cloud for Great Britain included:

Fig 2

Word cloud based on positive sentiment for vaccine-related tweets, Great Britain.

Source: Authors own compilation using word cloud software.

“Here it is, worth its weight in gold. My consent form for the covid vaccine next week, normality on the horizon hope”

“So excited to hear my mum, an NHS nurse, will be receiving the Pfizer Covid-19 vaccine today—a glimmer of hope af”

“Grandmother the vaccine, as you can see, absolutely delighted?? (all credit to my younger brother for this absolute”

Please note the above tweets were taken directly from Twitter and do not represent the views of the authors or their institutions.

From the positive sentiment vaccine-related tweets, we determined that more than 90 per cent were directly related to receiving the COVID-19 vaccine. We realise that the data carries limited noise, but we believe that this noise does not affect our results, especially considering the large number of tweets analysed.

As discussed in section 3.1.1, we use the NRC lexicon to calculate our VPAI by deriving the mean value of the positive sentiment coded tweets per day and standardising these values using the minimum-maximum method. The index is measured on a scale from 0 to 1, with 0 the lowest positive sentiment and 1 the highest. As a robustness test for the VPAI, we derive a similar index using Sentiment140. However, in this instance, we calculate the VPAI by expressing the number of positive tweets per day as a percentage of the sum of the number of positive and negative sentiment tweets (see the supporting information section for the graphs on the trends in positive attitudes using the VPAI based on Sentiment140).

3.2.2 Selection of covariates

To select the covariates in the analyses, we used several methods, including relevant literature (see section 2.2), theoretical models, topic modelling and the analysis of negative sentiment and negative emotion tweets.

By analysing the latter, we can determine the major issues that limit the uptake of vaccines and which are likely related to decreased positive attitudes. The reader should note that the negative sentiment associated with vaccine tweets is not the inverse of the positive sentiment. Our analyses find that negatively coded tweets primarily relate to anger, fear or sadness due to the procurement and the efficiency of the vaccine rollout and a lack of information about the side effects. Additionally, our topic modelling revealed that people are dissatisfied with vaccine passports and QR scanning. Concerns were expressed about the Delta variant and misinformation (COVID-19 vaccines and the virus is fake), and various conspiracy theories also came to light, especially for New Zealand. People also express their discontent with social distancing and wearing masks.

In terms of theoretical framework, we use a measure that captures relevant predictors of vaccination behaviour, called the 5C scale. The 5C scale measures the "psychological antecedents of vaccination" as designed by Betsch et al. [29] and is grounded in established theoretical models of vaccine hesitancy and acceptance (Thomson et al. [30], MacDonald [31], Larson et al. [32]) and relates these predictors to psychological models to explain health behaviour (Betsch et al. [33]). We note from the 5C that confidence, constraints, collective responsibility, complacency, and calculation are important when investigating vaccination behaviour.

For the analysis of the negative sentiment and negative emotion vaccine-related tweets, we follow the same process as described in section 3.2.1. For an example of the issues that cause negative sentiment and emotions in people, see the word cloud in Fig 3 generated from tweets extracted for South Africa.

Fig 3

Word cloud based on negative sentiment for vaccine-related tweets, South Africa.

Source: Authors own compilation using word cloud software.

Sample tweets that generated the word cloud for South Africa’s negative sentiment include, for example:

“With an incompetent government, a Minister of Health without a medical degree, NDZs dictatorial tendencies & a rural population still totally unaware of what a pandemic is added to a vaccine shortage, we are doomed”

We are bored about 1) corruption 2) poor vaccine strategy 3) terrible national government 4) incompetent cabinet 5) stealing during a pandemic!”

This vaccine rollout has been disastrous from the government, has cost lives, now another lockdown killing our already hurt economy. Massive change is needed in the running of this country

The rate with which people are dying every day should get SAHPRA concerned and energised to approve more vaccine even on trial basis, this apparent incompetence is really killing and destroying families.

Please note that the above tweets were taken directly from Twitter and do not represent the views of the authors or their institutions.

After conducting an in-depth analysis of the negative sentiment vaccine-related tweets for all ten countries, we discovered that the negative sentiment was mainly related to anger towards governments’ incompetence in procurement, the lack of procuring a sufficient number of (or wrong) vaccines and the execution of the vaccine rollout, fear regarding side effects, fear of people dying because they cannot get access to vaccines and people refusing to be vaccinated. Interestingly, the words prominent in the word cloud, such as ’death’, ’die’, and ’killing’, are related to not receiving the vaccine rather than fearing the side effects.

In analysing the negative sentiment tweets, we found that tweets expressing dissatisfaction with governments are false negatives. This means that, in reality, people have positive attitudes towards vaccines. However, they are negative about government incompetence related to issues such as the rollout process, the procurement and accessibility of vaccines etc., hence the negative sentiment of the tweets. To test this hypothesis, we also create a VPAI in which we add the tweets coded as false negatives to the tweets coded as positive. We name the index VPAI2. See the trends in the supporting information section. S4 to S6 Figs indicate a predominant upward trend. This suggests that if policymakers address the grievances of people related to the abovementioned government incompetence, they can turn around the downward trend in the VPAI.

Therefore, the selected covariates included in the regression analyses are:

Trust in the COVID-19 vaccine: as a proxy for how people perceive the vaccine’s safety. To construct this variable, we follow the method as explained in section 3.3.1. We use the NRC lexicon to return the emotion score for each COVID-19 vaccine-related tweet for ’trust’. We construct a daily time-series by averaging the measured value of ’trust’ per tweet per day (Greyling et al. [9], Betsch et al. [29]). We lag the variable to address possible endogeneity that might spread from confounding factors.

Anger towards the government: is included in our interaction variable (see point 8). However, to construct the ’anger towards the government’ variable, we first extract all tweets that include the following keywords: government, parliament, ministry, minister, senator, MPs, legislator, political, politics, prime minister. We use the same method to construct the time-series as for the ’trust in the COVID-19 vaccine’ variable (Greyling et al. [9], Betsch et al. [29]). We use the anger emotion as a proxy for dissatisfaction with the government. We also lag the variable.

Compliance: as a proxy for collective responsibility. We follow Sarracino et al. [34] and define compliance as the degree of association between people’s behaviours and COVID-19 containment policies to construct the compliance variable. We use information gathered from Google Mobility Reports (the change in duration from the residential category) (Google [10] [11]) and the Stringency Index, which consists of the following nine indicators: school closing, workplace closing, events cancelled, restriction of gatherings, closed public transport, staying at home requirements, restrictions of internal movements, international travel controls, and public information campaigns. The Stringency Index ranges from 0 to 100, with 100 being the most stringent, and we sourced it from Oxford’s COVID-19 Government Response Tracker (Hale [8]). Therefore, we estimate the following equation: where res is residential mobility in country c on day t; Country is a vector of dummies for each country included in the dataset; Day is a vector of dummies for the days from 1 February 2021 to 31 July 2021. We focus on this period because prior to February 2021, the vaccine rollout did not occur in all countries under investigation. Policy represents the stringency of containment policies in country c on day t. A vector of dummies is depicted by λ for each combination of month m and hemisphere s, to account for the different seasons and evolution of the pandemic among the Northern and Southern hemispheres. The coefficient β is our measure of compliance. It provides the correlation between policy stringency and mobility by country and day. We are aware that creating a daily compliance measure risks introducing noise in the correlation. However, to fulfil our aim of determining the daily evolution of positive attitudes towards vaccines, we need to assess daily changes in compliance.

All tweets related to vaccines (Greyling et al. [9]): this is a proxy for the prominence of vaccines as a conversation topic.

Daily COVID-19 vaccine doses administered per million people (Mathieu et al. [13], Betsch et al. [29]): a proxy for how well a country handles the vaccine rollout. We lag this variable to address possible endogeneity that might spread from confounding factors. The rollout or lack thereof also proxies various constraints such as problems with the physical availability of the COVID-19 vaccine, lack of geographical accessibility, or signalling a less than adequate appeal for vaccination services uptake. We find that the VPAI and the daily vaccines have an inversely proportional relationship; therefore, we transformed this variable using a hyperbolic function.

Daily total new cases: a proxy for the evolution of the COVID-19 pandemic across all ten countries (Hale et al. [8]). In our models, we lag new cases to capture people’s expectations of the trend of the pandemic.

Vaccine policy: we control for the vaccination policy across our ten countries. According to Hale et al. [8], a vaccination policy is classified as follows: 0—no vaccine available; 1 –vaccine available for one of the following groups: key workers / clinically vulnerable groups / elderly groups; 2—available for two of the abovementioned groups; 3—available for all the abovementioned groups; 4—available for all three groups plus partial additional availability (select broad groups/ages) and 5 –the vaccine is universally available.

To capture anger directed towards the government, we use an interaction variable ’government anger’ interacted with ’new daily vaccinations’. This variable captures the anger expressed towards the government given the number of new vaccinations per day. We use the above as a proxy for people’s dissatisfaction with the vaccination rollout, which also encapsulates procurement, capacity and corruption issues, and accessibility of the vaccines.

Table 2 provides summarised statistics for the variables included in our study.

Table 2

Descriptive statistics of the variables included in the estimations of attitudes against the COVID-19 vaccine.

Variable	Observations	Mean/ Frequency (%)	Std Dev.	Min	Max
VPAI	1,780	0.35	0.12	0.10	0.91
Lagged trust in the COVID-19 vaccine	1,780	0.37	0.09	0.16	0.91
Stringency index	1,780	60.88	17.48	22.22	87.96
Residential mobility	1,780	8.32	24.62	-29.67	50.85
Lagged compliance	1,780	1.07	0.273	0.621	2.37
Lagged anger towards the government # Daily vaccinations	1,780	0.84	0.54	0.00	2.52
Vaccine tweets*	1,780	106.32	108.20	6	690
Lagged new daily vaccinations*	1,780	231776.60	219928.20	0	873515
Lagged new daily cases*	1,780	148	154	0	701
Vaccine policy
0	605	25.80	-	-	-
1	119	5.88	-	-	-
2	444	21.13	-	-	-
3	574	22.08	-	-	-
4	353	13.98	-	-	-
5	305	11.13	-	-	-

Source: Authors’ calculations.

*Note: Vaccine tweets were logged, and the hyperbolic function of new daily vaccination was derived; the new daily cases were logged, and all variables were smoothed using a seven-day average.

3.3 Methodology

We first use descriptive statistics and graphs to analyse the trend in the VPAI over time and compare the results for the Northern and Southern hemispheres and across the ten countries in our sample. Our descriptive analysis includes topic modelling of the tweets per country. We explore the text corpus by applying NLP and statistical analysis. The main statistical procedure we use in the topic modelling is factor analysis. We attempt to uncover the text corpus’ hidden thematic structure (topics) using this method. Secondly, we use various econometric techniques to derive and test the robustness of the relationships between our selected covariates and the attitudes towards vaccines.

The correlation between the VPAI and the covariates over time is likely to be affected by confounding factors, such as the severity of the pandemic, exposure to different types of social media, emotional well-being (depression) of the people, accessibility of the vaccine, the prejudice built into the social-cultural environment and the seasons of the year. Therefore, we resort to various econometric techniques to address biases arising from the confounding effects of these variables.

Ideally, we would like to estimate the following equation: where VPAI is the vaccine positive attitude index as defined in section 3.2.1 for country c on day t; Vac_Trust (see section 3.3.1) is the average level of trust related to the COVID-19 vaccine for country c on day t − 1. Gov_Anger is the average level of anger towards government for country c on day t − 1; Compliance is the average level of compliance as defined in section 3.2.2 for country c on day t − 1 X is a vector of variables, λ are month effects capturing common effects across countries, such as seasonal effects (changes in seasons), the evolution of the pandemic and holiday seasons (July for the Northern hemisphere means Summer holidays and for the Southern hemisphere Winter holidays), while μ are country effects.

3.3.1 Pooled ordinary least squares (POLS)

As a baseline model, we use a POLS estimation. To address the bias that might spread from reverse causality, we lag ’trust in the COVID-19 vaccines’, ’anger towards government’, ’compliance’, ’the daily number of COVID-19 vaccinations’ and ’cases’. To address heteroscedasticity, we use robust standard errors in the estimated models.

3.3.2 Fixed effect (FE) estimation

Having the benefit of a panel dataset allows us to control for additional biases, particularly unobserved confounding factors. Specifically, the FE approach reduces the impact of confounding by time-invariant factors, such as the unobserved and, in this instance, observed characteristics of the countries.

We use the Haussmann test to test if the FE model rather than the Random Effects (RE) model is the most efficient estimator in the current study. We reject the null hypothesis that there is "no correlation between the unique errors and the regressors in the model", confirming that the FE will give the most robust estimations.

The country (individual) FE included in the model addresses the unobserved time-invariant heterogeneity between countries, which considerably reduces the risk of the confounding factors discussed above. Additionally, the FE model partly addresses bias originating from omitted observed variables (related to country characteristics). However, the FE model cannot address bias for unmeasured time-varying confounding factors or reverse causality. To further address reverse causality, we turn to Instrumental Variable regressions.

3.3.3 Instrumental variable (IV) regression

In addition to the lagged variables introduced in the POLS and the FE estimations, we also use an IV model to address possible endogeneity and reverse causality. We use the Generalised Method of Moments (GMM) estimation rather than the Two-Stage Least Square (2SLS) estimator, due to the efficiency gains derived from using the optimal weighting matrix. The efficient GMM estimator is robust to heteroscedasticity of unknown form.

We instrument ’lagged trust in the COVID-19 vaccine’ and ’lagged compliance’, with ’lagged fear of the vaccines’, ’lagged disgust with the vaccines’ and a two-day lag in ’compliance’. We use the Hansen’s J statistic to test for over-identifying restrictions. The joint null hypothesis is that the excluded instruments are valid instruments, i.e. uncorrelated with the error term and correctly excluded from the estimated equation. A rejection casts doubt on the validity of the instruments. However, in our specified model, serial correlation is present as the error term in one period is correlated with the errors in previous periods. This causes the estimated variances of the regression coefficients to be biased, leading to unreliable hypothesis testing. Therefore, we consider the IV estimations with the POLS and FE estimations.

4. Results

4.1 Results of the trend in the VPAI

We first focus on our descriptive analysis (graphs) to explain the trends in the VPAI towards the COVID-19 vaccines for the period 1 February 2021 to 31 July 2021. We describe the trends in our overall sample, the different hemispheres and lastly, each country. In all instances, we report the findings on the VPAI using NRC, and as a robustness test, we repeat the analyses using Sentiment140. We report the results using Sentiment140 in the supporting information section.

4.1.1 Trends in the VPAI

When we consider Fig 4, we see that the trend in the VPAI towards the COVID-19 vaccines across all countries is downwards; we note an almost 8 per cent decrease over time. Section 4.1.2 discusses possible explanations for this downward trend.

Fig 4

Trend in positive attitude from February 2021 to the end of July 2021 for the whole sample.

Source: Authors’ calculations.

Additionally, we note from Fig 5 that the downward trend in positive attitude holds across both the Northern and Southern hemispheres. However, the downward trend seems stronger in the Southern hemisphere than in the Northern Hemisphere.

Fig 5

Trend in positive attitude across the Northern and Southern hemispheres from February 2021 to the end of July 2021.

Source: Authors’ calculations.

4.1.2 Trends in positive attitude per country

If we consider the individual countries, Fig 6 shows the trend in the VPAI towards the COVID-19 vaccines for each of the ten countries and indicates that the attitude improved in only two countries, namely Belgium and the Netherlands. For the remaining eight countries, the trend was negative over time.

Fig 6

Trend in the positive attitude for each of the ten countries.

Source: Authors’ calculations.

To explain the trends in the VPAI for our individual countries, we relied on existing literature and our topic modelling.

Upon further investigation into Belgium, we found that the positive trend in the VPAI in Belgium was likely due to the steps taken to correct government failure that plagued the country in 2020. In 2020 (Villani et al. [35]), Belgium was the European country with the highest loss of life and hospitalisation rate relative to the size of the population in Europe. According to Vanham [36], Belgium was also hit with capacity issues, struggling to get vaccination centres up and running because of vaccine delivery delays. It seems that the Belgian people did not trust information coming from their government after reports of political favouritism in deciding who would get what little vaccine stock was available were leaked (Vanham [36]). The above events led to widespread anger towards the politicians for making COVID-19 a political game.

However, the government took many steps to correct the situation, likely turning the attitude towards the COVID-19 vaccine positive. The Belgium government set up a COVID-19 task force responsible for addressing logistics and capacity issues. According to Vanham [36], the high uptake of vaccinations could also result from lockdown regulation policies being relaxed during spring, which would depend on the vaccination rates rather than case numbers or hospitalisation rates. People wanting to return to ’normal’ reacted positively to the policy. Our topic modelling also revealed that Belgians was found to be optimistic about the effectiveness of the COVID-19 vaccines, including AstraZeneca. However, they were pleased that younger people could receive the Janssen vaccine.

The Netherlands was the last European country to start their vaccine rollout on 6 January 2021. Their rollout was hampered by a poor vaccination policy and a conservative strategy that kept more than 40 per cent of its vaccines from being used (Bahceli [37]). Additionally, during the beginning phase of the vaccine rollout, young Dutch adults (18 to 34 years, which constitutes approximately 25 per cent of the total population) who were willing to receive the COVID-19 vaccine constantly lagged about ten percentage points after the average percentage of the whole population (Vollmann and Salewski [38]). To encourage a positive attitude towards vaccines, the Dutch government spent around €6 million (at the time of the study) on information campaigns to increase the vaccine uptake by informing the public about the safety of the various COVID-19 vaccines (Bahceli [37]). Vollmann and Salewskis’ [38] results show that the relationship between information campaigns and a positive attitude towards vaccines leads to higher vaccine uptake. The Netherland’s fully vaccinated population increased significantly from 2,430 people on 31 January 2021 to 9,288,187 by 1 August 2021 (Mathieu et al. [13]). We also found (from topic modelling) that the Dutch believed the benefits of getting the COVID-19 vaccine outweighs any potential costs, which is why they expressed support for the elderly and the vulnerable to be prioritised and getting the younger generation vaccinated as soon as possible.

Literature and topic modelling revealed the following likely explanations for those countries that experienced downward trends in positive attitudes over time.

Australia follows a Federal system of government, and contradictory government-implemented regulations across the different states led to widespread confusion regarding COVID-19 vaccines and caused a downtrend in attitudes (Attwell et al. [39]). This was confirmed by our topic modelling as well. New South Wales (NSW) (Greater Sydney region), home to 32.33 per cent of the total Australian population, was at the time of writing this paper in lockdown, while other regions were not regardless of case numbers. These discrepancies in COVID-19 regulations resulted in a significant proportion of people living within NSW refusing to comply with government-imposed regulations. Topic modelling illustrated that there was anger over government-mandated lockdowns, a perceived incompetent government and fear of side effects. In July 2021, approximately 15,000 people (most not wearing masks) protested against the lockdowns, and they demanded their liberties be restored (Swain [40]). Fig 6 shows the downward trend in positive attitudes towards the COVID-19 vaccine.

Spain shows a downward trend in the VPAI. A possible explanation could be that the government delayed action and did not have workable contingency plans in place. This led to a failure containing the virus early on, nearly overwhelming the health system and causing Spain to experience one of the highest death rates attributable to COVID-19 (Casasnovas et al. [41]). This lack of action caused that during the first five months of 2021, Spain’s COVID-19 vaccination campaign progressed slowly and failed to reach marginalised populations (Lazarus et al. [42]). In mid-April, when 13 per cent of Great Britain’s citizens were fully vaccinated, only about 7 per cent of Spaniards were similarly protected (Mathieu et al. [13]). The above rhetoric was echoed through our topic modelling, where Spaniards expressed anger towards their government for failing to rollout the COVID-19 vaccine efficiently while at the same time expressing fear for the unknown long-term side effects of the vaccine.

According to Sprengholz et al. [43], the Germans responded with anger towards their government’s proposed policy to contain COVID-19. The policy stated that only vaccinated people would be allowed to enter venues like sports stadiums, movie theatres or restaurants because they deemed the residual risk high in such places. In July 2021, Chief of Staff Helge Braun announced that he did not expect another COVID-19 related lockdown in Germany (Schultheis [44]). However, this would mean that if there were a future outbreak, the liberties of the unvaccinated would be taken away with immediate effect. Additionally, according to Mario Czaja, head of the Berlin Red Cross, Germany has seen an increase in people not showing up for their vaccination appointments, with 5 to 10 per cent missing appointments daily since July 2021 (Reuters [45]).

The results of the topic modelling of tweets showed that the Germans were angry as they had to wait in queues at the vaccine centres. Furthermore, they were unhappy to receive AstraZeneca due to the news of potential blood clotting. They also mentioned concerns about side effects, but many tweets reassured that these were only mild. Lastly, Germany’s anti-lockdown movement, the Querdenken, has been very active in spreading conspiracy theories ranging from the idea that "masks are deadly" to "vaccines can alter your DNA" (BBC Trending [46]). These could all have contributed to the decreasing trend in the VPAI.

In Great Britain, the government faced criticism because of their vaccination policy, which at the time of writing this paper, was yet to approve the COVID-19 vaccine for 12–15-year-olds (Mason and Elgot [47]). This meant sending children back to schools with inadequate mitigations for COVID-19 in place, which could lead to widespread infections and more disruptions to learning. Additionally, trust in the ability of the government to see this pandemic through has decreased since the announcement of their so-called ’Freedom Day’ (Donovan [48]). Freedom Day brought with it a lifting of any remaining COVID-19 restrictions and came amidst 47,000 new cases of COVID-19 being reported in the previous 24 hours (Donovan [48]). The decision of Freedom Day brought with it 1,200 scientists worldwide criticising the decision to open up, saying it could pose a threat to the entire world if daily cases increased exponentially and vaccine-resistant mutations of the virus were allowed to develop (Ball [49]). In our topic modelling, we found that people were concerned about Freedom day and not wearing masks or applying social distancing rules. They were also concerned about increasing numbers of COVID-19 cases and the blood clotting side effect from AstraZeneca. The downward trend in the VPAI is likely (Fig 6) a product of all the accumulated issues.

In France, introducing a stringent vaccination policy known as ’COVID-19 vaccine passports’ has decreased the positive attitude towards the COVID-19 vaccine (The Economist [50]). The news that movie theatres, museums and sports venues have begun asking visitors to provide proof of a COVID-19 vaccination or a negative test has many French nationals angry but willing to take the vaccine simply to return to their once normal way of living (The Economist [50]). The topic modelling of tweets supported the above mentioned and highlighted the reluctance to accept the health pass and scan QR codes. They were also concerned about the side effects and mentioned that the "Covidliste" has a long waiting list. "Covidliste" is a voluntary and civic initiative allowing the connection between vaccination volunteers and health professionals who have vaccine doses.

Italy, the second-worst Northern hemisphere country concerning people not fully vaccinated (52 per cent at the time of writing the paper), has faced an uphill battle to increase the COVID-19 vaccine uptake since the start of their vaccine rollout on 27 December 2020 (Roberts [51]). They decided to take a tough stance, approving emergency legislation to make COVID-19 vaccines mandatory for all healthcare workers, including pharmacy staff (Roberts [51]). Individuals working in this industry who refused the COVID-19 vaccine would be transferred to another job or suspended without pay for up to a year. In addition, they introduced vaccine passports. From both the topic modelling and Paterlini [52], we saw that Italians demonstrated and protested against Italy’s use of these passports and the subsequent green passes.

Furthermore, the emergency legislation faced fierce resistance from Italy’s deeply rooted anti-vaccine movement, which has been fostered in part by populist political forces (Roberts [51]). These included the 5 Star Movement, which entered government in 2018 promoting vaccine hesitancy. Public trust in the vaccine has also taken a hit after the country temporarily decided to suspend the use of the Oxford/AstraZeneca vaccine after several deaths (Roberts [51]).

In New Zealand, the Aged Care Association (Wallis [53]) described the COVID-19 rollout as a ’shambles’. This is due to the government of the day being responsible for a slow rollout of the vaccine because they and the country as a whole became complacent (Vance [54], Thaker and Floyd [55]). At the time of writing this paper, New Zealand found itself in a level-4 lockdown (the most stringent level of lockdown), even though it did not have a positive COVID-19 case during the previous six months. New Zealand’s zero COVID-19 strategies were successful until the first Delta-variant positive case was announced. Soon, the government realised that they did not have enough vaccines to vaccinate everyone fully, as previously promised. This was partly because of the government’s strategy early in 2021 to reject the cheaper but potentially less effective vaccines like those made by AstraZeneca in favour of the high-performing vaccine made by Pfizer/BioNTech (Satherley [56]).

Additionally, the increased spread of misinformation about the COVID-19 vaccine has increased distrust towards the vaccine (Thaker and Subramanian [57]). Unfortunately, for New Zealand, the spread of misinformation and conspiracy theories are rampant among its ethnic minority and most vulnerable community (Tukuitonga [58]). The lack of trust in the government after failing in their vaccine rollout by not reaching the most marginalised first (topic modelling) is precisely what was needed for the spread of misinformation to take hold. Our results confirm the study done by Paul et al. [19], who found that distrustful attitudes towards vaccination were higher amongst individuals from ethnic minority backgrounds which isn’t a surprise since many minority groups have good reason not to trust the government given their historical mistreatment. All of the abovementioned likely contributed to the downtrend in the VPAI.

South Africa’ has faced problems such as capacity issues, mistrust in the government and anti-vaccination campaigns (Cocks [59]), which contributed to the decrease in a positive attitude towards the COVID-19 vaccines (see Fig 6). From as early as December 2020, it seemed that the COVID-19 strategy was haphazard apart from its dependency on their fragile COVAX arrangement. After receiving their first delivery of the AstraZeneca vaccine on 1 February 2021, it seemed that the government also did not have a clear vaccination policy (van den Heever et al. [60]). The Health Minister created confusion in the public arena when he announced that the AstraZeneca vaccine did not demonstrate efficacy against mild to moderate COVID-19 and placed the rollout of the vaccine on hold. The announcement by the Health Minister caused a decrease in trust in the COVID-19 vaccine and likely contributed to the downward trend in the VPAI. Local scientists criticised the decision, and the World Health Organisation did not support it (van den Heever et al. [60]). This decision left approximately 17 million high-risk population unvaccinated (van den Heever et al. [60]). During the winter months from June to September 2021, South Africa lost 25,660 lives to COVID-19 (Roser et al. [12]). According to van den Heever et al. [60], this probably could have been avoided if the South African government had not been plagued by corruption and mismanagement during its response to the pandemic. By August 2021, South Africa saw ’vaccine apathy’ or ’vaccine fatigue’, with the number of people coming forward to be vaccinated dropping below 200,000 a day, falling short of the set target of 300,000. According to a study conducted by the Human Sciences Research Council and the University of Johannesburg (Cooper et al. [61]), the vaccine-hesitant cite three primary concerns, which could contribute to the downtrend in positive attitudes: side effects, effectiveness, and distrust of the vaccine and institutions.

To summarise, the downward trend in positive attitudes is partly due to a fear of the side effects, but many other factors also contribute. These include dissatisfaction with governments’ rollout plan, procurement, corruption, resistance to mandatory vaccination and the use of COVID-19 passports.

4.2 POLS, FE and IV regression results

This section discusses the results of those covariates that are significantly related to the VPAI and, therefore, when addressed, could improve attitudes and the uptake of COVID-19 vaccines.

In Table 3, the results of the POLS estimation controlling for month and country fixed effects are similar to the results of the FE model and the IV regression. The covariate ’trust in the COVID-19 vaccine’ is statistically significant and positively related to the VPAI across all the estimated models. The methods have different advantages; thus, the joint results confirm that trust in the COVID-19 vaccine is a robust correlate of VPAI. We assume that when trust in the vaccine increases, the fear of adverse side effects decreases and that the positive attitude towards vaccines improves. This finding is in line with the works done by Akarsu et al. [14], Fisher et al. [15], Freeman et al. [16], Ward et al. [17], Seale et al. [18]), Paul et al. [19], Sallam [20] (refer to section 2).

Table 3

Results from POLS with FE and IV.

Variable	POLS		FE		IV
VPAI	Coefficient	SE	Coefficient	SE	Coefficient	SE
Lagged trust in the COVID-19 vaccine	0.2938***	(0.0281)	0.2938***	(0.0193)	0.3127***	(0.0999)
Lagged compliance	-0.0176***	(0.0060)	-0.0176***	(0.0037)	-0.0170***	(0.0064)
Lagged anger towards the government # Daily vaccinations	-0.0274***	(0.0073)	-0.0274***	(0.0073)	-0.0273***	(0.0072)
Lagged new daily vaccinations	0.0055***	(0.0013)	0.0055***	(0.0014)	0.0055***	(0.0014)
Lagged new daily cases	-0.0032***	(0.0080)	-0.0032***	(0.0011)	-0.00324***	(0.0080)
Log vacc tweets	-0.0049*	(0.0029)	-0.0049*	(0.0029)	-0.0047*	(0.0029)
Vaccine policy (Reference—Level 0)
Level 1	0.0598***	-0.0125	0.0598**	-0.0232	0.0596***	(0.0123)
Level 2	0.0451***	-0.0091	0.0451**	-0.0227	0.0454***	(0.0105)
Level 3	0.0476***	-0.0088	0.0476**	-0.0228	0.0480***	(0.0092)
Level 4	0.0536***	-0.0092	0.0536**	-0.0227	0.0542***	(0.0103)
Level 5	0.0473***	-0.0102	0.0473**	-0.0233	0.0477***	(0.0109)
Country FE	Yes		Yes		Yes
Month FE	Yes		Yes		Yes
N	1727		1727		1727
Adjusted R2	0.867		0.422		0.866
Hansen J-Statistic of overidentification					p = 0.6544

Source: Authors’ calculations.

Robust Standard errors in parentheses

* p< 0.10,

**p< 0.05,

*** p< 0.01

Compliance, the act of complying with government-mandated regulations to curb the spread of COVID-19, is statistically significant and negatively related to the VPAI. We interpret this as when people are reluctant unwilling (do not feel like) to comply with orders such as mask-wearing, staying at home etc., then those individuals would be more willing to receive the COVID-19 vaccine, hence a more positive attitude (thus an inverse relationship). A study conducted by Wright et al. [62] investigated the relationship between vaccinated individuals’ willingness to comply and the implemented behavioural regulations. The entire premise of the study is that vaccinated individuals believe they are less at risk because of their vaccination status. People think when vaccinated, they do not need to comply with, for example, mask-wearing, social distancing etc., therefore creating a more positive attitude towards vaccines. This finding is informative to policymakers as a message of "less strict regulations" after vaccination can increase vaccine uptake.

The variable ’anger towards the government’ interacted with the new daily vaccinations is statistically significant and negatively related to the VPAI. Therefore, when people’s dissatisfaction with the government increase, positive attitudes decrease. Analysing the tweets, we find that people are angry with governments due to the lack of procurement, procurement of the incorrect vaccine, the rollout of the vaccination plan and corruption within governments. This anger directed at governments due to a lack of access to vaccines sabotages the positive attitude towards vaccines and hinders the uptake of vaccines.

The relationship between the VPAI and the number of daily new vaccinations is inversely proportional, significant and positively related. This implies that when the daily number of vaccines administered are very low, the positive attitude is high, but as the number of vaccines administered per day increases, the positive attitude starts to plateau. Also, using the vaccine rollout or the lack thereof, as a proxy for constraints in information campaigns, the physical availability of vaccines or a lack of geographical accessibility, we can see how important it is to overcome any barriers which might impede the intention to be vaccinated (Cylus and Papanicolas [63]).

We find that daily new cases, a proxy for the evolution of the COVID-19 pandemic across all ten countries, is statistically significant and negative. If the daily cases are high, the positive attitude towards the vaccine is relatively low, but as the daily cases start decreasing, the positive attitude improves.

Controlling for the vaccination policy, thus the groups that can access the COVID-19 vaccine, we find that when more groups of people can access the vaccine, for example, all age groups compared to fewer groups, it is positively related to the VPAI. Once again, showing that when more people have access to the COVID-19 vaccine, positivity towards the vaccine is enhanced.

The number of vaccine-related tweets is statistically significant and negatively related to the VPAI in all the estimated models. This implies that, as the number of tweets related to vaccines increases, the positive attitude towards vaccines decreases. We found that the number of vaccine-related tweets increased over time; thus, vaccines became a "hot" topic of discussion. Topic modelling on the vaccine-related tweets indicates that negative sentiment is related to, among others: the long-term effects of the COVID-19 vaccine; dissatisfaction with vaccine passports and QR code scanning; concerns about the Delta variant; anger about procurement; struggling to make appointments (vaccine rollout) and conspiracy theories. People also express their discontent with social distancing and wearing masks. Our topic modelling results are in line with studies such as Küçükali et al. [23], Nuzhath et al. [24], Bonnevie et al. [2] and Thelwall et al. [25].

All of the above leads us to believe that increased tweets with negative connotations to the COVID-19 vaccine decrease positive attitudes towards vaccines.

In summary, we see that those variables that can improve the positive attitude towards vaccines are related to information about the safety and side effects of the vaccines (increased trust in vaccines) and a balance between the strictness of regulations and access to vaccines. Additionally, increased trust in the governments’ capabilities, honesty of governments and dealing with capacity constraints can decrease the dissatisfaction with governments and increase vaccine uptake. Precise information about the COVID-19 vaccines in general also disseminated via social media can increase positivity towards the COVID-19 vaccine. Misinformation about COVID-19 vaccines and social media should be monitored, and campaigns against this misinformation should be launched. Vaccines should also be made accessible to all groups of people.

5. Conclusions

In this study, we constructed a real-time Vaccine Positive Attitude Index (VPAI) derived from Big Data to illustrate the evolution of people’s positive attitudes toward the COVID-19 vaccine across ten countries. Our descriptive analysis showed that the VPAI generally indicates a decline in attitude over the time period investigated. When we consider the different hemispheres, the trend is downwards in the Northern and Southern hemispheres. Furthermore, considering the ten individual countries, only Belgium and the Netherlands experienced a positive trend in the VPAI, whereas the other countries experienced a negative trend.

Using POLS, FE and IV regression models, we determined which variables are significantly related to the VPAI and, therefore, could increase the uptake of COVID-19 vaccines if addressed by policy measures. We found that those variables that could improve people’s attitudes towards vaccines were: i) information related to the safety and side effects of the vaccines, ii) increased trust in governments in conducting the vaccine rollout and handling procurement and capacity issues, iii) cognisance of the compliance versus the vaccine up-take decision, and iv) better information about the COVID-19 vaccines in general, but primarily disseminated via social media.

These results give policymakers the necessary information to increase positive attitudes towards the COVID-19 vaccine. Policymakers should focus on improving trust in the COVID-19 vaccines. They could more openly disseminate information regarding the vaccine, do it in layman’s terms, and acknowledge people’s fears, anger, and other negative emotions. They can emphasise the stringent safety and efficacy standards of the COVID-19 vaccine development process, thus fostering trust in the vaccine. All of this may increase vaccine confidence. Additionally, countries can overcome the lack of accessibility to vaccination clinics by following examples set by the United States of America to introduce mobile vaccination clinics to reach people in remote areas.

Furthermore, policymakers should implement policies to increase people’s sense of collective responsibility. This can be done by raising awareness of emotional manipulations by anti-vaccine disinformation efforts and activating positive emotions such as altruism and hope as part of vaccine education endeavours. Another potential strategy is to elicit positive emotions toward helping one’s community restore health and well-being when deciding to vaccinate against what is called the most consequential disease of our time.

Lastly, the study has limitations. We only used Twitter in our analyses as other social media platforms, for example, Facebook, is highly protected. In the future, additional social media platforms should be analysed. A drawback of social media data, including Twitter, consists in their lack of representativeness of populations. However, a case can be made that the sample sizes using Big Data contributes to robust results. When translating the tweets from a non-English language, errors can occur at three levels: lexical, syntactical, and discursive. These errors inevitably can cause unintelligibility which means that the tweet will be disregarded. However, in the current study, we used lexicons in the original languages (if available) and compared the results to the translated text corpus. We find the time-series trends using the original language and translated text well correlated.

Furthermore, we only examined the tweets for a specific time period. Therefore, we cannot examine the tweet’s effects following the study. Another limitation is the timeline of research and the publication thereof. Publication rates cannot keep up with the tempo at which the pandemic evolves; therefore, by the time of publication, new COVID-19 variants might have emerged, and research on vaccines would have progressed. Nonetheless, the knowledge gained in this research contributes to more informed decision making in the future. Lastly, the countries used for this study are solely based on data availability, which can be considered a limitation, especially given a non-characterisation of how other countries could present attitudes towards vaccines. In future studies, with more resources, we can expand the text corpus to include more countries over a longer period with more sources of Big Data.

Trend in positive attitude from February 2021 to the end of July 2021 for the whole sample, using Sentiment140.

(TIF)

Click here for additional data file.

Trend in positive attitude from February 2021 to the end of July 2021 per hemisphere, using Sentiment140.

(TIF)

Click here for additional data file.

Trend in positive attitude from February 2021 to the end of July 2021 per individual country, using Sentiment140.

(TIF)

Click here for additional data file.

Trend in positive attitude including false negatives related to governments (VPAI2) for the whole sample from February 2021 to the end of July 2021.

(TIF)

Click here for additional data file.

Trend in positive attitude including false negatives related to governments (VPAI2) for the different hemispheres from February 2021 to the end of July 2021.

(TIF)

Click here for additional data file.

Trend in positive attitude including false negatives related to governments (VPAI2) for the individual countries from February 2021 to the end of July 2021.

(TIF)

Click here for additional data file.

(DTA)

Click here for additional data file.

7 Jan 2022

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Prof. Anat Gesser-Edelsburg, Ph.D.

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf  and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see http://journals.plos.org/plosone/s/data-availability.

Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories. Any potentially identifying patient information must be fully anonymized.

Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access.

We will update your Data Availability statement to reflect the information you provide in your cover letter.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

Reviewer #3: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: I Don't Know

Reviewer #3: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: The article uses Big Data to study a very important topic of COVID-19 vaccine hesitancy and factors that may enhance vaccine uptake. The authors provide an expansive picture of commonalities and diversities among varied countries, which contribute to the understanding of this vital issue. The article is well-written, maintaining clarity and logical flow in the description of the findings and their potential value.

Nonetheless, there are several elements that need to be improved in order for the article to be appropriate for publication, as follows:

1. The authors ‘mix’ between results and discussion. See section 4.1.2. the results should be presented as found, without the authors trying to provide an explanation to why each element was such found. These explanations (for example see lines 375-388 concerning Belgium or lines 389-396 concerning Netherlands) are not appropriate in the results section.

2. The authors should refrain from any judgmental comments (for example lines 395-396 – “It is safe to say that the information campaigns have paid off”). Scientific writing should state possible explanations to any phenomenon that is identified, usually based on previous publications or new theoretical hypotheses, but without any judgment of any kind (neither positive or negative). Furthermore, the authors have not shown any support that this is actually the case and that the information campaigns were the cause for the results identified; more cautious writing is recommended. The same relates to sentences such as line 467 “South Africa's woes are almost too many to count”. These types of phrases should be revised so that the sentence presents the finding as is and not referred to in ‘literary-type’ descriptions.

3. Whenever any assumption (or claimed ‘fact’) is written – it should be based on previously published data from scientific sources. Sentences such as “Australians have a deep-seated mistrust in their government officials, who seem to implement contradictory regulations across different states” (lines 398-399) have to be supported by such publications.

4. Lines 505-507 – “Compliance, the act of complying with government-mandated regulations to curb the spread of COVID506 19, is statistically significant and negatively related to the VPAI. Unwillingness to comply likely 507 motivates people to get vaccinated. “ is not clear. I would assume that compliance would follow positive attitudes (towards the vaccine or towards the authorized agencies). How do the authors explain the negative correlation?

5. The authors at times make claims that they do not show any support for based on their data. For example, their claim in lines 534-537 “This implies that, as the tweets related to vaccines increase, the attitude towards vaccines decreases. This may likely be because many of the tweets contain misinformation or conspiracy theories rather than campaigns and information to encourage being vaccinated - and, therefore, decreases positive attitudes towards vaccines.”. To support such a claim, they should check whether the tweets they investigated actually show this tendency.

Reviewer #2: The results reported in the paper are interesting as such, since the topic of public reaction to the COVID-19 pandemic and the related interventions is a timely one. However, to this referee it seems that the descriptive interest of the results dominates over the analytical one, due to various technical issues, as detailed below.

Why do authors call the analysis "panel analysis" when there is no panel of individuals followed at more than one time-point? Considering a country as an individual is a complicated assumption... In reality, the data is a sequence of cross-sectional analyses within each country without control of the relative media-related weight or publication frequency of pro- and anti-vax individuals in the data stream and therefore maybe biased as an estimate of the country specific sentiment...

The analysis methods, common in econometric panel data analyses, have been devised for individual data and, as also commented upon above, have to be motivated for use on aggregated, non-individual, data...

One could also suspect that people are more eager to announce an anti-vaccine stance than a neutral or pro-vaccine stance on social media, especially since the anti-vaxers are minorities, which therefore need more cohesion and solidarity...

The paper also seems to conclude that a part of the negative opinions are more due to vaccine availability and related problems than to an "ideological" anti-vaccine stance. A clearer separation of these issues in the analysis, in addition to the VPAI2 index, would help the understanding of vaccine-related sentiment...

In section 3.2.2 (Selection of covariates), many covariates are extracted from the same source as the response variable (the Tweet streams). Is there not a risk that this automatically induces correlations?

In the equation (2, p13), one wonders what common monthly effect (lambda_m) is expected in different hemispheres, in countries in different phases of epidemic spread...

The analysis in the paper is also to a large extent outdated, since so much , both in terms of restrictions and vaccinations, has happened since August '21...

In case of revision, the authors should incorporate some recent material, either in the analysis or the discussion.

Finally, regarding section 5 (Conclusions), a clearer distinction between conclusions strictly based on the analysis and authors opinions or suggestions for action should be made.

Small language errors or typos to be corrected, e.g. "illicit" instead of "elicit" (p2 l36) or "By analysing, the latter..." (p7 l162) instead of "By analysing the latter..."

This referee recommends a thorough revision and actualization of the paper before considering its possible publication.

Reviewer #3: Manuscript Identification

PONE-D-21-34544: This paper has two main objectives: 1) to conduct a cross-country panel analysis to investigate the trend in positive attitudes towards COVID-19 vaccines over time; and 2) lies with determining those variables which are significantly related to a positive vaccine attitude and can inform policymakers. The article has a good description of how previous studies analysed the emotions in vaccine-related tweets, and following this description, the authors present a considerable justification for why do this study, which is classified as the first study to use Big Data to determine the sentiment and emotions related to COVID-19 vaccines through a vaccine positive attitude index. The time series analysis is a good way to perform a point of view of this situation.

The authors' proposal is to present an analysis based on tweets from 10 countries, also dividing the analysis into countries in the southern hemisphere and countries in the northern hemisphere. This choice is based on data availability, which can be considered a limitation of the study, in view of a non-characterization of how other countries could present differences regarding the information provided.

Minor concerns

The text of the paper needs to be revised as it has some inconsistencies regarding the information.

1. About Figure 1 - it is necessary to modify the dates of the years presented in this figure, considering that the data were collected in 2021, the years change according to the countries. In addition, I also suggest a review of the percentages presented in terms of the number of people fully vaccinated, which differ from Table 1.

2. Section 3.2.1: the authors present some illustrations based on word clouds generated by tweets and inform that this was done for all country but in the paper, it's possible to find just 2-word clouds showing information by Great Britain and South Africa. It's important to inform why just these 2 countries were present in this analysis.

3. Section 3.2.2: Scale 5c. In the second paragraph, the word complacency is not present.

4. The authors do not present the limitations of the study, an important point for future studies on the subject. In view of some methodological characteristics, my suggestion is to highlight some points such as: the translation of tweets in a non-English language; choice of countries based on data convenience and availability.

5. There is no specific topic in the study for discussing the data, however, in the description of the results, the authors present some references that can be considered part of this discussion of the data. As this is an analytical study, it would be important to have a more in-depth discussion, comparing the results to other previous studies.

The manuscript has good methodological quality, is free of bias, needs some adjusts but the results are discussed based on the theoretical background properly of the manuscript theme.

So, the conclusions answered the aims of the study focused on the references and results.

The limitations of the study were not presented, and I suggest that this could be done.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: Yes: Leonardo Pestillo de Oliveira

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

14 Feb 2022

RESPONSE TO REVIEWERS

Manuscript ID PONE-D-21-34544

Positive attitudes towards COVID-19 vaccines: A cross-country analysis.

We thank the reviewers for the comments. Each has been comprehensively addressed, as set out below (authors' response shown in italics). Throughout the paper, we have also made improvements to enhance its quality and contribution to the literature. Therefore, we believe the paper to be significantly improved as a result of the comments received from the reviewers.

Editorial comments addressed by the authors

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf

and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

We thank the editor for pointing out this oversight. Subsequently, we have ensured that PLOS ONE's style requirements are followed throughout the revised submission.

2. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see http://journals.plos.org/plosone/s/data-availability.

Upon re-submitting your revised manuscript, please upload your study's minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories. Any potentially identifying patient information must be fully anonymized.

Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access.

We will update your Data Availability statement to reflect the information you provide in your cover letter.

We thank the editor for pointing out this requirement. Subsequently, we have uploaded our study's minimal underlying data set as supporting information files. Thank you for making the change online on our behalf.

Reviewer comments addressed by the authors

Reviewer #1:

The article uses Big Data to study a very important topic of COVID-19 vaccine hesitancy and factors that may enhance vaccine uptake. The authors provide an expansive picture of commonalities and diversities among varied countries, which contribute to the understanding of this vital issue. The article is well-written, maintaining clarity and logical flow in the description of the findings and their potential value.

We thank the reviewer for this kind and generous review of our study.

Nonetheless, there are several elements that need to be improved in order for the article to be appropriate for publication, as follows:

1. The authors' mix' between results and discussion. See section 4.1.2. the results should be presented as found, without the authors trying to provide an explanation to why each element was such found. These explanations (for example see lines 375-388 concerning Belgium or lines 389-396 concerning Netherlands) are not appropriate in the results section.

We thank the reviewer for this comment. Different disciplines have different approaches to the results and discussion section. As well-being economists, we follow the format set out in economic papers in which the results and discussion sections are combined. We prefer to keep it in this format. See similar well-being economics papers using the same structure below (please note that there are many more examples):

Consiglio V, Sologon D. The Myth of Equal Opportunity in Germany? Wage Inequality and the Role of (Non-)academic Family Background for Differences in Capital Endowments and Returns on the Labour Market. Social Indicators Research. 2021. Available from https://doi.org/10.1007/s11205-021-02719-2

Rözer J, Lancee B, Volker B. Keeping Up or Giving Up? Income Inequality and Materialism in Europe and the United States. Social Indicators Research. 2022; 159, 647–666. Available from https://doi.org/10.1007/s11205-021-02760-1

2. The authors should refrain from any judgmental comments (for example lines 395-396 – "It is safe to say that the information campaigns have paid off"). Scientific writing should state possible explanations to any phenomenon that is identified, usually based on previous publications or new theoretical hypotheses, but without any judgment of any kind (neither positive or negative). Furthermore, the authors have not shown any support that this is actually the case and that the information campaigns were the cause for the results identified; more cautious writing is recommended. The same relates to sentences such as line 467 "South Africa's woes are almost too many to count". These types of phrases should be revised so that the sentence presents the finding as is and not referred to in 'literary-type' descriptions.

We thank the reviewer for pointing this out. Subsequently, the entire manuscript has been edited to remove any judgemental comments. Additionally, we have followed a more cautious style of reporting the results.

3. Whenever any assumption (or claimed 'fact') is written – it should be based on previously published data from scientific sources. Sentences such as "Australians have a deep-seated mistrust in their government officials, who seem to implement contradictory regulations across different states" (lines 398-399) have to be supported by such publications.

We thank the reviewer for pointing this out. Subsequently, the entire manuscript has been edited to remove any assumptions reported as fact. Where facts are mentioned, the appropriate references are added.

4. Lines 505-507 – "Compliance, the act of complying with government-mandated regulations to curb the spread of COVID506 19, is statistically significant and negatively related to the VPAI. Unwillingness to comply likely 507 motivates people to get vaccinated. "is not clear. I would assume that compliance would follow positive attitudes (towards the vaccine or towards the authorized agencies). How do the authors explain the negative correlation?

We thank the reviewer for this comment. We wish to state that we tested this association multiple times and always reached the same outcome of a negative association. We interpret this as when people are unwilling (do not feel like) to comply with orders such as mask-wearing, staying at home etc., then those individuals would be more willing to receive the COVID-19 vaccine, hence a positive attitude (the relationship is negative).

A study conducted by Wright et al. (2022) investigated whether vaccinated individuals were less likely to comply with implemented behavioural measures. The entire premise of the study is that vaccinated individuals believe they are less at risk because of their vaccination status. People think they do not need to comply any longer with, for example, mask-wearing, social distancing etc., if they get vaccinated therefore creating a more positive attitude towards vaccines.

Wright L, Steptoe A, Mak HW, Fancourt D. Do people reduce compliance with COVID-19 guidelines following vaccination? A longitudinal analysis of matched UK adults. Journal of Epidemiology & Community Health. 2022; 76, 109-115.

5. The authors at times make claims that they do not show any support for based on their data. For example, their claim in lines 534-537 "This implies that, as the tweets related to vaccines increase, the attitude towards vaccines decreases. This may likely be because many of the tweets contain misinformation or conspiracy theories rather than campaigns and information to encourage being vaccinated - and, therefore, decreases positive attitudes towards vaccines.". They should check whether the tweets they investigated actually show this tendency to support such a claim.

We thank the reviewer for this comment. Subsequently, we added to our analysis and performed topic modelling on the extracted vaccine-related tweets. We expanded our explanation on the negative relationship between the number of tweets and the VPAI. Please see lines 169-177 and 590 onwards. We also include an example of the topic modelling in the word version of our response in Table 1 for your perusal. Please note we also added the topic modelling results to explain the trends in VPAI in section 4.1.2.

Reviewer #2:

The results reported in the paper are interesting as such, since the topic of public reaction to the COVID-19 pandemic and the related interventions is a timely one. However, to this referee it seems that the descriptive interest of the results dominates over the analytical one, due to various technical issues, as detailed below.

1. Why do authors call the analysis "panel analysis" when there is no panel of individuals followed at more than one time-point? Considering a country as an individual is a complicated assumption... In reality, the data is a sequence of cross-sectional analyses within each country without control of the relative media-related weight or publication frequency of pro- and anti-vax individuals in the data stream and therefore maybe biased as an estimate of the country specific sentiment...

We thank the reviewer for this comment. We wish to point out that the data are a country-level panel dataset. Thus, a macro panel dataset. We are analysing ten countries over time at a daily frequency. Panel datasets can be at a micro, regional or macro level. It only implies that the same unit of analysis (individual/region/country) was measured at different time periods. Please see the following textbooks in this regard:

Baltagi BH. Econometric analysis of panel data. 2005. John Wiley & Sons Ltd: England.

Gujarati DN. Basic Econometrics, (4th ed.). 2003. New York, n. y.: McGraw-Hill.

Additionally, similar to other data on happiness or subjective well-being (survey data), the survey is based on individuals but aggregated to country-level – please see the World Happiness Report.

Helliwell JF, Layard R, Sachs J, De Neve J, eds. World Happiness Report. 2021. New York: Sustainable Development Solutions Network.

With Big Data, each tweet is seen as a source which is then combined to create text corpus per day (thus aggregated to a country level per day) – similar to survey data (aggregated to country level). The number of tweets is vast. We extract 100000 tweets per day for the UK and NZ, the smallest country, 6000 tweets randomly. The sentiment of these tweets is then aggregated to derive a country level sentiment per day. This is a more representative sample than survey data.

2. The analysis methods, common in econometric panel data analyses, have been devised for individual data and, as also commented upon above, have to be motivated for use on aggregated, non-individual, data..

We thank the reviewer for this comment. However, as mentioned in point #1 above, panel analysis at a macro level is well recognised in econometrics. The panel analysis of macro data is included in econometric textbooks. Please see amongst other:

Breitung J. The Analysis of Macroeconomic Panel Data. 2015. Oxford handbooks. Available from https://www.oxfordhandbooks.com/view/10.1093/oxfordhb/9780199940042.001.0001/oxfordhb-9780199940042-e-15

Gujarati DN. Basic Econometrics, (4th ed.). 2003. New York, n. y.: McGraw-Hill

Specifically, please see page 660 "In panel data the same cross-sectional unit (say a family, a firm, a state or a country) is surveyed over time. In short, panel data have space as well as time dimensions."

3. One could also suspect that people are more eager to announce an anti-vaccine stance than a neutral or pro-vaccine stance on social media, especially since the anti-vaxers are minorities, which therefore need more cohesion and solidarity...

We thank the reviewer for this comment. Subsequently, we performed topic modelling on the extracted vaccine-related tweets in addressing point #5 from reviewer 1. We did not find what you are alluding to. Please see the example of the topic modelling in table 1 in the word version of our response for your perusal.

4. The paper also seems to conclude that a part of the negative opinions are more due to vaccine availability and related problems than to an "ideological" anti-vaccine stance. A clearer separation of these issues in the analysis, in addition to the VPAI2 index, would help the understanding of vaccine-related sentiment...

We thank the reviewer for this comment. We have made significant changes throughout the discussion of the results, and we trust we have addressed this comment.

5. In section 3.2.2 (Selection of covariates), many covariates are extracted from the same source as the response variable (the Tweet streams). Is there not a risk that this automatically induces correlations?

We thank the reviewer for this comment. The initial text corpus (extracted tweets) indeed are the same. However, we use different keywords to extract tweets from the initial corpus. Please see page 15 for the full description. We summarise below as well.

In our analyses, we make use of the following variables, which are constructed from tweets:

1) VPAI: on page 12, we explain that the index is constructed from corpus data extracted from vaccine-related keywords. To construct the index, we use sentiment analysis, in which an algorithm is used to determine the sentiment of the tweets.

2) Trust in the COVID-19 Vaccine: on page 15, we explain that this variable was constructed using the same corpus of extracted tweets as in 1) (keywords related to vaccines), although, in this instance, we do not use sentiment analysis but the underlying coded trust emotion. Two different algorithms are used based on different principles to analyse sentiment and emotion. The output of the sentiment and emotional analysis differ vastly. Only certain tweets will be coded as containing the trust emotion – whereas almost all tweets are coded as having a sentiment of different degrees.

3) Lagged anger towards the government: on page 15, we explain that the corpus of extracted tweets is based on government-related keywords. Therefore, the corpus tweets differ from the prementioned variables (1) and (2).

Therefore, even if the original corpus of tweets is the same, the coded tweets used to construct the variables are often not the same tweets. Even where tweets may overlap, the algorithms used to derive sentiment (Sentiment140 or Syuzhet, AFINN and Bing for robustness test) or emotion (NRC) differs. Please see the VIF in Table 2 in the word version of our response to show no multicollinearity.

The VIF value (test for multicollinearity) of lagged trust in COVID-19 vaccines is 2.70, and lagged anger towards governments is 4.51. Both are well within the accepted ranges of less than 5.

6. In the equation (2, p13), one wonders what common monthly effect (lambda_m) is expected in different hemispheres, in countries in different phases of epidemic spread...

We thank the reviewer for this comment. The month fixed effect eliminates bias from unobservable variables that change over time but are constant over the countries. Therefore, we control for the heterogeneity in months.

We extended our explanation in line 314 onwards to clarify this matter. However, we summarise it here as well:

1) seasonal effects (changes in seasons): occur in all countries in the panel in the same months. During March in the Northern hemisphere, the season changes to Spring and in the Southern hemisphere to Autumn. It has been shown that changes in seasons can affect attitudes. Similar in June, the change in season is to Summer and Winter, respectively.

2) evolution of the pandemic: as the pandemic progressed from one month to the next, people's attitudes might have changed. They might have gotten used to the idea of pandemics and vaccines. Or they might show increased pandemic or vaccine fatigue which could also affect their attitudes.

3) holiday seasons: July is a holiday season (Northern hemisphere – Summer holidays and Southern hemisphere- winter holiday).

We find that each month has a different Y-intercept. This implies that certain characteristics of months are the same across countries, but there is heterogeneity between the months. Please see the results in Table 3 in the word version of our response.

From Table 3, we can see that as the year progressed, compared to February, the Y-intercept per month decreased, and these differences were significant. Therefore, the months are not the same (there is heterogeneity), and these differences are unobserved. The Y-intercept for February is 0.24. Ceteris paribus, we notice that in March, the positive attitude is significantly different to February, which is lower with 0.04.

7. The analysis in the paper is also to a large extent outdated, since so much , both in terms of restrictions and vaccinations, has happened since August '21...

In case of revision, the authors should incorporate some recent material, either in the analysis or the discussion.

We thank the reviewer for this comment. However, we respectfully disagree. Our analysis is for a specific period for which we attempt to understand what happened to VPAI in specific circumstances. If we understand the past, we are also more likely to understand the implications of future actions. Therefore, our findings based on the past can inform policymakers on the effect their actions might have on attitudes towards vaccines. For example, we find that people unwilling to comply are more positive towards vaccines. This information informs policymakers what to expect from people's attitudes when regulations are implemented.

8. Finally, regarding section 5 (Conclusions), a clearer distinction between conclusions strictly based on the analysis and authors opinions or suggestions for action should be made.

We thank the reviewer for this comment. Subsequently, addressing points #2 and #3 raised by reviewer 1, we believe this point has been addressed. Please note we also added the topic modelling results to explain the trends in VPAI in section 4.1.2. Thereby eliminating any opinions and only reporting results.

9. Small language errors or typos to be corrected, e.g. "illicit" instead of "elicit" (p2 l36) or "By analysing, the latter..." (p7 l162) instead of "By analysing the latter..."

We thank the reviewer for pointing out these errors. Subsequently, the manuscript was proofread, and any other mistakes were corrected.

Reviewer #3:

This paper has two main objectives: 1) to conduct a cross-country panel analysis to investigate the trend in positive attitudes towards COVID-19 vaccines over time; and 2) lies with determining those variables which are significantly related to a positive vaccine attitude and can inform policymakers. The article has a good description of how previous studies analysed the emotions in vaccine-related tweets, and following this description, the authors present a considerable justification for why do this study, which is classified as the first study to use Big Data to determine the sentiment and emotions related to COVID-19 vaccines through a vaccine positive attitude index. The time series analysis is a good way to perform a point of view of this situation.

The authors' proposal is to present an analysis based on tweets from 10 countries, also dividing the analysis into countries in the southern hemisphere and countries in the northern hemisphere. This choice is based on data availability, which can be considered a limitation of the study, in view of a non-characterization of how other countries could present differences regarding the information provided.

Minor concerns

The text of the paper needs to be revised as it has some inconsistencies regarding the information.

1. About Figure 1 - it is necessary to modify the dates of the years presented in this figure, considering that the data were collected in 2021, the years change according to the countries. In addition, I also suggest a review of the percentages presented in terms of the number of people fully vaccinated, which differ from Table 1.

We thank the reviewer for pointing out this oversight and the suggestion. Subsequently, figure 1 and table 1 were updated with the fully vaccinated population percentage (31 July 2021).

2. Section 3.2.1: the authors present some illustrations based on word clouds generated by tweets and inform that this was done for all country but in the paper, it's possible to find just 2-word clouds showing information by Great Britain and South Africa. It's important to inform why just these 2 countries were present in this analysis.

We thank the reviewer for this comment. On page 12, we state that word clouds were constructed for all countries. However, we use Great Britain and later South Africa purely as examples of the word clouds. The manuscript already has twelve figures and three tables, so adding an additional eight word cloud figures becomes impractical.

3. Section 3.2.2: Scale 5c. In the second paragraph, the word complacency is not present.

We thank the reviewer for pointing out this oversight. This has been corrected.

4. The authors do not present the limitations of the study, an important point for future studies on the subject. In view of some methodological characteristics, my suggestion is to highlight some points such as: the translation of tweets in a non-English language; choice of countries based on data convenience and availability.

We thank the reviewer for this suggestion. Subsequently, the study's limitations have been added at the end of the conclusion section.

5. There is no specific topic in the study for discussing the data, however, in the description of the results, the authors present some references that can be considered part of this discussion of the data. As this is an analytical study, it would be important to have a more in-depth discussion, comparing the results to other previous studies.

We thank the reviewer for this comment. We wish to point out that, as you state in point #6, our results are discussed based on the theoretical background provided in section 3.2.2 and the literature review provided in section 2. Furthermore, we have made changes throughout the discussion of our results referring to previous studies.

6. The manuscript has good methodological quality, is free of bias, needs some adjusts but the results are discussed based on the theoretical background properly of the manuscript theme. So, the conclusions answered the aims of the study focused on the references and results. The limitations of the study were not presented, and I suggest that this could be done.

Submitted filename: Response to reviewers.docx

Click here for additional data file.

22 Feb 2022

Positive attitudes towards COVID-19 vaccines: A cross-country analysis.

PONE-D-21-34544R1

Dear Dr. Rossouw,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Prof. Anat Gesser-Edelsburg, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: (No Response)

Reviewer #3: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: (No Response)

Reviewer #3: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: (No Response)

Reviewer #3: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: (No Response)

Reviewer #3: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: The article adds important input to readers; the authors appropriately modified the manuscript and addressed all comments

Reviewer #2: The authors replied to all the reviewer's comments and edited the manuscript in various areas. The paper may be published as is, though some weaknesses remain, for example the debate on the representativeness of tweets to reflect the feelings of the general population, including those who do not use tweets.

Reviewer #3: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

2 Mar 2022

PONE-D-21-34544R1

Positive attitudes towards COVID-19 vaccines: A cross-country analysis.

Dear Dr. Rossouw:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Prof. Anat Gesser-Edelsburg

Academic Editor

PLOS ONE