Longitudinal Analysis of Behavioral Factors and Techniques Used to Identify Vaccine Hesitancy among Twitter Users: Scoping Review

"Digital data can help portray the dynamics of public health surveillance systems and allow public health professionals to pinpoint the general concerns or needs of the public during infectious disease events to create location-specific campaigns."

Attitudes toward vaccination - specifically, vaccine hesitancy - pose a potential threat to achieving sufficient coverage and community immunity. Anti-vaccine content is prevalent across social media platforms; while it may be posted by a minority of users, such content frequently generates greater user engagement than neutral or pro-vaccine content. The consequences of these online peers are unclear and my cause a behavioural shift in user sentiment toward vaccination. This scoping review aims to identify the community and individual factors that longitudinally influence public vaccination behaviour. The secondary aim is to gain insight into techniques and methodologies used to extract these factors from Twitter data.

Longitudinal analysis involves performing in-depth analyses of users' health and behavioural modifications with time. During this process, the researchers analyse the data collected from the same subjects or participants over a period of time. It is often used to examine changes, trends, or relationships within a specific group or population. This study focuses on longitudinal vaccine-related content on Twitter. Twitter debuted on March 21 2006, so the search was limited from 2006 to the present. A search was performed on 10 databases using relevant keywords and search queries. English-language studies that considered not only Twitter data but also other social media data were included in this study. Twenty-eight articles, out of 705 relevant studies, were selected for inclusion.

Most of the studies (N = 11 ~ 39%) were published in 2022, while (N = 9 ~ 32%) studies were published in 2021. Among the 28 studies, 20 (~72%) of the longitudinal studies on user behaviour analysis toward vaccination are most recently reported. Out of the 28 finalized longitudinal studies, only one-third encompassed long-term analysis, exceeding one year in duration. Long-term analysis offers enriched evidence to gain insights about the attitude patterns of the population that disseminate information about vaccines on Twitter and other social media tools.

Table 1 in the paper presents the overall taxonomy of the 28 articles. It contains information about the number of data samples, type of vaccine studied, and range of years followed for users' behaviour longitudinal analysis. It also contains information about the techniques or methodologies used to perform this longitudinal analysis.

Two main themes emerged, including: (i) individual and community factors influencing public attitudes toward vaccination and (ii) techniques employed to identify these factors. These broad themes are further divided into sub-themes; for example, in the case of (i), individual factors are classified into contextual factors, individual and group factors, or vaccine-specific factors, and community factors are classified into community-specific factors like politicians, religious and other influential activists, and the media. The methods/techniques to identify vaccine hesitancy - that is, (ii) - are divided into community detection methods, which are further divided into machine learning methods and statistical methods. The machine learning methods are then further dissected into shallow and deep architectures. The shallow architectures use binary patterns for classification and identification, while the deep architectures use neural network-based models for classification and identification purposes.

As an example, the capability, opportunity, motivation - behaviour (COM-B) model is presented to identify the factors influencing Twitter users' behaviour toward COVID-19 vaccination. The included studies identified different individual and community-based features that directly or indirectly affect public behaviour toward vaccination. Community factors encompass social norms, values, beliefs, social support networks, local institutions, community engagement, and economic opportunities within a community. Individual factors include sentiment, emotions, user demographics, beliefs, attitudes, prevention, knowledge/awareness, and vaccine-specific factors such as misinformation and vaccine campaigns. The findings reveal the significance of mass media in influencing information-seeking behaviour.

In the included studies, numerous machine learning-based methods are reported for vaccine hesitancy and calculation of users' vaccination sentiment. These data collection strategies and methods are discussed in the paper.

Based on the analysis, some recommendations are suggested in an effort to open new gates for the research community to explore:

• User-generated content on Twitter is often subject to bias, as it tends to reflect information that individuals feel comfortable sharing, which may not accurately represent the full range of their emotions and experiences. Among the 28 studies analysed, no longitudinal study was found that linked the findings with users' subjective experiences, whether self-reported or not, using text, image, or video data types. Therefore, there is a significant gap in research that can identify and address content biases that impact the collection and analysis of digital data for studying vaccination-related behaviours.
• The anonymity provided by the internet allows individuals with stigmatised attributes to benefit from supportive communication on Twitter. However, the challenge of accurately determining user demographics raises unresolved questions about the population biases present among internet users with diverse cultural backgrounds or socioeconomic statuses. Demographic data for most digital platforms are not representative at a national level and tend to be skewed toward younger age groups and users with higher levels of education. This topic remains underreported by the research community.
• The review found no studies that assessed digital media utilisation for vulnerable populations (e.g., low-income, older adults, or people with a disability) who are underrepresented on different digital platforms. Studies on detecting social bots are scarce.
• For longitudinal analysis, a considerable timeframe is required to perform an enriched analysis of different individual and community factors associated with vaccine hesitancy among Twitter users. However, in the included studies, only 3 studies selected a range of years greater than 2 years. Similarly, only one study of the included 28 articles used a dataset greater than one million tweets.
• Among the studies included in the analysis, a mere 4 studies (0.14%) took into account social media platforms other than Twitter for behavioural analysis. To conduct a more comprehensive and in-depth psychological analysis of user behaviour, the researchers suggest considering other scientific and nonscientific platforms as well.
• There is no significant contribution toward longitudinal analysis from low-income countries like Pakistan, Bangladesh, India, and many others. This dearth requires attention from the research community.
• The lack of studies reporting the explainability aspect of different approaches used for Twitter users' longitudinal behaviour analysis underscores the need for further research in this area.

In conclusion: "The findings of this scoping review will provide valuable insights for healthcare administrators and policymakers to understand the factors associated with vaccine hesitancy among different cohorts engaging in Twitter discussions. This understanding will facilitate the planning of vaccination campaigns and help improve the uptake rates of various vaccines."