Social media

Social media has become an area of interest around the world because of the marketing potential, business opportunities, and information dissemination potential [5]. Social media is also a form of socialization between members of the public in digital spaces using platforms to engage around a particular topic or within a particular discussion area around published media [6]. There are two types of engagements on social media: private engagements and public engagements. Private engagements are held between selected individuals in a closed system, and only the chosen participants have access to the dialogue. Public engagement on the other hand is conducted in virtual spaces where everyone with access to a network, such as the internet, can view the dialogue and potentially contribute toward the conversation [7]. Social media is seen as a two-way communication space where people may both send and receive information. However, social media alone is not enough to measure public discourse, but rather an addition to the discourse analysis of public dialogue. Social media data can have some biases due to an increase in social stigmatisation, either through social media users being more reluctant to share their true opinions or by being more radical to gain social approval [8]. Despite the biases, more users are engaging, and the rate of engagement is more rapid than before the 2020 global pandemic [9]. This means that the adoption and use of social media is widespread, global, and increasing. Social media, by its very nature, creates an ecosystem where users engage in continuous social comparison, ideological conflicts, and exposure to emotionally charged content [10].

Social media and social stress

Social stress, a psychological and emotional burden experienced in response to social interactions and societal pressures, has become increasingly prominent in digital spaces [11]. Unlike traditional forms of social interaction, where stressors may be more immediate and localized, social stress on digital platforms is persistent, amplified, and far-reaching [12]. This stress emerges not only from direct interactions—such as cyberbullying, harassment, or negative feedback—but also from the broader influence of social norms, misinformation, performative behavior, and the fear of missing out (FOMO) [13]. At present, some social media platforms have policies regarding the type of content that it will allow—and if the content falls outside of the prescribed policy, the comment may be removed by the platform owners or content moderators, depending on the policy [14]. Additionally, some social media platforms may not have this type of content moderation, exacerbating the potential for harmful content which will in turn potentially increase social stress. Globally, there is a gap in research related to the identification, monitoring and evaluation, and impact assessment of social media and social stress emanating from social media platforms [15].

A key driver of social stress in online environments is the pressure to conform [16]. Public engagement on social media often leads to self-censorship, performance-based interactions, and an excessive need for validation through likes, shares, and comments. This creates an inauthentic self-presentation, where users curate idealized versions of their lives, contributing to collective stress and unrealistic societal expectations. The need to engage, debate, or defend opinions in real-time, often in polarizing or hostile environments, further escalates stress levels among users. Moreover, misinformation, propaganda, and algorithm-driven content amplification contribute to collective anxiety, particularly in times of crisis or uncertainty [17]. While content moderation policies on social media platforms attempt to regulate harmful interactions, they often fail to capture the nuanced and evolving nature of social stressors. Sarcasm, passive-aggressive remarks, exclusionary behavior, and ideological attacks may fall outside explicit moderation guidelines yet still contribute significantly to social stress. Moreover, the anonymity and virality inherent in social media exacerbate these stressors, as users may feel emboldened to engage in hostile interactions without immediate consequences. The unfiltered, rapid-paced nature of public discourse can create a digital environment where stress compounds quickly, affecting both individuals and entire online communities [18].

Sentiment analysis (SA) and topic modeling (TM) have been used to measure online anxiety and stress [19]. On the one hand, SA deciphers the emotional tone of user-generated content, distinguishing between positive, negative, and neutral sentiments. This analysis provides insights into the collective emotional state of users, which is essential for assessing social stress levels. For instance, a study analyzing Reddit discussions used SA to gauge stress and anxiety expressions, highlighting its effectiveness in understanding user emotions in social media contexts [20]. On the other hand, TM identifies prevalent themes within large datasets. By uncovering dominant topics, researchers can pinpoint specific stressors or concerns affecting the online community. However, SA and TM alone do not provide evidence into the life cycle of a specific social stress concept. Given the complexity of social stress in digital spaces, traditional research methods have struggled to quantify its effects. The challenge lies in the scale of differentiating between temporary emotional reactions and sustained social stress patterns that contribute to poor mental health outcomes. While SA provides a general emotional snapshot, and TM helps identify dominant discourse themes, these methods alone are insufficient to track the life cycle of social stress [21]. To bridge this gap, social stress must be measured dynamically factoring in the frequency, intensity, and evolution of stress-related discussions over time. Given that social media is complex, and a form of public discourse, it remains challenging to study dialogue once the volume of a conversation reaches millions of interactions [22].

Another vital component of measuring social stress for social media is the incorporation of the information seeking behavior of people [23]. If a conversation is topical at that particular point in time, then the context of the conversation relevance can be added. If we can measure the economy in real time, why can’t we measure the rising waves of social stress that affect millions daily? Therefore, the purpose of this study was to design and evaluate a social stress indictor for capturing the real-time manifestation of social stress on social media platforms.

Research design

This study used a multistep methodological approach to develop and validate a Social Stress Indicator (SSI) by comparing it with an expert-labeled social stress score (SSS). The SSI consisted of SA, SUB, and ISB. The selection of these wer theoretically and empirically grounded based on their demonstrated ability to capture distinct yet complementary aspects of social stress as manifested on social media platforms. SA is widely used in affective computing to infer users’ emotional states from textual data, with negative sentiment consistently associated with heightened stress, anxiety, and distress in online communities [19,20]. SUB was included to measure the degree to which posts express personal opinions, beliefs, or perceptions—an essential proxy for emotional reactivity and interpretive bias, both of which are known stress amplifiers in digital discourse. Finally, ISB was incorporated to account for the temporal salience and societal relevance of stress-inducing topics. Drawing from information science and infodemiology, spikes in search interest or digital engagement with specific themes are often indicative of collective concern or uncertainty, particularly during crises or unfolding events [17,23]. Together, these three components were selected for their ability to provide a multi-dimensional, real-time representation of evolving stress narratives online. They serve not only as quantitative markers of affective and cognitive states but also as indicators of broader discursive trends, enabling the SSI to function as both a diagnostic and predictive tool for identifying stress-laden content across social media ecosystems. A synthetic data set containing 1000 low-stress, 1000 medium-stress and 1000 high-stress posts was used to conduct the experiments. To establish an independent ground truth measure, an expert labeled SSS was created using a rubric that evaluates each post in five dimensions: negative sentiment, anxiety expression, level of engagement, misinformation content and help seek behavior. Experts assign a score of 0-5 for each dimension, and the final SSS per post is calculated as the equally weighted average of these scores. The SSS is then aggregated over time windows to facilitate a direct comparison with the SSI. The threshold for low social stress was set to below 20% (0.2), moderate to 21%–40% and high above 40%.

Ethics statement

All permission to use synthetic data and ethics for this study was obtained under the ethical clearance numbers UFS-HSD2020/1846/2601 and UFS-HSD2023/2095. Because the data was synthetic, informed consent was not needed, as the data was generated using a large language model. The prompts to generate the data encouraged that no demogrtaphic slurs or any demographic isolation was used. Additionally, each data point was evaluated by the researcher to ensure that no profanities of any kind featured within the data. Furthermore, data collection had a synthetic time stamp attached to it, to represent what real-world data could look like, but no real-world data were used, only synthetic data. The microblogs themselves did not contain profanity or hate speech. Data were generated in-house and no human participants were involved in the generation of data.

Data

Synthetic microblogs were generated using ChatGPT based on prompting. The prompts focused on emulating ’social media like’ microblogs, which means that the microblogs included hashtags as well as emojis. No human subjects were used in this study, and 3000 microblogs were generated using various prompts to ensure the uniqueness of the microblogs. All the microblogs contained a variety of different information. There were a different number of microblogs generated for each time stamp over a 7-day period for 24 hours per day. All the data is in a public repository [24]. In this study, no human data was used, only the synthetic data generated using a large language model.

Social stress score

The SSS was designed to quantify social stress at the microblog level by evaluating key psychological and behavioral indicators and was adapted from the standard stress scale [25]. It was constructed using a rubric that scores each post across five dimensions: sentiment negativity, anxiety expression, engagement level, misinformation content, and help-seeking behavior, with each dimension rated on a scale from 0 to 5. The final SSS for a post is calculated as the equally weighted average of these scores, providing a standardized measure of social stress. Additionally, five expert raters trained in social sciences applied the rubric to synthetic microblogs generated to reflect diverse stress-inducing contexts. Each of the five experts independently scored each microblog and the average score was aggregated and used in the SSS rubric. The adaptation also included iterative refinements based on inter-rater reliability assessments, ensuring consistency and validity in measuring social stress across varying content types and linguistic expressions. The average SSS scores were used for the statistical comparisons between the SSS and SSI.

Social stress indicator

The polarity of a person’s emotions or opinions on a certain topic or event was measured using SA, as either positive (to a maximum value of 1 or 100%) or negative (to a minimum value of −1 or ) number. SA measure was thus defined as (Eq 1):

(1)

where SA is the sentiment, N is the sample size of the number of microblogs measured and is the sentiment of each microblogs per time stamp. The average of the sentiment is then used as the SA for that time interval of the dataset. For this indicator there will thus be an average sentiment score ranging between −1 and 1 for each time stamp in the series. A social media lexicon was used for the SA. Subjectivity analysis was conducted using a lexical-based approach. In this study, each microblog was processed to extract subjective lexical units from a predefined subjectivity lexicon. For each microblog, the set L denoted the top subjective words identified from the text. The subjectivity score, denoted by for a given microblog t, was computed as follows (Eq 2):

(2)

where S(l_i) represented the subjectivity rating of the word l_i as provided by the lexicon (used in Text Blob and VADER), was the total number of subjective words identified in the microblog, and ε was a small smoothing constant added to avoid taking the logarithm of zero. The logarithmic ratio measured the relative strength of subjective language by comparing the probability of subjectivity to its complement (objectivity). This formulation was conceptually adapted from the coherence measure used in topic modeling where instead of assessing word co-occurrence, the focus was on the degree of subjectivity inherent in the text [26]. Subjectivity scores were calculated for each microblog individually, thereby quantifying the extent to which each text conveyed personal opinions and emotional content. These scores, in conjunction with sentiment polarity measures and other components, were averaged for each time interval for the overall indicator. ISB in the context of digital platforms like social media was defined as the measurement of the number of people that searched for a related topic, over a certain period of time on that platform. The ISB measure was calculated as (Eq 3):

(3)

where β is the frequency of searched topic at a specified time, and is the highest frequency of the same topic at a specific point in time, expressed as a percentage. The measurement of ISB is always in the context of the series, and for this, Google trends were used. Once the SA, SUB, and ISB has was determined, the SSI indicator was calculated for each timestamp in the series. SSI was defined as (Eq 4):

(4)

where SA is seen as sentiment analysis (Eq 1) with a negative designation due to the emphasis on sentiment negativity, SUB is seen as subjectivity rating subtracted from 1 (Eq 2), and ISB is seen as information seeking behavior (Eq 3). In the case of the initial SSI indictor, the weightings are assumed as equal.

Comparison between SSS and SSI

To compare the SSI and the expert-labeled SSS, Bland-Altman analysis and Granger causality testing were conducted. The Bland-Altman analysis assessed the level of agreement between the two measures by computing the mean bias, standard deviation of differences, and limits of agreement (LoA). This method was used to determine whether SSI and SSS were interchangeable in measuring social stress. Additionally, Granger causality tests were performed to examine whether past values of one variable could predict future values of the other, helping to identify whether SSI led or followed expert evaluations (SSS). Confidence intervals were calculated for both tests, with the Bland-Altman analysis focusing on measurement agreement and the Granger causality test identifying predictive relationships over time. In both cases, the ranges with the 95% confidence intervals were compared and not the absolute values.