This was a mixed methods study using participatory methods [15,16] with carers of people with dementia. We held 3 focus groups with carers to identify search terms for data collection and generated an initial framework of misconceptions. Search terms (Figure 1) from carers, the literature, researchers’ Twitter searches, and dementia awareness campaigns were used to extract tweets (described next in “Tweet Collection and Screening”) and carers’ feedback iteratively refined the framework. Carers then individually categorized tweets into the framework and their interrater reliability was examined. The final framework was used by service user researchers (researchers with lived experience of using mental health services) to analyze tweets that carers categorized as negative.

Participants were recruited if they had experience caring for someone with a diagnosis of dementia. We recruited from (1) a research advisory group, MALADY [17], made up of dementia carers, and (2) Join Dementia Research, a United Kingdom–wide web-based platform hosted by the National Institute for Health Research (NIHR). Participants were included if they were at least 18 years old and were dementia carers who could give capacity to consent. A total of 7 carers were recruited and invited to take part in as many of the research activities as possible.

Publicly available tweets originating from across the world were extracted in real time between February 4 and 7, 2020, using Twitter’s streaming application programming interface (API). The connection to Twitter’s API was made via Python’s open source Tweepy library (Python Software Foundation). Tweets were captured if they contained any occurrence of the English dementia search terms identified by carers, those previously cited by Oscar et al [8], or words identified in tweets from patient advocacy groups or awareness campaigns. Most words were directly associated with dementia (see Figure 1), but some words or phrases not specific to dementia were also included because carers thought they related to negative aspects of dementia. Through a discussion with carers, there was a lack of agreement on what differentiated a positive term from a neutral term; therefore, we asked the carers to simply categorize words as either negative or neutral (which included positive) search terms. All search terms were then defined by carers as negative or neutral. The stages of analysis are shown in Figure 2. A total of 48,211 tweets were collected, 35,704 using neutral search terms and 12,507 using negative search terms (see Multimedia Appendix 1 for a breakdown of tweet collection). To manage this data set, 10,000 neutral and 10,000 negative tweets were randomly selected. From these 20,000, we selected 2000 tweets (1000 negative and 1000 neutral) that met the following criteria: (1) written in English, (2) made clear reference to dementia, (3) had a comprehensible meaning (ie, not a Uniform Resource Locator [URL] or a random string of words generated by a bot), and (4) were neutral or negative. These 2000 tweets were given to carers to carry out 2 tasks. First, carers coded a subsample of tweets (n=500) and subsequently refined their initial framework. Then, carers were given the remaining 1500 tweets to code into the final categories (see Figure 2 for an overview). This number and the types of tweets were defined through discussion with the carers on the burden tweet rating would place on them.

The study was granted ethical approval from the King’s College London Psychiatry, Nursing and Midwifery Research Ethics Committee on December 4, 2019 (HR-19/20-14565). The procedure consisted of 2 stages.

An interrater reliability analysis using the kappa statistic was performed to determine consistency among carer raters in the final development of the framework. We report the average kappa score across all carers and the range.

Service user researchers carried out a thematic framework approach [19] for the qualitative analysis of the tweets. All tweets that the carers categorized into negative framework categories as well as those categorized as “other” or “I don’t know,” and the tweets the carers omitted, were thematically analyzed as one data set, employing an inductive, holistic methodology. This process involved (1) familiarization with the data by reading through the tweets, (2) coding the tweets, (3) combining the relevant tweets together, (4) examining the codes (ie, the framework categories) to identify themes, (5) reviewing and refining the themes, and (6) defining and naming the themes. Two researchers independently conducted this analysis using NVivo 12 for Windows (QSR International), identifying themes and subthemes within the framework created by the carers. The researchers compared their coding, and any discrepancies were resolved through discussion between the researchers. A guidance document (Multimedia Appendix 4) was written with the criteria used by the 2 researchers to categorize the tweets, which was used by a third researcher to ensure consistency of the coding process when resolving any coding disagreement.

Carer participant characteristics are summarized in Table 1. See Multimedia Appendix 5 for the breakdown of carer attendance at each focus group. We found 25.94% (12,507/48,211) of our total data set contained misconceptions or stigmatizing language originating from our negative search terms.

Carers’ feedback from focus groups 1 to 3 was used to construct 6 finalized categories: 3 neutral categories (lived experience, organizational and community group statements, and individual comments on dementia-related topics) and 3 negative categories (minimizing or underestimating words/statements; dehumanizing, weaponizing, or outdated words/statements; and incorrect or questionable words/statements).

In step 3, 6 carers categorized 250 tweets each, but 3 tweets were not categorized, leaving 1497 categorized tweets. See Table 2 for the number of tweets falling into each category.

There was fair agreement between carers across 6 categories (3 neutral and 3 negative) on average in the framework (κ=0.43; range 0.067-0.7). Agreement was better when we aggregated the data to investigate agreement between neutral and negative categories, but there was still evidence that carer views differed (κ=0.92; range 0.5-1).

A total of 863 tweets were thematically analyzed from the 3 negative categories (minimizing or underestimating words/statements; dehumanizing, weaponizing, or outdated words/statements; and incorrect or questionable words/statements), as well as those categorized as “other” and “I don’t know.” All the coding discrepancies were resolved between the service user researchers. The summary of the final framework of themes is shown in Table 3 and Multimedia Appendix 6 with example tweets.

The majority of tweets were specifically insults targeted towards politicians (469/863, 54.3%), and a large portion contained general dehumanizing, weaponizing, or outdated words/statements (n=143, 16.6%). Dehumanizing language featured heavily in the tweets about politics (63/863, 7.3%), and the most frequently found words in the tweets featured American politicians alongside the words “senile” and “demented.”

There is limited qualitative research investigating dementia misconceptions on Twitter, with most literature focusing on content relating to dementia awareness [12,20] or supporting people with dementia [13,21]. To our knowledge, this is the first participatory study focusing on dementia misconceptions on Twitter to develop a framework to categorize misconceptions. We found that dementia misconceptions and weaponizing terms are prevalent and problematic on Twitter.

From the tweets extracted on dementia, 25.94% (12,507/48,211) were negative. We then extracted a sample representing half negative and half neutral tweets and validated this categorization by carers’ ratings. They rated just over half of the tweets (777/1497, 51.90%) as displaying negative attitudes, which is slightly over the 50% (750/1500) of these tweets extracted using negative search terms. Most negative tweets were insults targeted towards politicians. Our prevalence of negative tweets (12,507/48,211, 25.94%,) is similar to previous work by Oscar et al [8], who found 21% of their Alzheimer disease–related tweets (N=6583) used Alzheimer disease–related words to perpetuate stigma. Their analysis was carried out by 2 researchers manually coding only 311 tweets across 6 broad categories (metaphorical, personal experience, informative, joke, ridicule, organization). Our participatory work focuses on the end-user views—the carers’ ratings and views of misconceptions. We found an overlapping theme in “jokes,” but through our qualitative analysis, we were able to highlight that jokes manifest as minimizing or underestimating words or statements. This high prevalence of misconceptions and stigma in tweets is mirrored in research investigating other neurological conditions. For example, McNeil et al [22] found 41% of tweets using the word “seizure” were derogatory in nature, and likewise found ridicule or jokes were common in these tweets. These misconceptions towards dementia are also widespread in the general population and are not exclusive to views disseminated on social media. Crisp et al [23] found that over half of the UK adults surveyed expressed negative attitudes towards people with dementia, including that they were unpredictable, hard to talk to, and feel things in a different way than other people.

We employed an inductive methodology to categorize each tweet into 1 theme. This approach has also been applied in previous qualitative research [12], but others adopted deductive approaches (with categories decided a priori) to categorize almost 70% of tweets to multiple dimensions [8]. We made the conscious decision to involve carers from the very beginning to develop a framework based on their experiences, and then employ an inductive approach for our qualitative analysis. This was important as this is the first piece of research to focus specifically on dementia misconceptions on Twitter, but it also ensured that we captured the meaning of the tweet from the recipient’s viewpoint (taking an emic perspective [24]), particularly given that tweets are short snippets of text which can lack context.

This study has significant public health implications. We provide terms that carers of people with dementia consider to be misconceptions or stigmatizing towards dementia. Therefore, social media platforms should incorporate these terms into their algorithms to enable users to filter out any tweets containing these negative terms. As these terms have been generated by carers after conducting Twitter searches, their validity is reinforced as they have been rated as negative by the people they affect the most.

Additionally, these terms could be used to identify Twitter users who propagate these attitudes and target them in an awareness campaign to reduce their misconceptions. This would aim to promote awareness of the use of words which can perpetuate stigma around mental illness, benefiting the reduction of stigma related to any mental illness [25].

Understanding what constitutes stigmatizing or weaponizing language on Twitter requires the incorporation of personal perspectives, but this approach is rare. Previous studies investigating misconceptions or stigma in mental health have rarely consulted with service users or carers [8,26,27]. Our participatory methods ensured that our framework is grounded in the personal perspective of those who will be affected by the poor use of language.

Our sample of tweets thematically analyzed by researchers (n=863) is larger than those in previous studies, such as Cheng et al [12] (n=398) and Oscar et al [8] (n=311), and this broader sample provides a better understanding of the prevalence and forms of dementia misconceptions on Twitter. However, many of our tweets were related to American politics; therefore, future work should consider using a broader time period to understand whether this effect is one of time (an election period) or one of American politics in general. The timing of tweet collection will have affected the prevalence of tweets relating to politicians and the rate may be lower if tweets are collected at other times.

Additionally, we extracted tweets during UK office hours and, therefore, overnight events would have been captured the following morning. This may not have allowed us to capture the initial conversations surrounding controversial events. This work only focuses on Twitter and Twitter users, who may not represent the general population [28] or users of other social media platforms. Future work should investigate misconceptions on other social media platforms and in the wider general public.

Our carer group was small and consisted predominately of White British participants, and there was mixed agreement by carers on what constitutes misconceptions and stigma. We found that agreement about tweet categories was greater when assessing whether a tweet was negative or neutral, rather than its individual category; some tweets could be interpreted as stigmatizing by one person, but not by another. Our findings reflect the heterogeneity in neurological and mental health conditions, combined with societal and cultural factors, which shape how individuals communicate and understand their mental health [29]. We propose that future work ensures not only a larger group, but also a more diverse group of carers, patients, and members of the public classify tweets, and that clinical, social, and cultural data are used to understand some of their personal reactions.

This study demonstrates the importance of coproduction in assessing dementia misconceptions. Contributions from people with lived experience and carers can provide a perspective that may be overlooked by researchers. We highlight the high frequency of misconceptions or weaponizing language used in dementia-related tweets. The most commonly used terms are “demented” and “senile” to disparage American politicians including Nancy Pelosi, Donald Trump, and Joe Biden. These findings may prove to be useful to inform a campaign aiming to reduce these misconceptions, correct people’s misunderstandings of dementia, and highlight the effect their words have on carers of, and people with, dementia.

This work was supported by the NIHR Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London (IS-BRC-1215-20018), and Alzheimer’s Research UK’s Inspire Fund awarded to SJ.

Data available upon request from SJ.