“My ADHD Hellbrain”: A Twitter Data Science Perspective on a Behavioural Disorder

Tweets related to ADHD and 99 other conditions were gathered in parallel from Twitter from July 9, 2019 to February 3, 2020 using a curated set of queries. This date range gave a sufficiently large volume of ADHD-related tweets (58,893) and spanned summer holidays and school term time, which present different challenges for children with ADHD. The quoted query “my ADHD” (not case sensitive) was used to identify tweets that were likely to be discussing the tweeter's condition. Omitting the “my” would match tweets from researchers, parents, and the media, which are not relevant here. It captures an incomplete sample of disorder discussions because there are many ways in which a person can discuss their condition without using the phrase “my ADHD”. A person claiming to have ADHD may not have been diagnosed with it and therefore the tweets are from Twitter users that implicitly claim to have the disorder.

For the set of related tweets needed for RQ2, 99 other queries for doctors or common disorders and diseases were submitted as phrase searches starting with the word “my” (e.g. “my depression”, “my doctor”, “my allergies”, “my flu”, “my acne”, “my cancer”) (see Appendix for a list). Tweets matching these were combined to create a reference set of personal health-related tweets (n=1,341,442) for comparison with the ADHD tweets. Whilst this is not an exhaustive list, it encompasses a wide range of common conditions.

Two analyses were carried out and written up independently and blinded from each other to give method (and investigator) triangulation. The results were compared only after both had been completed (see the Discussion section for the comparison).

For analysis 1, three researchers (the last three authors) conducted a standard thematic analysis (Braun & Clarke, 2006, 2013) on a random set of 200 tweets. The tweets were analysed separately rather than in the context of users’ timelines because each tweet may appear separately on the timeline of readers rather than forming a coherent whole with any prior tweets. When a tweet was a reply to a previous message, Twitter was searched to find that message and add context, when necessary.

Rooted within social constructionism, this thematic analysis followed an inductive approach based upon Braun and Clarke's (2006) six-phase method. In phase 1, the three researchers read and re-read the dataset to become familiar with the content. In phase 2, the researchers independently generated initial codes, paying close attention to patterns of similarities and variation. At this stage, emerging topics and sub-topics were discussed. For instance, medication and coping tools were considered topics within a broader theme: “managing my ADHD”. This coding and classification of topics into themes is subjective and relies on the researchers’ interpretation of each tweet and their knowledge of the general subject area. Initial coding was informed by current literature on ADHD research (e.g. Frigerio & Montali, 2016; Ringer, 2020) and then guided by the tweets. In phase 3, the three researchers independently assigned one or more tentative themes to each tweet. In phase 4, the complete set of codes across all data were reviewed and combined, disagreements among the researchers were discussed until reaching a consensus on the themes and sub-themes assigned to each tweet. In phase 5, themes and sub-themes were further defined and named, here the aim was to identify the core message of each theme; according to Braun and Clarke (2013), this exercise allows for the development of “a concise, punchy, and informative name for each theme”. In phase 6, thematic narratives and examples (rephrased quotes to respect Twitter users’ anonymity) were discussed among the researchers and written up.

For the comparison between ADHD tweets and other health-related tweets, a word frequency test was used with the social media data analysis software Mozdeh to identify words that were more common in ADHD tweets than in tweets about the other conditions. This is a qualitative big data method that avoids the face validity and understanding limitations of quantitative big data approaches (Mills, 2018) with a final qualitative stage. A 2x2 chi-square test was used for each word to check if it was in a higher proportion of ADHD tweets than non-ADHD tweets. For example, the word blame was in 0.8% (443) of ADHD tweets compared to 0.3% of the remaining medical tweets. Although this difference of 0.5% is probably too small to be noticeable by humans reading the tweets, a chi-square test revealed it to be highly statistically significant. In this example, the percentage is small and might be thought irrelevant but given that these tweets can discuss any aspect of the disorder, low prevalence overall does not equate to irrelevance.

Since the above procedure gives a huge number of simultaneous statistical tests, false positives (words that seem to have an association with ADHD but do not) are highly likely. A Benjamini-Hochberg (Benjamini & Hochberg, 1995) adjustment (in Mozdeh) was therefore used to protect the familywise statistical false positive rate. Terms were retained only if they were judged to be statistically significant with p≤0.001 after the Benjamini-Hochberg procedure. Thus, the probability that there are no errors in the set of terms found to have a word association is at least p=0.999.

Over 1,000 words occurred statistically significantly more often in ADHD tweets than in the other health-related tweets. These words were put into themes using thematic analysis with a single coder (the first author), different from the coders used for the first analysis to give investigator triangulation. The coder read a random sample of tweets containing each term to check its context for the thematic classification. This continued until saturation was judged to have been achieved (i.e. no new themes were emerging), after 200 terms. After this, the codes were revisited, clustered into major themes and reclassified when necessary to produce a set of meaningful themes.

A loophole in this process is that a term might be judged statistically significant because it had been used repeatedly by one or more users, violating the assumptions behind the chi-squared tests. To check for this this, the test was run again after filtering out multiple posts from the same user (choosing one at random). Although some words failed this second test (@usernames and event-related #hashtags), all themes were still supported. Thus, the results are robust against repeated use of a term by one or more users.

Four main themes were generated by the thematic analysis of a random sample of 200 tweets matching the query “My ADHD”: (1) My ADHD feels like; (2) Managing my ADHD; (3) Understanding, Support & Awareness of ADHD; and (4) Embracing my ADHD. Four subthemes were also identified for the second theme. Quotes given below are substantial paraphrases of the original tweets.

The word association thematic analysis is based on 200 words that are in a statistically significantly higher percentage of tweets matching the query “My ADHD” than in the set of the comparable health-related tweets, and 19 themes were identified from them. Fifteen of these themes are listed below, together with an associated substantially modified quote, a description of the theme and examples of the terms that were classified as part of that theme. The final four themes are of little interest and are named but not described.

Medication, “the dosage was upped on my meds”: Medication in general is a feature of ADHD-related tweets (e.g. meds, medicine: medicine names were recorded in a separate theme).

The themes identified from Analysis 1 confirm, but sometimes with a different perspective, the results of prior studies. Previous research noticed the need for people with ADHD to discuss their condition (Theme 1: My ADHD feels like), and has mentioned several subthemes of this, such as ADHD-related behaviour (Coppersmith et al., 2015; Guntuku et al., 2019; Wilens et al., 2009).

Coping mechanisms (Theme 2: Managing my ADHD) are clearly important, however, and these are shared or explained on Twitter. The strategy of using humour to cope with ADHD has been previously mentioned (Gajaria et al., 2011; Sedgwick, Merwood, & Asherson, 2019) and it is a natural communication device in stressful situations (e.g. Tsukawaki et al., 2019). As in the case of people experiencing pain, distress or a disability (Csordas, 1997; Leder, 1990), the strategy of constructing the self and ADHD separately may help Twitter users to make sense of their own identity and their everyday life experience of ADHD. Nevertheless, ADHD is closely associated with autism, and autism spectrum disorders seem to be linked to a lack of sense of humour and reduced humour processing (Silva et al., 2017). Medication and seeking medical help were both commented on, seeming to be important to the patient perspective, as previously noted (Gajaria et al., 2011). For ADHD, this theme seems to reflect the idea that medication can be effective in changing a person's behaviour (Charach et al., 2014). This appeared to be the reason why there were tweets about the effect of medicines and whether they had been taken.

For Theme 3, Understanding, support, and awareness of ADHD, previous studies have found evidence of teens with ADHD feeling that family, friends, and professionals do not understand the disorder (Ringer, 2020) and that teachers do not understand ADHD or how to teach students with ADHD (Walker-Noack et al., 2013). Whilst teachers may have reasonable knowledge about ADHD symptoms, they can also have misconceptions about the causes and treatment of ADHD (Ohan et al., 2008; Sciutto, Terjesen, & Frank, 2000). Misconceptions regarding ADHD among healthcare professionals (Julivia Murtani et al., 2020) and society (Gilmore, 2010; Pescosolido et al., 2008) have also been reported, although Twitter itself seems to be overall supportive towards mental illness (Budenz et al., 2020). This lack of understanding from others is often perceived as the most negative thing about having ADHD (Gajaria et al., 2011). Therefore, it is not surprising that feelings of acceptance and support from family, friends, healthcare professionals and society has been reported to be essential by teens with ADHD (Enggaard et al., 2020), in common with other long-term conditions (Lambert & Keogh, 2015). This lack of understanding tends to lead to ADHD being perceived as stigmatising (Mueller et al., 2012), for people with ADHD at all stages of their life span (Lebowitz, 2013).

Embracing my ADHD (Theme 4) was the least common theme within the sample but overlapped with the positivity instances within Theme 1. Previous research has shown that whilst some young people with ADHD do not perceive their condition in a positive light (Walker-Noack et al., 2013), others see some features of ADHD as part of their self-identity and not as impairments (Charach et al., 2014; Ringer, 2020). Positive attributes of ADHD are three times more likely to be discussed on ADHD-related Facebook groups than negative attributes (Gajaria et al., 2011) so Facebook groups worked as supportive networks where people with ADHD share their personal experiences and establish a positive ADHD collective identity. Whilst there is a tendency for adults with ADHD to be overly optimistic (Knouse & Mitchell, 2015), being positive about having the disorder seems to be desirable.

Themes that are non-trivial and different from Analysis 1, or giving a different perspective, are discussed here. Not all have been noted in previous research. The need for people to understand the Brain does not seem to have been noticed in previous studies. For example, the brain is not mentioned in one recent meta-analysis of qualitative studies of children (Ringer, 2020). Nevertheless, the functioning (rather than perception) of the brain is the focus of much research into the disorder (Brown, 2019; Wehmeier, Schacht, & Barkley, 2010). For example, studies have found relationships between ADHD and executive dysfunction (Seidman, 2006) or emotional dysregulation (Beheshti, Chavanon, & Christiansen, 2020; Shaw et al., 2014).

Two previous studies have noted formation issues related to ADHD that could be used to explain the My ADHD brain theme (this phrase was also noticed in Analysis 1) including phrases like “my ADHD ass”. These report some diagnosed children having problems with understanding their own identity, given that their behaviour is partly controlled by the disorder (Grant, 2009; Leyland, 2016; O’Connor et al., 2018; Ringer, 2020). In this context, the phrases might introduce some distance between the tweeter and the action, also expressing that they do not have the choice to avoid it. The phrase “ADHD brain” is used in popular explanations of ADHD, such as a video and article in ADDitude magazine about “neuroscience of the ADHD brain” (Silver, 2020). Thus, this or similar explanations seem to have resonated with those having the disorder to the extent that it guides how they explain themselves to others on Twitter.

The importance of a diagnosis is presumably tacitly or explicitly understood by professionals, for example as a gateway to mental health care (Shepherd et al., 2015). Tweeters also seemed to be concerned that people without diagnosed ADHD were claiming ADHD as an excuse for apparently bad behaviour, perhaps undermining the claims of people with ADHD.

The Responses theme is perhaps unsurprising given prior studies of ADHD online discussion groups (Gajaria et al., 2011), and has broad agreement with investigation of Facebook addiction that found ADHD to be a contributing factor (Gul et al., 2018). In partial contrast, technology use has been found to be similar between ADHD and other teens (Dawson et al., 2019), whereas Facebook users with ADHD seem to have fewer friends (Mikami et al., 2015). Complex sentence constructions might be related to the need to explain behaviour, hence leading to multi-part sentences (e.g. “I did X because of Y”), but does not seem to have been discussed before in the literature.

The likelihood of Co-morbidity is in line with previous research showing that comorbidity is a distinct clinical feature of both childhood and adult ADHD, with mood and anxiety disorders, substance use disorders, and personality disorders being the most frequent comorbid psychiatric conditions (Katzman et al., 2017; Kessler et al., 2006; Sobanski, 2006; Spencer et al., 2007). The new information here is that people with ADHD also talk about this online more than those with other disorders and diseases.

The themes emerging from the two analyses overlap to some extent but not completely (Figure 1). From Analysis 1, the broad Theme 1 (My ADHD feels like) matches several symptom-related Analysis 2 themes: Focus/distraction; Fidgeting; Other symptoms. It also matches the theme Self to some extent (Figure 1). Both analyses identified articulations of separation from the disorder. This was characterised as detachment from ADHD within Theme 1 (My ADHD feels like), matching the distancing characteristic of the Analysis 2: My ADHD brain theme.

Figure 1

From Analysis 1, the broad Theme 2 (Managing my ADHD) only primarily matched in terms of the Medication subtheme, which matched Analysis 2, Medication and Medicines. Coping tools and mechanisms were not identified as a theme in Analysis 2, but there is some commonality with the Accommodation and Psychiatrists themes. Coping with humour was not identified as a theme in the second analysis, although the My ADHD brain theme contained non-serious tweets. Seeking medical help was also not identified in Analysis 2, but matched Diagnosis to some extent.

The third and fourth Analysis 1 broad themes seemed to match some of the Analysis 2 themes in all respects. The term neurodivergent seemed to be used in the context of accepting ADHD in a positive sense.

Several of the remaining Analysis 2 themes did not match Analysis 1 themes. The relatively trivial themes can be ignored (Disorder name, Twitter usernames, ADHD-event hashtags). The Brain theme, focusing on the mechanisms behind the disorder, is a theoretical dimension not found in Analysis 1. In fact, the Analysis 1 coders had discussed a possible “brain talk” theme but did not include it in the final set of themes. It could therefore have been identified by the Analysis 1 methods. The Complex sentence construction linguistic theme is something that seems unlikely to be identifiable from a standard thematic analysis because it would not be the type of pattern that a coder would be looking for. The likelihood of co-morbidities had been noticed in Analysis 1, but not selected for a theme. The Analysis 1 random sample identified 13 other disorders, with depression (13% of the random sample tweets), anxiety (8%), and autism (3%) being the most common. The Responses theme might also have been identified by a standard thematic analysis but seems easy to overlook in the absence of benchmark data from other conditions.

Based on the above comparison, the standard thematic analysis method is able to identify some themes that do not translate into language use differences, for example due to language use variations within the theme (e.g. perhaps many different words are used to create humour or express positivity). In addition, the standard thematic analysis method may identify themes that are important but not unique to the topic. For example, the two coping themes found (humour and other tools/mechanisms) seem likely to be important for many health conditions and may therefore not produce words unique to ADHD. Thus, word association thematic analysis may overlook generic issues. In contrast, the word association method is better able to identify linguistic themes and can identify themes that may be overlooked without benchmark data to compare against.

Comparing the word association thematic analysis to the standard thematic analysis from a methodological perspective may shed light on any differences between the two methods. Whilst only one standard theme or subtheme (Medication) was identical to a word association theme, most themes overlapped to some extent. In general, the first method found more general themes and the second found more themes. In principal a standard thematic analysis could have produced more themes, making them more specific, perhaps by analysing additional texts. In contrast, the focus on words in the second method probably makes it difficult to generate larger themes because each word-based theme is relatively discrete, based on a set of words, and therefore relatively different from other themes based on different sets of words. Thus, standard thematic analysis may be inherently better at detecting general themes. In contrast the word association thematic analysis found more specific themes that could easily be overlooked without a benchmark collection or that may have been overlooked because they did not occur often enough to be recognised from reading texts. An advantage of the word association method here is in finding differences that are statistically significant and need to be explained (put into a theme), drawing the attention of the analyst to information that might otherwise be overlooked.

Despite the above argument, the differences between the two analyses cannot be conclusively shown to be methodological rather than accidental or due to different coders. It would be unreasonable to use the same coders with the two methods to compare the methods directly, however, since they could not forget the first method results when applying the second. Thus, the results suggest, but do not prove, that the novel method (in this context) can give novel insights compared to existing methods.