In an Absolute State: Elevated Use of Absolutist Words Is a MarkerSpecific to Anxiety, Depression, and Suicidal Ideation
Absolutist thinking is considered a cognitive distortion by most cognitivetherapies for anxiety and depression. Yet, there is little empirical evidence ofits prevalence or specificity. Across three studies, we conducted a textanalysis of 63 Internet forums (over 6,400 members) using the Linguistic Inquiryand Word Count software to examine absolutism at the linguistic level. Wepredicted and found that anxiety, depression, and suicidal ideation forumscontained more absolutist words than control forums (ds >3.14). Suicidal ideation forums also contained more absolutist words thananxiety and depression forums (ds > 1.71). We show thatthese differences are more reflective of absolutist thinking than psychologicaldistress. It is interesting that absolutist words tracked the severity ofaffective disorder forums more faithfully than negative emotion words. Finally,we found elevated levels of absolutist words in depression recovery forums. Thissuggests that absolutist thinking may be a vulnerability factor.
Absolutist thinking underlies many of the cognitive distortions (Beck, 1979; Burns, 1989) and irrational beliefs (A. Ellis & Harper, 1975)that are purported to mediate the core affective disorders. Words, phrases, and ideasthat denote totality, either of magnitude or probability, are often referred to as“absolute.” Absolutist thoughts are independent of context and unqualified by nuance. Inthis observational study, we aimed to measure absolutist thinking in a specific andecologically valid manner. We then compared its relative association between a varietyof affective and nonaffective groups.
Absolutist thinking has strong empirical links to three distinct mental health groups:suicidal ideation, borderline personality disorder (BPD), and eating disorder (ED).Regarding suicidal ideation, structured response formats have shown more extreme valuejudgments by suicidal patients than controls (e.g., Neuringer, 1961, 1964). Thematic analysis by independent ratersalso deemed the stories and poetry of suicidal individuals as highly “polarized” (Litinsky & Haslam, 1998;Wedding, 2000). Inaddition, dichotomous thinking, cognitive rigidity, and problem-solving deficits havebeen repeatedly found to co-occur in suicidal individuals (for review, see T. E. Ellis & Rutherford,2008). This is supported by a series of empirical studies from Pollock and Williams (1998,2001, 2004; J. M. G. Williams & Pollock, 2008).
BPD patients also make more extreme responses on structured response formats thancontrols (e.g., Moritz et al.,2011; Napolitano &McKay, 2007; Sieswerda,Barnow, Verheul, & Arntz, 2013; Veen & Arntz, 2000). Some scholars haveused “spontaneous reactions” or short interviews to identify extreme or dichotomousthinking styles (e.g., Arntz &ten Haaf, 2012; Arntz& Veen, 2001).
With respect to ED, the Dichotomous Thinking in Eating Disorders Scale (Byrne, Allen, Dove, Watt, &Nathan, 2008) is widely used in ED studies (e.g., Antoniou, Bongers, & Jansen, 2017; Palascha, van Kleef, & van Trijp,2015). Although obesity and anorexia are often studied separately, they bothlink to absolutist thinking. For obesity, several reviews have found that avoidingabsolutist dichotomous thinking improves weight loss maintenance (e.g., Ohsiek & Williams, 2011).Absolutism often takes the form of perfectionism in anorexia, as identified throughclinical observations (e.g., Fairburn, Cooper, & Shafran, 2003; Garner, Garfinkel, & Bemis, 1982),structured response formats (e.g., Feixas i Viaplana, Montebruno, Dada, Castillo, & Compañ, 2010; Zotter & Crowther, 1991),and interviews (e.g., Johnson &Holloway, 1988).
Despite the inclusion of absolutist thinking into many cognitive therapy models foranxiety and depression (Beck,1979; Burns, 1989;C. Williams & Garland,2002), this association remains mostly neglected in the empirical literature(A. Ellis, 1987). In anotable exception, Teasdale et al.(2001) found that an “absolutist, dichotomous thinking style” predictedfuture depressive relapse, over and above the content of responses. This was evidencedby both positive and negative “extreme responses” on Likert-type scales.
Attempts to investigate absolutist thinking have mostly employed some type of structuredresponse format. Ertel (1985)was the first to use quantitative text analysis to measure dogmatism with the manualDogmatism Text Analysis Tool. More recently, with the advent of automated text analysis,Cohen (2012) measured“cognitive rigidity” in the “spontaneous autobiographical narratives” of undergraduatestudents and found correlations with negative emotionality. Unlike structured responseformats, these natural language text analysis studies have more ecological validity.
With the growth of social media, Internet forums are increasingly being used as a sourceof naturalistic writing for research in depression and other affective disorders (e.g.,Fekete, 2002; Griffiths, Calear, & Banfield,2009; Houston, Cooper,& Ford, 2014). It is believed that insights into the cognitive processesassociated with particular affective disorders can be gleaned from how people with thosedisorders write about their experiences. In three connected studies, we investigated thefrequency of absolutist words contained in different affective and nonaffective Internetforum groups (Table 1; formore details, see TableS1 in the Supplemental Material available online). In the first study wecompared anxiety, depression, and suicidal ideation (test) groups with general, asthma,diabetes, and cancer (control) groups. We had two specific hypotheses:
aGeneral forums = Mumsnet (Women), Ladies Lounge (Women), Gentlemen’s Club(Men), Ask Men (Men), Pensioners Forum (Elderly), Student Room (Young), WorkProblems. bNumber of Internet forums in each group.cNumber of members who contributed to that group’s corpus.
Hypothesis 1 (H1): The percentage ofabsolutist words in anxiety, depression, and suicidal ideation test forum groupswill be significantly greater than in Study 1 control forum groups.
Hypothesis 2 (H2): The percentage ofabsolutist words in the suicidal ideation forum group will be significantlygreater than in both anxiety and depression forum groups.
Our second hypothesis is partly based on the strong association between suicidal ideationand absolutist thinking (for review, see Arffa, 1983). But also, as suicidal ideation isthe more severe mental health concern, it could be hypothesized that absolutist thinkingwill be correspondingly more extreme.
In Study 2, our aim was to show that absolutist words reflect absolutist thinking, ratherthan psychological distress. We attempted to control for psychological distress bycomparing groups believed to have similar levels of negative emotions but differentlevels of absolutist thinking (Table 1 and TableS1). We compared mental health groups strongly associated with absolutistthinking (BPD and ED, cited above) with mental health groups less associated withabsolutist thinking (posttraumatic stress disorder [PTSD] and schizophrenia). Althoughwe recognize that PTSD and schizophrenia may also have some links to absolutistthinking, the literature suggests these links are likely to be much weaker than those ofBPD and ED. Relatively few researchers have examined absolutist thinking in PTSD andschizophrenia, and these have often been limited or produced mixed results (e.g., Colbert, Peters, & Garety,2010; Joseph & Gray,2011). Conversely, there is a widespread consensus, based on a multitude ofstudies, that BPD and ED are firmly linked to absolutist thinking (e.g., Alberts, Thewissen, & Raes,2012; Napolitano &McKay, 2007; Veen &Arntz, 2000). We also measured the frequency of negative emotion terms tofurther support the assumption that the four mental health groups had comparable levelsof negative emotions.
Hypothesis 3 (H3): The percentage ofabsolutist words in BPD and ED test forum groups will be significantly greaterthan in PTSD and schizophrenia control forum groups.
In Study 3, we aimed to determine the extent to which absolutist thinking could be acognitive vulnerability factor for depression and suicidal ideation. In a subset ofdepression and suicidal ideation forums, there are “recovery” subforums (Table 1 and TableS1). These subforums are visited by members who feel they are currentlyout of depression. They often write very positive posts about their progress and wordsof encouragement to other members. Theoretically, a cognitive vulnerability factorshould not only be present during an episode of depression but also persist duringrecovery. Therefore,
Hypothesis 4 (H4): The percentage ofabsolutist words in the recovery forum group will be significantly greater thanin Study 1 control forum groups.
Previous text analysis research has examined many different dictionary “dimensions.” Whenanalyzing written samples from anxious, depressed, or suicidal individuals, an increaseduse of personal pronouns and negative emotion words has commonly been found (Bucci & Freedman, 1981;Fekete, 2002; Lorenz & Cobb, 1952; Rude, Gortner, & Pennebaker,2004; Stirman &Pennebaker, 2001; Weintraub, 1981). In particular, pronouns have been identified as having astronger relationship with affective disorder than negative emotions (Pennebaker & Chung, 2013).Like pronouns, absolutist words are functional; they help determine our style ofwriting, not its contents. Moreover, functional words are ordinarily outside ofconscious control (Pennebaker &Chung, 2013); therefore, they can serve as implicit markers. We believe ashift in focus to how we think rather than what we think can provide greater insightinto possible cognitive mechanisms underlying affective disorders.
From the outset, we identified and validated a single dictionary of interest, as thisstudy was motivated by specific a priori hypotheses. This is in contrast to previoustext analysis studies that have used a subset of already constructed dictionaries oridentified features of interest based on the data itself (e.g., using an iterativeprocess with cross-validation and feature reduction; Mladenić, 2005). The large data set in thisstudy, from 12 different groups, representing 63 different Internet forums and more than6,400 members, afforded a degree of ecological validity not achievable in experimentalstudies. However, as with many observational studies, these benefits come with inherentcosts. We had limited information about the members posting in the forums, and for themost part, their true identities and motivations were unknowable. Recognizing thislimitation, we hope that follow-up studies, using alternative experimental designs, willextend the findings presented here.
We used English-language Internet forums as a source of naturalistic writing forour test and control categories. For all three studies, representative websiteswere located through a Google search (search words: e.g., “suicide forums,”“asthma forums”). Forums were selected for inclusion into the study on the basisof Google rank (Table1 and Table S1), were popular (thus yielding sufficient data foranalysis), and were actively moderated with clearly written moderation policies.Each group in the test and control categories was composed of between four toseven separate forums, as determined by forum availability. For Study 1, controlgroups were carefully selected to provide the broadest level of control. The“general” group provides a gender control with two forums for female members(Mumsnet and Ladies Lounge) and two for male members (Askmen and Gentlemen’sClub). The general group also controls for age, with a designated forum foryoung members (Student Room) and older members (Pensioners Forum). The asthmaand diabetes groups control for chronic physical illness, and the cancer groupcontrols for severe physical and psychological distress. Study 3 recovery forumswere located within Study 1 depression and suicidal ideation test forums.
Forum members can either introduce a new topic (“first posts”) or contribute toan ongoing discussion (“replies”). In the interest of simplicity andinterpretability, only first posts were collected. Posts were copied and pastedinto a text document ready for subsequent text analysis. Where an individualmember contributes multiple posts, these were combined into a single textdocument. All text files used in this study are hosted on Figshare(doi:10.6084/m9.figshare.4743715). If a forum was further divided intosubforums, only the single most appropriate subforum was used (Table S1). For each test and control forum, we aimed to collect30,000 words. Seven out of the 63 forums were not large enough to provide a30,000-word corpus but were nevertheless retained in the study as they surpassed10,000 words. Posts were only collected if they met our selection criteria: (a)contain a minimum of 100 words, (b) be authored by a representative member ofthat online community (i.e., not written on behalf of someone else/news articleetc.), and (c) be written in continuous prose (i.e., not lists, poems). Postsfrom all test and control forums which met the selection criteria were collectedsequentially as presented by the respective forum website (usually by dateorder). Posts were collected between April and May 2015 and December and January2016. All data in this study was collected from the public domain; therefore,although ethical consideration is still important, informed consent is notrequired. This complies with the University of Reading research ethicsguidelines and the ethical guidance for internet-mediated research set out byThe British Psychological Society (British Psychological Association,2013). The aggregate data used in this study are hosted on Figshare(doi:10.6084/m9.figshare.4743547.v1).
Word count text analysis
Word counting text analysis was conducted using validated dictionaries thatcharacterize a particular linguistic dimension (i.e., negative words, auxiliaryverbs, family related words). For this study, we validated an absolutist and anonabsolutist words dictionary using independent expert judges.
Absolutist and nonabsolutist words indicate magnitudes or probabilities; absolutewords do so without nuance (i.e., always, totally, entire), whereas nonabsolutewords indicate some degree of nuance (i.e., rather, somewhat, likely). Bothdictionaries are composed of functional words devoid of valence, mostlyadverbial intensifiers or modal verbs. A subclass of nonabsolutist words, whichwe have termed “extreme words,” indicate extreme (but not absolute) magnitudesor probabilities (i.e., “very”). Although the terms extreme andabsolute have previously been used interchangeably (e.g.,Teasdale et al.,2001), we treat them here as qualitatively distinct.
To construct these dictionaries, we initially brainstormed more than 300absolutist words and 200 nonabsolutist words (including extreme words). Testingon pilot data (control and test groups) revealed that many of the words on theseoriginal lists were too obscure to register with sufficient frequency foranalysis. Consequently, the original dictionaries were reduced to the mostprevalent 22 absolutist words and 43 nonabsolutist words (including 21 extremewords). Although this was based on a mostly arbitrary cutoff, it was intendedthat the lists be large enough to produce representative dictionary percentages,but small enough to facilitate independent validation by experts. The 22absolutist words and 43 nonabsolutist words were combined into a single list of65 words. Five independent expert judges were asked to categorize them asabsolute, nonabsolute, and/or extreme. Two of the judges are clinicalpsychologists from the University of Reading Charlie Waller Institute and threeare linguists from the University of Reading School of Clinical LanguageSciences. Judges were permitted to place words into more than one category(i.e., extreme and absolute). The agreement between our original categorizationof the words (absolutist/nonabsolutist) and that of the judges ranged between83% and 94%, whereas the interjudge agreement was 96%. Words were consideredabsolute, extreme, or nonabsolute on the basis of a majority decision by thejudges. Three words, anything, need, andneeded, were moved from the absolutist dictionary to thenonabsolutist dictionary as they were not categorized as absolute by themajority of judges. All the words on our nonabsolutist dictionary were judgednonabsolute. Judges showed almost no agreement on extreme words, this categorywas consequently removed from the analysis (collapsed into the nonabsolutistcategory).
The resulting 19-word absolutist dictionary is shown in Table S2 in the Supplemental Material. Both dictionaries wereused in the text analysis of test and control groups. We also ran dictionariescontained within the Linguistic Inquiry and Word Count program (LIWC; Pennebaker, Booth, Boyd, &Francis, 2015). This program provides 73 validated dictionariescovering a wide range of “dimensions” (i.e., questioning words, affectiveprocesses, auxiliary verbs). All dictionaries, other than the absolutistdictionary, were run purely for the benefit of comparison.
The LIWC text analysis software was used to test our absolutist and nonabsolutistdictionaries as well as the LIWC dictionaries. It calculates the prevalence of agiven dictionary as a percentage of the total number of words analyzed.Throughout, we have referred to this percentage measure of a dictionary’sprevalence as its “index.” In each forum, we calculated an index for 75dictionaries (1 absolute, 1 nonabsolute, and 73 LIWC).
For the absolutist index we have endeavored to account for false positives. Thereare three principal types of false positives: a negation before the absolutistword (i.e., “not completely”), a qualifier before the absolutist word (i.e.,“almost completely”), and a salutation (i.e., “hello everyone”). These wouldordinarily register on our absolutist index and distort our measure ofabsolutism. Fortunately, the LIWC (2015 version) can also count phrases, so weran a second version of our absolutist dictionary composed of the most commonfalse positives (as described). The absolutist false positive index wassubtracted from the absolutist index to provide a better estimate of absolutism.We nevertheless rely on the assumption that any remaining false positives areequally distributed between groups.
The control and test category forums were subdivided into groups as shown inTable 1. Toanalyze the data, a multilevel mixed-effects modeling approach was adopted(the SPSS syntax script can be found in the Supplemental Material). This isthe recommended analysis method for this type of data structure (Baayen, Davidson, &Bates, 2008). Members were nested within forums, and forums werenested within groups (i.e., depression). Mixed-effects models consider bothfixed and random effects and can be used to assess the influence of thefixed effects on the dependent variables after accounting for some outsiderandom effects. Residuals were weighted by the word count of each text fileand all the analysis was conducted using IBM SPSS software (version 21). Tocorrect for positive skew in the data, we used a log10(x + 1)transformation, adding 1 to deal with 0 values (cf. Yamamura, 1999). We report rawvalues for descriptive statistics to facilitate a more intuitiveunderstanding. The bootstrap procedure was also used to produce betterestimates of p values and confidence intervals (CIs). Thismethod is often recommended because it does not assume normally distributeddata (Cumming,2014). Bootstrapped CIs (95%; bias-corrected and accelerated)were computed through 1,000 random resamples (with replacement) using thestratified sampling method, with forums as the strata variable.
There was no significant omnibus effect among the control groups asdetermined by a multilevel mixed effects model, F(7, 11) =0.754, p = .635 (Table 1 and Table S1). Consequently, they were combined into a single“control group.” It is important that this suggests that the absolutistindex is largely independent of content, as it demonstrates remarkablylittle variance across a wide range of very different discussion topics.
Multilevel mixed-effects model for the absolutist index
There was a large, significant difference in the absolutist index between theStudy 1 groups, as determined by a multilevel mixed-effects model,F(3, 29) = 71.549, p < .001. Usingpaired comparisons in the mixed-effects model, we compared the control groupwith each of the Study 1 test groups to assess our first hypothesis. We alsocompared the suicidal ideation forum group with the remaining two testgroups (anxiety and depression forums) to assess our second hypothesis. Themean absolutist index for the control forum group (M =0.97%, SD = 0.85) was significantly lower than anxiety(M = 1.45%, SD = 0.97,p < .001, d = 3.24, 95% CI = [0.07,0.11]), depression (M = 1.45%, SD = 1.0,p < .001, d = 3.14, 95% CI = [0.08,0.11]), and suicidal ideation (M = 1.80%,SD = 1.04, p < .001,d = 4.56, 95% CI = [0.14, 0.18]) test forum groups.Moreover, the suicidal ideation group was significantly greater than boththe anxiety (p < .001, d = 1.74, 95% CI= [−0.09, −0.05]) and depression (p < .001,d = 1.71, 95% CI = [−0.09, −0.05]) groups (Fig. 1a). Theseresults are consistent with both of our Study 1 hypotheses. Post hoccomparisons with a Bonferroni correction revealed that there was nosignificant difference between anxiety and depression forum group means(p = 1.00).
Multilevel mixed-effects model for the comparison dictionaries
Using the LIWC software, we produced indices for our nonabsolutist dictionaryand all 73 LIWC dictionaries. We were interested in determining whichcomparison dictionary index would produce comparable significance levels andeffect sizes to that of our absolutist index. We again conducted amultilevel mixed-effects model and pairwise comparisons for each of the 74comparison dictionary indices. Table 2 displays the teststatistics and effect sizes for the 16 dictionaries with the largesteffects. Notably, our absolutist index has satisfied the study hypothesesbetter than any of the linguistic dimensions previously linked to affectivedisorder (negative emotions, personal pronouns etc.). Although “negativeemotion” words were predictably more prevalent in test group forums thancontrol forums, they paradoxically were less prevalent in suicidal ideationforums than anxiety or depression forums. This was also the case for othercontent dictionaries like “sad,” affect,” and “feel.”
|Control <anxiety||Control <depression||Control < suicidalideation||Anxiety < suicidalideation||Depression < suicidalideation|
Note: Displayed are 16 dictionaries with the largest effects. Foreach dictionary, three t tests compared thetransformed data for the control group index (dictionary %prevalence) to each of the test groups (anxiety, depression, andsuicidal ideation forums) to address Hypothesis 1(H1). Two ttests also compared the suicidal ideation forum group with theremaining two test groups (anxiety and depression) to addressHypothesis 2 (H2). LIWC dictionariesare ordered according to average Cohen’s deffect size. Neg. emo = negative emotions; I = first-personsingular pronouns (e.g., I, me, my).
*p < .05. **p < .001.
Analysis of covariance
We ran an analysis of covariance (ANCOVA) to measure the unique predictivevalidity of absolutist words after partialling out the effects of thenegative emotion words, pronouns, and certainty words. Negative emotions andpronouns have previously been identified as strong linguistic markers ofaffective disorder, and the certainty words index is the most conceptuallyrelated to our absolutist index. We found that there was still a significantmain effect for the absolutist index between groups, after controlling forthe certainty index, negative emotions index, and the pronoun’s index,F(3, 3860) = 20.575, p < .001.Paired comparisons reveal that all contrasts remained significant top < .01.
Confirmatory factor analysis
For Study 1 forums, we calculated indices for each individualword in the absolutist and nonabsolutist dictionariesusing an in-house python script (full python code is available in theSupplemental Material) and the Natural Language Tool Kit (Bird, Klein, & Loper,2009). This means that we had the percentage prevalence of eachword rather than each dictionary. Using these data, weconducted a confirmatory factor analysis on the combined list of 65absolutist and nonabsolutist words with a direct oblimin rotation and aloadings cutoff > 0.55. We found that the highest loading words on thefirst factor were all absolutist except for really (whichis an adverbial intensifier) and anything, which we hadoriginally categorized as absolutist but, because of a lack of independentexpert validation, was moved to the nonabsolutist dictionary. The highestloading words on Factor 2 were all nonabsolutist except for the absolutistword definitely. Other than definitely, noabsolutist word loaded outside of Factor 1. The factor analysis was not ableto separate “extreme words” from nonabsolutist words (see Table S3 in the Supplemental Material). To examine theabsolutism factor further, we used structural equation modeling to test themodel fit of the seven highest loading words on Factor 1 from the factoranalysis. Model fit was assessed using AMOS version 24 (SPSS). A seven-item,one-factor model adequately fit the data (χ2 = 14.461,df = 14, goodness of fit index = .912, comparative fitindex = .996, normed fit index = .903). Including more words in the modelreduced the model fit below generally accepted levels.
The smallest group in this study is the suicidal ideation group. Inferencesabout this group are based on data from 368 members in four separatesuicidal ideation forums. Moreover, these forums may be perceived as lessconventional than others used in this research. For this reason, weconducted a sensitivity analysis to ensure the results obtained from thisgroup are robust. The multilevel mixed-effects model for the absolutistindex was recalculated after sequentially excluding all data from each ofthe suicidal ideation forums in turn. This produced four sets of teststatistics, each with one suicidal ideation forum excluded. Pairedcomparisons showed that the absolutist index for the suicidal ideation groupremained significantly greater than the control group (ps< .001, ds = 3.85–4.41), the anxiety group(ps < .001, ds = 1.39–1.71), andthe depression group (ps < .001, ds =1.37–1.69). The narrow range of effect sizes for each comparison confirmsthat these findings are robust, and not driven by a forum outlier in thesuicidal ideation group.
Multilevel mixed-effects model for the absolutist index
Our third hypothesis predicted that mental health forum groups stronglyassociated with absolutist thinking (BPD and ED) would use more absolutistwords than mental health forum groups less associated with absolutistthinking (PTSD and schizophrenia). A multilevel mixed-effects analysis foundthat there was a significant difference in the absolutist index betweenStudy 2 groups, F(3, 16) = 5.515, p =.009. Paired comparisons revealed that the mean absolutist index for the BPDforum group (M = 1.47, SD = 1.01) wassignificantly greater than the PTSD (M = 1.13,SD = 0.83, p < .001,d = 0.36, 95% CI = [−0.07, −0.02]) and theschizophrenia forum groups (M = 1.14, SD =0.86, p < .001, d = 0.35, 95% CI =[−0.07, −0.03]). They also revealed that the absolutist index of the EDforum group (M = 1.25, SD = 0.95) wassignificantly greater than the schizophrenia (p = .009,d = 0.12, 95% CI = [−0.04, −0.001]) but not PTSD(p = .081, d = 0.13, 95% CI = [−0.03,0.01]) forum groups (Fig.1b). A critical assumption in this contrast, is that the controland test groups have similar levels of psychological distress. We sought toverify this assumption using the LIWC negative emotions dictionary. A pairedcomparison found no significant difference in the mean negative emotionsindex between the Study 2 control (M = 3.51,SD = 2.02) and test (M = 3.71,SD = 1.76, p = .335) forum groups(Fig. 1c).Therefore, it seems that absolutism is associated with certain types ofpsychopathology forums and not psychological distress forums perse.
Comparison of Study 1 with Study 2
In comparing the absolutist index of Study 1 and 2 groups, post hoccomparisons with a Bonferroni correction revealed that the suicidal ideationforum group had an index significantly greater than ED and BPD forum groups(p < .001). ED but not BPD had an indexsignificantly lower than anxiety and depression forum groups(ps = .001). Study 2 control forum groups PTSD andschizophrenia had an index significantly lower than all Study 1 test forumgroups (ps < .001).
The smallest group in this study is the BPD group. Inferences about thisgroup are based on data from 326 members in four separate BPD forums. Thisgroup also produced the most extreme absolutist index scores. Once again, weconducted a sensitivity analysis to ensure the results obtained from thisgroup are robust. The multilevel mixed-effects model for the absolutistindex was recalculated after sequentially excluding all data from each ofthe BPD forums in turn. This produced four sets of test statistics, eachwith one BPD forum excluded. Paired comparisons show that the absolutistindex for the BPD group remained significantly greater than the PTSD group(ps < .026, ds = 1.25–1.91) and theschizophrenia group (ps < .008, ds =1.56–2.24). Once again, the positive findings from the smallest group in thestudy appear to be robust and not dependent on any single forum outlier.
Multilevel mixed-effects model for the absolutist index
Our final hypothesis predicted that the recovery forum group would usesignificantly more absolutist words than the Study 1 control forum group.Paired comparisons in a multilevel mixed-effects model found that the meanabsolutist index of the recovery forum group (M = 1.31%,SD = 0.95) was significantly greater than the Study 1control forum group (p < .001, 95% CI = [−0.09, −0.05],d = 0.37). Paired comparisons also found a significantdifference in the absolutist index between the recovery forum group and theanxiety group (p = .018, 95% CI = [−0.001, 0.04],d = 0.15) and depression group (p =.018, 95% CI = [−0.001, 0.04], d = 0.15). Like the anxietyand depression groups, the recovery group also had a significantly lowerabsolutist index than the suicidal ideation group (p <.001, 95% CI = [−0.06, −0.12], d = 0.50). Although theabsolutist index of the recovery group was significantly different fromanxiety and depression groups, the more accurate bias-corrected CIs revealthat the differences are marginal; relative effect sizes reveal that therecovery group absolutist index is closer to anxiety and depression(d = 0.15) than to the control group(d = 0.37; Fig. 2a). We noted earlier that thecontents of the recovery forums were very positive. To illustrate this fact,we ran the LIWC positive emotions dictionary on the above groups (Fig. 2b). There wasindeed a very large difference in the prevalence of positive emotions.Paired comparisons found that the recovery forum group contained morepositive emotion words than all the remaining groups (ps< .001).
Although the recovery group is relatively large, with 558 members in 7different forums, this group is somewhat unconventional and the number ofmembers in each forum were somewhat unequal (see Table S1). We therefore deemed it appropriate to conductanother sensitivity analysis to ensure the results obtained from this groupare robust. The multilevel mixed-effects model for the absolutist index wasrecalculated after sequentially excluding all data from each of the recoveryforums in turn. This produced seven sets of test statistics, each with onerecovery forum excluded. Paired comparisons show that the absolutist indexfor the recovery group remained significantly greater than the control group(ps < .001, ds = 1.88–2.30). Thisagain confirms that the positive findings from this group are robust and notdependent on any single forum outlier.
The data we have presented confirm that the use of absolutist words is elevatedin the natural language of various affective disorder forum groups. As expected,in Study 1 we found that the percentage of absolutist words in anxiety,depression, and suicidal ideation test groups was significantly greater than incontrol groups (H1), and that the percentage ofabsolutist words in the suicidal ideation forum group was significantly greaterthan in both the anxiety and depression forum groups(H2). These findings have support from aprevious study, Fekete(2002) used an adapted Weintraub text analysis method on fourInternet forums (suicide, depression, anxiety, and a journalism control). Theyfound significant results for 13 language variables including negations anddichotomous expressions. Our first study has built on this preliminary finding,using a wider range of more rigorous controls, a larger corpus of data, and ahypothesis-driven study design.
In Study 2, consistent with our expectations, we found the absolutist index wasgreater for BPD and ED forums than PTSD and schizophrenia forums, although thisdid not reach significance between ED and PTSD. All four mental health groupscontained similar amounts of negative emotion terms, but only BPD and ED arestrongly associated with absolutist thinking. This suggests that our index ismore sensitive to absolutism than psychological distress.
In Study 3, we proposed that if the absolutist index for the recovery forums wassimilar to depression forums, this would suggest that absolutist thinking hassome trait-like qualities that persist outside of depressive episodes. This isindeed what we observed. Even though the recovery forums were largely verypositive, the percentage of absolutist words in the recovery group hadoverlapping CIs with both the anxiety and depression forum groups, and wassignificantly greater than the control forum group. It is widely acknowledgedthat an episode of depression increases the risk of future depressive episodes(Teasdale et al.,2000). In many ways, preventing this recurrence is the focus of mosttreatments. Consequently, there is keen interest in identifying potentialcognitive vulnerability factors which are observed during episodes of depressionand persist even after the episode has ended. Our findings indicate thatabsolutism may be such a vulnerability factor. The “scar hypothesis” (Lewinsohn, Steinmetz, Larson,& Franklin, 1981) provides a different explanation. Here thedepressive episode itself alters the linguistic style/vocabulary of theindividual, this then persists as a “scar” after the depressive episode hasabated.
Comparison with other dictionaries
Text analysis research on written data from depressed and suicidal individualshas repeatedly shown elevated use of negative emotion words and pronouns (forreview, see Tausczik &Pennebaker, 2010). We also found these to be strong markers ofaffective disorder in the present study. However, we have paradoxically foundthat “negative emotions,” “sad,” “affect,” and “feel” dictionaries were moreprevalent in anxiety and depression than the suicidal ideation group. This isinconsistent with the belief that suicidal individuals have a greater amount ofnegative emotions (de Klerket al., 2011; Orbach, Mikulincer, Gilboa-Schechtman, & Sirota, 2003; Stein, Apter, Ratzoni,Har-Even, & Avidan, 1998), and some research has previously shownthat “negative emotion [words] use tends to increase approaching suicide” (Pennebaker & Chung,2013). These mixed findings only reaffirm that “function” words are abetter gauge of thinking processes than “content” words (Chung & Pennebaker, 2007). Ourabsolutist dictionary also produced larger effects than pronouns (and itsfirst-person singular subcategory), which had previously been identified asbetter markers of affective disorder than negative emotion words (Pennebaker & Chung,2013).
The LIWC “certainty” index (Table 2) is the most closely related to our absolutist index,comprising words that denote high or total certainty. Although indeed similar,the certainty index does not include some words that are absolutist (i.e.,“nothing”) and contains others that are not (i.e., “frankly”). Moreover, unlikeour absolutist dictionary, many of its component words are not neutrallyvalenced (i.e., perfect).
Finally, we found that “swear” words produced a similar significance pattern toabsolutist words (Table2). Swear words are commonly used as adverbial intensifiers (Peters, 1994; Romero, 2013). Forexample, instead of writing “I’m completely sick of this,”depressed/suicidal individuals may write something akin to “I’mfucking sick of this,” replacing the absolutist word“completely” with something even more forceful, both functionally serving asadverbial intensifiers of the strongest kind.
Absolute versus extreme
Previous studies have often used the terms absolute and extreme interchangeably(e.g., Teasdale et al.,2001). A central assumption in the present research is thatabsolutist words are uncorrelated with extreme words; this assumption wastested. We found that only 25% of absolutist words were also deemed extreme bysome of the independent expert judges. Moreover, none of the words we hadcategorized as extreme were deemed absolutist, with the single exception ofreally, which was categorized as absolutist by one out ofthe five judges. This was reaffirmed by the confirmatory factor analysis(Table S3), in which only words we had categorized as absolutistloaded onto Factor 1, with the single exception, once again, of the adverbialintensifier really. We believe that a clear distinction shouldbe made between these two concepts in future research; and that the terms shouldnot be used interchangeably.
Anxiety and depression within control groups
Individuals with cancer, PTSD, and schizophrenia have high levels of comorbidanxiety and depression. This might lead us to expect a higher absolutist indexfor these forum groups. However, the cancer group produced an absolutist indexidentical to the other Study 1 control groups; and the PTSD and schizophreniagroups had a significantly lower absolutist index than all Study 1 test groups.This may be because symptoms of anxiety and depression in cancer, PTSD, andschizophrenia have a known specific cause, namely, having cancer, PTSD, orschizophrenia. One does not have to be absolutist, or even disposed to affectivedisorder, to experience feelings of anxiety or depression about a brain tumor, atraumatic event, or hallucinations. In contrast, anxiety and depressiondisorders often have multiple vague or even unknown causes. Predisposedindividuals are pushed into anxiety and depression by circumstances that bynecessity would not have the same effect in the general population.
The maladaptive status of absolutist thinking is a recognized part of cognitivetherapy (CT; C. Williams& Garland, 2002). To date, theoretical and anecdotal support hasmostly served as the basis for its inclusion; we hope the findings from ourstudies will add empirical justification. The extent to which absolutistthinking is currently addressed by CT depends on the form of CT used and thepreferred methods of each practitioner. For example, combatting absolutistthinking is at the very core of rational-emotive behavioral therapy (David, Lynn, & Ellis,2009), whereas reducing negative thoughts takes primacy in otherforms of CT. Recently, research into treating cognitive vulnerabilities andpreventing relapse has migrated toward the new “third-wave” therapies (Teasdale et al., 2000).These therapies, such as mindfulness-based cognitive therapy and acceptance andcommitment therapy, are largely geared toward increasing cognitive flexibility(e.g., Kahl, Winter, &Schweiger, 2012). Our findings are therefore in step with the recenttrend toward cultivating adaptive cognitive processes (i.e.,flexibility) as distinct from changing the content of thoughts(i.e., negativity).
Limitations and future directions
Because this study had large samples from multiple sources, and a naturalisticobservational design, it consequently had low experimental control. For example,we could only infer general demographic characteristics from different forums(e.g., women post on Mumsnet and young people post on Student Room). Usernamesserved to distinguish members, however it is possible that some members mightpost using more than one profile or use different usernames for differentforums. Fundamentally, the identities and motivation of users are largelyunknowable, and this is an inevitable limitation in this study. As outlined inthe methods, we did check that the authors of posts were at least purporting tobe a representative of the relevant online community, but we had no power to gobeyond this basic check. Follow-up studies could use an experimental studydesign, and perhaps alternative methodologies, to replicate and extend thefindings initially presented here. Despite likely being limited to a smallersample size and perhaps lacking ecological validity, such studies would be ableto control participant characteristics, writing topics and the setting.
Our findings in this study relate to differences between groups, such an analysisprovides important insights into the markers associated with affective disorder.However, in this research, we have not addressed within-person variation inabsolutist thinking and how that relates to changes in affective symptoms at anindividual level (cf. Molenaar & Campbell, 2009). For example, are individual changesin suicidal ideation over time reflected in changes in use of absolutist words?Future research could seek to track absolutist thinking (and affective disorder)in individuals over time. This could have even greater utility for clinicalpractice.
In measuring aggregate differences in absolutist words between groups we have notexamined the specific nature of the relationship. Although we present data thatmay point to absolutism as a possible cognitive vulnerability factor, the extentand mechanism of any causal role are not addressed here. Future interventionstudies could examine the causal status of absolutist thinking; one possibilitywould be to use a cognitive bias modification paradigm (Hallion & Ruscio, 2011). The aimwould be to introduce some manipulation of absolutist thinking in participantsand then examine the subsequent effects. Alternatively, a narrow form ofcognitive behavioral therapy that focuses on targeting absolutist thinking couldbe clinically trialed.
Supplemental material,Al-MosaiwiMultilevel_linear_mixed_effects_model_syntax_Supplemental_Material forIn an Absolute State: Elevated Use of Absolutist Words Is a Marker Specific toAnxiety, Depression, and Suicidal Ideation by Mohammed Al-Mosaiwi and TomJohnstone in Clinical Psychological Science
Supplemental material, Al-Mosaiwi_Open_Practices_Disclosure for In an AbsoluteState: Elevated Use of Absolutist Words Is a Marker Specific to Anxiety,Depression, and Suicidal Ideation by Mohammed Al-Mosaiwi and Tom Johnstone inClinical Psychological Science
Supplemental material, Code_Supplemental_Material for In an Absolute State:Elevated Use of Absolutist Words Is a Marker Specific to Anxiety, Depression,and Suicidal Ideation by Mohammed Al-Mosaiwi and Tom Johnstone in ClinicalPsychological Science
Supplemental material, Table_S1_Supplemental_Material for In an Absolute State:Elevated Use of Absolutist Words Is a Marker Specific to Anxiety, Depression,and Suicidal Ideation by Mohammed Al-Mosaiwi and Tom Johnstone in ClinicalPsychological Science
Supplemental material, Table_S2_Supplemental_Material for In an Absolute State:Elevated Use of Absolutist Words Is a Marker Specific to Anxiety, Depression,and Suicidal Ideation by Mohammed Al-Mosaiwi and Tom Johnstone in ClinicalPsychological Science
Supplemental material, Table_S3_Supplemental_Material for In an Absolute State:Elevated Use of Absolutist Words Is a Marker Specific to Anxiety, Depression,and Suicidal Ideation by Mohammed Al-Mosaiwi and Tom Johnstone in ClinicalPsychological Science
Author Contributions: M. Al-Mosaiwi created the research design from an initial idea contributed by T.Johnstone. M. Al-Mosaiwi collected, analyzed, and interpreted the data under thesupervision of T. Johnstone. M. Al-Mosaiwi drafted the manuscript, and T.Johnstone provided critical revisions. Both authors approved the final versionof the manuscript for submission.
Declaration of Conflicting Interests: The author(s) declared that there were no conflicts of interest with respect tothe authorship or the publication of this article.
Funding: This research was supported by the United Kingdom Medical Research Council.
Supplemental Material: Additional supporting information may be found at http://journals.sagepub.com/doi/suppl/10.1177/2167702617747074
All data and materials have been made publicly available via Figshare and can beaccessed at https://doi.org/10.6084/m9.figshare.4743547.v1. The completeOpen Practices Disclosure for this article can be found at http://journals.sagepub.com/doi/suppl/10.1177/2167702617747074.This article has received badges for Open Data and Open Materials. Moreinformation about the Open Practices badges can be found at https://www.psychologicalscience.org/publications/badges.
Special thanks to Ian Cunnings, Allan Laville, Sara Fincham-Majumdar, Helen Marlow,and Hannah Whitney who helped validate the study dictionaries. We also thank ShirleyReynolds as well as the reviewers for their insightful comments.
- 1. (2012). Dealing with problematiceating behaviour: The effects of a mindfulness-based intervention on eatingbehaviour, food cravings, dichotomous thinking and body imageconcern. Appetite, 58,847–851.
- 2. (2017). The mediating role ofdichotomous thinking and emotional eating in the relationship betweendepression and BMI. Eating Behaviors,26, 55–60.
- 3. (1983). Cognition and suicide: Amethodological review. Suicide and Life-ThreateningBehavior, 13,109–122.
- 4. (2012). Social cognition inborderline personality disorder: Evidence for dichotomous thinking but noevidence for less complex attributions. BehaviourResearch and Therapy, 50,707–718.
- 5. (2001). Evaluations of others byborderline patients. Journal of Nervous and MentalDisease, 189,513–521.
- 6. (2008). Mixed-effects modeling withcrossed random effects for subjects and items.Journal of Memory and Language, 59,390–412.
- 7. (1979). Cognitive therapy and the emotionaldisorders. London, England:Penguin.
- 8. (2009). Natural language processing withPython. Sebastopol, CA:O’Reilly Media.
- 9. British Psychological Association. (2013).Ethics guidelines for internet-mediated research.Leicester, UK: BritishPsychological Society.
- 10. (1981). The language ofdepression. Bulletin of the MenningerClinic, 45,334–358.
- 11. (1989). The feeling good handbook: Usingthe new mood therapy in everyday life. New York,NY: WilliamMorrow.
- 12. (2008). The reliability and validityof the Dichotomous Thinking in Eating Disorders Scale.Eating Behaviors, 9,154–162.
- 13. (2007). The psychological functionsof function words. In (Ed.), Frontiers of social psychology: Socialcommunication (pp. 343–359).New York, NY: PsychologyPress.
- 14. (2012). Construction and preliminaryvalidation of a dictionary for cognitive rigidity: Linguistic markers ofoverconfidence and overgeneralization and their concomitant psychologicaldistress. Journal of PsycholinguisticResearch, 41,347–370.
- 15. (2010). Delusions and beliefflexibility in psychosis. Psychology andPsychotherapy: Theory, Research and Practice, 83,45–57.
- 16. (2014). The new statistics: Why andhow. Psychological Science,25, 7–29.
- 17. (Eds.). (2009). Rational and irrationalbeliefs: Research, theory, and clinical practice.Oxford, England: OxfordUniversity Press.
- 18. (2011). Prevalence and correlatesof lifetime deliberate self-harm and suicidal ideation in naturalisticoutpatients: The Leiden Routine Outcome Monitoring study.Journal of Affective Disorders, 133,257–264.
- 19. (1987). A sadly neglected cognitiveelement in depression. Cognitive Therapy andResearch, 11,121–145.
- 20. (1975). A new guide to rationalliving. Upper Saddle River, NJ:Prentice Hall.
- 21. (2008). Cognition and suicide: Twodecades of progress. International Journal ofCognitive Therapy, 1,47–68.
- 22. (1985). Content analysis: Analternative approach to open and closed minds. HighSchool Journal, 68,229–240.
- 23. (2003). Cognitive behaviour therapyfor eating disorders: A “transdiagnostic” theory andtreatment. Behaviour Research and Therapy,41, 509–528.
- 24. (2010). Self construction, cognitiveconflicts and polarization in bulimia nervosa.International Journal of Clinical and Health Psychology,10, 445–457.
- 25. (2002). The Internet—A new source ofdata on suicide, depression and anxiety: A preliminarystudy. Archives of Suicide Research,6, 351–361.
- 26. (1982). A multidimensionalpsychotherapy for anorexia nervosa. InternationalJournal of Eating Disorders, 1,3–46.
- 27. (2009). Systematic review onInternet support groups (ISGs) and depression (1): Do ISGs reduce depressivesymptoms?Journal of Medical Internet Research, 11,e40.
- 28. (2011). A meta-analysis of theeffect of cognitive bias modification on anxiety anddepression. Psychological Bulletin,137,940–958.
- 29. (2014). Internet support groups fordepression: A 1-year prospective cohort study.American Journal of Psychiatry, 159,2062–2068.
- 30. (1988). Conceptual complexity andobsessionality in bulimic college women. Journal ofCounseling Psychology, 35,251–257.
- 31. (2011). The utility of measuringexplanatory flexibility in PTSD research. CognitiveTherapy and Research, 35,372–380.
- 32. (2012). The third wave of cognitivebehavioural therapies: What is new and what is effective?Current Opinion in Psychiatry, 25,522–528.
- 33. (1981). Depression-relatedcognitions: Antecedent or consequence?Journal of Abnormal Psychology, 90,213–219.
- 34. (1998). Dichotomous thinking as asign of suicide risk on the TAT. Journal ofPersonality Assessment, 71,368–378.
- 35. (1952). Language behavior in manicpatients. AMA Archives of Neurology andPsychiatry, 67,763–770.
- 36. (1998). Feature subset selection intext-learning. In
Proceedings of the EuropeanConference on Machine Learning(pp.95–100). London, England: Springer.
- 37. (2009). The new person-specificparadigm in psychology. Current Directions inPsychological Science, 18,112–117.
- 38. (2011). Psychotic-like cognitivebiases in borderline personality disorder. Journalof Behavior Therapy and Experimental Psychiatry,42, 349–354.
- 39. (2007). Dichotomous thinking inborderline personality disorder. Cognitive Therapyand Research, 31,717–726.
- 40. (1961). Dichotomous evaluations insuicidal individuals. Journal of ConsultingPsychology, 25,445–449.
- 41. (1964). Rigid thinking in suicidalindividuals. Journal of ConsultingPsychology, 28,54–58.
- 42. (2011). Psychological factorsinfluencing weight loss maintenance: An integrative literaturereview. Journal of the American Association of NursePractitioners, 23,592–601.
- 43. (2003). Mental pain and itsrelationship to suicidality and life meaning.Suicide and Life-Threatening Behavior, 33,231–241.
- 44. (2015). How does thinking in blackand white terms relate to eating behavior and weight regain?Journal of Health Psychology, 20,638–648.
- 45. (2015). Linguistic Inquiry and Word Count:LIWC2015. Austin, TX:Pennebaker Conglomerates.
- 46. (2013). Counting little words in bigdata. In (Eds.), Social cognition and communication (pp.25–42). New York,NY: PsychologyPress.
- 47. (1994). Degree adverbs in earlymodern English. Studies in Early ModernEnglish, 13,269–288.
- 48. (1998). Problem solving and suicidalbehavior. Suicide and Life-ThreateningBehavior, 28,375–387.
- 49. (2001). Effective problem solving insuicide attempters depends on specific autobiographicalrecall. Suicide and Life-ThreateningBehavior, 31,386–396.
- 50. (2004). Problem-solving in suicideattempters. Psychological Medicine,34, 163–167.
- 51. (2013). This is so cool! Acomparative corpus study on intensifiers in British and AmericanEnglish. Perspectives, 21,526–542.
- 52. (2004). Language use of depressedand depression-vulnerable college students.Cognition & Emotion, 18,1121–1133.
- 53. (2013). Neither dichotomous norsplit, but schema-related negative interpersonal evaluations characterizeborderline patients. Journal of PersonalityDisorders, 27,36–52.
- 54. (1998). Association between multiplesuicide attempts and negative affects in adolescents.Journal of the American Academy of Child and AdolescentPsychiatry, 37,488–494.
- 55. (2001). Word use in the poetry ofsuicidal and nonsuicidal poets. PsychosomaticMedicine, 63,517–522.
- 56. (2010). The psychological meaning ofwords: LIWC and computerized text analysis methods.Journal of Language and Social Psychology,29, 24–54.
- 57. (2001). How does cognitive therapyprevent relapse in residual depression? Evidence from a controlledtrial. Journal of Consulting and ClinicalPsychology, 69,347–357.
- 58. (2000). Prevention ofrelapse/recurrence in major depression by mindfulness-based cognitivetherapy. Journal of Consulting and ClinicalPsychology, 68,615–623.
- 59. (2000). Multidimensional dichotomousthinking characterizes borderline personality disorder.Cognitive Therapy and Research, 24,23–45.
- 60. (2000). Cognitive distortions in thepoetry of Anne Sexton. Suicide and Life-ThreateningBehavior, 30,140–144.
- 61. (1981). Verbal behavior: Adaptation andpsychopathology. New York, NY:Springer.
- 62. (2002). A cognitive–behaviouraltherapy assessment model for use in everyday clinicalpractice. Advances in PsychiatricTreatment, 8,172–179.
- 63. (2008). The psychology of suicidalbehaviour. In (Eds.), The international handbook of suicide andattempted suicide (pp. 79–93).New York, NY: JohnWiley.
- 64. (1999). Transformation using (x+0.5) to stabilize the variance of populations.Researches on Population Ecology, 41,229–234.
- 65. (1991). The role of cognitions inbulimia nervosa. Cognitive Therapy andResearch, 15,413–426.