Archives of Suicide Research
ISSN: 1381-1118 (Print) 1543-6136 (Online) Journal homepage: https://www.tandfonline.com/loi/usui20
Public Response to Suicide News Reports as Reflected in Computerized Text Analysis of Online Reader Comments
G. Rosen, H. Kreiner & Y. Levi-Belz
To cite this article: G. Rosen, H. Kreiner & Y. Levi-Belz (2019): Public Response to Suicide News Reports as Reflected in Computerized Text Analysis of Online Reader Comments, Archives of Suicide Research, DOI: 10.1080/13811118.2018.1563578 To link to this article: https://doi.org/10.1080/13811118.2018.1563578
Accepted author version posted online: 13 Jan 2019. Published online: 26 Feb 2019.
Submit your article to this journal
Article views: 42
View Crossmark data
Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=usui20 Archives of Suicide Research, 0:1–17, 2019 # 2019 International Academy for Suicide Research ISSN: 1381-1118 print/1543-6136 online DOI: 10.1080/13811118.2018.1563578 Public Response to Suicide News Reports as Reflected in Computerized Text Analysis of Online Reader Comments
G. Rosen , H. Kreiner, and Y. Levi-Belz
Previous research has documented the rise in rates of suicidal behaviors following media reports of celebrity suicide. Whereas most research has focused on documenting and analyzing suicide rates, little is known about more subtle psychological effects of celebrity suicide on the public, such as despair and feelings of abandonment. The Internet has revolu- tionized the responses to news reports, enabling immediate and anonymous responses potentially reflecting these psychological processes. Thus, the current study explored the unique psychological impact of a celebrity suicide on the public by analyzing the big data of readers' comments to suicide news reports, using computational linguistics meth- ods. Readers’ comments (N ¼ 14,506) to suicide news reports were retrieved from 4 leading online news sites. The comments were posted in response to 1 of 1 types of reports: a celebrity suicide (Robin Williams), a non-celebrity suicide, and general reports of suicide as a social phenomenon. LIWC software for computerized linguistic analysis was used to calculate the frequency of the various types of words used. Comparison of the responses to the 3 types of suicide reports revealed higher frequency of first-person pronouns and for emotionally charged wordsoncommentstoacelebritysuicide,comparedwithcommentsto the other types of suicide reports. The findings suggest that celebrity sui- cide news reports evoke the expression of positive emotions, possibly related to the venerated celebrity, alongside negative, internalized emo- tions, and feelings of social isolation. Theoretical, practical, and meth- odological implications are discussed.
Keywords big data linguistic analysis, celebrity suicide, LIWC, online reader comments, Robin Williams
1 Public Response to Suicide News Reports
INTRODUCTION According to the vertical identification theory (Stack, 1990), individuals tend to Suicide is one of the leading causes of death identify with those of higher status, such worldwide, with more than 800,000 people as celebrities. Thus, celebrities' actions dying by suicide annually (WHO [World may modulate cognitive processes and Health Organization], 2014). One of the arouse emotional responses among mem- major proximal risk factors for suicide is bers of society (see also Bandura & imitation (Mann et al., 2005;Cheng, Huston, 1961). However, the public’s Hawton, Lee, & Chen, 2007). The rise in expression of subtle social, emotional, and rates of suicidal behaviors following media cognitive psychological responses to reportsofsuicidehasbeentermedthe reports of celebrity suicide in the media is Werther effect (Phillips, 1974), alluding to unclear. The current study addressed this the high rates of suicidal attempts following issue by analyzing users' responses, as the publication of Goethe’srenowned reflected in online readers’ comments to book, The Sorrows of Young Werther. suicide news reports. Our goal was to Imitation of suicidal behavior has been cal- explore the subtle emotional effects that culated to be 14.3 times more likely to celebrity suicide may have on the public occur following reports of celebrity suicide, and its relationship to the Werther and compared with a non-celebrity suicide Papageno effects. (Stack, 2000), and Niederkrotenthaler et al. The Internet has revolutionized media (2012) found this surge to be especially consumption, opened new channels for characteristic of the month immediately fol- news reporting, and created opportunities lowing the suicide. While most of the for the public to express opinions and celebrity suicide studies have focused on the emotions instantly, through readers’ com- increasing suicide rates subsequent to the ments (Bergstr€om & Wadbring, 2015). suicide (Koburger et al., 2015; The vast volume of online comments can Niederkrotenthaler et al., 2012;Pirkis, be considered the big data of the public's Burgess, Francis, Blood, & Jolley, 2006), response to any transpiring event. These more subtle psychological processes, such as raw data waiting to be harvested offers mental pain, despair, and feelings of aban- novel opportunities for a wide spectrum of donment that may be linked to these research fields. Recent studies suggest that reports have been only rarely investigated. the immediacy and anonymity of online Recently, Niederkrotenthaler et al. reader comments facilitate attaining (2010) asserted that media stories about authentic insight into people's opinions suicide may also have a protective impact, and emotions (Back et al., 2010; Henrich conceptualized as the Papageno effect, a ref- &Holmes,2013). Hence, several studies erence to a character from Mozart’s opera, have investigated the public’s reaction to The Magic Flute (Sisask & V€arnik, 2012). suicide by analyzing readers’ comments to This effect was found in media stories online suicide news reports. Most studies with a focus on suicidal ideation unaccom- used human raters to evaluate readers’ panied by a suicide attempt or completed responses (Sisask, Mark, & V€arnik, 2012; suicide. Thus, it would be useful to also Sisask, Varnick, & Wasserman, 2005). examine the possible protective impact of While this approach may help detect spon- media reports on celebrity suicide on the taneous public cognitions and emotions general population. toward suicide, the use of subjective
2 VOLUME 0 NUMBER 0 2019 G. Rosen et al.
ratings to analyze and classify readers’ Psychological analysis of the text is based comments to specific content and emo- on interpretation of the relative counts of tional themes may constrain the objectivity various word categories. For example, a of the results. Novel algorithms based on high percentage of singular compared with computational linguistics offer a more plural first-person pronouns may be indica- objective tool for text analysis. These com- tive of social isolation and lack of social puterized tools also allow big-data analysis support (Cohn, Mehl, & Pennebaker, of online users’ behavior that may be more 2004). Based on analyses of different texts, informative, economical, and reliable. including personal diaries, college assign- ments, chat rooms, and online commun- Linguistic Analysis ities (Pennebaker & King, 1999; Nguyen et al., 2017;Stone&Pennebaker,2002), The notion of linguistic analysis as a LIWC was evaluated and further developed window on psychological processes can be to expand the number of categories and traced back to Freud's notion of the slip- attain a better understanding of the rela- of-the-tongue phenomenon (Freud, 1917). tionship between word categories and spe- This notion was later developed into for- cific psychological dimensions (for a mal projective tests (e.g., Rorschach Test), review, see Tausczik & Pennebaker, 2010). derived from the assumption that con- Most relevant to the current research scious and subconscious psychological are studies utilizing LIWC to analyze texts processes, including thoughts, intentions, written by individuals who died by suicide emotions, and motivations are projected (Barnes, Lawal-Solarin, & Lester, 2007; onto neutral, ambiguous stimuli (Mihura Egnoto & Griffin, 2016; Gunn & Lester, & Meyer, 2015; Murray, 1943) and can 2012; Handelman & Lester, 2007;Li, be interpreted by analyzing the verbal con- Chau, Yip, & Wong, 2014; Lightman, tent. Aiming to develop a more objective McCarthy, Dufty, & McNamara, 2007; way to analyze verbal content, Weintraub Stirman & Pennebaker, 2001). These (1989) used word counts and demon- studies revealed typical linguistic patterns strated the association between counts of of suicidal individuals. For example, different pronouns and depression. Stirman and Pennebaker (2001) showed Building on developments in compu- that texts of poets who died by suicide are tational linguistics, Tausczik and characterized by higher rates of first-person Pennebaker (2010) developed the singular words, such as “I” and “me,” Linguistic Inquiry and Word Count compared with texts of non-suicidal poets (LIWC), a computerized software for psy- over their careers. Interestingly, online dis- chological-linguistic analysis of large cussions of celebrity suicide cases show amounts of text. The LIWC software is some similarity to these patterns. In a based on a dictionary, in which every word study examining such discussions, Kumar, was classified by human judges into one or Dredze, Coppersmith, and De Choudhury more categories, with each category relating (2015) analyzed posts published by users to a different psychological dimension in an online suicide support group, com- (Schwartz et al., 2013). LIWC output dis- paring the responses culled during the two plays the percentage of words from each weeks prior to and the two weeks follow- category out of the total word count of the ing the suicidal act. The authors analyzed text (Tausczik & Pennebaker, 2010). approximately 66,000 posts relating to 10
ARCHIVES OF SUICIDE RESEARCH 3 Public Response to Suicide News Reports
different celebrity suicide cases and meas- person pronouns,”“emotions,”“death ured the change in posting activity and words,” and “word count.” These catego- their content. Their findings showed that ries reflect psychological dimensions that posts during the two weeks following the have been shown to be associated with sui- celebrity suicide reflected a significant cidal behavior and with reactions of indi- increase in the use of first-person singular viduals who were exposed to such pronouns, and in the total word count behavior. Specifically, the use of pronouns per post. may reflect how someone relates to others and interacts with them (Tausczik & The Current Study Pennebaker, 2010). A study investigating married couples showed that the use of The current study aimed to expand on various types of pronouns may be indica- the work of Kumar et al. (2015)intwo tive of the quality of marital interaction important aspects. First, it examined com- (Simmons, Gordon, & Chambless, 2005). ments of the general public posted in For example, higher frequency of the use popular news websites, rather than those of second-person pronouns is associated posted in a closed suicide support group. with more negativity in couple interac- Thus, this study comprises an ecological tions, and the use of first-person pronouns big-data sample of reader comments on is related to better problem-solving abilities suicide reports in online news websites. as a couple (Simmons et al., 2005). Second, to further understand the special Positive and negative emotion words characteristics of celebrity suicide, we com- categories were found to be reliable in pared readers’ comments to a celebrity sui- detecting human emotions (Holmes et al., cide report (CSR) with comments on two 2007; Kahn, Tobin, Massey, & Anderson, types of non-celebrity suicide reports 2007). For example, a study using LIWC (NSR), and general reports of suicide as a software revealed that depressed college social phenomenon (social suicide students used more negative words than reports; SSR). did non-depressed or formerly depressed For CSR, we focused on the case of students (Rude, Gortner, & Robin Williams, whose suicide was one of Pennebaker, 2004). the celebrity suicides examined in the The death words category comprises Kumar et al. (2015) study. Williams, a words relating to death, such as “coffin.” In renowned actor and comedian, who per- their published works, suicidal poets used formed in 103 movies and was awarded more death-related words, compared with numerous prizes, died by suicide at the age non-suicidal poets (Stirman & Pennebaker, of 63 on August 11, 2014. We selected 2001). Narratives containing death words Robin Williams’s case because of his indis- of female assault victims predicted worse putable status as a worldwide celebrity and post-treatment functioning and lower over- the fact that his death was a clear case of all well-being (Alvarez-Conrad, Zoellner, & suicide that was publicly known Foa, 2001). In addition, using more death and reported. words may signify distress (Pennebaker, Although LIWC software has multiple Mayne, & Francis, 1997). categories, for the current study, we ana- Concerning the word count category, lyzed readers' comments as manifested in it seems that when team members use a four major linguistic categories: “first- greater number of words, the chances of
4 VOLUME 0 NUMBER 0 2019 G. Rosen et al.
success on a group task increase (Sexton & commits suicide”), and 18 articles regard- Helmreich, 2000). The use of more words ing SSR (e.g., “Number of suicides in UK indicates increased involvement that can increases, with male rate highest lead to more success on a task. Moreover, since 2001”). total word count has been shown to be a good predictor of self-disclosure levels of Reader Comments the text’s author (Levi-Belz & Kreiner, 2016). An abundance (14,506) of spontan- In sum, the words people use may dis- eous comments were collected from the close their current mental state, feelings, three kinds of suicide news reports: (a) and cognitions. By analyzing online com- 7,501 reader comments to a celebrity sui- ments in the wake of suicide news reports, cide report (CSR); (b) 1,074 reader com- we aim to shed more light on their ments to several non-celebrity suicide authors’ emotions and cognitions concern- reports (NSR); and (c) 5,931 reader com- ing the suicide cases. Thus, in this study, ments to a social suicide report (SSR). we hypothesized that readers’ comments to the CSR will be characterized by higher levels of the linguistic categories of first- Instruments person pronouns, emotions, and death words, as well as by higher levels of total LIWC computer program, 2015 ver- word count, compared with readers’ com- sion (Pennebaker, Boyd, Jordan, & ments regarding non-celebrity suicide Blackburn, 2015) was used to calculate the reports and general reports of suicide as a percentage of each specific word type in social phenomenon. specific categories included in the full text of the comment. Eight linguistic categories were analyzed: first-person singular pro- METHOD nouns, first-person plural pronouns, nega- tive emotion words (e.g., nasty, ugly, Suicide News Reports hurt), positive emotion words (e.g., love, nice, sweet), as well as more specific cate- Forty-nine online suicide news reports gories of anger (e.g., hate, kill), sadness were retrieved from the New York – (e.g., grief, crying), anxiety (e.g., worried, Times U.S. Edition, Yahoo! News, Google fearful), and death words (e.g., coffin, kill). News (news search engine), and The Following several studies employing Guardian, from April 2016 to July 2016. LIWC analysis (e.g., del Pilar Salas-Z arate These news sites were selected because et al., 2014; Holtgraves, 2011; Sexton & they support reader comments and are 1 Helmreich, 2000; Tausczik & Pennebaker, ranked among the top 10 news sites on 2010;), we also included each comment’s the Web. The 49 reports included 14 total word count as a ninth category. articles regarding Robin Williams’ suicide, published within two weeks of his suicide (August 11–24); 17 articles regarding NSR Procedure (e.g., “Harassed by seniors, banker The comments were extracted from the various news sites, either manually or 1Alexa.com—top sites by category: News, 2016 imported from the webpage script. From
ARCHIVES OF SUICIDE RESEARCH 5 Public Response to Suicide News Reports
the raw comments data, we retrieved the relating to the role of celebrity status when comments’ text and date (when possible) analyzing comments. The Statistical and sorted them onto a spreadsheet con- Package for the Social Sciences (SPSS, ver- taining the comment serial number, type sion 20.0 for Windows) was used for all of news report, article number, comment analyses. The level of statistical significance placement within the specific article, date, was set at p =.05. and content. The sorted comments were analyzed using LIWC software. The pro- gram scanned each comment separately RESULTS and classified each word into the relevant linguistic categories, yielding an output of Factor Analysis the percentage of words used in each selected linguistic category per each com- At first, we examine the relationships ment. Subsequently, the full data were between the study variables, using the statistically analyzed. Pearson correlation test. The results are pre- sented in Table 1. The matrix shows that most variables were significantly and weakly Data Analysis related to each other. A relatively high posi- The analysis was carried out on a data tive association was found between the set of 14,506 reader comments, such that death and anger categories. A relatively each comment represented a single case, high negative association was found one independent variable, namely the between the word count and positive emo- news report type (CSR, NSR, SSR), and tions categories, as well as between the first- the nine linguistic categories noted in the person singular and plural word categories. Instruments section. In the first step, a ser- Following that, we conducted an ies of Pearson correlation tests were con- exploratory factor analysis with Varimax ducted, followed by an exploratory factor rotation of all the reader comments, ena- analysis on the linguistic characteristics to bling us to reveal categories loading onto reveal the main linguistic categories emerg- the same factor. Our analysis revealed a ing from the data. Based on these results, four-factor structure explaining 56.7% of we calculated weighted factors to be used the variance of all variables. Table 1 shows in subsequent analyses. the variance explained by each factor and In the second step of the analysis, a the variable loadings to the factors. The generalized linear model (GLM) was per- minimal loading value to each of the fac- formed to determine group differences in tors was 0.25. the original LIWC parameters (the propor- As shown in Table 2, the LIWC cate- tion of each variable), with a Poisson gories were loaded on one factor each, log-linear function appropriate for propor- with one exception: the words charged tions. Moreover, GLM was also conducted with positive emotion category, which for the word-count category and the loaded onto Factors 1 and 2, with a low weighted factors that emerged from the negative loading ( .30) on the former. factor analysis. We also compared two spe- Four specific factors were extracted: Factor cific groups’ differences: CSR vs NSR 1 comprised three LIWC categories: words comments and NSR vs SSR comments, expressing anger, words concerning death, enabling us to identify the differences and words charged with positive emotion
6 VOLUME 0 NUMBER 0 2019 G. Rosen et al.
(negatively). This factor represents the expression of negative feelings and a higher score on Factor 1 indicated a higher fre- quency of negative emotions. .011 1 1 Factor 2 comprised two LIWC catego- ries: words charged with positive emotion and word count per comment, which were both negatively correlated, indicating that .04 .18