- Email: [email protected]
Tweeting celebrity suicides: Users' reaction to prominent suicide deaths on Twitter and subsequent increases in actual suicides
Accepted Manuscript Tweeting celebrity suicides: Users' reaction to prominent suicide deaths on Twitter and subsequent increases in actual suicides Michiko Ueda, Kota Mori, Tetsuya Matsubayashi, Yasuyuki Sawada PII:
S0277-9536(17)30408-2
DOI:
10.1016/j.socscimed.2017.06.032
Reference:
SSM 11296
To appear in:
Social Science & Medicine
Received Date: 29 March 2017 Revised Date:
9 June 2017
Accepted Date: 25 June 2017
Please cite this article as: Ueda, M., Mori, K., Matsubayashi, T., Sawada, Y., Tweeting celebrity suicides: Users' reaction to prominent suicide deaths on Twitter and subsequent increases in actual suicides, Social Science & Medicine (2017), doi: 10.1016/j.socscimed.2017.06.032. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Tweeting Celebrity Suicides: Users' Reaction to Prominent Suicide Deaths on Twitter and Subsequent Increases in Actual Suicides
RI PT
Michiko Ueda1*, Kota Mori2, Tetsuya Matsubayashi3, Yasuyuki Sawada4
M AN U
SC
1. Faculty of Political Science and Economics, Waseda University, Tokyo, Japan 2. McCann Erickson, Tokyo, Japan 3. Osaka School of International Public Policy, Osaka University, Osaka, Japan 4. Faculty of Economics, University of Tokyo, Tokyo, Japan
AC C
EP
TE D
* Corresponding author: Michiko Ueda, Faculty of Political Science and Economics, Waseda University; Building No.3 1-6-1 Nishiwaseda, Shinjuku-ku, Tokyo 169-8050; tel: +81-3-32081176
ACCEPTED MANUSCRIPT Tweeting Celebrity Suicides: Users' Reaction to Prominent Suicide Deaths on Twitter and Subsequent Increases in Actual Suicides
Abstract
RI PT
A substantial amount of evidence indicates that news coverage of suicide deaths by
celebrities is followed by an increase in suicide rates, suggesting a copycat behavior.
However, the underlying process by which celebrity status and media coverage leads to
SC
increases in subsequent suicides is still unclear. This study collected over 1 million
individual messages (“tweets”) posted on Twitter that were related to 26 prominent
M AN U
figures in Japan who died by suicide between 2010 and 2014 and investigated whether media reports on suicide deaths that generated a greater level of reactions by the public are likely to be followed by a larger increase in actual suicides. We also compared the number of Twitter posts and the number of media reports in newspaper and on television to understand whether the number of messages on Twitter in response to the
TE D
deaths corresponds to the amount of coverage in the traditional media. Using daily data from Japan’s national death registry between 2010 and 2014, our analysis found an increase in actual suicides only when suicide deaths generated a large reaction from
EP
Twitter users. In contrast, no discernible increase in suicide counts was observed when
AC C
the analysis included suicide deaths to which Twitter users did not show much interest, even when these deaths were covered considerably by the traditional media. This study also found suicides by relatively young entertainers generated a large number of posts on Twitter. This sharply contrasts with the relatively smaller volume of reaction to them generated by traditional forms of media, which focuses more on the deaths of nonentertainers. The results of this study strongly suggest that it is not sufficient to examine only traditional news media when investigating the impact of media reports on actual suicides.
1
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
Keywords: Suicide; Twitter; social media; media; Japan; celebrity suicide; imitation
2
ACCEPTED MANUSCRIPT
1. Introduction
According to the World Health Organization (WHO, 2014), over 800,000 people die by
RI PT
suicide every year, which translates to one person every 40 seconds. Each of these
suicide deaths entails negative externalities, including significant psychological and
financial burdens on bereaved family members (Chen et al., 2009). Particularly, media
SC
reports on suicides by celebrities, which receive considerable coverage, may induce a series of related suicides once they are publicized. A substantial amount of evidence
M AN U
indicates that news coverage of suicide deaths by celebrities is followed by an increase in suicide rates, suggesting a copycat behavior. Coined the “Werther effect”, the phenomenon is claimed by several studies to be one of the main sources of negative externalities of suicide (e.g., Phillips, 1974; Wasserman, 1984; Stack, 1987; Ueda, Mori, and Matsubayashi, 2014). A recent meta-analysis by Niederkrotenthaler et al. (2012)
TE D
found that, on average, the suicide rate increases by 0.26 after a celebrity suicide, confirming the significance of the Werther effect.
EP
Although the underlying process by which celebrity status and media coverage leads to
AC C
increases in subsequent suicides is still unclear, social learning theory is often used to explain the observed change in suicide rates after media coverage of prominent suicides (Pirkis and Blood, 2010). The theory implies that media reports reinforce the notion that suicide is acceptable, and that other people may regard the deceased as a “model” either by his or her celebrity status (vertical integration) or shared characteristics (horizontal integration). Because such integration with a model hinges on the status and characteristics of the deceased, the theory also implies that media reports on the suicide deaths of certain celebrities have a greater impact than other celebrities, depending on the manner by which people perceive and react to the death of a potential model.
3
ACCEPTED MANUSCRIPT
Researchers have tried to understand such differential increases in actual suicides by the celebrity status of the deceased, or by the number of media reports on suicide deaths (e.g. Pirkis et al. 2006; Niederkrotenthaler et al. 2009). However, not all celebrities are
RI PT
revered in a similar way, and the extent to which people are affected by their deaths should vary. In addition, it is unclear whether the level of media publicity given to
SC
suicide deaths captures the level of people’s reactions to these deaths.
This study explicitly considered the level of reaction by the public to prominent suicide
M AN U
deaths and examined how it is associated with the subsequent change in actual suicides. It measured the level of public reaction to prominent suicide deaths by using comments posted on Twitter, a microblogging platform. The use of large-scale data from social media including Twitter for academic research has recently increased. Recent suiciderelated studies have used Twitter content to identify those at risk of suicide (Jashinsky
TE D
et al., 2014; O’Dea et al., 2015; Colombo et al., 2016), or gauge negative emotional feeling of the public such as depression (Cavazos-Rehg et al., 2016) and suicide-related sentiments (Woo et al., 2015). In other areas, activities on Twitter have been monitored
EP
to measure the level of national happiness (Dodd et al., 2011), or to predict heart disease
AC C
mortality (Eichstaedt et al. 2015).
Twitter is suitable for studying the public’s reaction to celebrity suicide for several reasons. First, Twitter has 313 million monthly active users worldwide (Twitter Inc., 2016) and its reach to the public is extensive. Second, Twitter users are likely to express their feelings and reactions without being subjected to censorship or restrictions on contents. Previous studies have shown that people demonstrate a higher degree of selfdisclosure and self-expression in computer-mediated settings (Suler 2004), such as social media (Ma et al. 2016), than in face-to-face settings. Thus, posts on Twitter are
4
ACCEPTED MANUSCRIPT assumed to contain the public’s genuine sentiment and reaction on suicide deaths by prominent figures.
Third, the volume of posts is expected to indicate the level of people’s reaction to the
RI PT
death that might have not been captured adequately by news outputs in traditional media, including newspapers and televisions. This is particularly true if the celebrity
was popular among the general public. Further, it has been reported that many young
SC
adults seek information about a celebrity’s suicide, career, and depression following a
celebrity’s suicide by using social media outlets such as Twitter and Facebook (Dillman
M AN U
Carpentier & Parrott 2016) and vulnerable individuals may be adversely affected if they are exposed to sensational content online.
Using the collection of over 1 million individual messages (“tweets”) posted on Twitter that were related to 26 prominent figures in Japan who died by suicide between 2010
TE D
and 2014, this study sought to understand the reaction by the public to media reports of celebrity suicides and its association with actual suicides. More specifically, we examined (1) how the number of tweets on celebrities changed before and after the
EP
media reports on the prominent suicide deaths; (2) whether the number of tweets on the suicide deaths corresponded to the amount of coverage in the traditional media such
AC C
as newspaper and television; and (3) whether a greater level of reactions to the celebrity deaths by Twitter users and by the traditional media led to a larger increase in actual suicides, using a method employed by Ueda et al. (2014) that studied the impact of media reporting on 109 celebrity suicides in Japan using daily suicide counts between 1989 and 2010.
2. Data and Method
5
ACCEPTED MANUSCRIPT To answer the three research questions, this study employed four sets of data: the number of prominent suicide deaths, Twitter posts, media reports, and suicide death records.
RI PT
2.1 Data on prominent suicide deaths
The authors identified a list of prominent Japanese who died by suicide by consulting
SC
newspaper articles and information on Wikipedia, an Internet encyclopedia. A page titled “well-known figures in Japan who died by suicide” on Wikipedia in Japanese
M AN U
(http://ja.wikipedia.org/wiki; accessed in March 2015) was first consulted. The authors decided to start with Wikipedia’s list because it contained all the 109 celebrity suicides that were included in the analysis by Ueda et al. (2014). These authors employed a manual coding of newspaper articles from 1989 to 2010. During our study period from July 2010 to December 2014, 43 individuals were included in the list. This list, and thus
TE D
our analysis, did not include non-Japanese figures who died by suicide.
As Wikipedia’s list contains individuals who are not well-known to the public, we
EP
searched the full names of these 43 individuals in the online database of the largest national newspaper in Japan, Yomiuri Shimbun, to determine whether their deaths were
AC C
reported in the national media. The online database of the Yomiuri Shimbun contains all articles published by the newspaper since 1986. Individuals were considered “(nationally) well-known” if their deaths were reported in the national edition of the Yomiuri Shimbun. These individuals were included in this study’s analysis, and those whose deaths were reported only in the local edition were excluded. In short, this study included “well-known’” figures in Japan whose suicide deaths were reported both in Wikipedia and the Yomiuri Shimbun.
6
ACCEPTED MANUSCRIPT In the end, 26 prominent figures (22 males and 4 females) who died by suicide between July 2010 and December 2014 were identified. The list included 12 entertainers (including two TV anchors), 6 business owners, 3 public officials (a cabinet member, a cabinet minister, and a judge), 3 journalists, a former athlete, and a scientist. The date of
RI PT
their death and the date of the first report on their death by the Yomiuri Shimbun were recorded.
SC
2.2 Data on Twitter posts
M AN U
Twitter started its service in April 2008 in Japan. As of December 2015, the number of monthly active users in Japan was estimated to be 35 million (Twitter Inc., 2016). The younger generation is more likely to use Twitter; According to one survey, 52.8% of individuals aged less than 30 use Twitter, compared with 24.3% of aged 50-59 and 15.5 % of those aged 60 years and older, respectively (Ministry of Internal Affairs and
TE D
Communications, 2015).
For each of the 26 individuals identified above, the authors searched tweets related to
EP
his or her death through an online platform provided commercially by Crimson Hexagon, a social media analytics company. Crimson Hexagon is a Twitter Official
AC C
Partner with access to full public tweets since July 2010 via the Twitter Firehose. Thus, this study’s analysis could utilize its full collection of Twitter posts dating back to July 2010, subject to certain restrictions described below. In contrast, if tweets were accessed through an application programming interface (API) provided by Twitter, comprehensive historical tweets would not be available for retrieval. The search was limited to messages made between July 2010 and December 2014.
7
ACCEPTED MANUSCRIPT The full name of the deceased was used as a search term. In certain cases, the title or a word that describes the position of the deceased was used as a search term. For example, for politicians, the search term “Representative+ last name” was used, in
RI PT
addition to their full name.
For each individual, the search date was set to begin seven days prior to the first
newspaper report of his or her suicide death, and end 15 days after the first report,
SC
inclusive of the day that their deaths were first reported in the Yomiuri Shimbun. Thus,
the search was conducted for a total duration of 22 days. The search was limited to posts
March 2015 and May 2015.
M AN U
in Japanese. The search of tweets through the Crimson Hexagon was conducted between
Data collection began with the recording of the total number of posts made during the 22-day study period for each well-known figure. Individual posts using the bulk export
TE D
function of the platform were then extracted. However, not all posts were available for bulk exports for the following two reasons. First, if an account had been deleted by the time of this study’s data retrieval, the messages made by the account were not available
EP
for download. Second, the platform does not allow users to download more than 10,000 posts at once, and the smallest time increment that can be specified is one day. Thus, if
AC C
there is a high volume of posts in one day, only a subset of the randomly-selected posts can be downloaded. When this happened, search terms were modified to narrow the results. For example, we narrowed the results by dividing the posts into two by the presence of the retweet sign (RT), or the reference to the method of suicide (e.g. “hanging”) in the main text. For certain high-volume cases, all the posts could not be retrieved even with this method. On average, 82.3% of the total posts were available for data extraction and analysis.
8
ACCEPTED MANUSCRIPT Information contained in the individual posts includes time stamp with date and time in GMT, contents, author, user name of the author, and the number of followers. The date and time were converted to the Japan Standard Time (JST) to make them comparable to
RI PT
the timing of media reports.
2.3 Data on media reports
SC
Newspaper articles and television programs that featured the 26 individuals after their suicide deaths were collected. For newspaper articles, the Yomiuri Shimbun database
M AN U
described above was used. For each individual, articles that contained the name of the deceased were retrieved, limiting the search to articles published within 14 days after the first newspaper report. The total number of Japanese characters in the articles was then counted. To collect information on television programs, this study used information compiled by the M-data, a Japanese media analytics company. The company
TE D
keeps record of the detailed contents of all major network television programs in the three largest metropolitan areas in Japan (Tokyo, Osaka, and Nagoya) using professional human coders. The number of seconds that TV programs featured the death of each of
EP
the 26 individuals was counted for analysis. Because the television programs reported their deaths earlier than the newspapers did in certain cases, the search period was set
AC C
to start one day before and end 14 days after the first newspaper report. We used the number of characters (newspapers) and seconds (television) as opposed to the number of media items because we think the former captures the level of media attention more accurately than the latter, as there are considerable variations in the length of media items.
2.4 Data on suicide deaths
9
ACCEPTED MANUSCRIPT The mortality data used in this study were obtained from death records in the Vital Statistics Report compiled by the Ministry of Health, Labour and Welfare (2010-2014). The Vital Statistics data used in this study were collected for administrative purposes, and anonymized and de-identified by the Ministry prior to analysis. Individual death
RI PT
records between July 2010 and December 2014 were made available for research purposes under the approval of the Ministry.
SC
The data in the Vital Statistics Report were based on death certificates issued by
physicians and subsequently reported to the local government by family members of the
M AN U
deceased. Deaths coded as X60–X84 under the ICD10 standard were classified as deaths by suicide and were thus included in this study’s dataset. For each calendar day, the total number of suicides was calculated. The total number of observations was 1,614.
TE D
2.5 Methods
We employed the following three sets of methods to achieve our three research aims. First, we examined how the number of tweets on the 26 prominent figures changed as a
EP
result of media reports on their suicide deaths by plotting the total number of tweets per day during the 22-day period for each celebrity. Second, we examined whether the
AC C
number of tweets on the suicide deaths corresponded to the amount of coverage in newspapers and television programs by comparing the number of tweets with the number of seconds spent by TV programs and the total number of Japanese characters in newspaper articles when the suicide deaths were reported. Our comparison was based on inspections of data and correlation coefficients.
Third, we examined whether a greater level of attention by Twitter as well as traditional media on these prominent deaths was associated with a larger increase in actual
10
ACCEPTED MANUSCRIPT suicides. We employed the same method as in Ueda et al. (2014), because their model allowed us to estimate the daily changes in the number of suicides following media reporting on prominent suicides compared to the baseline period with no suiciderelated reports. The model that we estimated was the same Poisson regression model
=∑
where
+
,
+
+
+
,
SC
log
RI PT
employed by Ueda et al (2014):
is the Poisson rate of suicides on day d in month m of year y, and
,
is a
M AN U
dummy variable that equals one if a prominent figure’s suicide was reported for the first time on day d in month m of year y. The subscript k shows the time difference from the onset of each suicide. The dummy variables were lagged forward up to 20 days and added separately to the model as
,
to
,
. The dummy variables should
capture the lagged influence of the news report on the number of suicides in the post-
TE D
reporting period. The variable of the first-reporting of celebrity suicide was also lagged backward up to 20 days as coefficients on
,
and
,
,
to
to
,
,
and included in the model. The were expected to be positive if the media
EP
reports were associated with an increase in actual suicides, whereas the coefficients on
AC C
the pre-reporting period were expected to be zero.
This study’s model also included year, month, day, and day of the week fixed effects. The year fixed effects,
, and the month fixed effects,
, were expected to capture the
influences of any year-specific factors, such as the national-level unemployment rate and population characteristics, and any month-specific factors, such as temperature. The day of the month fixed effect,
, and the day of the week fixed effects,
, were included in
the model because the number of suicides could experience a surge or drop on a particular day of the month or the week.
11
ACCEPTED MANUSCRIPT
We divided the 26 individuals into two groups by the volume of tweets and also by the amount of traditional media reporting, and then estimated the Poisson model separately for each of the two groups. The purpose of this exercise was to see if an increase in
RI PT
population suicides are more likely to happen after suicide deaths were followed by a larger number of tweets or greater media attention. We used Stata 13 by StataCorp for
SC
our statistical analysis.
M AN U
3． ．Results
3.1 The number of tweets on 26 prominent suicide deaths
We first summarized the number of tweets related to the 26 prominent suicide deaths before and after the media reports on their suicide deaths . The total number of relevant
TE D
posts during the 22-day period was 985,735. Among them, 159,891 (22,842 per day) were made before the first newspaper reports, and 825,844 (55,056 per day) were posted on or after the day that Yomiuri Shimbun first reported their deaths. As discussed
EP
below, television programs announced their deaths prior to newspaper reports, and thus the number in the pre-reporting period also included posts made after their deaths.
AC C
The gross sum of followers of the users who posted these messages was approximately 1.2 billion. In other words, tweets related to the suicide deaths were delivered potentially to over 1.2 billion Twitter feeds. The total number of retweets (forwarded posts) during the study period was 393,264, among which 184,639 were made on the day of the first newspaper report.
During the 22-day period, the minimum number of tweets per calendar day was 0, whereas the maximum was 120,463. The left column in Table 1 displays a list of
12
ACCEPTED MANUSCRIPT individuals with the total number of tweets during the 22-day period in descending order. Table 1 shows that the number of relevant tweets varied considerably across well-known figures. This means that the public was more familiar with and interested in the deaths of certain individuals than the others. The average number of posts over the
RI PT
22-day data collection period for the 26 individuals was 36,590 and the standard deviation was 54,324. The minimum total number of posts per day was 59, and the
maximum was 195,296. On average, the number of tweets was higher for entertainers
SC
(N = 12, Mean = 51,154, SD = 36,726), than for non-entertainers (N = 14, Mean = 24,793, SD = 49,284). The major exception was the death of a scientist at the top of the list; he
M AN U
was a supervisor of a high-profile project, but at the time of his death, the fabrication of research results by a junior scientist had been suspected and investigated. When this outlier was excluded from the non-entertainer group, the mean number of posts for non-entertainers was reduced to 12,614 (SD = 19,432). Thus, the public’s level
TE D
of reaction to the deaths was far from uniform.
Figure 1 displays trajectories in the number of tweets related to the 26 individuals within three weeks of the first newspaper report on their death. The x-axis shows the
EP
number of days from the first report, and the value of 0 corresponds to the day of the first report in the Yomiuri Shimbun. The y-axis shows the number of tweets on each day.
AC C
The well-known figures in this study were categorized into two groups. The first one included those whose names were mentioned more than 10,000 times during the 22day period; the second group contained individuals with less than 10,000 posts related to them.
The top panel in Figure 1 presents the number of tweets related to the 14 individuals in the first group (those mentioned more than 10,000 times on Twitter). This group included nine entertainers, one politician (a cabinet minister), one athlete, one scientist,
13
ACCEPTED MANUSCRIPT one business owner, and one journalist. The bottom panel shows the number of tweets related to the 12 individuals in the second group (those mentioned less than 10,000 times on Twitter). This group included five business owners, three entertainers, two journalists, one politician, and one judge. The ranges of the y-axes in the top and bottom
RI PT
panels are different.
For most of the deceased individuals, the number of tweets that mentioned them
SC
increased sharply on the day their deaths were reported. For example, prior to their
deaths, the numbers of posts related to the two individuals listed at the top of Table 1
M AN U
over a seven-day pre-reporting period were 52 and 156. These figures increased to 113,123 and 120,463 on the day when their death was reported, corresponding to a 2,174-fold and a 771-fold increase, respectively.
According to Figure 1, the surge started before day 0 for certain individuals. In most
TE D
cases, this is because the dates of the first media report were established by consulting newspaper articles. Although the Yomiuri Shimbun and other major national newspapers are published twice a day (in the morning and the late afternoon), they can
EP
sometimes lag behind other media such as television and online news outlets. Moreover, in the cases of two individuals, the media started reporting their suspected or attempted
AC C
suicides prior to their deaths while they remained missing or in critical condition; this circumstance explains the observed early surges.
Excluding these two cases, the average total daily number of posts related to the individuals was 340 during the pre-reporting period. The total number of posts surged to 106,763 one day before the Yomiuri Shimbun published its first article on their deaths, and to 464,243 on the day that it reported their deaths, a 1,363-fold (i.e., 136,300%) increase in the number of posts in comparison to the pre-reporting period.
14
ACCEPTED MANUSCRIPT
The number of Twitter posts remained elevated after the initial surge. In several cases, Twitter users kept mentioning the names of the deceased even after two weeks had passed since their suicides were reported. The average total number of posts on the 24
RI PT
individuals (after excluding the above two cases) was 13,566 and 5,898 on the 7th and 14th day, respectively, after the first day of the newspaper report.
SC
3.2. Comparison of Twitter activity and media coverage
M AN U
Next, we examined whether Twitter users and traditional media focused on the same set of prominent suicide deaths. The second and third columns of Table 1 list the number of seconds on television programs and the number of characters on the Yomiuri Shimbun (newspaper) that reported their deaths. The table indicates that Twitter posts were concentrated on the deaths of entertainers, whereas newspaper articles focused less on
TE D
these individuals. The age profiles of the entertainers were also distinct; Twitter posts focused on relatively young entertainers, but newspaper articles were centered on older
EP
ones.
Furthermore, Table 1 shows that the deaths of public officials (two politicians and a
AC C
judge) and business owners were featured more prominently in newspaper articles. Out of these individuals, only one politician’s death (a cabinet minister) was discussed widely on Twitter. The television programs’ focus was somewhat similar to that of Twitter users, but they also tended to draw attention to relatively older entertainers.
The correlation coefficient between the number of Twitter posts and the number of characters in newspaper articles was 0.587, and the correlation coefficient between the number of tweets and the total duration of television programs was 0.703 (N = 26).
15
ACCEPTED MANUSCRIPT Thus, the relationship between Twitter activity and traditional media coverage is not strong, confirming the distinct focus of Twitter users and traditional media.
RI PT
3.3. Twitter activity, media coverage, and population suicides
Our final task was to examine whether increased Twitter activity in the post-reporting period was associated with an increase in actual suicides. We first replicated the
SC
analysis by Ueda et al. (2014) to check whether there was a similar overall increase in
M AN U
actual suicides following media reports during the study period.
Using data on the deaths of 26 well-known figures, we found an increase in the number of actual suicides during the post-reporting period. For example, the estimated IRRs on day 4 and 5 were 1.09 (95% CI: 1.037-1.148) and 1.133 (95% CI: 1.076-1.192), respectively. No statistically-significant increase was observed from day 0 to day 3, and
TE D
the increase started from day 4. In addition, our estimation results indicate that the elevation in the number of suicide lasted more than 10 days. The estimated IRRs for day 15 and day 20 were 1.101 (95% CI: 1.054-1.151) and 1.078 (95% CI: 1.058-1.097),
EP
respectively, suggesting that the number of suicides was 8-10 percent higher even after
AC C
several weeks had passed since the first media reports.
We then examined the association between the volume of tweets and the number of suicides. The 26 individuals were split into the two groups described above by the number of posthumous tweets, and each group was estimated separately using the same model. The top panel of Figure 2 presents the estimated incidence rate ratios (IRRs) when 14 individuals with a relatively large number of posthumous posts (more than 10,000) were included in the analysis. The bottom panel of Figure 2 presents the
16
ACCEPTED MANUSCRIPT estimated IRRs when 12 individuals with less than 10,000 posts were included in the analysis.
Figure 2 indicates that a statistically-significant increase in total suicides occurred when
RI PT
this study’s analysis included individuals that generated large posthumous reactions on Twitter, but such an increase was almost non-existent when it included individuals with a relatively small number of tweets after their deaths. The top panel shows a clear
SC
elevated trend in the number of suicides after the media reports, and the number of suicides increased about 19 percent on day 5 (IRR 1.196, 95% CI: 1.063-1.344),
M AN U
compared to the baseline period. In contrast, the bottom panel shows no visible change in the number of suicides after the media reports. In other words, when suicide deaths did not generate much reaction in the social media, there was no associated increase in actual suicides.
TE D
The top panel of Figure 3 reports the estimation results when ten individuals with the highest number of newspaper articles (measured by the number of Japanese characters) were included in the analysis. In contrast to Figure 2, no discernible increase was
EP
observed after the media reports of their deaths. The bottom panel of Figure 3 shows the estimated IRRs when ten individuals who gathered most attention by TV programs
AC C
were included in the analysis. The pattern reported here is similar to the one reported in the top panel of Figure 2, but the increase started on day 0, not day 4 as reported in the top panel of Figure 2.
4． ．Discussion
The present study used messages posted on Twitter to understand the extent to which people react to the news on suicide deaths by prominent individuals. It also examined
17
ACCEPTED MANUSCRIPT the association between the level of reaction by Twitter users and actual suicides. To our knowledge, the present study constitutes the first attempt to systematically examine the reaction by social media users to prominent suicide deaths. Our analysis offered several notable findings. First, the number of tweets on the deceased individuals surged sharply
RI PT
when the news of their suicides was circulated. Second, Twitter users and traditional media focused on different sets of individuals. For example, suicides by relatively young entertainers generated a large number of posts on Twitter, whereas they generated a
SC
relatively smaller volume of reaction on traditional forms of media, such as the Yomiuri Shimbun newspaper. Traditional media tended to focus more on the deaths of public
M AN U
officials, business owners, and older entertainers. Third, prominent suicide deaths were followed by an increase in population suicides only when they generated a large reaction from Twitter users. In contrast, no discernible increase in suicide counts was observed when Twitter users did not show much interest, even when these deaths were
TE D
covered considerably by the traditional media.
The strength of this study includes the pioneering use of Twitter data in examining the public’s reaction to prominent suicide incidences. The Twitter data were used to
EP
measure how users of the social media reacted to the news on suicides by prominent figures in Japan. In contrast to past studies, this study did not make any arbitrary
AC C
assumption regarding the significance of those suicide deaths for the public.
This study has several limitations. First, we did not examine the content of the tweets to understand the types of reactions. Analyzing the contents of the tweets should allow us to distinguish whether users were merely spreading news on prominent suicides, or expressing sentiment in response to their deaths. Moreover, content analysis should help us explore whether prominent suicide deaths trigger particular emotional reactions. It should also elucidate whether the connectivity between suicidal Twitter
18
ACCEPTED MANUSCRIPT users (Colombo et al., 2016) produces a potential contagion effect following the death of a celebrity. Thus, in order to fully understand the mechanisms underlying the Werther effect, it is crucial to employ topic or sentiment analysis of the posts (e.g., Blei 2012; Ravi
RI PT
and Ravi 2015; Sykora et al. 2013).
The second limitation of our study concerns the representativeness of the sample and the issue of selection bias in the use of social media data (Bright et al. 2014). The
SC
reactions and sentiments expressed by Twitter users may not accurately represent the reactions of the general population, especially those of the older generations who are
M AN U
less likely to use social media (Ministry of Internal Affairs and Communications, 2015; Sloan et al., 2015). In addition, if there was a high-volume of tweets in one day, we were not able to download all posts, thus our collection of tweets contains fewer posts than the actual number of posts for high-volume cases (see section 2.5). However, because the actual number of tweets did not enter our estimation model when we investigated
TE D
the association between the level of reaction by Twitter users and actual suicides, the consequence of the missing data should be minimal.
EP
Third, this study did not analyze the deaths of celebrities who died of other causes. Thus, it is not clear if the observed large volume of tweets for certain individuals was due to
AC C
the nature of their death, namely suicide, or deaths by certain types of prominent figures tend to generate a large volume of comments on Twitter, regardless of their cause of death. Distinguishing these two possibilities warrants future research.
There are several important implications of this study. First, this study’s findings suggest that people are interested in the news on celebrity suicide. At the same time, they also indicate that the reaction to prominent suicides is far from uniform. Some of the suicide deaths included in our study did not generate a large reaction from Twitter
19
ACCEPTED MANUSCRIPT users, even though some of the deceased can be considered “celebrities” judging from the amount of media reporting by traditional media. Thus, people may not regard the deceased as a model merely by their celebrity status, as presumed in the “vertical integration” hypothesis in social learning theory. Our study also indicates that the media
RI PT
reports on prominent suicides may be a necessary condition for the Werther effect, but not a sufficient condition. In order for the Werther effect to happen, people should react
SC
strongly to their deaths.
Second, the results of this study strongly suggest that it is not sufficient to examine only
M AN U
traditional news media when investigating the impact of media reports on actual suicides. Such was the focus of previous studies (Pirkis and Blood, 2010). As we have demonstrated, the focus of traditional media can be vastly different from the interests of the general population. In light of these findings, one may speculate that the response from traditional media is not representative of the public’s reaction to suicides by well-
TE D
known figures.
Third, the vast amount of circulated social media posts on celebrity suicides found in
EP
this study may have a detrimental effect on vulnerable individuals. Although the content of the tweets was not analyzed formally, it was evident from casual inspection of
AC C
posts that a number of these posts were sensational and speculative. Thus, those on social media are likely to be exposed to a large amount of information on suicide, some of which can be sensational and unfounded. In addition, comments related to their deaths, especially the sympathetic ones, may further reinforce the notion that suicide is acceptable, resulting in “social learning” of acceptable behavior. It is known that suicidal individuals use social media to express their suicidal thoughts (O’Dea et al., 2015), and that adolescents who engage in self-harm are more exposed to the Internet than other adolescents (Ybarra and Mitchell, 2007; Daine et al., 2013). Thus, with ever-increasing
20
ACCEPTED MANUSCRIPT users of social media, vulnerable individuals may be exposed to suicide-related
AC C
EP
TE D
M AN U
SC
RI PT
information that may threaten their well-being.
21
ACCEPTED MANUSCRIPT
References
Blei, D.M., 2012. Introduction to Probabilistic Topic Modeling. Commun. ACM 55, 77–84.
RI PT
Bright, J., Margetts, H., Hale, S., Yasseri, T., 2014. The Use of Social Media for Research and Analysis : A Feasibility Study. Leeds, Corporate Document Services.
Cavazos-Rehg A., Krauss M.J., Sowles S., Connolly S., Rosas C., Bharadwaj M., Bierut L.J.,
SC
2016. A content analysis of depression-related tweets. Comput. Hum. Behav. 54, 351357.
M AN U
Chen, J. et al., 2009. Those who are left behind: an estimate of the number of family members of suicide victims in Japan. Social Indicators Research, 94(3), pp.535–544. Colombo G.B., Burnapa P., Hodoroga A., Scourfield J., 2016. Analysing the connectivity and communication of suicidal users on Twitter. Comput. Commun. 73, 291-300. Daine K., Hawton K., Singaravelu V., Stewart A., Simkin S., Montgomery P., 2013.
TE D
The power of the web: a systematic review of studies of the influence of the internet on self-harm and suicide in young people. Plos One 8, e77555. Dillman Carpentier, F.R. & Parrott, M.S., 2016. Young adults’ information seeking
EP
following celebrity suicide: Considering involvement with the celebrity and emotional distress in health communication strategies. Health Communication,
AC C
31(11):1334-1344.
Dodds P.S., Harris K.D., Kloumann I.M., Bliss C.A., Danforth C.M., 2011. Temporal patterns of happiness and information in a global social network: hedonometrics and Twitter. Plos One 6, e26752.
Eichstaedt J.C., Schwartz H.A., Kern M.L., Park G., Labarthe D.R., Merchant R.M., Jha S., Agrawal M., Dziurzynski L.A., Sap M., Weeg C., Larson E.E., Ungar L.H., Seligman M.E.P., 2015. Psychological language on Twitter predicts county-level heart disease mortality. Psychol. Sci. 26, 159-169.
22
ACCEPTED MANUSCRIPT Jashinsky J., Burton S.H., Hanson C.L., West J., Giraud-Carrier C., Barnes M.D., Argyle T., 2014. Tracking suicide risk factors through Twitter in the US. Crisis 35, 51-59. Ma, X., Hancock, J., Naaman, M., 2016. Anonymity, intimacy and self-disclosure in social media, Proceedings of the 2016 CHI Conference on Human Factors in Computing
RI PT
Systems. pp. 3857–3869. Ministry of Health, Labour and Welfare, 2010-2014. The Vital Statistics Report.
Ministry of Internal Affairs and Communications, 2015. White paper on information and
SC
communication (in Japanese).
Niederkrotenthaler T., Fu K., Yip P., Fong D., Stack S., Cheng Q., 2012. Changes in suicide
Commun. H. 66, 1037–1042.
M AN U
rates following media reports on celebrity suicide: a meta-analysis. J. Epidemiol.
Niederkrotenthaler, T., Till, B., Kapusta, N.D., Voracek, M., Dervic, K., Sonneck, G., 2009. Copycat effects after media reports on suicide: A population-based ecologic study. Soc. Sci. Med. 69, 1085–1090.
TE D
O'Dea B., Wan S., Batterhamc P.J., Calear A.L., Paris C., Christensen H., 2015. Detecting suicidality on Twitter. Internet Interventions 2, 183-188. Phillips D., 1974. The influence of suggestion on suicide: substantive and theoretical
EP
implications of the Werther effect. Am. Sociol. Rev. 39, 340–354. Pirkis, J.E., & Blood, W., 2010. Suicide and the news and information media: A critical
AC C
review. Barton, ACT: Commonwealth of Australia.
Pirkis, J.E., Burgess, P.M., Francis, C., Blood, R.W., Jolley, D.J., 2006. The relationship between media reporting of suicide and actual suicide in Australia. Soc. Sci. Med. 62,
2874–2886.
Ravi, K., Ravi, V., 2015. A survey on opinion mining and sentiment analysis: Tasks, approaches and applications. Knowledge-Based Syst. 89, 14–46.
23
ACCEPTED MANUSCRIPT Sloan L., Morgan J., Burnap P., Williams M., 2015. Who tweets? Deriving the demographic characteristics of age, occupation and social class from Twitter user meta-data. Plos One 10, e0115545. Stack S., 1987. Celebrities and suicide: a taxonomy and analysis, 1948-1983. Am. Socio.
RI PT
Rev. 52, 401–12. Suler, J., 2004. The online disinhibition effect. CyberPsychology & Behavior, 7(3), 321– 326.
SC
Sykora, M.D., Jackson, T.W., O’Brien, A., Elayan, S., 2013. Emotive ontology: Extracting
fine-grained emotions from terse, informal messages. IADIS Int. J. Comput. Sci. Inf.
M AN U
Syst. 8, 106–118.
Twitter.com, 2016. About Twitter, Inc. https://about.twitter.com/company (accessed 17.02.24).
Ueda M., Mori K., Matsubayashi T., 2014. The effects of media reports of suicides by wellknown figures between 1989 and 2010 in Japan. Int. J. Epidemiol. 43, 623-629.
TE D
Wasserman I., 1984. Imitation and suicide: a reexamination of the Werther effect. Am. Sociol. Rev. 49, 427-436.
Woo H., Cho Y., Shim E., Lee K., Song G., 2015. Public trauma after the Sewol ferry
EP
disaster: the role of social media in understanding the public mood. Int. J. Environ. Res. Public Health 12, 10974-10983.
AC C
World Health Organization. (2014) Preventing suicide: A global imperative. Ybarra, M.L. & Mitchell, K.J., 2007. Prevalence and frequency of Internet harassment instigation: Implications for adolescent health. Journal of Adolescent Health,
41(2),189–195.
24
ACCEPTED MANUSCRIPT Table 1: List of prominent individuals who died by suicide in Japan with the volume of Twitter posts and the amount of media reporting by the types of media: 2010-2014.
Newspaper
Television
Scientist/52/M
195296
18130
122728
Entertainer/24/F
167444
180
23837
Entertainer/62/F
135001
1247
189416
Entertainer/35/M
132439
309
27624
Politician/73/M
74362
7916
43013
Entertainer/45/M
57957
249
551
Entertainer/78/M
39533
Athlete/42/M
34654
Artist/34/M
26221
Journalist/52/M
19819
Entertainer/79/M
15779
Entertainer/57/M
SC
M AN U 53988
2155
19372
269
136
252
1823
290
52
15203
692
5049
10791
245
728
10168
195
3576
EP
Entertainer/44/M
1271
TE D
Business owner/64/M
Entertainer/42/F
9076
249
39738
Business owner /61/M
6981
274
596
Entertainer/34/F
6544
142
8121
Journalist/66/M
6482
152
565
Business owner/73/M
5068
2776
10001
Journalist/56/M
4742
461
1593
Business owner /48/M
3025
122
288
Entertainer/36/M
2890
170
26
Politician/60/M
2655
718
2167
AC C
RI PT
Twitter
25
ACCEPTED MANUSCRIPT Business owner /72/M
1309
129
43
Business owner /68/M
1271
242
541
Politician/57/M
59
606
0
Note: The first column presents the occupation/age/sex of the deceased prominent
RI PT
individuals who died by suicide in Japan between July 2010 and December 2014. The second column shows the total number of posts on Twitter that contained the name of the deceased during the 3-week study period around their death. The third and fourth
SC
columns contain data on the total number of characters in newspaper articles and the total seconds on TV that were related to the deceased during the same study period,
AC C
EP
TE D
M AN U
respectively.
26
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
Figure 1: The number of tweets before and after the media report on celebrity suicide
27
ACCEPTED MANUSCRIPT Note: The top panel shows the number of tweets that mentioned the name of the 14 deceased individuals with a relatively large number of tweets on their death. They had more than 10,000 total posthumous tweets over the 3-week study period around their death. The bottom panel shows the number of tweets for 12 individuals with less than
AC C
EP
TE D
M AN U
SC
was first reported in the Yomiuri newspaper.
RI PT
10,000 tweets on their deaths. The zero on the x-axis indicates the day that their death
28
ACCEPTED MANUSCRIPT Figure 2: Incidence rate ratios of total suicides before and after the media reports on
AC C
EP
TE D
M AN U
SC
RI PT
celebrity suicide by the volume of posthumous tweets
29
ACCEPTED MANUSCRIPT Note: The top panel reports the estimated IRRs when 14 individuals with more than 10,000 posthumous tweets were included in the analysis, and the bottom panel shows estimation results when 12 individuals with less than 10,000 were included in the
AC C
EP
TE D
M AN U
SC
RI PT
analysis. The vertical bars indicate 95% confidence intervals.
30
ACCEPTED MANUSCRIPT Figure 3: Incidence rate ratios of total suicides before and after the media reports on
AC C
EP
TE D
M AN U
SC
RI PT
celebrity suicide for individuals with the largest media items
31
ACCEPTED MANUSCRIPT Note: The top panel reports the estimated IRRs when 10 individuals with the largest number of newspaper articles (measured by the number of Japanese characters) were included in the analysis, and the bottom panel shows estimation results when 10 individuals with the largest number of TV programs were included in the analysis. See
RI PT
Table 1 for lists of individuals included in the analysis. The vertical bars indicate 95%
AC C
EP
TE D
M AN U
SC
confidence intervals.
32
ACCEPTED MANUSCRIPT
Media reports on celebrity suicides are known to result in imitational behavior. This study examined people’s reaction on prominent suicides using Twitter. Twitter users tend to react a lot to reports on suicides by young entertainers.
RI PT
Celebrity suicides that received many tweets were followed by increased suicides.
AC C
EP
TE D
M AN U
SC
Suicides that generated little interest were not followed by increased suicides.
S0277-9536(17)30408-2
DOI:
10.1016/j.socscimed.2017.06.032
Reference:
SSM 11296
To appear in:
Social Science & Medicine
Received Date: 29 March 2017 Revised Date:
9 June 2017
Accepted Date: 25 June 2017
Please cite this article as: Ueda, M., Mori, K., Matsubayashi, T., Sawada, Y., Tweeting celebrity suicides: Users' reaction to prominent suicide deaths on Twitter and subsequent increases in actual suicides, Social Science & Medicine (2017), doi: 10.1016/j.socscimed.2017.06.032. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Tweeting Celebrity Suicides: Users' Reaction to Prominent Suicide Deaths on Twitter and Subsequent Increases in Actual Suicides
RI PT
Michiko Ueda1*, Kota Mori2, Tetsuya Matsubayashi3, Yasuyuki Sawada4
M AN U
SC
1. Faculty of Political Science and Economics, Waseda University, Tokyo, Japan 2. McCann Erickson, Tokyo, Japan 3. Osaka School of International Public Policy, Osaka University, Osaka, Japan 4. Faculty of Economics, University of Tokyo, Tokyo, Japan
AC C
EP
TE D
* Corresponding author: Michiko Ueda, Faculty of Political Science and Economics, Waseda University; Building No.3 1-6-1 Nishiwaseda, Shinjuku-ku, Tokyo 169-8050; tel: +81-3-32081176
ACCEPTED MANUSCRIPT Tweeting Celebrity Suicides: Users' Reaction to Prominent Suicide Deaths on Twitter and Subsequent Increases in Actual Suicides
Abstract
RI PT
A substantial amount of evidence indicates that news coverage of suicide deaths by
celebrities is followed by an increase in suicide rates, suggesting a copycat behavior.
However, the underlying process by which celebrity status and media coverage leads to
SC
increases in subsequent suicides is still unclear. This study collected over 1 million
individual messages (“tweets”) posted on Twitter that were related to 26 prominent
M AN U
figures in Japan who died by suicide between 2010 and 2014 and investigated whether media reports on suicide deaths that generated a greater level of reactions by the public are likely to be followed by a larger increase in actual suicides. We also compared the number of Twitter posts and the number of media reports in newspaper and on television to understand whether the number of messages on Twitter in response to the
TE D
deaths corresponds to the amount of coverage in the traditional media. Using daily data from Japan’s national death registry between 2010 and 2014, our analysis found an increase in actual suicides only when suicide deaths generated a large reaction from
EP
Twitter users. In contrast, no discernible increase in suicide counts was observed when
AC C
the analysis included suicide deaths to which Twitter users did not show much interest, even when these deaths were covered considerably by the traditional media. This study also found suicides by relatively young entertainers generated a large number of posts on Twitter. This sharply contrasts with the relatively smaller volume of reaction to them generated by traditional forms of media, which focuses more on the deaths of nonentertainers. The results of this study strongly suggest that it is not sufficient to examine only traditional news media when investigating the impact of media reports on actual suicides.
1
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
Keywords: Suicide; Twitter; social media; media; Japan; celebrity suicide; imitation
2
ACCEPTED MANUSCRIPT
1. Introduction
According to the World Health Organization (WHO, 2014), over 800,000 people die by
RI PT
suicide every year, which translates to one person every 40 seconds. Each of these
suicide deaths entails negative externalities, including significant psychological and
financial burdens on bereaved family members (Chen et al., 2009). Particularly, media
SC
reports on suicides by celebrities, which receive considerable coverage, may induce a series of related suicides once they are publicized. A substantial amount of evidence
M AN U
indicates that news coverage of suicide deaths by celebrities is followed by an increase in suicide rates, suggesting a copycat behavior. Coined the “Werther effect”, the phenomenon is claimed by several studies to be one of the main sources of negative externalities of suicide (e.g., Phillips, 1974; Wasserman, 1984; Stack, 1987; Ueda, Mori, and Matsubayashi, 2014). A recent meta-analysis by Niederkrotenthaler et al. (2012)
TE D
found that, on average, the suicide rate increases by 0.26 after a celebrity suicide, confirming the significance of the Werther effect.
EP
Although the underlying process by which celebrity status and media coverage leads to
AC C
increases in subsequent suicides is still unclear, social learning theory is often used to explain the observed change in suicide rates after media coverage of prominent suicides (Pirkis and Blood, 2010). The theory implies that media reports reinforce the notion that suicide is acceptable, and that other people may regard the deceased as a “model” either by his or her celebrity status (vertical integration) or shared characteristics (horizontal integration). Because such integration with a model hinges on the status and characteristics of the deceased, the theory also implies that media reports on the suicide deaths of certain celebrities have a greater impact than other celebrities, depending on the manner by which people perceive and react to the death of a potential model.
3
ACCEPTED MANUSCRIPT
Researchers have tried to understand such differential increases in actual suicides by the celebrity status of the deceased, or by the number of media reports on suicide deaths (e.g. Pirkis et al. 2006; Niederkrotenthaler et al. 2009). However, not all celebrities are
RI PT
revered in a similar way, and the extent to which people are affected by their deaths should vary. In addition, it is unclear whether the level of media publicity given to
SC
suicide deaths captures the level of people’s reactions to these deaths.
This study explicitly considered the level of reaction by the public to prominent suicide
M AN U
deaths and examined how it is associated with the subsequent change in actual suicides. It measured the level of public reaction to prominent suicide deaths by using comments posted on Twitter, a microblogging platform. The use of large-scale data from social media including Twitter for academic research has recently increased. Recent suiciderelated studies have used Twitter content to identify those at risk of suicide (Jashinsky
TE D
et al., 2014; O’Dea et al., 2015; Colombo et al., 2016), or gauge negative emotional feeling of the public such as depression (Cavazos-Rehg et al., 2016) and suicide-related sentiments (Woo et al., 2015). In other areas, activities on Twitter have been monitored
EP
to measure the level of national happiness (Dodd et al., 2011), or to predict heart disease
AC C
mortality (Eichstaedt et al. 2015).
Twitter is suitable for studying the public’s reaction to celebrity suicide for several reasons. First, Twitter has 313 million monthly active users worldwide (Twitter Inc., 2016) and its reach to the public is extensive. Second, Twitter users are likely to express their feelings and reactions without being subjected to censorship or restrictions on contents. Previous studies have shown that people demonstrate a higher degree of selfdisclosure and self-expression in computer-mediated settings (Suler 2004), such as social media (Ma et al. 2016), than in face-to-face settings. Thus, posts on Twitter are
4
ACCEPTED MANUSCRIPT assumed to contain the public’s genuine sentiment and reaction on suicide deaths by prominent figures.
Third, the volume of posts is expected to indicate the level of people’s reaction to the
RI PT
death that might have not been captured adequately by news outputs in traditional media, including newspapers and televisions. This is particularly true if the celebrity
was popular among the general public. Further, it has been reported that many young
SC
adults seek information about a celebrity’s suicide, career, and depression following a
celebrity’s suicide by using social media outlets such as Twitter and Facebook (Dillman
M AN U
Carpentier & Parrott 2016) and vulnerable individuals may be adversely affected if they are exposed to sensational content online.
Using the collection of over 1 million individual messages (“tweets”) posted on Twitter that were related to 26 prominent figures in Japan who died by suicide between 2010
TE D
and 2014, this study sought to understand the reaction by the public to media reports of celebrity suicides and its association with actual suicides. More specifically, we examined (1) how the number of tweets on celebrities changed before and after the
EP
media reports on the prominent suicide deaths; (2) whether the number of tweets on the suicide deaths corresponded to the amount of coverage in the traditional media such
AC C
as newspaper and television; and (3) whether a greater level of reactions to the celebrity deaths by Twitter users and by the traditional media led to a larger increase in actual suicides, using a method employed by Ueda et al. (2014) that studied the impact of media reporting on 109 celebrity suicides in Japan using daily suicide counts between 1989 and 2010.
2. Data and Method
5
ACCEPTED MANUSCRIPT To answer the three research questions, this study employed four sets of data: the number of prominent suicide deaths, Twitter posts, media reports, and suicide death records.
RI PT
2.1 Data on prominent suicide deaths
The authors identified a list of prominent Japanese who died by suicide by consulting
SC
newspaper articles and information on Wikipedia, an Internet encyclopedia. A page titled “well-known figures in Japan who died by suicide” on Wikipedia in Japanese
M AN U
(http://ja.wikipedia.org/wiki; accessed in March 2015) was first consulted. The authors decided to start with Wikipedia’s list because it contained all the 109 celebrity suicides that were included in the analysis by Ueda et al. (2014). These authors employed a manual coding of newspaper articles from 1989 to 2010. During our study period from July 2010 to December 2014, 43 individuals were included in the list. This list, and thus
TE D
our analysis, did not include non-Japanese figures who died by suicide.
As Wikipedia’s list contains individuals who are not well-known to the public, we
EP
searched the full names of these 43 individuals in the online database of the largest national newspaper in Japan, Yomiuri Shimbun, to determine whether their deaths were
AC C
reported in the national media. The online database of the Yomiuri Shimbun contains all articles published by the newspaper since 1986. Individuals were considered “(nationally) well-known” if their deaths were reported in the national edition of the Yomiuri Shimbun. These individuals were included in this study’s analysis, and those whose deaths were reported only in the local edition were excluded. In short, this study included “well-known’” figures in Japan whose suicide deaths were reported both in Wikipedia and the Yomiuri Shimbun.
6
ACCEPTED MANUSCRIPT In the end, 26 prominent figures (22 males and 4 females) who died by suicide between July 2010 and December 2014 were identified. The list included 12 entertainers (including two TV anchors), 6 business owners, 3 public officials (a cabinet member, a cabinet minister, and a judge), 3 journalists, a former athlete, and a scientist. The date of
RI PT
their death and the date of the first report on their death by the Yomiuri Shimbun were recorded.
SC
2.2 Data on Twitter posts
M AN U
Twitter started its service in April 2008 in Japan. As of December 2015, the number of monthly active users in Japan was estimated to be 35 million (Twitter Inc., 2016). The younger generation is more likely to use Twitter; According to one survey, 52.8% of individuals aged less than 30 use Twitter, compared with 24.3% of aged 50-59 and 15.5 % of those aged 60 years and older, respectively (Ministry of Internal Affairs and
TE D
Communications, 2015).
For each of the 26 individuals identified above, the authors searched tweets related to
EP
his or her death through an online platform provided commercially by Crimson Hexagon, a social media analytics company. Crimson Hexagon is a Twitter Official
AC C
Partner with access to full public tweets since July 2010 via the Twitter Firehose. Thus, this study’s analysis could utilize its full collection of Twitter posts dating back to July 2010, subject to certain restrictions described below. In contrast, if tweets were accessed through an application programming interface (API) provided by Twitter, comprehensive historical tweets would not be available for retrieval. The search was limited to messages made between July 2010 and December 2014.
7
ACCEPTED MANUSCRIPT The full name of the deceased was used as a search term. In certain cases, the title or a word that describes the position of the deceased was used as a search term. For example, for politicians, the search term “Representative+ last name” was used, in
RI PT
addition to their full name.
For each individual, the search date was set to begin seven days prior to the first
newspaper report of his or her suicide death, and end 15 days after the first report,
SC
inclusive of the day that their deaths were first reported in the Yomiuri Shimbun. Thus,
the search was conducted for a total duration of 22 days. The search was limited to posts
March 2015 and May 2015.
M AN U
in Japanese. The search of tweets through the Crimson Hexagon was conducted between
Data collection began with the recording of the total number of posts made during the 22-day study period for each well-known figure. Individual posts using the bulk export
TE D
function of the platform were then extracted. However, not all posts were available for bulk exports for the following two reasons. First, if an account had been deleted by the time of this study’s data retrieval, the messages made by the account were not available
EP
for download. Second, the platform does not allow users to download more than 10,000 posts at once, and the smallest time increment that can be specified is one day. Thus, if
AC C
there is a high volume of posts in one day, only a subset of the randomly-selected posts can be downloaded. When this happened, search terms were modified to narrow the results. For example, we narrowed the results by dividing the posts into two by the presence of the retweet sign (RT), or the reference to the method of suicide (e.g. “hanging”) in the main text. For certain high-volume cases, all the posts could not be retrieved even with this method. On average, 82.3% of the total posts were available for data extraction and analysis.
8
ACCEPTED MANUSCRIPT Information contained in the individual posts includes time stamp with date and time in GMT, contents, author, user name of the author, and the number of followers. The date and time were converted to the Japan Standard Time (JST) to make them comparable to
RI PT
the timing of media reports.
2.3 Data on media reports
SC
Newspaper articles and television programs that featured the 26 individuals after their suicide deaths were collected. For newspaper articles, the Yomiuri Shimbun database
M AN U
described above was used. For each individual, articles that contained the name of the deceased were retrieved, limiting the search to articles published within 14 days after the first newspaper report. The total number of Japanese characters in the articles was then counted. To collect information on television programs, this study used information compiled by the M-data, a Japanese media analytics company. The company
TE D
keeps record of the detailed contents of all major network television programs in the three largest metropolitan areas in Japan (Tokyo, Osaka, and Nagoya) using professional human coders. The number of seconds that TV programs featured the death of each of
EP
the 26 individuals was counted for analysis. Because the television programs reported their deaths earlier than the newspapers did in certain cases, the search period was set
AC C
to start one day before and end 14 days after the first newspaper report. We used the number of characters (newspapers) and seconds (television) as opposed to the number of media items because we think the former captures the level of media attention more accurately than the latter, as there are considerable variations in the length of media items.
2.4 Data on suicide deaths
9
ACCEPTED MANUSCRIPT The mortality data used in this study were obtained from death records in the Vital Statistics Report compiled by the Ministry of Health, Labour and Welfare (2010-2014). The Vital Statistics data used in this study were collected for administrative purposes, and anonymized and de-identified by the Ministry prior to analysis. Individual death
RI PT
records between July 2010 and December 2014 were made available for research purposes under the approval of the Ministry.
SC
The data in the Vital Statistics Report were based on death certificates issued by
physicians and subsequently reported to the local government by family members of the
M AN U
deceased. Deaths coded as X60–X84 under the ICD10 standard were classified as deaths by suicide and were thus included in this study’s dataset. For each calendar day, the total number of suicides was calculated. The total number of observations was 1,614.
TE D
2.5 Methods
We employed the following three sets of methods to achieve our three research aims. First, we examined how the number of tweets on the 26 prominent figures changed as a
EP
result of media reports on their suicide deaths by plotting the total number of tweets per day during the 22-day period for each celebrity. Second, we examined whether the
AC C
number of tweets on the suicide deaths corresponded to the amount of coverage in newspapers and television programs by comparing the number of tweets with the number of seconds spent by TV programs and the total number of Japanese characters in newspaper articles when the suicide deaths were reported. Our comparison was based on inspections of data and correlation coefficients.
Third, we examined whether a greater level of attention by Twitter as well as traditional media on these prominent deaths was associated with a larger increase in actual
10
ACCEPTED MANUSCRIPT suicides. We employed the same method as in Ueda et al. (2014), because their model allowed us to estimate the daily changes in the number of suicides following media reporting on prominent suicides compared to the baseline period with no suiciderelated reports. The model that we estimated was the same Poisson regression model
=∑
where
+
,
+
+
+
,
SC
log
RI PT
employed by Ueda et al (2014):
is the Poisson rate of suicides on day d in month m of year y, and
,
is a
M AN U
dummy variable that equals one if a prominent figure’s suicide was reported for the first time on day d in month m of year y. The subscript k shows the time difference from the onset of each suicide. The dummy variables were lagged forward up to 20 days and added separately to the model as
,
to
,
. The dummy variables should
capture the lagged influence of the news report on the number of suicides in the post-
TE D
reporting period. The variable of the first-reporting of celebrity suicide was also lagged backward up to 20 days as coefficients on
,
and
,
,
to
to
,
,
and included in the model. The were expected to be positive if the media
EP
reports were associated with an increase in actual suicides, whereas the coefficients on
AC C
the pre-reporting period were expected to be zero.
This study’s model also included year, month, day, and day of the week fixed effects. The year fixed effects,
, and the month fixed effects,
, were expected to capture the
influences of any year-specific factors, such as the national-level unemployment rate and population characteristics, and any month-specific factors, such as temperature. The day of the month fixed effect,
, and the day of the week fixed effects,
, were included in
the model because the number of suicides could experience a surge or drop on a particular day of the month or the week.
11
ACCEPTED MANUSCRIPT
We divided the 26 individuals into two groups by the volume of tweets and also by the amount of traditional media reporting, and then estimated the Poisson model separately for each of the two groups. The purpose of this exercise was to see if an increase in
RI PT
population suicides are more likely to happen after suicide deaths were followed by a larger number of tweets or greater media attention. We used Stata 13 by StataCorp for
SC
our statistical analysis.
M AN U
3． ．Results
3.1 The number of tweets on 26 prominent suicide deaths
We first summarized the number of tweets related to the 26 prominent suicide deaths before and after the media reports on their suicide deaths . The total number of relevant
TE D
posts during the 22-day period was 985,735. Among them, 159,891 (22,842 per day) were made before the first newspaper reports, and 825,844 (55,056 per day) were posted on or after the day that Yomiuri Shimbun first reported their deaths. As discussed
EP
below, television programs announced their deaths prior to newspaper reports, and thus the number in the pre-reporting period also included posts made after their deaths.
AC C
The gross sum of followers of the users who posted these messages was approximately 1.2 billion. In other words, tweets related to the suicide deaths were delivered potentially to over 1.2 billion Twitter feeds. The total number of retweets (forwarded posts) during the study period was 393,264, among which 184,639 were made on the day of the first newspaper report.
During the 22-day period, the minimum number of tweets per calendar day was 0, whereas the maximum was 120,463. The left column in Table 1 displays a list of
12
ACCEPTED MANUSCRIPT individuals with the total number of tweets during the 22-day period in descending order. Table 1 shows that the number of relevant tweets varied considerably across well-known figures. This means that the public was more familiar with and interested in the deaths of certain individuals than the others. The average number of posts over the
RI PT
22-day data collection period for the 26 individuals was 36,590 and the standard deviation was 54,324. The minimum total number of posts per day was 59, and the
maximum was 195,296. On average, the number of tweets was higher for entertainers
SC
(N = 12, Mean = 51,154, SD = 36,726), than for non-entertainers (N = 14, Mean = 24,793, SD = 49,284). The major exception was the death of a scientist at the top of the list; he
M AN U
was a supervisor of a high-profile project, but at the time of his death, the fabrication of research results by a junior scientist had been suspected and investigated. When this outlier was excluded from the non-entertainer group, the mean number of posts for non-entertainers was reduced to 12,614 (SD = 19,432). Thus, the public’s level
TE D
of reaction to the deaths was far from uniform.
Figure 1 displays trajectories in the number of tweets related to the 26 individuals within three weeks of the first newspaper report on their death. The x-axis shows the
EP
number of days from the first report, and the value of 0 corresponds to the day of the first report in the Yomiuri Shimbun. The y-axis shows the number of tweets on each day.
AC C
The well-known figures in this study were categorized into two groups. The first one included those whose names were mentioned more than 10,000 times during the 22day period; the second group contained individuals with less than 10,000 posts related to them.
The top panel in Figure 1 presents the number of tweets related to the 14 individuals in the first group (those mentioned more than 10,000 times on Twitter). This group included nine entertainers, one politician (a cabinet minister), one athlete, one scientist,
13
ACCEPTED MANUSCRIPT one business owner, and one journalist. The bottom panel shows the number of tweets related to the 12 individuals in the second group (those mentioned less than 10,000 times on Twitter). This group included five business owners, three entertainers, two journalists, one politician, and one judge. The ranges of the y-axes in the top and bottom
RI PT
panels are different.
For most of the deceased individuals, the number of tweets that mentioned them
SC
increased sharply on the day their deaths were reported. For example, prior to their
deaths, the numbers of posts related to the two individuals listed at the top of Table 1
M AN U
over a seven-day pre-reporting period were 52 and 156. These figures increased to 113,123 and 120,463 on the day when their death was reported, corresponding to a 2,174-fold and a 771-fold increase, respectively.
According to Figure 1, the surge started before day 0 for certain individuals. In most
TE D
cases, this is because the dates of the first media report were established by consulting newspaper articles. Although the Yomiuri Shimbun and other major national newspapers are published twice a day (in the morning and the late afternoon), they can
EP
sometimes lag behind other media such as television and online news outlets. Moreover, in the cases of two individuals, the media started reporting their suspected or attempted
AC C
suicides prior to their deaths while they remained missing or in critical condition; this circumstance explains the observed early surges.
Excluding these two cases, the average total daily number of posts related to the individuals was 340 during the pre-reporting period. The total number of posts surged to 106,763 one day before the Yomiuri Shimbun published its first article on their deaths, and to 464,243 on the day that it reported their deaths, a 1,363-fold (i.e., 136,300%) increase in the number of posts in comparison to the pre-reporting period.
14
ACCEPTED MANUSCRIPT
The number of Twitter posts remained elevated after the initial surge. In several cases, Twitter users kept mentioning the names of the deceased even after two weeks had passed since their suicides were reported. The average total number of posts on the 24
RI PT
individuals (after excluding the above two cases) was 13,566 and 5,898 on the 7th and 14th day, respectively, after the first day of the newspaper report.
SC
3.2. Comparison of Twitter activity and media coverage
M AN U
Next, we examined whether Twitter users and traditional media focused on the same set of prominent suicide deaths. The second and third columns of Table 1 list the number of seconds on television programs and the number of characters on the Yomiuri Shimbun (newspaper) that reported their deaths. The table indicates that Twitter posts were concentrated on the deaths of entertainers, whereas newspaper articles focused less on
TE D
these individuals. The age profiles of the entertainers were also distinct; Twitter posts focused on relatively young entertainers, but newspaper articles were centered on older
EP
ones.
Furthermore, Table 1 shows that the deaths of public officials (two politicians and a
AC C
judge) and business owners were featured more prominently in newspaper articles. Out of these individuals, only one politician’s death (a cabinet minister) was discussed widely on Twitter. The television programs’ focus was somewhat similar to that of Twitter users, but they also tended to draw attention to relatively older entertainers.
The correlation coefficient between the number of Twitter posts and the number of characters in newspaper articles was 0.587, and the correlation coefficient between the number of tweets and the total duration of television programs was 0.703 (N = 26).
15
ACCEPTED MANUSCRIPT Thus, the relationship between Twitter activity and traditional media coverage is not strong, confirming the distinct focus of Twitter users and traditional media.
RI PT
3.3. Twitter activity, media coverage, and population suicides
Our final task was to examine whether increased Twitter activity in the post-reporting period was associated with an increase in actual suicides. We first replicated the
SC
analysis by Ueda et al. (2014) to check whether there was a similar overall increase in
M AN U
actual suicides following media reports during the study period.
Using data on the deaths of 26 well-known figures, we found an increase in the number of actual suicides during the post-reporting period. For example, the estimated IRRs on day 4 and 5 were 1.09 (95% CI: 1.037-1.148) and 1.133 (95% CI: 1.076-1.192), respectively. No statistically-significant increase was observed from day 0 to day 3, and
TE D
the increase started from day 4. In addition, our estimation results indicate that the elevation in the number of suicide lasted more than 10 days. The estimated IRRs for day 15 and day 20 were 1.101 (95% CI: 1.054-1.151) and 1.078 (95% CI: 1.058-1.097),
EP
respectively, suggesting that the number of suicides was 8-10 percent higher even after
AC C
several weeks had passed since the first media reports.
We then examined the association between the volume of tweets and the number of suicides. The 26 individuals were split into the two groups described above by the number of posthumous tweets, and each group was estimated separately using the same model. The top panel of Figure 2 presents the estimated incidence rate ratios (IRRs) when 14 individuals with a relatively large number of posthumous posts (more than 10,000) were included in the analysis. The bottom panel of Figure 2 presents the
16
ACCEPTED MANUSCRIPT estimated IRRs when 12 individuals with less than 10,000 posts were included in the analysis.
Figure 2 indicates that a statistically-significant increase in total suicides occurred when
RI PT
this study’s analysis included individuals that generated large posthumous reactions on Twitter, but such an increase was almost non-existent when it included individuals with a relatively small number of tweets after their deaths. The top panel shows a clear
SC
elevated trend in the number of suicides after the media reports, and the number of suicides increased about 19 percent on day 5 (IRR 1.196, 95% CI: 1.063-1.344),
M AN U
compared to the baseline period. In contrast, the bottom panel shows no visible change in the number of suicides after the media reports. In other words, when suicide deaths did not generate much reaction in the social media, there was no associated increase in actual suicides.
TE D
The top panel of Figure 3 reports the estimation results when ten individuals with the highest number of newspaper articles (measured by the number of Japanese characters) were included in the analysis. In contrast to Figure 2, no discernible increase was
EP
observed after the media reports of their deaths. The bottom panel of Figure 3 shows the estimated IRRs when ten individuals who gathered most attention by TV programs
AC C
were included in the analysis. The pattern reported here is similar to the one reported in the top panel of Figure 2, but the increase started on day 0, not day 4 as reported in the top panel of Figure 2.
4． ．Discussion
The present study used messages posted on Twitter to understand the extent to which people react to the news on suicide deaths by prominent individuals. It also examined
17
ACCEPTED MANUSCRIPT the association between the level of reaction by Twitter users and actual suicides. To our knowledge, the present study constitutes the first attempt to systematically examine the reaction by social media users to prominent suicide deaths. Our analysis offered several notable findings. First, the number of tweets on the deceased individuals surged sharply
RI PT
when the news of their suicides was circulated. Second, Twitter users and traditional media focused on different sets of individuals. For example, suicides by relatively young entertainers generated a large number of posts on Twitter, whereas they generated a
SC
relatively smaller volume of reaction on traditional forms of media, such as the Yomiuri Shimbun newspaper. Traditional media tended to focus more on the deaths of public
M AN U
officials, business owners, and older entertainers. Third, prominent suicide deaths were followed by an increase in population suicides only when they generated a large reaction from Twitter users. In contrast, no discernible increase in suicide counts was observed when Twitter users did not show much interest, even when these deaths were
TE D
covered considerably by the traditional media.
The strength of this study includes the pioneering use of Twitter data in examining the public’s reaction to prominent suicide incidences. The Twitter data were used to
EP
measure how users of the social media reacted to the news on suicides by prominent figures in Japan. In contrast to past studies, this study did not make any arbitrary
AC C
assumption regarding the significance of those suicide deaths for the public.
This study has several limitations. First, we did not examine the content of the tweets to understand the types of reactions. Analyzing the contents of the tweets should allow us to distinguish whether users were merely spreading news on prominent suicides, or expressing sentiment in response to their deaths. Moreover, content analysis should help us explore whether prominent suicide deaths trigger particular emotional reactions. It should also elucidate whether the connectivity between suicidal Twitter
18
ACCEPTED MANUSCRIPT users (Colombo et al., 2016) produces a potential contagion effect following the death of a celebrity. Thus, in order to fully understand the mechanisms underlying the Werther effect, it is crucial to employ topic or sentiment analysis of the posts (e.g., Blei 2012; Ravi
RI PT
and Ravi 2015; Sykora et al. 2013).
The second limitation of our study concerns the representativeness of the sample and the issue of selection bias in the use of social media data (Bright et al. 2014). The
SC
reactions and sentiments expressed by Twitter users may not accurately represent the reactions of the general population, especially those of the older generations who are
M AN U
less likely to use social media (Ministry of Internal Affairs and Communications, 2015; Sloan et al., 2015). In addition, if there was a high-volume of tweets in one day, we were not able to download all posts, thus our collection of tweets contains fewer posts than the actual number of posts for high-volume cases (see section 2.5). However, because the actual number of tweets did not enter our estimation model when we investigated
TE D
the association between the level of reaction by Twitter users and actual suicides, the consequence of the missing data should be minimal.
EP
Third, this study did not analyze the deaths of celebrities who died of other causes. Thus, it is not clear if the observed large volume of tweets for certain individuals was due to
AC C
the nature of their death, namely suicide, or deaths by certain types of prominent figures tend to generate a large volume of comments on Twitter, regardless of their cause of death. Distinguishing these two possibilities warrants future research.
There are several important implications of this study. First, this study’s findings suggest that people are interested in the news on celebrity suicide. At the same time, they also indicate that the reaction to prominent suicides is far from uniform. Some of the suicide deaths included in our study did not generate a large reaction from Twitter
19
ACCEPTED MANUSCRIPT users, even though some of the deceased can be considered “celebrities” judging from the amount of media reporting by traditional media. Thus, people may not regard the deceased as a model merely by their celebrity status, as presumed in the “vertical integration” hypothesis in social learning theory. Our study also indicates that the media
RI PT
reports on prominent suicides may be a necessary condition for the Werther effect, but not a sufficient condition. In order for the Werther effect to happen, people should react
SC
strongly to their deaths.
Second, the results of this study strongly suggest that it is not sufficient to examine only
M AN U
traditional news media when investigating the impact of media reports on actual suicides. Such was the focus of previous studies (Pirkis and Blood, 2010). As we have demonstrated, the focus of traditional media can be vastly different from the interests of the general population. In light of these findings, one may speculate that the response from traditional media is not representative of the public’s reaction to suicides by well-
TE D
known figures.
Third, the vast amount of circulated social media posts on celebrity suicides found in
EP
this study may have a detrimental effect on vulnerable individuals. Although the content of the tweets was not analyzed formally, it was evident from casual inspection of
AC C
posts that a number of these posts were sensational and speculative. Thus, those on social media are likely to be exposed to a large amount of information on suicide, some of which can be sensational and unfounded. In addition, comments related to their deaths, especially the sympathetic ones, may further reinforce the notion that suicide is acceptable, resulting in “social learning” of acceptable behavior. It is known that suicidal individuals use social media to express their suicidal thoughts (O’Dea et al., 2015), and that adolescents who engage in self-harm are more exposed to the Internet than other adolescents (Ybarra and Mitchell, 2007; Daine et al., 2013). Thus, with ever-increasing
20
ACCEPTED MANUSCRIPT users of social media, vulnerable individuals may be exposed to suicide-related
AC C
EP
TE D
M AN U
SC
RI PT
information that may threaten their well-being.
21
ACCEPTED MANUSCRIPT
References
Blei, D.M., 2012. Introduction to Probabilistic Topic Modeling. Commun. ACM 55, 77–84.
RI PT
Bright, J., Margetts, H., Hale, S., Yasseri, T., 2014. The Use of Social Media for Research and Analysis : A Feasibility Study. Leeds, Corporate Document Services.
Cavazos-Rehg A., Krauss M.J., Sowles S., Connolly S., Rosas C., Bharadwaj M., Bierut L.J.,
SC
2016. A content analysis of depression-related tweets. Comput. Hum. Behav. 54, 351357.
M AN U
Chen, J. et al., 2009. Those who are left behind: an estimate of the number of family members of suicide victims in Japan. Social Indicators Research, 94(3), pp.535–544. Colombo G.B., Burnapa P., Hodoroga A., Scourfield J., 2016. Analysing the connectivity and communication of suicidal users on Twitter. Comput. Commun. 73, 291-300. Daine K., Hawton K., Singaravelu V., Stewart A., Simkin S., Montgomery P., 2013.
TE D
The power of the web: a systematic review of studies of the influence of the internet on self-harm and suicide in young people. Plos One 8, e77555. Dillman Carpentier, F.R. & Parrott, M.S., 2016. Young adults’ information seeking
EP
following celebrity suicide: Considering involvement with the celebrity and emotional distress in health communication strategies. Health Communication,
AC C
31(11):1334-1344.
Dodds P.S., Harris K.D., Kloumann I.M., Bliss C.A., Danforth C.M., 2011. Temporal patterns of happiness and information in a global social network: hedonometrics and Twitter. Plos One 6, e26752.
Eichstaedt J.C., Schwartz H.A., Kern M.L., Park G., Labarthe D.R., Merchant R.M., Jha S., Agrawal M., Dziurzynski L.A., Sap M., Weeg C., Larson E.E., Ungar L.H., Seligman M.E.P., 2015. Psychological language on Twitter predicts county-level heart disease mortality. Psychol. Sci. 26, 159-169.
22
ACCEPTED MANUSCRIPT Jashinsky J., Burton S.H., Hanson C.L., West J., Giraud-Carrier C., Barnes M.D., Argyle T., 2014. Tracking suicide risk factors through Twitter in the US. Crisis 35, 51-59. Ma, X., Hancock, J., Naaman, M., 2016. Anonymity, intimacy and self-disclosure in social media, Proceedings of the 2016 CHI Conference on Human Factors in Computing
RI PT
Systems. pp. 3857–3869. Ministry of Health, Labour and Welfare, 2010-2014. The Vital Statistics Report.
Ministry of Internal Affairs and Communications, 2015. White paper on information and
SC
communication (in Japanese).
Niederkrotenthaler T., Fu K., Yip P., Fong D., Stack S., Cheng Q., 2012. Changes in suicide
Commun. H. 66, 1037–1042.
M AN U
rates following media reports on celebrity suicide: a meta-analysis. J. Epidemiol.
Niederkrotenthaler, T., Till, B., Kapusta, N.D., Voracek, M., Dervic, K., Sonneck, G., 2009. Copycat effects after media reports on suicide: A population-based ecologic study. Soc. Sci. Med. 69, 1085–1090.
TE D
O'Dea B., Wan S., Batterhamc P.J., Calear A.L., Paris C., Christensen H., 2015. Detecting suicidality on Twitter. Internet Interventions 2, 183-188. Phillips D., 1974. The influence of suggestion on suicide: substantive and theoretical
EP
implications of the Werther effect. Am. Sociol. Rev. 39, 340–354. Pirkis, J.E., & Blood, W., 2010. Suicide and the news and information media: A critical
AC C
review. Barton, ACT: Commonwealth of Australia.
Pirkis, J.E., Burgess, P.M., Francis, C., Blood, R.W., Jolley, D.J., 2006. The relationship between media reporting of suicide and actual suicide in Australia. Soc. Sci. Med. 62,
2874–2886.
Ravi, K., Ravi, V., 2015. A survey on opinion mining and sentiment analysis: Tasks, approaches and applications. Knowledge-Based Syst. 89, 14–46.
23
ACCEPTED MANUSCRIPT Sloan L., Morgan J., Burnap P., Williams M., 2015. Who tweets? Deriving the demographic characteristics of age, occupation and social class from Twitter user meta-data. Plos One 10, e0115545. Stack S., 1987. Celebrities and suicide: a taxonomy and analysis, 1948-1983. Am. Socio.
RI PT
Rev. 52, 401–12. Suler, J., 2004. The online disinhibition effect. CyberPsychology & Behavior, 7(3), 321– 326.
SC
Sykora, M.D., Jackson, T.W., O’Brien, A., Elayan, S., 2013. Emotive ontology: Extracting
fine-grained emotions from terse, informal messages. IADIS Int. J. Comput. Sci. Inf.
M AN U
Syst. 8, 106–118.
Twitter.com, 2016. About Twitter, Inc. https://about.twitter.com/company (accessed 17.02.24).
Ueda M., Mori K., Matsubayashi T., 2014. The effects of media reports of suicides by wellknown figures between 1989 and 2010 in Japan. Int. J. Epidemiol. 43, 623-629.
TE D
Wasserman I., 1984. Imitation and suicide: a reexamination of the Werther effect. Am. Sociol. Rev. 49, 427-436.
Woo H., Cho Y., Shim E., Lee K., Song G., 2015. Public trauma after the Sewol ferry
EP
disaster: the role of social media in understanding the public mood. Int. J. Environ. Res. Public Health 12, 10974-10983.
AC C
World Health Organization. (2014) Preventing suicide: A global imperative. Ybarra, M.L. & Mitchell, K.J., 2007. Prevalence and frequency of Internet harassment instigation: Implications for adolescent health. Journal of Adolescent Health,
41(2),189–195.
24
ACCEPTED MANUSCRIPT Table 1: List of prominent individuals who died by suicide in Japan with the volume of Twitter posts and the amount of media reporting by the types of media: 2010-2014.
Newspaper
Television
Scientist/52/M
195296
18130
122728
Entertainer/24/F
167444
180
23837
Entertainer/62/F
135001
1247
189416
Entertainer/35/M
132439
309
27624
Politician/73/M
74362
7916
43013
Entertainer/45/M
57957
249
551
Entertainer/78/M
39533
Athlete/42/M
34654
Artist/34/M
26221
Journalist/52/M
19819
Entertainer/79/M
15779
Entertainer/57/M
SC
M AN U 53988
2155
19372
269
136
252
1823
290
52
15203
692
5049
10791
245
728
10168
195
3576
EP
Entertainer/44/M
1271
TE D
Business owner/64/M
Entertainer/42/F
9076
249
39738
Business owner /61/M
6981
274
596
Entertainer/34/F
6544
142
8121
Journalist/66/M
6482
152
565
Business owner/73/M
5068
2776
10001
Journalist/56/M
4742
461
1593
Business owner /48/M
3025
122
288
Entertainer/36/M
2890
170
26
Politician/60/M
2655
718
2167
AC C
RI PT
25
ACCEPTED MANUSCRIPT Business owner /72/M
1309
129
43
Business owner /68/M
1271
242
541
Politician/57/M
59
606
0
Note: The first column presents the occupation/age/sex of the deceased prominent
RI PT
individuals who died by suicide in Japan between July 2010 and December 2014. The second column shows the total number of posts on Twitter that contained the name of the deceased during the 3-week study period around their death. The third and fourth
SC
columns contain data on the total number of characters in newspaper articles and the total seconds on TV that were related to the deceased during the same study period,
AC C
EP
TE D
M AN U
respectively.
26
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
Figure 1: The number of tweets before and after the media report on celebrity suicide
27
ACCEPTED MANUSCRIPT Note: The top panel shows the number of tweets that mentioned the name of the 14 deceased individuals with a relatively large number of tweets on their death. They had more than 10,000 total posthumous tweets over the 3-week study period around their death. The bottom panel shows the number of tweets for 12 individuals with less than
AC C
EP
TE D
M AN U
SC
was first reported in the Yomiuri newspaper.
RI PT
10,000 tweets on their deaths. The zero on the x-axis indicates the day that their death
28
ACCEPTED MANUSCRIPT Figure 2: Incidence rate ratios of total suicides before and after the media reports on
AC C
EP
TE D
M AN U
SC
RI PT
celebrity suicide by the volume of posthumous tweets
29
ACCEPTED MANUSCRIPT Note: The top panel reports the estimated IRRs when 14 individuals with more than 10,000 posthumous tweets were included in the analysis, and the bottom panel shows estimation results when 12 individuals with less than 10,000 were included in the
AC C
EP
TE D
M AN U
SC
RI PT
analysis. The vertical bars indicate 95% confidence intervals.
30
ACCEPTED MANUSCRIPT Figure 3: Incidence rate ratios of total suicides before and after the media reports on
AC C
EP
TE D
M AN U
SC
RI PT
celebrity suicide for individuals with the largest media items
31
ACCEPTED MANUSCRIPT Note: The top panel reports the estimated IRRs when 10 individuals with the largest number of newspaper articles (measured by the number of Japanese characters) were included in the analysis, and the bottom panel shows estimation results when 10 individuals with the largest number of TV programs were included in the analysis. See
RI PT
Table 1 for lists of individuals included in the analysis. The vertical bars indicate 95%
AC C
EP
TE D
M AN U
SC
confidence intervals.
32
ACCEPTED MANUSCRIPT
Media reports on celebrity suicides are known to result in imitational behavior. This study examined people’s reaction on prominent suicides using Twitter. Twitter users tend to react a lot to reports on suicides by young entertainers.
RI PT
Celebrity suicides that received many tweets were followed by increased suicides.
AC C
EP
TE D
M AN U
SC
Suicides that generated little interest were not followed by increased suicides.