Sample of media reports

We obtained all reports from the 11 largest Austrian nationwide newspapers that included the term suicide between 1 January and 30 June 2005 from the Austrian Press Agency. Each of these newspapers has coverage rates exceeding 1% of the population. Articles using ‘self-murder’ (original search term: selbstmord) or ‘free death’ (freitod) were also included, because of their frequent use as German synonyms for suicide. The search yielded 1055 articles. We excluded items that used suicide only as a metaphor (n = 92), were about suicide bombing (n = 381), mentioned suicide briefly (<50% of the article) but were on topics not related to suicide (n = 113) or were located in the movie or television listings or in the classified advertisements (n = 58). The final sample comprised 497 items.

Content analysis of media items

Quantitative analysis

The quantitative analysis aimed at measuring associations between item contents and suicide rates. Statistics Austria (www.statistik.at/) provided daily data on suicides for each Austrian federal state for the time period December 2004 to July 2005 as well as data on the population sizes. The outcome variable (suicide) was defined as the difference between the suicide rates (suicides per 100 000 total population) in the week preceding the publication date and in the week after publication. More specifically, days –14 to –8 (where day 0 marks the publication date) were defined as the pre-publication period and days 0 to 6 were the corresponding post-publication period. We did not use the week immediately preceding the publication date as the pre-period because of the possibility that other media (e.g. television or internet) may have reported the suicidal event slightly earlier than the newspapers. This, given the corresponding impact of these media, might have led to contamination of the data. Only suicides that occurred in the key market of the respective media item, i.e. the area where more than 10% of the population were exposed to the item, were used for the calculation of the outcome variable. In total, there were 18 key markets. Nine out of the 11 newspapers analysed had just one key market each; the 2 remaining newspapers each had several overlapping key markets. These newspapers produced regionally differing editions of their papers. The population sizes of the key markets ranged from 362 000 to 3 492 000, with each market comprising one or more of the nine Austrian federal states. The sociodemographic profile of the population was similar across the key markets with regard to age and gender. ²³ The proportion of people living in urban areas ranged from 51 to 100% for the different key markets. ²³

We assigned each media item to its key market, and the change in the suicide rate was assigned to the item. Seven items (1.4%) were not published in an area with item coverage of over 10%. These items were excluded. For an overview of the suicide rates and article counts in the ten most frequent key markets (altogether containing 477 items (97.3% of all items)), see Fig. 1. Data on the circulation rates of newspapers across the country were obtained from Media Analysis Austria (2005). ²⁴ Data on the local editions including individual items were obtained from the Austrian Press Agency.

Point-biserial and Pearson correlation coefficients were calculated to quantify the associations between the qualitative characteristics of the items and suicide rates. The codes in the family ‘basic item characteristics’ were tested alongside significant qualitative variables in a multivariate model. This model included controls for the newspaper in which the item was published, because local factors relevant to the specific key markets of newspapers (e.g. degree of urbanisation) may have an impact on media effects. In addition, we controlled for month of publication, thus accounting for seasonal variation in suicide rates, which might also influence media effects. ^{Reference Stack and Leenars2}

To assess whether post-item changes in suicide rates varied with the circulation rates of the media items we compared the absolute value of the change in the suicide rate in key markets with the change in the rate in those parts of the country where coverage rates were under 10% using an independent groups t-test for unequal variances (Welch's test). Data analysis was performed in SPSS 15 for Windows XP.

Latent class analysis

Potentially protective and harmful item characteristics may form classes based on the reporting strategy applied in the production of the report. Latent class analysis (LCA), a method with manifold applications in psychiatry and related fields, ^{Reference Breslau, Reboussin, Anthony and Storr25–Reference Magidson and Vermunt27} postulates that associations among a set of observed variables are the result of a common underlying latent class structure. The goal of LCA is to identify the smallest number of latent classes that adequately describes the associations between observed variables, starting with a one-class model and proceeding to successive models with increasing class numbers. ^{Reference Breslau, Reboussin, Anthony and Storr25,Reference Magidson and Vermunt27} Information about the class structure is given by the latent class probabilities, which indicate the proportion of the sample assigned to each class, and by response probabilities, which are the proportions of class members reporting each item characteristic. ^{Reference Breslau, Reboussin, Anthony and Storr25,Reference Garson26} In addition to the indicator variables that define classes, covariates, which control and describe the classes, can be included in the model. We used all the item characteristics listed as potentially protective, and also the harmful characteristics that were significantly associated with suicide rates, as indicators in the bivariate analyses. As a result of their practical relevance in describing suicide reports, we controlled for several basic item characteristics, namely for the quantity of reporting (variable: ‘repetitive reporting’), for the focus on a completed suicide, and for article length. Consistent with the literature, ^{Reference Magidson and Vermunt27} we utilised two LCA modules in the Latent Gold 4.5 software for Windows XP, namely latent class cluster analysis and latent class factor analysis, to identify the best fitting solution.

Fig. 1

Key markets: suicides rates and article counts.

Box plots of weekly suicide rates (deaths/100 000 population and article counts for each of the ten most frequent key markets of the Austrian newspapers in the observation period (January to June 2005). Boxes represent values between 25th and 75th percentiles; whiskers, the upper and lower adjacent values; vertical lines, the median; *, the arithmetic mean.

The latent class cluster algorithm differs from traditional cluster algorithms, which group cases near each other according to some definition of distance. It defines one cluster per latent class, using model-based probabilities to classify cases. ^{Reference Garson26} Latent class factor analysis treats the latent classes as factors, but avoids problems inherent in traditional factor analysis, such as having to rotate factors and having to assume continuous observed variables. ^{Reference Garson26}

The fit statistics served as a tool for model selection. The observed frequencies of item characteristics were compared with the expected frequencies predicted by the model by calculating a likelihood-ratio goodness of fit value. ^{Reference Breslau, Reboussin, Anthony and Storr25,Reference Garson26} Because there were sparse contingency tables, the likelihood-ratio statistic did not follow a chi-squared distribution; hence, we calculated bootstrapped P-values. Models were compared using the Bayesian information criterion and via the percentages of classification error. ^{Reference Breslau, Reboussin, Anthony and Storr25,Reference Garson26} Further, we analysed bivariate residuals between pairs of observed variables. The initial assumption of local independence of pairs of variables with high residuals (>3.84) was relaxed to improve model fit. ^{Reference Breslau, Reboussin, Anthony and Storr25,Reference Garson26} Finally, we allowed for correlations among identified factors. This strategy is recommended whenever theoretical expectation suggests that the dimensions of an underlying structure are not completely mutually independent. ^{Reference Fabrigar, Wegener, MacCallum and Strahan28} Based on the most probable class assignment of media items, we calculated the correlation coefficients of classes with post-report changes in suicide rates.

Bivariate analysis

Multivariate analysis

In the multivariate model, repetitive reporting, reports about suicides by jumping, and reporting on public myths about suicide predicted post-item increases in suicide rates. Media items with a main focus on suicidal ideation predicted decreases in suicide rates. The model was significant (F = 2.82, d.f. = 34,455, P<0.001; Table 4). Model fit in terms of R ² was 0.17, indicating that 17% of the variance in changes in suicide rates were explained by the model, which is acceptable. ^{Reference Nau29} The corresponding effect size f ² of 0.21 was medium sized according to the benchmarks proposed by Cohen. ^{Reference Cohen30} Part correlation coefficients, which represent the unique contributions of individual variables to the variance explained, are provided in Table 4.

Table 4

Multivariate regression coefficients for post-report changes in suicide rates for ‘basic item characteristics’ and significant qualitative characteristics^a

Code and subcode	b b	95% CI	P	Part correlation
Quantity of reporting
Repetitive reportingc	0.01	0.003 to 0.02	0.005	0.10
Density of reportingc	0.002	–0.003 to 0.01	0.44	0.04
Main focus
Completed suicide	–0.02	–0.10 to 0.07	0.70	–0.02
Attempted suicide	–0.01	–0.11 to 0.10	0.93	0.001
Homicide suicide	–0.01	–0.11 to 0.08	0.78	–0.02
Suicidal ideation	–0.15	–0.27 to –0.03	0.02	–0.11
Prevention programme	0.01	–0.14 to 0.15	0.98	0.01
Suicide research	0.12	–0.01 to 0.25	0.07	0.07
Reported method
Hanging	–0.09	–0.18 to 0.004	0.06	–0.07
Jumping	0.13	0.04 to 0.21	0.004	0.13
Shooting	–0.03	–0.11 to 0.05	0.49	–0.02
Drowning	0.07	–0.06 to 0.20	0.28	0.06
Poisoning	0.06	–0.04 to 0.16	0.22	0.05
Other method	0.07	–0.03 to 0.17	0.15	0.06
Prominence
Title page	0.04	–0.10 to 0.18	0.56	0.03
Fictionality
Content is fictional	0.09	–0.03 to 0.21	0.15	0.07
Causes of suicidality
Societal problems reported to have increased	0.03	–0.11 to 0.17	0.67	0.02
Style and wording
Public myths reported	0.20	0.03 to 0.37	0.02	0.10
Reference to a suicide ‘epidemic’	0.08	–0.07 to 0.24	0.28	0.05
Several independent suicidal acts reported	0.02	–0.05 to 0.09	0.54	0.02
Dichotomous thinking	0.07	–0.01 to 0.16	0.08	0.08
Background information
Expert opinion reported	0.03	–0.08 to 0.14	0.63	0.02
Contact support service	0.14	–0.003 to 0.28	0.06	0.08

Analysis of residuals did not indicate any violations of the assumptions of linear regression, such as non-linearity, nonnormality of error terms, or heteroscedasticity. Autocorrelation plots (ACF, pACF) and Durbin–Watson tests (DW = 1.89) did not suggest autocorrelation, and variance inflation factors did not indicate any multicollinearity problems (largest variance inflation factors: 3.05).

t-test

Post-item changes in suicide rates were larger in regions with high circulation rates of the media items (mean 0.22, s.d. = 0.18) than in regions with low circulation rates (mean 0.09, s.d. = 0.09; t(710.48) = 14.18, P<0.001).

Latent class analysis

The LCA identified a (two-factor) four-class model as the best fit. The model-fit indices of several models tested are reported in Table 5. Of the 14 indicator variables used initially 12 were included in the final model. The variables ‘pathology reported’, which showed only weak discrimination, and ‘suffering due to suicidal act’, which had a very low level of endorsement across classes, were omitted. The correlation between the two factors was r = 0.30. Table 6 and Fig. 3 display the endorsement probabilities for each class.

Table 5

Model-fit indices for latent classes in suicide-related media reports

Model	L 2	Bootstrap P	BIC, L 2	d.f.	Classification error, %
1-cluster	3693	0.26	818.22	463	0.00
2-cluster	3156	0.35	380.27	447	0.02
3-cluster	3016	0.41	340.17	431	0.09
4-cluster	2969	0.35	392.83	415	0.11
5-cluster	2824	0.31	346.75	399	0.13
6-cluster	2792	0.20	414.57	383	0.05
1-factor	3156	0.34	380.27	447	0.02
2-factor	2955	0.34	279.14	431	0.02
3-factor	2834	0.33	257.52	415	0.08
2-factora	2876	0.29	243.51	424	0.02

Table 6

Latent class analysis: response probabilities

			Joint probability (s.e.)a
	Frequency (%)	Mean	Class 1: suicide case class	Class 2: mastery of crisis class	Class 3: epidemiological facts class	Class 4: expert opinion class
Class assignment probabilities			0.72 (0.03)	0.09 (0.02)	0.12 (0.02)	0.08 (0.01)
Protective indicators
Focus: suicidal ideation	26 (5.2)		<0.001 (<0.001)	0.60 (0.13)	0 (0)	<0.001 (0.01)
Focus: prevention programme	31 (6.2)		50.001 (<0.001)	0.07 (0.04)	0.002 (0.01)	0.75 (0.08)
Focus: suicide research	26 (5.2)		<0.001 (<0.001)	0 (<0.001)	0.36 (0.07)	0.14 (0.06)
Expert opinion reported	42 (8.4)		0.006 (0.004)	0.06 (0.03)	0.18 (0.05)	0.71 (0.08)
Myths debunked	33 (6.6)		<0.001 (<0.001)	0.06 (0.04)	0.03 (0.02)	0.75 (0.08)
Contact support service	28 (5.6)		<0.001 (<0.001)	0.07 (0.04)	< 0.001 (0.002)	0.63 (0.09)
Suicide-related statistical data reported	51 (10.2)		<0.001 (0.002)	<0.001 (<0.001)	0.63 (0.08)	0.38 (0.08)
Harmful indicators
Several independent suicidal acts reported	94 (18.9)		0.12 (0.02)	0.07 (0.04)	0.72 (0.07)	0.11 (0.05)
Reference to a suicide ‘epidemic’	16 (3.2)		<0.001 (<0.001)	<0.001 (0.001)	0.24 (0.06)	0.05 (0.04)
Societal problems reported to have increased	23 (4.6)		<0.001 (<0.001)	<0.001 (<0.001)	0.19 (0.06)	0.32 (0.08)
Public myths reported	11 (2.2)		0.01 (0.01)	<0.001 (<0.001)	0.13 (0.05)	<0.001 (0.002)
Dichotomous thinking		0.16	0.13 (0.01)	0.17 (0.04)	0.25 (0.05)	0.28 (0.05)
Covariates
Focus: completed suicide	234 (47.0)		0.57	0.03	0.43	0.11
Repetitive reporting		4.44	4.97	2.31	4.21	2.15
Article length		160.7	135.8	126.4	249.6	301.5

Class 1 (72.2% of media items), which we conceptualised as the suicide case class, had low probabilities of reporting potentially protective or harmful item characteristics. Media items in this class tended to report on a case of completed suicide. As compared with the other classes, items were more likely reported repeatedly and in short articles (Table 6, Fig. 3).

Fig. 3

Latent class analysis: response probabilities.

Estimated probabilities of showing potentially protective and harmful item characteristics for each of the four latent classes of media items.

a. Continuous variable: the value is the class-specific arithmetic mean mapped onto the interval 0 to 1, where 0 is identified with the lowest and 1 with the highest observed value across classes.

Class 2 (8.8% of media items), which we conceptualised as the mastery of crisis class, tended to report individual suicidal ideation not accompanied by suicidal behaviour. Repetitive reporting and long articles were rather uncommon in this class. Compared with the other classes, members of this class had the smallest mean number of harmful item characteristics (Table 6, Fig. 3). This was the only class with a negative mean change in the suicide rate after publication (Table 7).

Table 7

Latent class analysis: mean number of item characteristics in media items in each class by potentially protective and harmful characteristics and totals

		Mean (s.e.)
	All media items	Class 1: suicide case class	Class 2: mastery of crisis class	Class 3: epidemiological facts class	Class 4: expert opinion class
Protective item characteristics	0.68 (1.20)	0.18 (0.42)	1.44 (0.62)	1.58 (0.94)	3.94 (1.26)
Harmful item characteristicsa	0.29 (0.59)	0.13 (0.34)	0.09 (0.30)	1.40 (0.77)	0.50 (0.70)
Totalb	3.42 (2.11)	2.85 (1.80)	3.75 (1.48)	5.40 (1.96)	6.19 (1.89)
Suicide ratesc	0.04 (0.28)	0.03 (0.28)	–0.08 (0.27)	0.14 (0.29)	0.13 (0.23)

Class 3 (7.6% of media items), which we conceptualised as the epidemiological facts class, comprised media items that tended to report on suicide research and suicide statistics. The reporting of several independent suicidal acts and suicide myths, references to a suicide epidemic, repetitive reporting and long articles were rather common in this class (Table 6, Fig. 3). On average, these articles had the highest number of harmful item characteristics (Table 7).

Class 4 (11.5% of media items), which we conceptualised as the expert opinion class, was likely to report on suicide prevention programmes and expert opinions, to give contact information on public support services, and to debunk public suicide myths. Articles in this class tended to include the statement that societal problems are increasing, and were characterised by a higher degree of dichotomous thinking. Repetitive reporting was uncommon, and articles were longer than for the other classes (Table 6, Fig. 3). On average, articles in this class had the highest number of potentially protective item characteristics (Table 7, Fig. 3).

Mean posterior assignment probabilities were well over 90% for all classes (Class 1: 94.8%; Class 2: 95.7%; Class 3: 97.7%; Class 4: 95.5%), indicating high classification certainty. ^{Reference Nagin31}

Correlations of latent classes with suicide rates

Assignment to the mastery of crisis class correlated negatively with suicide rates (r = –0.111, n = 490, P = 0.014). Assignments to the epidemiological facts class (r = 0.124, n = 490, P = 0.006) and the expert opinion class were positively correlated with suicide rates (r = 0.095, n = 490, P = 0.035). Assignment to the suicide case class was not significantly associated with suicide rates (r = –0.084, n = 490, P = 0.062).

Key findings

The possibility of a suicide-protective effect of media items on positive coping in adverse circumstances, which has been discussed in the literature on a purely theoretical basis, was empirically supported by the present findings. Reporting of individual suicidal ideation (not accompanied by attempted or completed suicide) was associated with a decrease in suicide rates. The LCA suggested that media items on suicidal ideation formed a distinctive class of articles, which had a low probability of containing potentially harmful item characteristics. By contrast, media items reporting public suicide myths and repetitive suicide-related reporting were associated with increases in suicide rates. Items with a high probability of containing expert opinions or epidemiological facts were identified as distinctive latent classes of media reports, which tended to be intertwined with unfavourable, sensationalist contents. These classes were associated with increases in suicide rates.

Meaning of the key findings

The study suggests that prevention of suicidal behaviour by media reports is possible. It expands the research base on the possible impact of suicide-related media content on suicide rates to also include a preventive effect, which we conceptualised as a Papageno effect. In particular, items with a focus on suicidal ideation not accompanied by a suicide attempt or completed suicide may have a protective impact. The reporting strategy underlying these media items seems to deviate from the general tendency to present potentially protective suicide-related information together with potentially harmful item characteristics. Evidence to date suggests the preventive potential of quantitative reductions in sensationalist reporting. ^{Reference Sonneck, Etzersdorfer and Nagel-Kuess10,Reference Niederkrotenthaler and Sonneck32} To this, the present findings add that the actual reporting of suicidal ideation itself may contribute to preventing suicide. Media items that focus on individual suicidal ideation may enhance identification with the reported individual and thus highlight the reported outcome as ‘going on living’. Health-promoting activities may be most effective when they encourage the publication of articles on individuals who refrained from adopting suicidal plans, and instead adopted positive coping mechanisms in adverse circumstances.

The framing of media items reporting on epidemiology and expert opinions with potentially harmful item characteristics, such as a reference to a suicide epidemic or a statement that suicide-related societal problems are becoming more pressing, cast some doubt on the role of these types of information in media discourse. Consistent with findings in the field of mass communication, ^{Reference Krüger33} expert opinions seem to be used as a means of giving an air of seriousness to reporting, but at the same time tend to be embedded in an unfavourable, sensationalist context. Similar conclusions were reached in an analysis of suicide-related print media reports in Australia concerning statistical media items, with the researchers noting that, in their opinion, statistical items were written in a more unfavourable, sensationalist style than media items on completed suicides. ^{Reference Pirkis, Francis, Blood, Burgess, Morley, Stewart and Putnis34} More generally, the tendency to present news in a negative way in the tabloid media has been assumed to result in a negative world image among recipients (the so-called ‘scary world hypothesis’). ^{Reference Heitmayer and Hagan35} The effects of such ‘mixed messages’ on individuals with suicidal constriction clearly warrants further scrutiny. The correlations of items in the expert opinion class and in the epidemiological facts class with increases in suicide rates suggest that these items may have a harmful effect.

Other research

Dose–response relationships between the quantity of reporting on completed suicide and subsequent suicides, as suggested in the present study, have been repeatedly identified in the literature. ^{Reference Phillips1,Reference Pirkis, Burgess, Francis, Blood and Jolley4} Also, this study replicates the finding that changes in suicide rates following the publication of media items are more pronounced in regions where a higher proportion of the population is exposed rather than in regions with a lower exposure. ^{Reference Etzersdorfer, Voracek and Sonneck11} A further finding in line with the literature is the association of reports on one particular suicide method (i.e. jumping) with post-item increases in suicide. ^{Reference Sonneck, Etzersdorfer and Nagel-Kuess10} Of note, suicides by jumping are more likely to be reported by the print media than suicides by any other method. ^{Reference Niederkrotenthaler, Till, Herberth, Voracek, Kapusta and Etzersdorfer36} Even after controlling for a range of variables likely to influence the probability of reporting (e.g. sociodemographic characteristics of the suicide victim), jumping as a suicide method was found to predict subsequent suicide reporting in the Austrian print media. ^{Reference Niederkrotenthaler, Till, Herberth, Voracek, Kapusta and Etzersdorfer36}

Study strengths and limitations

A strength of the present study is its coverage of all print media items over a period of 6 months that included the word ‘suicide’, which is likely to have succeeded in the identification of a large proportion of all suicide reports. Furthermore, the coding system applied showed high intercoder reliability, thereby demonstrating high intersubjective analytic consensus. Qualitative approaches are still sparse in suicide research, but are deemed fruitful, and even vital, for research progress, because of their potential to generate new hypotheses. ^{Reference Hjemeland and Knizer37} Mixed-method (i.e. quantitative–qualitative) approaches, as implemented here, are increasingly undertaken in the social sciences and complement each other. ^{Reference Creswell and Clark38} Further strengths of the present study are the comparatively large number of media items analysed qualitatively, the assessment of repetitive reporting on the same story and the adoption of an ecological study design to fit variations in the circulation rate of the analysed media items across the country.

Limitations of the study concern its ecological design, which is appropriate for identifying associations, but not for establishing causality. Methodological approaches using individual-level data are necessary to deal with problems such as the ecological fallacy (i.e. cross-level bias), and further potential confounders beyond seasonality (which was accounted for in the multivariate analysis). Furthermore, the unique contributions of single variables to the variance explained in the multivariate analysis are, albeit nominally significant, small. This may be partly attributable to the definition of the outcome variable as the difference between post-item and pre-item suicide rates. ^{Reference Nau29} However, completed suicide is a rare event per se, and thus even small reductions in suicide prevalence seem to be relevant to collaborations with the mass media. Finally, the present study is confined to an analysis of print media. This appears justified by the high exposure of the population to the set of newspapers analysed. In 2005, 71.9% of the population were regularly exposed to one or several of these newspapers. ²⁴ A recent meta-analysis found that studies of print media effects are more likely to find evidence of a post-report increase in suicide than studies of other media types, ^{Reference Stack3} but did not take into account the internet. Future research should also focus on this rapidly emerging medium. ^{Reference Curah39}

Suggestions for future research

This study is the first, to our knowledge, to use extensive measures of media item quality. It enabled us to generate more fine-grained hypotheses regarding the effects of suicide-related reporting. Research in other countries is needed to assess the generalisability of the findings. The hypotheses raised should be further examined in qualitative settings, such as through analyses of interviews and focus groups with suicide attempters. Furthermore, increases in suicide rates after the publication of media items have been shown to peak within 1 week of publication. ^{Reference Stack and Leenars2} Future research should also investigate potential patterns of decay in protective media effects, that is, the time pattern of any suicide-protective effect of media reporting.

The present findings add a new perspective to internationally ongoing collaborations between suicide researchers and media professionals. The traditional print and television media, owing to their general lack of space or time, sometimes hesitate to report stories on individuals who adopted coping strategies other than suicidal behaviour in adverse circumstances. The possible role of media reports in preventing suicide may make it worthwhile for journalists of both traditional and online news platforms to follow media guidelines on the reporting of suicide.