We obtained the number of completed suicide events in South Korea from January 1 2008 to December 31 2010. The data were thoroughly examined and verified by the Korea National Statistical Office (KNSO, http://kostat.go.kr/portal/english). Data for those years were considered because contemporaneous demographic and social media data were available. Data were extracted from death records defined as suicides according to the International Classification of Diseases-10 (ICD-10) codes X60–X84, which include suicides from all causes, including intentional self-poisoning and self-harm. [27] Monthly five year averages of suicide number from January 2003 to December 2007 also were computed so as to allow adjustment for seasonal variation. [28].

Daumsoft, one of the leading social media analysis and consulting firms in Korea, provided the social media data for the current study. The data were drawn from weblog posts in Naver blog (http://section.blog.naver.com), a weblog service offered by the biggest portal site in South Korea. A set of filtering operations was applied to exclude advertisements and other noisy texts for weblog posts written during the period between January 1, 2008 and December 31, 2010. The weblog service processed 153,107,350 posts on 5,093,832 registered weblogs during the above 3-year period.

To effectively simplify and quantify the enormous amount of social media data, we defined two measures: ‘suicide weblog count’ and ‘dysphoria weblog count’. The suicide weblog count was defined as the daily document frequency mentioning the Korean word jasal ‘suicide’ at least once. Similarly the ‘dysphoria weblog count’ was defined as the daily document frequency mentioning the Korean word himdeulda, which conveys the negative meanings ‘be tired’, ‘be painful’, or ‘be exhausted’ at least once. The word himdeulda was specifically chosen since it has co-occurred most frequently with jasal among the words expressing subjective psychological status of a writer or a speaker. The actual measures were obtained using SOCIALmetrics™, a social media analysis system offered by Daumsoft (http://www.daumsoft.com/eng/index.html). This system provides deep level keyword analysis and opinion mining for social media texts and other web documents.

Economic and meteorological variables identified in previous studies of suicide were also considered. The economic data, [29], [30] including consumer price index, unemployment rate, and stock index valuations (Korea Composite Stock Price Index, KOSPI), were extracted from the KNSO. The meteorological data (sunlight hours and temperature) [9], [27] were obtained from the Korea Meteorological Administration (KMA, http://web.kma.go.kr/eng). Measurements from the observation station in Seoul were chosen as representative data.

Prior to model construction, we divided the data into a 2-year training set (2008–2009) for identifying significant predictor variables and constructing a prediction model, and a 1-year validation set (2010) for evaluating the model. Most variables, including suicide counts and social media data, were summed in discrete 3-day epochs. There were 243 3-day epochs in the training set and 121 in the validation set. All computations were performed using these 3-day binned numbers. The end-of-week and holiday closing values of the KOSPI stock index were carried forward to the next active trading day. The most recent monthly data for the consumer price index and the unemployment rate were used each day, and these data were averaged for each 3-day epoch (see Text S1, Table A). For the meteorological variables we recorded averages of the 3-day data for temperature and sunlight hours in each epoch. For the purpose of illustration in Figure 1c, the dysphoria weblog count was divided by 5.

In order to control for the influence of celebrity suicides, we noted the periods following these events. We defined celebrity suicide as a suicide exposed during more than two weeks in news programs of the three major national television networks (KBS, MBC and SBS). Six celebrity suicides met this definition during the 3 years of this study: actor (Jae-Hwan Ahn, 9/8/2008), actress (Jin-Sil Choi, 10/2/2008), actress (Ja-Yeon Jang, 3/7/2009), president (Moo-Hyun Roh, 5/23/2009), actor (Jin-Young Choi, 3/29/2010) and actor (Yong-Ha Park, 6/30/2010). Based on the study of Phillips, [31] the affected period was defined as a month (30 days) after the first report of the celebrity suicide. Prediction time points (3-day epochs) within or partly within this 30-day window were coded 1, while all others were coded 0 on the celebrity variable.

Our research analyzes existing data and documents that are publicly available in a manner that does not allow individual subjects to be identified, therefore ethics approval was deemed unnecessary.

We identified significant variables by testing them individually in univariate linear regression analyses using the training set. The dependent variable (suicide numbers) was logarithm-transformed to satisfy a normal distribution assumption in the regression analysis. To avoid redundancy among the predictor variables, we chose the one variable with significance (P<0.05) and the highest adjusted R-squared value in each set of candidate variables that examined multiple time periods (t-1) to (t-5) (Text S1, Table A). In each case, this proved to be the most recent time period (t-1).

The multivariate regression model was constructed using these selected variables identified in the training set. Variables that became nonsignificant (P>0.05) were removed stepwise, so the final model included only significant variables. For all 3 years, we predicted suicide numbers by 3-day epoch using the ‘predict’ function with ‘prediction interval level’ set at 0.85 in the R software, which means that the observed number is expected to fall within the upper and lower boundaries of the prediction interval with 85% probability. Then, we compared the predicted numbers with the observed numbers. We regarded predictions as correct if the observed numbers fell within the prediction interval, and we defined the prediction accuracy as the ratio of correct predictions to total predictions. All statistical analyses including variable selection and model construction were performed using the R 2.9.1 public statistics software (http://www.r-project.org).

Over the 3 years of our study, national suicide numbers trended upwards (daily average: 105.7 in 2008, 126.9 in 2009, 127.7 in 2010), reflecting a general trend over 20 years preceding 2010 (Text S1, Figure A). Total suicides in 2010 were approximately 4.3 times the number in 1991. Daily suicide numbers varied substantially within each year, with a 5.4-fold range from a daily low of 16 to a daily high of 87 in 2008, a 4.6-fold range (18–83) in 2009 and a 3.5-fold range (20–69) in 2010. Spiking of completed suicides was apparent in proximity to celebrity suicides (Figure 1a, b).

Social media data also displayed marked variation. The dysphoria weblog count consistently exceeded the suicide weblog count. Daily suicide weblog counts ranged from a low of 104 to a high of 2,444 in 2008, 147 to 4,516 in 2009 and 186 to 3,588 in 2010; the daily dysphoria weblog counts ranged from 1,214 to 3,716 in 2008, 1,749 to 5,402 in 2009 and 2,400 to 7,668 in 2010. The suicide variable was more volatile or fluctuating than the dysphoria variable as evidenced by the coefficients of variation, defined as the ratio of the standard deviation to the mean. These were higher in daily suicide weblog counts than in dysphoria weblog counts (0.71 vs. 0.13 in 2008, 0.85 vs. 0.13 in 2009, and 0.53 vs. 0.16 in 2010). Spiking of suicide weblog counts was apparent in proximity to celebrity suicides, whereas no such short term spiking of the dysphoria weblog count was seen (Figure 1c).

Based on prior studies as reviewed above, we examined several candidate predictors of suicide numbers (Text S1, Table A). Univariate linear regression analyses confirmed the significance of all categories of variables except for sunlight duration. With the exception of the consumer price index, all the nominally significant variables displayed in Text S1 (Table A) remained highly significant (P<0.0002 or lower) after Bonferroni correction for multiple testing. Among the economic variables, the most significant were the unemployment rate and the stock market index. Rising unemployment and falling stock prices were each associated with higher suicide rates. The temperature data also showed a significant association with suicide data: increasing temperature was associated with higher suicide rates.

Among all the variables, ‘suicide (t-1)’, indicating the suicide number over the previous 3-day epoch, showed the most significant association, indicative of short term trending. This pattern is also seen extending back to 5 epochs (effectively 2 weeks). The significance of ‘suicide_5yr_avg’ suggests that there might be a monthly pattern concordant with a seasonal effect on suicides. [28] The ‘celebrity’ variable was confirmed as highly significant (Text S1, Table A and Figure 1a, b).

The social media variables also proved to be strongly associated with suicide numbers (Text S1, Table A). Among the social media variables, the most recent suicide weblog count had the strongest association with completed suicide (P = 3.32×10^–15). This variable displayed marked short term trends, with prominent spiking related to celebrity suicides (Figure 1c). By contrast, the dysphoria weblog count showed longer term secular trends and was not so obviously responsive to celebrity suicides in the short term (Figure 1c). Five variables showed a recency effect in the form of a declining strength of association extending back 2 weeks (t-5). These were suicide number, stock index, temperature, and the two social media variables (Text S1, Table A). The most significant variables from each set of candidate predictors are displayed in Table 1. These variables, that were the most significant in the univariate analyses, were used to develop the prediction model.

Using these selected variables, we constructed a prediction model for nation-wide suicide numbers (Table 2). Two classical economic variables, the consumer price index and the unemployment rate, were excluded in the final multivariate model as they became nonsignificant (P = 0.13 and 0.44, respectively). Within the final multivariate model, the ‘big three’ variables were the most recent suicide event rate, then the most recent dysphoria weblog count, then the most recent KOSPI stock market data. The celebrity variable and the most recent suicide weblog count were less powerful (Table 2). The final model showed an increased adjusted R-squared value (0.66) compared with the univariate models (0.03–0.53) and when compared with the multivariate model excluding social media data (0.59).

When fitted to the training data set, the model showed a prediction accuracy of 0.88 (the observed numbers for 213 out of 243 3-day epochs ranged within the prediction interval), and in the validation set the prediction accuracy was 0.79 (96 out of 121 3-day epochs) (Figure 1a). The correlations of observed with predicted suicide numbers were 0.82 in the training set and 0.74 in the validation set (Figure 1b). Because of the obvious short term effect of celebrity suicides (Figure 1a–c), we conducted a sensitivity analysis of the model’s performance outside the celebrity suicide periods. No significant difference in performance of the model was seen in this secondary analysis: the prediction accuracy in the training set was 0.87 compared to the original accuracy of 0.88, and in the validation set the prediction accuracy was 0.83 compared with the original accuracy of 0.79; correlations of observed with predicted suicide numbers outside the celebrity suicide periods were 0.79 for the training data set and 0.72 for the validation data set. The corresponding original correlations were 0.82 and 0.74.