According to the World Health Organization (WHO) projections, by 2030, there will be 1,007,000 deaths by suicide, making suicide the 15th leading cause of death globally and accounting for 1.4% of all deaths [1]. Despite the common idea that suicide is more prevalent in high-income countries, about 75% of suicides worldwide occur in low- and middle-income countries. In general, suicide rates are lower among people aged <15 years and >70 years [2].

With a rate of 10 cases every day, suicide is the leading cause of unnatural death in Spain, producing more than twice as many deaths than traffic accidents, 7 times more deaths than workplace accidents, and 70 times more deaths than domestic violence. It is also the leading cause of death among men aged 20 to 24 years [3].

The incidence of suicide in a society depends on a range of factors, of which clinical depression is a particularly common cause [4]. Substance abuse, severe physical disease, and disability are also recognized causes of suicide. Countries in Eastern Europe and East Asia have the highest suicide rate in the world. The region with the lowest suicide rate is Latin America. Gender differences also play a significant role: Among all age groups in most parts of the world, females tend to show higher rates of reported nonfatal suicidal behavior and males have a much higher rate of completed suicides.

The study has two objectives: (1) It evaluates the correlation between suicidal rates released by the INE and internet search trends in Spain reported by Google Trends for 57 suicide-related terms representing major known risks of suicide; these terms have already been tested in previous scientific studies systematized by Mok et al [7] (topic “Suicide-related internet search trends can provide an indicator of suicide risk in a population”). (2) It examines the differential association between male and female suicide rates published by the INE and internet searches related to the aforementioned 57 terms. The study included data from 2004 to 2013, as this was the maximum period for which relevant data were available from the INE and Google Trends.

Google Trends provides a time-series index of the volume of queries users entered into Google in a given geographic area. Wikipedia explains it as follows [38]:

Although Google Trends does not show the absolute number of searches, it calculates a query share for a search term. This means that Google calculates the number of searches for a given term as a proportion of the total number of searches in each location at a given time. These calculations are then normalized to a Google Trends Relative Search Volume (RSV) index between 0 and 100, where an RSV index of 100 designates the date when there was the highest amount of search activity for that given term. Thus, a search index of 40 equates to 40% of the most intense search activity in the selected country at a given period.

Thus, the RSV index is a way to normalize (from 0 to 100) the query share that is the total volume of queries of the search term in question within a particular geographic region divided by the total number of searches in that region for the period under review. The maximum percentage of consultation in the specified time period is normalized to 100, and the other measures for that period of time are calculated relative to this value.

Google Trends also allows for the comparison of the relative volumes of blocks of searches for up to 5 terms or phrases. In this case, the RSV of other terms that did not reach the peak of 100 is normalized to the 100 value of the term with the highest search volume of the 5 terms of phrases in the block. However, in our work, terms were consulted one by one.

It is interesting to point out that although, according to Google Scholar, more than 10,000 scientific papers used or mentioned Google Trends service, we did not find any mathematical formulation of how the RSV value was calculated or operationalized by Google Trends. Therefore, we proposed a tentative mathematical formulation of how this value is calculated (Figure 1).

In short, Google Trends calculates the number of searches as percentages (formula 2 of Figure 1) based on the total searches in a month (formula 1 of Figure 1), normalizes the series allocated to the highest value (ie, the value of 100), and scales all other values accordingly (formula 3 of Figure 1).

Data were analyzed using the statistical package IBM SPSS Statistics, version 22 (IBM Corporation). The Pearson correlation coefficient was used to assess a possible monthly correlation between suicidal rates and Google Trends RSV data for the search terms that we defined. Next, we performed a multiple linear regression analysis to propose an explanatory-predictive model of the variance of the suicide rates variable.

With regard to the first objective of our study, the values for correlation between suicide-related terms and suicide rates for Spain from Google Trends data are shown in Table 3 after calculating the Pearson correlation coefficient. We centered moderate or superior results of correlation values according to Evans’ study [37], as detailed in Table 4, with a significance value P<.05 (in italics). Since the study terms are Spanish translations of ones already studied in English, Japanese, and Mandarin in the mentioned studies, we devised a Reference column indicating the previous study and a Correlation column to indicate the presence of a correlation according to the original study, with values yes, insufficient, or no.

A linear regression analysis (steps forward) was performed; predictors included all variables (search terms) that demonstrated a significant correlation with previous suicide rates collected by the INE and had an r>0.2. These are the terms in Table 3 that have at least one P value with a significant correlation in the men, women, or total columns. Table 4 presents the explanatory-predictive model.

Overall, the model predicts a significant percentage of variance (adjusted r²=0.387) of the suicide variable. The term “unemployment” translated as “paro” in Spanish has a high beta value and a positive sign, whereas the term “unemployment” translated as “desempleo” has a lower and negative value. This may seem contradictory because both terms are, a priori, synonyms. However, searches of the term “desempleo” could be carried out, in greater proportion, by people seeking information related to the official term “unemployment benefits and aid” offered by the Spanish Government, while the term “paro,” which is used more colloquially, may be associated with searches carried out by people who are suffering due to “unemployment.” This may explain why searches for that term are positively associated with the incidence of suicide committed in Spain between 2004 and 2013.

Regarding the beta value for “headache,” which has a high and negative correlation with the variable “incidence of suicide,” it could be argued that people who search for headache (a condition that can be associated with a wide variety of medical conditions) do so with an intent of self-care, which is contrary to the intention of committing suicide.

With regard to the second objective of our work, we found correlations between the terms of study and suicide rates between women and men (Table 5). To describe the strength of the correlation between our variables, we have used the interpretation by Evans [37]; as it can be seen in absolute terms, there is an important difference between the correlation of men and women.

The significant difference between male and female correlations can be explained by women’s use the internet for searching for health and lifestyle information. In contrast, men tend to focus on information about investment, purchase, and personal interests [40]. Moreover, this would be consistent with the idea that women have higher emotional intelligence and more communication skills than men [41].

It is not clear whether the information found on the internet contributes to the promotion of suicide and inspires suicidal thoughts or reduces the risk of suicidal behavior. The causal relationship between suicide and the use of the internet to search for topics related to self-harm or suicide is difficult to prove; however, the results of our study suggest a significant correlation for a number of the search terms that we have studied. This is consistent with previous studies that have been outlined throughout this paper that mostly state that there is indeed some association between certain searches and social phenomena in the economics, health, and other sectors.

Suicide rates in Spain for the 2004-2013 period were examined for their association with search volume on Google for 41 suicide-related searches already tested in scientific studies in other countries and languages. For the general population, suicide rates in Spain were positively associated with the search volume (r>0.2) for 7 terms and negatively associated with the search volume for 1 of the 41 terms. Our interpretation of the results is that they corroborate the hypothesis that certain searches on Google may serve as an indicator of a country’s suicide rate, perhaps even of its social well-being.

The negative correlations that we would call “protective” (for searches aimed at finding a solution to the problem) are interesting; the 41 searches in the original studies were supposed to be “risk related” in relation to suicide incidence, but one of them, Stock Market (r=–0.231; P=.006), correlated negatively along with some others that may have a "protective" significance (Social welfare, Religious belief, etc). As a preliminary explanation, we believe that this is due to the social and cultural translocation of the search terms, eg, in Spain unlike Taiwan—where the previous study for the search Stock Market was performed—only wealthy people are concerned about the topic.

In the case of Drunkenness (r=0.211; P=.01) versus the negative correlation for Alcohol (r=–0.113; P=.11), the latter could be interpreted as a protective search to find a solution to the problem, while the former may be used in a leisurely way (ie, without a problematic consciousness). This led us to an important point: Google Trends includes data from subjects with suicidal behavior searching in Google in addition to searches by other people concerned about the issue. We called these two perspectives as “first person” and “third person.” We then realized that Google Trends data includes “first person” searches from subjects with suicidal ideation and “third person” searches from their relatives, social surroundings, and institutions; therefore, it is crucial to try to segregate one from the other for future studies. Perhaps, this can be done with the help of linguistics differentiating the denotative aspects of words from their connotative aspects. 

The term that correlates more strongly with the overall rate of suicide is Allergy (r>0.5 and P<.001), which is consistent with other studies linking depression and allergy [42,43].

However, the overlap between the terms that correlate in our study and those that correlate in other studies is only about half. This could be due to cultural differences between the regions of the study subjects. It could also be due to semantic changes lost (or gained) in translation: Although the studies that we used to build our research were written in English, the original language of the study was Japanese or Mandarin Chinese in several cases, which resulted in two nested translations in this study.

Comparing our results for Spain with some of the results from a study by Yang et al [33] for Taiwan, there are differences that we can charge to cultural or sociological variances between the subjects of each study. Although in the Taiwanese study, the term Divorce correlates with suicide rates, in our study, it does not correlate with suicide rates. Unemployed does not correlate with suicide rates, according to Yang et al [33], but it does so in Spain, where unemployment is approximately 4 times higher than it is in Taiwan. We interpret these as indicators of issues that have a different significance and social context in these two regions. It should be noticed that while the search term Divorce (r=0.064; P=.24) does not correlate with suicide rates in Spain, Relationship Breakup (r=0.268; P=.002) does and this is the second strongest correlation among women. Further consideration on gender differences are made in the following section.

Another reason for these disparities might be deficiencies in the study methodology since using Google Trends as a diagnostic indicator of a society’s well-being is still fairly new.

Other interesting evidence our study demonstrates is that well-known risk factors (eg, depression) and explicit searches (eg, suicide) are not correlated with suicide rates; this could be interpreted as follows: the better the knowledge of the risk situation, the less likely it is that this risk of suicide will materialize.

Although there is no gender segregation in Google Trends RSV data, as Spanish suicide rates from the INE are segregated by gender, we were able to find differing gender-based correlations: We have found 5 terms that correlate for suicide rates among men and 12 terms in the case of women. Search terms that correlate with suicide rates of women are consistent with previous studies, showing that the incidence of depression is higher in women than in men [44].

In short, we have obtained more than twice the correlations between suicide rates for women compared with those obtained for men. We understand that this is due to the fact that patterns of internet usage among women are more oriented toward searches on health or lifestyle [40], which is also very much in line with the idea that women have more emotional intelligence than men [41]. This would explain why their suicide rate is much lower than that of men in Spain.

Owing to the limitations of evidence, we cannot actually predict increases of suicidal mortality using web search data. Rather, we undertook a preliminary investigation using the entire available dataset to establish a statistical association between search term usage and actual suicides in Spain. Further studies should compute time-lagged correlations between Google searches and suicides to help prevent suicide-related deaths.

The practical implication of our results are as follows: It is desirable that competent authorities establish agreements with Google to facilitate suicide prevention by monitoring searches in Google for any of the terms that have been shown to correlate with suicide statistics and other terms that are proven to be significant in future studies.

We also hope that our research will help design and maintain websites that provide better education for suicide prevention, focusing on the treatment of depression and management of labor or emotional problems, as these fields show greater explanatory-predictive value in the incidence of suicide according to our regression model.

An interesting avenue for future research on suicide-related searches is obtaining data from large social networks such as Facebook or Twitter, rather than just metasearch engines like Google.

In addition, the results of this study suggest the feasibility of using the Google search volume to predict other social risk behaviors such as traffic accidents, domestic violence, and bullying, and for the epidemiological monitoring of the evolution of emotional disorders in society. In general, we believe that tracking the search volumes of certain terms (eg, ones related to suicide) represents satisfaction in and well-being of a society. Hence, there may even be an application to the field of politics.

We want to point out some interesting facts that we have come across in our research and that we consider to be significant in correctly interpreting the world of big data and metasearch engines.

First, as mentioned earlier, it is interesting to note that despite the fact that more than 10,000 scientific papers used or mentioned the Google Trends service, according to Google Scholar, we did not find any mathematical formulation of how Google Trends operationalizes the values that it returns, which is the reason why we have developed it ourselves (Figure 1).

Another fact that seems significant is the case of GISI [20] or the Google Price Index (GPI), which are Google initiatives from 2010, that disappeared despite evidence of good results for forecasting phenomena. According to comments on web-based forums, Google chief economist, Dr Hal Varian, said that GPI was never intended be a project or public source of data; it was simply an internal Google project made visible by the press [45].

In any case, the opportunities and risks of using information from internet metasearches are yet to be determined; with this work, we hope to have contributed some clarity to this field of study.