Wohlford [23] measured the time perspective by examining responses to the Thematic Apperception Test (TAT). In concrete terms, when measuring the time perspective using this test, the participant writes a story that features past, present, or future viewpoints, in accordance with the TAT cards provided. Then, based on the outcome of that story, a score is generated that reflects the participant’s degree of preference for the past versus the future [24]. Furthermore, the Zimbardo Time Perspective Inventory (ZTPI), a questionnaire method to improve subjectivity of TAT, was developed to measure time perspective [19]. ZTPI measures the individual’s time perspective (future, present, and past) through a questionnaire. In summary, the methods of measuring the time perspective center on finding one’s preference vis-à-vis time orientation among the past, present, and future through the questionnaire method.

Similarly, Preis et al [14] quantified the future orientation index of a country by utilizing search query data. To measure the future orientation index, Preis et al [14] calculated the ratio of the future year (eg, ‘‘2010” in 2009) phrase of the Google search queries index to that of the past year (eg, ‘’2008’’ in 2009). We measured the past orientation of state residents each year and calculated the ratio of the past year (eg, ‘‘2008’’ in 2009) phrase of the Google search queries index to that of the future year (eg, ‘‘2010’’ in 2009) by US state. For each year, “past year” and “future year” phrases were changed according to the base year. For example, past orientation in 2010 is the ratio of the “2009” phrase of the Google search queries index to the “2011” phrase of the Google search queries index. For another example, past orientation in 2006 is the ratio of the “2005” phrase of the Google search queries index to the “2007” phrase of the Google search queries index. In this way, we calculated the past orientation of residents of the 50 US states and the District of Columbia between 2004 and 2012. While Preis et al [14] conducted an international comparison of 45 countries for 3 years, we performed an intranational analysis of 50 US states and the District of Columbia for 9 years. As a result, although we controlled variables such as gross state product (GSP), other factors that can affect suicide rate (eg, cultural difference) also needed to be controlled. In addition, if we analyze international countries, aggregating different sources of suicide rate data in different countries is unavoidable; however, this can cause a data quality problem. Based on this criterion, we were able to measure past orientation index by using Google search query data and comparing yearly search volumes in past and future years among residents of the 50 US states and the District of Columbia. In addition, although Preis et al [14] excluded countries with population less than 5 million, because of the possible inaccessibility of search query data due to the low number of search queries, search query data in US states are sufficiently accessible without any exception. Also, in the US, Internet penetration and Google market share are sufficiently high to utilize search query data. According to internetlivestats, penetration rates of Internet in the US in 2004 and 2012 are 64.76% and 81.03%, respectively.

Returning to the model, Equation (1) is the formula we used to quantify past orientation. The method for calculating the Google search query index for each numerator and denominator is identical to that with regard to Google Trends.

Past orientation _it = 100 × {[Number of Google search – queries for “past year”] _it /[Number of Google search – queries for “past year”]_max}/{[Number of Google search – queries for “future year”] _it /[Number of Google search – queries for “future year”]_max} (1)

The numerator in Equation (1), that is, is the index of Google search queries in the past year for state i during the year t. Specifically, [Number of Google search queries for “Past year”]_max is the “past year” search volume that is the largest “past year” search volume among the 50 US states and the District of Columbia in the year t. We then calculated the standardized relative proportion of search volume for the “past year” phrase by state i.

Furthermore, the denominator is a Google search query index for “future year” for state i during the year t. The rest of the calculation is identical to that of the previous denominator for “past year.” In more concrete terms, we utilized Google search query data from Google Trends, which provides Google search query data over time and by region, such as by country or state. Finally, the measured past orientation was found to vary from 0.775 to 1.517 by state during the analysis period from 2004 to 2012. Figure 1 shows state-specific average past orientation and average suicide rates.

In Figure 1, Montana, Maine, and Oregon showed a high past orientation value. By contrast, Maryland, California, and Georgia showed a relatively low past orientation value. We examined the past orientation differences among states to determine how they may affect state-specific suicide rates.

Suicide rates for each of the 50 US states and the District of Columbia for the 2004-2012 period were obtained from the Centers for Disease Control and Prevention’s deaths data. The suicide rate unit is the number of suicides per 100,000 population. These data are originally recorded on death certificates and filed on states registration offices, and the suicide statistic is processed by Vital Statistics Cooperative Program of Centers for Disease Control and Prevention [25].

There was some variation among the suicide rates of the US states. Figure 1 shows state-specific average suicide rates. Wyoming, Montana, Nevada, Oregon, and Maine recorded higher average suicide rates; in particular, Montana, Maine, and Oregon had high past orientation values. By contrast, the average recorded suicide rates of Massachusetts, New York, and California were relatively low.

In this study, we used independent variables that were mainly used in previous studies. These variables also served as control variables in pinpointing the determinants of suicide in the US states. Detailed descriptions of the variables are presented in Table 1.

Figure 2 shows the monthly time variance in Google search queries index values of the past year and future year. In Figure 2, the Google search query index values of the past year and future year showed an interesting pattern, depending on the season. The Google search query index for any given past year was highest at the beginning of the year, and it decreased gradually as the year passed. A similar phenomenon was observed in another large-scale big-data study [40]. By contrast, the Google search query index value for the future year showed a gradual increase from the beginning of the year. It then reached its peak when the current year changed to the future year (eg, future search query [“2013”] on December 31, 2012). Additionally, while it can be seen that people gradually change their time perspective focus from the past year to the succeeding year, past-oriented people find it difficult to keep up with a future or present time perspective; indeed, it is not easy for them to depart from the past to which they cling.

Following many previous studies about suicide rate [28,37,41], we utilized a regression model to investigate the relationship between past orientation and suicide rate. In addition, for using the linear model, we already completed the linearity test through scattering variables. We then calculated the AIC, BIC, and R ² for all possible variable combinations for testing the goodness-of-fit and explanatory power of models and variables. Through this process, we determined whether the past orientation variable increases the goodness-of-fit. As such, we were able to investigate whether past orientation is an important variable of the regression model. The following formula is the multiple OLS regression equation that we used:

Suicide rate _it = α _t + β₁ (unemployment rate _it ) + β₂ (Gini coefficient _it ) + β₃ (GSP _it ) + β₄ (population growth rate _it ) + β₅ (past orientation _it ) + ∈ _i (2)

where i represents a state, t the year, and ϵ _i the error term.

We obtained four regression models by combining independent variables. In Model 1, we used only the unemployment rate and the Gini coefficient. In Model 2, we added past orientation to Model 1. We also investigated changes to the AIC and BIC of Models 1 and 2. Through this process, we were able to examine the degree to which the model vis-à-vis past orientation and the suicide rate improved.

The independent variables in Model 3 consist of the unemployment rate, the Gini coefficient, GSP, and population growth rates. Finally, Model 4 added the past orientation variable to Model 3; we also investigated changes to the AIC and BIC in Models 3 and 4.

Table 4 shows our regression results by model. For both Models 2 and 4, we see that past orientation had a positive relationship with the suicide rate whenever we included the past orientation variable (P≤.001). Based on Model 4, when a past orientation value of 1 increases, the suicide rate increases to 8.5 people per 100,000 population (P≤.001). These results are consistent with our assumption that past orientation has a positive relationship with the suicide rate. These results indicate that past orientation is often related to life regrets [21] and that it can be a suicide risk factor. In addition, we looked at improvement in the model’s goodness-of-fit with respect to past orientation, and found that in Model 1, the AIC and BIC are 2513 and 2525, respectively, whereas in Model 2, these values are 2468 and 2485, respectively. The AIC and BIC in Model 2 are smaller than those in Model 1, indicating that Model 2 has better goodness-of-fit than Model 1. The AIC and BIC in Model 4 are 2355 and 2379, respectively, whereas these in Model 3 are 2385 and 2405, respectively. As is the case for Models 1 and 2, the AIC and BIC in Model 4 are lower than those in Model 3. In other words, given our AIC and BIC results, it can be said that the use of the past orientation variable improves the goodness-of-fit.

Next, we find that the unemployment rate has a statistically significant positive relationship with the suicide rate in Models 3 and 4 (P≤.001 in Models 3 and 4). Based on Model 4, when the unemployment rate increases to 1%, the suicide rate increases by 0.265 people per 100,000 population (P≤.001).

In addition, the Gini coefficient has a statistically significant and positive relationship with the suicide rate (P≤.001). Based on Model 4, when the Gini coefficient increases to 1, the suicide rate in a state would increase by about 28.5 people per 100,000 population (P≤.001).

The GSP variable in Model 3 was found to have a statistically significant and negative relationship with the suicide rate (P≤.001). A high GSP state tends to have a statistically significantly lower suicide rate. Based on Model 4, when the GSP increases by 1 million dollars, the suicide rate decreases by 4.122 people per 100,000 population (P≤.001).

Next, the population growth rate variable was found to have a statistically significant and positive relationship with the suicide rate (P≤.001). Based on Model 4, when the population growth rate increases by 1%, the suicide rate increases by 1.14 people per 100,000 population (P≤.001).

Furthermore, we verified the goodness-of-fit and explanation power for all possible variable combinations. Table 5 shows the goodness-of-fit of regression results. Model 4 in Table 4, including past orientation and all independent variables, has the smallest AIC and BIC, and the largest R ² and adjusted R ². In addition, R ² of the model, which only has past orientation, is about .079. Its explanation power ranking is third between the independent variables and it is almost similar to the second ranking explanation power. Thus, past orientation also can be a significant factor in a regression model of suicide rate. Lastly, we verified our regression model again through a stepwise regression test with 1% significance level. To conclude, the result is same as that of the goodness-of-fit test, and hence, Model 4 is the best model, consistently.

Finally, we mapped the average past orientation and the average suicide rate by state (Figure 3). Although there are many factors that affect a state’s suicide rate, we plot between past orientation and suicide rate to focus on past orientation. Notably, the past orientation and suicide rate values vary widely among the US states. As seen in our regression results, there is a tendency where the higher the past orientation of the residents of the state, the higher the state’s suicide rate. In particular, many states with high suicide rates (eg, Oregon, Colorado, and New Mexico) have high past orientation; the past orientation of Georgia, California, and Maryland is low, and their suicide rates are also relatively low. As a result, past orientation—including attachment to the past—can be seen as a suicidal risk factor, and it is found to have a positive correlation with the suicide rate.

We investigated the potential impact of past orientation on suicide rates; we measured past orientation through the use Google search query data. We also found that suicide rates vary widely by state. We built a yearly panel dataset by considering the categorized control variables for the US states in the 2004-2012 period. We were then able to reveal the relationship between past orientation and the suicide rate through multiple OLS regression: past orientation was found to have a positive relationship with suicide rate in a statistically significant manner (P≤.001). In addition, through AIC and BIC analyses, past orientation was confirmed as being an important variable of suicide rate in the US states. Ultimately, we were able to pinpoint the relationship between suicide rate and its risk factors in the US states.

We have made three salient contributions to the suicide literature. First, we were able to empirically identify the relationship between past orientation and suicide rate. At the individual level, the risk of suicide tends to increase when one faces a divorce or the death of a loved one, is dismissed from work, or experiences health problems, inter alia. These situations are worse than the aforementioned situations (eg, economic status, job stability, and health status) and are specific to a situation where one has lost a relationship with the people around him or her. When one falls into such a situation, he/she tends to focus on the past and may fall into obsession. These also can be one of reasons why past orientation has a positive relationship with the suicide rates. Second, we were able to measure the past orientation of the residents of US states by applying big data-driven Google search query to the phenomenon of suicide. Finally, we were able to verify clearly that the unemployment rate, Gini coefficient, GSP, and population growth rate are the determinants of the suicide rate in the United States.

More specifically, the unemployment rate finding accords with our expectation that it affects the suicide rate positively. This is consistent with the results of previous studies—such as those of Yang [28] and Neumayer [29], who examined national suicide rate determinants. In particular, vulnerable social groups face unemployment and life hardships whenever the unemployment rate is high; therefore, the unemployment rate has a positive relationship with the suicide rate.

In addition, Gini coefficient was found to have a statistically significant and positive relationship with the suicide rate. This result is consistent with the findings of Gunnell et al [27] in England and Wales. This can be interpreted as follows: deepening wealth inequality has a positive association with the high suicide rate. As a result, not only the GSP but also wealth inequality is an important factor of suicide rate.

By contrast, GSP was found to have a negative relationship vis-à-vis the suicide rate. There are strong links between GSP and suicide rate. While income or GDP variables have been frequently considered in many previous studies [29-36], the results thereof have not been consistent. Our results align with those of previous studies that found GSP to have a negative correlation with the suicide rate [29,33-36]. However, they are inconsistent with some studies that found income or economic boom to have a positive correlation with the suicide rate [30-32]. Although it is possible to explain modernization factor as the reason for income or economic factors affecting the suicide rate in these studies [30-32], our study investigated an already sufficiently modernized region and period, and hence, the results may differ.

Finally, the result of population growth rate can be interpreted as follows: states with a high population growth rate can be more changeable and unstable because immigrants and temporal job opportunities are critical reason of population growth [39]. This instability could contribute to a high suicide rate.

Although this study considered many aspects, it nonetheless has some limitations. First, Google search queries data are accessible only from 2004. In addition, the initial stage of Google search queries data, such as data in 2004, is of relatively low reliability because of relatively low Internet users (penetration rate of Internet in the United States is about 64.76% in 2004, but 81.03% in 2012).

In addition, although Google trends provide data only when they have sufficient search query data, a low sample error may occur because of the relatively small population in some states such as Wyoming or Vermont. Finally, Google search query data reflect only the views of people who can access the Internet and Google; for this reason, we cannot reflect on people with no access to the Internet.

Future studies need to investigate the causal relationship between past orientation and suicide rate. This can boost our results and bridge the gap of interventions directed at influencing behavior and attitude. Furthermore, next studies will be conducted on the development of tools by which Internet users can request medical help; on the basis of these findings, such tools would leverage past orientation. It can also be valuable to examine how the government can effectively intervene in suicide risk situations using big-data analysis. We will also try to investigate the relationship between suicide or disorders and other specific attitudes by undertaking big-data analysis made possible by the provision of search query or social network data.