Using Real-Time Suicide Monitoring Systems to Inform Policy and Practice

Many countries have now established, or are establishing, suicide monitoring systems (Baran et al., 2021). These systems typically use data from police reports and death certificates to identify suspected suicides, with key information about the deceased entered into a register soon after death. Information entered into the register may include the location of death and the deceased’s place of residence, the geocoordinates of these locations, the age and sex of the deceased, and information about the method of suicide. Because information is entered soon after death, these registers act as real-time or near real-time surveillance systems.

Real-time suicide registers resolve the problem of timeliness that has hampered the use of vital statistics systems for suicide monitoring. Vital statistics are usually based on an investigation by a coroner, medical examiner, or other authority. These investigations are often lengthy – in Australia, it takes between 12 and 18 months for a determination by a coroner – meaning these data are too old to be used for responding to increases in suicide rates, increases at specific locations, or increases by emerging suicide methods.

As countries establish real-time suicide registers, there is now the potential to use these data sets to create opportunities for suicide prevention that have not previously existed. Specifically, the availability of real-time data means these registers have the potential to be used to detect any increases in suicides and that information can then be used to inform a rapid response. For this goal to be realized, however, we also need to undertake analysis of the registers in real time using tools appropriate for the detection of rare events like suicide. In this editorial, we discuss these issues in detail, drawing on our own experience in Australia.

Can Registers Be Used to Determine if Suicide Trends Are Changing?

Many registers now produce reports of monthly suspected suicide counts that are published online. In Australia, monthly counts of total suicides in New South Wales (population 8.17 million) and Victoria (population 6.68 million) are published by the Australian Institute of Health and Welfare (2023). An example of these data is shown in Figure 1. The first point to make about the data is that they are contemporary: from January 2019 to June 2023 (in New South Wales) and to July 2023 (in Victoria). The second point to make it that it is very difficult to discern a trend in these data. There is considerable variation in both series and a peak in one month is often followed by a trough the next month. Presenting the data in this way does not shed light on whether suicides are increasing, decreasing, or remaining stable.

One way to add more meaning into the presentation of suicide counts is to take the statistical process control-perspective. These ideas were developed in the 1920s to monitor and control for defects in the manufacturing process and have since been applied to infectious disease epidemiology. The originator of the field, Walter Shewhart, identified two key processes that lead to manufacturing defects: special cause variation and common cause variation (Shewhart, 1930). Special cause variation refers to sudden changes that may only impact some people or be confined to a specific location. The reporting of a celebrity suicide in the media is a good example of special cause variation in suicide prevention (Niederkrotenthaler et al., 2020). Common cause variation refers to variation due to an abnormal system, for example, a small but persistent change that affects everyone in the system. In suicide prevention, this could refer to changes in the economy associated with the risk of suicide such as increasing unemployment (Thor & Hans, 2015).

Special cause variation can be detected using Shewhart charts, while common cause variation can be detected using exponentially weighted moving average (EWMA) charts. A Shewhart chart for New South Wales and Victoria is shown in Figure 2. The data have been detrended (using first differencing) and then plotted with their 95% and 99% confidence limits. In New South Wales, there is 1 month where suicides fall below the 95% limit and another month where suicides are above the same limit. In Victoria, there are 2 months where suicides are above the 95% limit (including the most recent month in the series) and 1 month where it is below. The months where the number of suicides is above the limit may signal that there has been a shock to the system with further investigation warranted, for instance, to identify which population groups may be most affected and to identify the specific cause.

An EWMA chart for the two states is shown in Figure 3. The graph shows four lines: the observed data (solid gray lines and the same as presented in Figure 1), the weighted trend line (solid black line), and the 95% and 99% upper limit lines (dashed lines). The black trend line that is of primary interest here and whether that line crosses the upper limit lines. In New South Wales, the trend line never exceeds the 95% limit line, although there is some weak evidence that the trend has been elevated since March/April 2022. In Victoria, the trend line exceeded the 95% limit in October 2022 and remained above this threshold until May 2023. This finding suggests that a broad, population-level risk factor for suicide may have influenced suicide rates during this period.

A final point to make about detecting changes in suicide rates in real time is that the development of graphs such as these is only part of the challenge. It is equally important to work with end users to refine the graphs so that the results are meaningful for them. This codesign work should extend to developing the language used to describe any increase, for instance, when describing results to Ministers, the media, or directly to the community. Using phrases such as evidence corresponding to values between the 95% and 99% limits and strong evidence if above the 99% limit may be helpful. Other terms could be used for values between other limits (e.g., between the 90% and 95% limits).

Can We Use Registers to Detect Suicide Clusters?

One type of suicidal behavior that can have far-reaching consequences is the occurrence of suicides in clusters. Suicide clusters are commonly defined as a group of suicides that occur closer together in space, time, or space and time than would be expected given the underlying population size. Examples of space-only clusters are clusters in specific settings such as schools, psychiatric units, or at locations such as bridges and cliffs. Conceptually, time-only clusters are similar to special cause variation, mentioned earlier (i.e., a sudden external change affecting some groups). Space–time clusters refer to an increase in suicides in a confined geographic area and over a period of time (e.g., in a community over several weeks or months).

The detection of suicide clusters is a complex task, and there are several reasons for this. One reason is that the statistical tool commonly used to detect suicide clusters, the scan statistic, is a sophisticated and complex method. As it does not appear in epidemiology textbooks, there is no easy pathway to understanding the theory underlying the technique. A second reason is that the software used to implement the scan statistic, SaTScan (Kulldorff, 1997), has a number of default settings. Their meaning is not always clear (e.g., the percentage of the population-at-risk setting), and these default values may not be appropriate for the detection of suicide clusters. A third reason is that small adjustments to the way data are entered into the software can result in different cluster locations being identified. The clearest example of how changing the data can change the results is in the choice of geographic level for analysis – the geographic area that contains both the suicide counts and the data on population size.

The complexity of the task could partially explain the slow uptake in the use of real-time data to detect suicide clusters. The exception to this is the work by Benson and colleagues, who have used the scan statistic as the basis for prospective monitoring of suicide clusters in Cork County, Ireland (Benson et al., 2022). They developed a prototype real-time cluster surveillance system that presented summary results for dissemination in a way that supports decision-making in suicide prevention.

We have undertaken similar work using data from the Victorian Suicide Register (Sutherland et al., 2018) to identify possible clusters and summarize the findings for end users. In the prototype system we have developed, we use a rolling 24-month window of data for cluster detection. When a new month of data is added to the data set, the oldest month drops out and the models are rerun. Models are fit at three different geographic levels as well as for the whole population and for young people only (25 years or younger). We have also developed a range of tools for presenting the data. We can plot the location of clusters and of individual cases on a map that users can interact with in a similar way to Google Maps. Further information can be added to the map (e.g., location of schools, health services) to give a full picture of the community and the resources available. We can also produce time series plots for each possible cluster showing the number of suicides at that location over time.

Once operational, this system of cluster detection is likely to be very useful for coroners, health departments, and other agencies that have a prevention mandate. First, rather than being reactive to community concern (as is currently the case), this system will be proactive because a new scan will be undertaken each month. If a cluster occurs, it is possible that this system will detect the cluster sooner than would otherwise happen, meaning services can then be deployed more quickly to prevent further harm. Second, this system will be immensely valuable to coroners and health departments when there is unfounded speculation or inaccurate media reporting about clusters that inadvertently causes harm. Being able to provide advice that there is no evidence of a cluster will enable them to make clear statements to the community to counter these narratives.

Can Registers Be Used to Detect Novel Suicide Methods?

The emergence of new suicide methods presents a challenge for suicide prevention, particularly if the new method is highly lethal, easily accessible, and acceptable to a large proportion of at-risk individuals (Gunnell et al., 2015). The availability of a new method with these characteristics also has the potential to increase the overall suicide rate, as happened in Taiwan and Hong Kong in the early 2000s with charcoal burning (Chang et al., 2014).

In this context, real-time suicide registers are a valuable resource for detecting new methods. The Victorian Suicide Register, like many similar registers around the world, includes detailed coded information extracted from police, autopsy, and toxicology reports as it becomes available. Taken together, this information can be used to identify whether new methods are emerging, for example, by plotting the incidence of different methods over time. If coronial staff observe new patterns in their caseload, then older cases with similar characteristics can be identified from the register to further gauge changes in the use of that method over time. The rarity of new methods means that tools like the Shewhart or EWMA charts are unlikely to be helpful, but other charts like cumulative sum (CUSUM) charts may be informative. Determining the best charts for identifying new methods is an open area of research.

Can Registers Be Used to Evaluate Interventions as They Are Implemented?

The discussion to date has focused on suicide monitoring and surveillance. A final way that registers may be useful is in the evaluation of interventions, particularly as they are implemented. We illustrate this with an example from our recent work.

The Victorian Government established a project in 2015 to remove 85 level crossings in Melbourne. Level crossings occur where the road (and footpaths) intersects with the rail track; thus, they provide access to the rail track and a means of suicide. By removing these crossings, the rail tracks are relocated under roads or on bridges that are inaccessible to the public. While the primary purpose of the removals is to reduce traffic congestion and improve travel time by train, by restricting access to the rail track, we thought it likely that an additional benefit would be a reduction in rail suicides (Clapperton et al., 2022). We therefore used real-time data from the Victorian Suicide Register to test the effectiveness of level-crossing removal on suicide rates, and we were able to do this as the level-crossing removal project is ongoing. We used a difference-in-difference design to compare the number of rail suicides before and after level-crossing removal at intervention sites with the number of suicides at control sites (matched sites where crossings were planned but had not yet been removed). We found a large reduction in the number of rail suicides at intervention sites but not control sites, suggesting this is an effective suicide prevention strategy.

The removal of level crossings will continue until at least 2030. However, the use of real-time data means that we have been able to provide good evidence of the effectiveness of this intervention from a suicide prevention perspective before the project is complete. It is possible that other interventions, particularly those related to restricting access to means, can be successfully evaluated in real time, thus providing decision-makers with evidence of effectiveness sooner than might otherwise happen.

Conclusion

Real-time suicide registers have the potential to provide timely information on current trends. This was starkly illustrated by the COVID-19 pandemic, where there was widespread concern that the lockdowns that began in 2020 would lead to increased suicide rates. Two studies using real-time data from 21 and 33 countries showed that this was not the case; rather, suicides either remained stable or declined in most countries during this period (Pirkis et al., 2021, 2022).

We think that the successful use of real-time data requires several key ingredients. The first key ingredient is reliable and timely data. This requires ongoing and appropriate investments by government. The second ingredient is a strong relationship between those who have responsibility for data custodianship and those who analyze and interpret the data. Real-time data are often highly sensitive, especially data giving the geocoordinates of the address of the person who has died, and this information can only be shared with researchers when there is strong trust, governance, and procedures for the secure storage of data. Finally, analysis of real-time data – especially for the detection of suicide clusters – requires substantial computing resources. Our work required access to secure servers with a large number of processors and significant amounts of memory. These costs should be factored into any research project. Good work has been done to establish real-time registers in many countries. Now we need to realize the potential of these registers to answer questions that can have a real and immediate impact in terms of suicide prevention.