Journal of Transport & Health 20 (2021) 100976
Available online 17 December 2020
2214-1405/© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
The situational conditions of suicide in transit environments: An
analysis using CCTV footage
Vania Ceccato
a,*, Douglas J. Wiebe
b, Katerina Vrotsou
c, Ullakarin Nyberg
d,
Arne Grundberg
daKTH Royal Institute of Technology, Sweden bUniversity of Pennsylvania, United States cLink¨oping University, Sweden dRegion Stockholm, Sweden
A R T I C L E I N F O Keywords: Suicidal behavior Metro Transit Space-time budget VISUAL-TimePAcTS Logistic regression A B S T R A C T
Introduction: We explore the use of CCTV footage to map suicidal self-injurious behavior on a
subway platform to better understand the settings and the situational conditions of individuals just before they attempt suicide.
Methods: We use footage from CCTV cameras for gaining new insight into the situational
con-ditions that relate to suicidal self-directed violence in the transit system in Stockholm, Sweden. We adopt a space-time budget template to record, step-by-step, what happens over time as in-dividuals on the platform wait for an incoming train. The analysis applies visualization tools (VISUAL-TimePAcTS) and uses a cross-over design to identify risk factors associated with suicide.
Results: Findings show that suicide risk varies both temporally and spatially. Among all types of
possible behaviors and places, being close to the edge of the platform of the opposite direction of the train and crossing the security line – this behavior and place combined – are associated with increased risk of suicide.
Conclusions: We confirm that using CCTV footage as data source provides valuable insight into
relevant situational conditions in which suicides take place, which can be useful to inform pre-vention strategies, particularly information about behavior and place combined. The article concludes by reflecting upon the importance of these results for future research.
1. Introduction
One of the key issues in suicide prevention is to understand the importance of situational conditions that lead individuals to take their own life. Traditionally, efforts to identify these conditions were carried out by assessing past suicide instances and conditions in which they took place (e.g. Dinkel et al., 2011; Too et al., 2016; Uittenbogaard and Ceccato, 2015). Methods for obtaining information and conducting interventions in suicide prevention have recently expanded to include ICT - Information and Communication Tech-nologies. This is particular true for suicides in public places, such as transit environments including subways, railways, bridges, and roads, where ICT are often an inherent part of the infrastructure of the transit systems. In particular, the use of closed-circuit television * Corresponding author. Royal Institute of Technology, Department of Urban Planning and Environment, Teknikringen 10A, 100 444, Stockholm, Sweden.
E-mail address: vania.ceccato@abe.kth.se (V. Ceccato).
Contents lists available at ScienceDirect
Journal of Transport & Health
journal homepage: http://www.elsevier.com/locate/jthhttps://doi.org/10.1016/j.jth.2020.100976
(CCTV) images in research has increased for many uses (Goldman et al., 2014; Haw and Hadfield, 2011; Lee et al., 2017; Zelniker et al., 2008) and is argued to be a valuable resource to identify the conditions in which suicides, suicide attempts and interrupted attempts take place (Mishara et al., 2016). We build on these previous studies by exploring the use of CCTV footage to map suicidal self-injurious behavior on the subway platform to better understand the situational conditions of individuals just before they attempt suicide.
The aim of this article is to gain new insight into the situational conditions that relate to suicidal self-directed violence in the transit system in Stockholm, Sweden. This is achieved by using a space time budget template to record, step-by-step, what happens over time as individuals on the platform wait for an incoming train. The analysis applies visualization tools (VISUAL-TimePAcTS) and uses a cross-over design to identify risk factors associated with suicide relative to for example, behavior and location. In order to report the locations and behaviors prior to a suicide attempt, the subway platform is divided into sections in such a way that data collection and subsequent analyses are linked to these platform sections.
1.1. Terminology
The following suicide related definitions are used in this article based on Crosby et al. (2011) and Posner et al. (2014); Uitten-bogaard and Ceccato (2014):
Self-Directed Violence (analogous to self-injurious behavior): Behavior that is self-directed and deliberately results in injury or the
potential for injury to oneself.
Suicidal self-directed violence: Behavior that is self-directed and deliberately results in injury or the potential for injury to
oneself. There is evidence, whether implicit or explicit, of suicidal intent.
Suicide: Death resulting from intentional self-injurious behavior, associated with any intent to die as a result of the behavior. Suicide attempt: A non-fatal self-directed potentially injurious behavior with any intent to die as a result of the behavior. A suicide
attempt may or may not result in injury. Interrupted attempt – by self or by other:
By self (Aborted attempt): A person takes steps to injure self but is stopped by self-prior to fatal injury.
By other: A person takes steps towards making a suicide attempt but is stopped by another person prior to fatal injury. 2. Theoretical background
2.1. Suicide in transit environments: situational conditions 2.1.1. Temporal patterns
Suicides do not happen at random in time. In terms of temporal patterns, Dinkel et al. (2011) found that the incidence of jumping from the train station platform was more common during 9:01 a.m. to 6:00 p.m., and at night, lying on the tracks was most frequent. van Houwelingen and Beersma (2001) found that in the Netherlands suicide levels were lower at night-time than day-time. Moreover peak levels of suicide occurred just after sunset and at the end of the morning. In terms of weekly variations, most studies indicate that weekends (Saturday and Sunday) have fewer suicides as compared to weekdays (De Leo and Krysinska, 2008; Erazo et al., 2004) although De Leo and Krysinska (2008) found that in Australia, most acts of suicide at railways happen on Thursdays and Fridays, whereas in Germany most suicides occur on Monday to Wednesday (Emmerson and Cantor, 1993; Erazo et al., 2004; Ladwig and Baumert, 2004). With regards seasonality, in both the northern and southern hemispheres (represented by Germany and Australia, respectively), spring has shown high suicide levels (De Leo and Krysinska, 2008; Erazo et al., 2004) and winter and summer months experience the fewest suicide events (Erazo et al., 2004; De Leo and Krysinska, 2008).
In Sweden, previous studies (Radbo and Andersson, 2012; Radbo et al., 2005) showed a concentration of suicides in the day-time (afternoons) and light seasonal variation in railway suicides, with more in the warmer months of the year (April–September). In Stockholm, Uittenbogaard and Ceccato (2015) showed that suicides in the underground were concentrated during the day (mostly off-peak hours, when fewer passengers are around) and in the spring. When studying suicide cases in commuting trains in Stockholm, Ceccato and Uittenbogaard (2016) found no sign of seasonality, but winter months accounted for a larger share of events. Moreover, suicides did not occur evenly throughout the day and tend to take place more often in weekdays.
2.1.2. Spatial patterns
Suicides follow distinct spatial patterns. In Germany, for example, Erazo et al. (2004) suggested that larger, main-line stations and stations through which fast train lines run account for more suicides than other station types. They also pointed out the importance of the high speed trains in certain lines. They suggested that faster trains do have a longer braking stretch and will hit a person at a higher speed, making it more probable for the attempted suicide act to be fatal.
In the UK, studies in London found that within stations, most acts occur at the beginning of platforms (Clarke and Poyner, 1994), where fewer passengers are likely to be found than at the middle of the platform. This provides a more isolated place where a suicidal person can remain unnoticed and evade possible interruption of the suicide attempt, resulting in higher rates of fatal outcomes.
The surrounding neighborhood of stations and transit systems may also be an influencing factor in the selection of suicide place. In recent evidence from Canada, Mishara et al. (2016) showed that more than one-third of suicides occur on rails near psychiatric fa-cilities. Long-term economic and social problems may cause certain people to be depressed. De Leo and Krysinska (2008, page 774) suggest that individuals with low socioeconomic status show a higher vulnerability to suicide.
2.1.3. Situational conditions and behavior
In terms of behavioral patterns of individuals showing signs of suicidal self-directed violence, Dinkel et al. (2011) found that “jumping” happened predominantly in the station area, while “lying” and “wandering” took place when the individual was on open track. Fatality was highest in people who lay on the track and lowest in people who jumped. In a more recent study, Mishara et al. (2016) found that people who attempt suicide in the metro have specific behavioral patterns that can distinguish them from ‘regular passengers’. Although a fourth of future attempters did not show any non-suicidal behavior, the presence of two or more risky be-haviors is associated with an increased likelihood of suicide, such as frequently looking down the tunnel, standing for long periods on the security line, continually walking on the yellow line, psychomotor agitation, and staring at the tracks (p. 9). In that study, they found out that:
•83% of individuals attempting suicide on a platform had a behavior that could be identified as there being ‘something wrong’. •the probability of being at risk for suicide increased significantly when 61% of the cases expressed two or more of these behaviors (a
total of 60 cases over 2 years). Risky behaviors were also present in ‘non-attempters’ (regular passengers) but having two or more of these indicated a higher likelihood of being at risk of attempting suicide than not showing these behaviors.
•75% of the cases had at least one abnormal behavior, indicating a possible ambivalence, such as waiting for several trains to pass before attempting suicide, leaving an object on the platform, and going back and forth from the yellow line, adjacent to the edge of the platform. Going back and forth from the yellow line (adjacent to the edge of the platform) was observed in 24% of cases. Instead of focusing on individuals, an opportunity-reduction approach, we argue, introduces the opportunity for transport oper-ators to implement discrete managerial and environmental changes to reduce opportunities (see also Clarke, 1994; Clarke and Poyner, 1994). In Stockholm specifically, Ceccato and Uittenbogaard (2016) found that suicide rates increased with speed commuting trains and decreased where barriers along tracks are installed. However, little evidence is found about how the internal environment of the stations affect individuals’ behavior, in particular in the platform of the stations. Below we turn our attention to suicide prevention initiatives discussing specifically opportunity-reduction studies using CCTV footage.
2.2. CCTV images and suicide prevention
The use of materials from CCTV (closed-circuit television) has recently increased in many areas of safety research (Cerezo, 2013; Goldman et al., 2014; Lee et al., 2017; Mishara et al., 2016; Taylor, 2010; Zelniker et al., 2008) given the ability of video to capture phenomena of interest that are observable as recorded material and/or in real time. With the increased attempts to use data from CCTV, methods for analysis of this ‘data’ have also been developed. For a systematic review of the literature on safety research see Havˆarneanu et al. (2015) and (Barker et al., 2017).
In suicide prevention, in particular, Mishara et al. (2016) present 2 studies that report a framework for detecting abnormalities in human behavior when observed over time in transit environments using CCTVs material. The overall aim was to assess the feasibility of using live CCTV images to identify at risk people on platforms, based on their behaviors. Study 1 aimed at identifying the behaviors present in suicide attempters on the metro. The underlying hypothesis was that people in stations who are going to attempt suicide have recognizable behaviors that may be indications that a suicide attempt is imminent. Study 2 aimed at assessing the specificity of behavior patterns that can be observed within 5 min of the attempt, by comparing videos of attempters with videos where no one attempted suicide. In particular, they used several camera views to show that people who attempt or die by suicide in the subway have specific behavioral patterns that can be distinguished from other passengers. These behaviors may reflect ambivalence, anxiety, or planning before the suicide act (3/4 of attempters). Mishara et al. (2016) suggest that some behavior may be easier to observe by watching videos or using an automated surveillance system, such as leaving an object and walking back and forth to the security line while other behaviors are more complex and require a certain level of analysis, such as determining if a person “seems depressed”. These behaviors demand interpretation and may be reliably identified by intensively trained observers. They conclude that real time observations of CCTV monitors in combination with other suicide preventive measures can be useful for suicide prevention in transit environments.
2.3. Research questions
Based on previous national and international literature, this work sets out to investigate the following research questions (RQ) relating to the temporal, behavioral and situational aspects of suicide:
RQ1 – Does the occurrence of suicide vary temporally, by day, week, and month? Are there observable temporal variations relating to the time spent at platform before an attempted suicide act?
RQ2 – Do the behavior(s) of an individual immediately prior to a suicide act affect the outcome, such as waiting for the train beyond the security line? As suggested by Dinkel et al. (2011), “jumping” happened predominantly in the station area, while “lying” and “wandering” took place when the individual was on the open track.
RQ3 – Are there locations that are considered of higher risk to be in prior to a suicide attempt (is associated with an increased risk for suicide), such as those at the end/corner of the platform (Dinkel et al., 2011; Mishara et al., 2016).
3. Case study description
3.1. Study area
This study uses Stockholm’s subway station network and commuting train network as its setting, with a selected commuting train station as a case study. While Stockholm’s subway has 100 stations (divided into three lines: green, red and blue), the commuter train system has 50. Within the commuting train system, a total of 15 stations were included in the analysis since they also contained data on suicides. Yearly around 300 people die of suicides in the Stockholm county (a rate of 12.4 per 100 000 inhabitants, 2014–2018) about 10% occur in the track-bound public transport in the county (The Public Health Agency of Sweden, 2020). For an in-depth analysis of suicides by types of stations in subway and commuting train system in Stockholm, see the study by Ceccato and Uittenbogaard (2016); Uittenbogaard and Ceccato (2015).
3.2. Data
CCTV video recordings for train stations, stored by Stockholm transportation, were used to extract data exhibiting suicidal self- directed violence (suicides, suicide attempts and persons with other suicidal behavior). In total, video recordings of 118 cases covering the period from March 29, 2015 to 6th September 2017 (namely data over 2 years, 5 months, 13 days) were evaluated, of
which 24 are categorized as suicides, 29 as suicide attempts, and 65 as individuals with suicidal behavior.
3.3. Methods
3.3.1. Data preparation and systematization
CCTV footage episodes were first pre-selected from the full recordings of the stations over the whole transit system. They were composed of cases of suicides, suicide attempts (including interrupted attempt), a-typical behavior, in other words, potentially suicidal behavior. In order to create a template to collect systematically information from these video recordings, a sample of these CCTV episodes (20–25 events, from February to March 2017) were observed by the research team (multiple people code the behaviors to identify potential suicidal behavior vs. other behavior) using as a reference the study by Mishara et al. (2016). Using principles of space-time budget (see e.g., Wikstr¨om et al. (2010) and Ceccato et al. (2017)), the most common patterns of behavior observed in place and time were used to create a template for data collection. This template contains also basic characteristics of the subjects studied (e.g. gender, age) and more importantly, features of the station, the course of events (behavior) from the time the individual came to the station/platform, as well as the places the individuals spent time. Experts in the transportation company used the template to classify the events in Excel. The analytical template was created to facilitate a categorization of events in time and space, with the purpose of detecting commonly occurring conditions and periods of time associated with these events and their likelihood associated with suicide. Consequently, each identified case of suicidal self-directed violence in the CCTV data, was divided into a sequence of observed events each of which was characterized by a time, behavior/activity, and location. The types of behaviors/activities were classified into four
Table 1
Situational suicidal behavior – identifiable by CCTV footages
Behavior category Type of behavior
Signs of ambivalence o Uncertain if he/she wants to enter the station
o Begin to jump (down the track) but stops, hesitate to jump o Walks between the platforms
o Goes back to another train o Walks the entire platform’s length
o Waiting for one or more trains to pass before initiating an attempt o Leaves one or more objects on the platform
o Removes items (for example, a bag) o Jumps but tries to protect from the train Lethargic signs o Detached, stays for a long time with no movements
o Enters the station and does not pay for the ticket o Leaves one or more objects on the platform Odd/strange* behaviors o Scream/without control
o Shakes hands/talks with oneself o Crying, signs of stress
o Looking depressed (lethargic condition)
o Fight with someone who he/she knows/or is unfamiliar Suicidal behavior o Walks over the white line to another location on the platform.
o Trying to jump in front the train o Sits on platform edge
o Sits on the floor near the white line o Laying down on the tracks
o Looking into the tunnel several times where the train comes o Goes back and forth over the white line(at border of platform) *see Mishara et al. (2016, pp. 4–5).
categories based on the behavior categories derived from Mishara et al. (2016) (Table 1).
We assessed the variation of behavior-activity over time and place. The employed method is based on time-geography principles (H¨agerstrand, 1970) and has already been applied to various areas of research (e.g. Ceccato et al., 2017; Qvistr¨om et al., 2020). For simplicity the method is illustrated here by an example of interaction of place and behavior, as follows:
1. Entrance (alone) = fare dodges→ talks on the phone→ gesticulates
2. Platform (alone) = looks at the tunnel when train comes→walks back and forth in the platform/"white line"→ waits for several trains→ sits on the edges of the platform
3 Platform (bystander) = is rescued by another passenger
The resulting template consists of an Excel file divided into 6 parts, identification number, time, behavior, place in the platform and elsewhere, socialization/indication if the person was alone or with someone else (Fig. 1).
Out of 118 cases, 93% of the individuals were by themselves versus 7% that were there with an acquaintance, family member or a friend (Fig. 2(a)). As expected, most of the time spent in transit was in the platform (not at the entrance) (Fig. 2(b)). In terms of behavior, most frequent types of behavior are: Walk in the white line (security line), rises/stand up, the person is helped by rescue services, guards, someone in the platform helps, comfort her/him, sits on the bench (Fig. 2(c)).
Out of 118 cases, 24 are categorized as suicides (e.g., individual jumps from the platform in front of the train and dies from the injuries) and 29 as suicide attempts or interrupted attempts (cases when an individual jumps but survives, or shows signs of suicidal behavior, but is stopped either by self or by other). Since it is not always possible from the video recordings to determine the nature and cause of one’s behavior, we included 65 individuals with potentially suicidal behavior (when an individual behaves strangely and his/ her behavior can lead to an injury, e.g. he is upset and looks like he is ‘training to jump’). Note however that the fact that a person behaves atypically does not necessarily have to be linked with suicidal behavior. Therefore, we assessed this group separately from the rest, and highlight as often as possible when they are included in the analysis.
3.3.2. Visualization
Based on the information from the surveillance cameras gathered using the template, we were able to extract event sequences (as described previously) revealing different types of suicidal behavior. An event sequence is an orderly list of activities that have occurred at a location at a certain time. We classified and visualized these, using the graphs and visualization tool VISUAL-TimePAcTS, and we identified the most vulnerable environments, all types of behaviors (also called activities or events in this study) and times with regard to the frequency of suicide (or suicide attempts) at each section of the platform. VISUAL-TimePAcTS provides functionality to interactively explore different aspects of the collected event sequences, apply data mining techniques to identify patterns across them and cluster the sequences based on the identified patterns (Ceccato et al., 2017; K. Vrotsou, Ellegård and Cooper, 2009; Katerina Vrotsou, Ynnerman and Cooper, 2013). Inspecting patterns of activities with regard to their nature, timing in the sequence of events and duration gives rise to hypotheses about the temporal relations of these activities regarding the outcome that can be investigated further through modeling. Visualization outputs from VISUAL-TimePAcTS are also expected to provide a descriptive overview of the collected data and serve as an insight for suggesting best practices in suicide prevention in time and space.
3.3.3. Modeling
We first investigated which places are most associated with a higher/lower risk (odds) of suicide. Second, we looked at the be-haviors that occurred just prior to the suicide associated act (i.e. suicide, suicide attempt or interrupted suicide). We accomplished each of these analyses by framing the data in the context of a case-crossover study design, which was designed specifically to investigate whether intermitted, abrupt-onset exposures appear to “trigger” the onset of an acute outcome of interest (Malcolm Maclure, 1991; M. Maclure and Mittleman, 2000). Motor vehicle crashes are the prototypical outcome for this study design, which compares each driver
at the time they crash to themselves a short time (e.g., minutes) before they crashed. Conditioning on subject in this way enables a test of whether using a cell phone, for example, poses a risk to crash. Finding that the prevalence of cell use is significantly more common at the time of the crash than a few minutes earlier is evidence that cell phone use while driving poses a risk for crashing. Importantly, comparing each individual to only themselves, rather than to others in the sample, serves to control for characteristics of each
Fig. 3. (a) Suicide per hour and Other (attempt, interrupted, suicidal behavior) per hour. (b) Spending time at a place just before suicide (or just
individual that may be hard or not even possible to measure, including intermittent exposures like the amount of alcohol consumed that day, and inherent exposures like degree of impulsivity.
Following this convention, we used conditional logistic regression to treat each individual as a case at the time of their suicide or suicide attempt, and compare their location and behavior at that moment to their locations and behaviors minutes earlier when they each served as their own control, to thereby identify factors that constitute being ‘risky’ or ‘protective’ from suicide. We assessed the variation of behavior/activity over time and place and social interaction (see an example in section 3.3 for more details). Hence our unit of analysis (observations) was the number of behaviors/activities (or places) from the 118 individuals. While the case-crossover study design provides a control for potential confounders such as impulsivity and previous suicide attempts, as a related matter it is of interest to consider whether a certain location or behavior poses a suicide risk for all individuals or only for subgroups by sex, age group, and so on. Our sample size of 118 was not large enough to enable the stratified analysis that otherwise could have accomplished analyses within subgroups.
4. Results
4.1. Main situational patterns
In the total 118 cases we reported from the database, 51% of individuals were identified as men and 46% as women (note that in 3% of cases, it was not possible to identify the gender of the individual based on those records). However, when suicides and suicide attempts only were analyzed (excluding those expressing suicidal behavior), we confirmed the expected; more males were present in the sample: 60% were men and 40% women. The results from this selection indicate that the individual’s characteristics, such as gender, relate to the outcome, but not always in the expected way. Although the number of males who died by suicide was larger, half of them were younger than 30 years old, which differs from national average age on suicide statistics that indicates that middle aged and older men are overrepresented among suicidal individuals. Findings also show that CCTV captured 55% of individuals at the entrance to the station before attempting suicide on the platform while less than 10% were brought to that particular station by another train and the rest of the individuals were registered on the platform (for 35% of the sample, this information was lacking). Of those 55% of in-dividuals captured by CCTV at the entrance to the station, 12% did not pay the fare, for example, they jumped over the barriers or smashed through without paying. The incidents occurred in 66 of the 100 metro stations with several stations having more than one attempt. The majority happens in the central station, followed by Slussen, Stadshagen, Hornstull and Fridhemsplan. Suicides are in this sample concentrated in the commuting stations of Flesmingsberg and Rågsved.
4.2. Temporal patterns
To answer our first research question, we investigated whether suicide varies temporally. For the suicides, results show a few peaks at certain hours of the day and evenings (7:00–8:00, 10:00–11.00, 15:00–16:00 and 19:00–20:00) while for all recorded cases (sui-cides, suicide attempts and interrupted suicides), the pattern shows a more homogenous pattern across the day, with a peak in the late evening (22:00–23:00), see Fig. 3(a).
Among the recorded suicides, weekends (defined as Friday, Saturday, Sunday) have twice as many cases as weekdays (defined as Monday, Tuesday, Wednesday, Thursday). No major significant differences were found over the months of the year (but they were slightly higher in January and June). Our results also show that the time of day did not impact on where an individual spent time at the station, namely, at platform, tunnel or track (χ2(4, N = 118) = 9.16, p < 0.165).
In terms of types of stations, results show that among the suicides and suicide attempts in the sample, 75% of cases occurred in the subway system (compared to commuting train), mainly on the green and red lines, 31% and 27% respectively (these lines have also more stations), while 25% occurred at stations within the commuting train system (note that our study did not include all registered suicides in commuter train systems, only 15 cases, 13% of the total sample we analyzed. For details on suicides in commuter trains, see Ceccato and Uittenbogaard (2016).
4.3. Spatial patterns
To investigate whether the individuals’ behavior and location on the platform relates to the outcome, we looked more closely at different spatial patterns; that is to say, where an individual spent time before the suicide attempt.
If we look at the total sample, individuals were located in 1 out of 4 places (A, B, T, D) just before the suicidal self-directed violence act. We include all individuals in the analysis to identify risky places since at this stage we cannot differentiate places that are used by attempters (suicide and attempt/interrupted suicide). Fig. 3(b) shows the places where the individuals spent the most time imme-diately prior to the outcome. At this point in time, 44% of all cases were on the platform edge, that is, on the security white line painted on the platform or in B areas (B1, B2, B3); 16% went on the track, 14% in the tunnel, and some of them came back to the platform, indicating ambivalence while 20% were found at the start of the platform. Note that the "C" position was not relevant because this position regards few seconds before the attempt only.
Individual’s risky location has also attracted attention of people who were at the platform so that rescue services could come in time. This fact flags the difficulty to statistically distinguish between those who intend to die by suicide and those who show suicidal behavior in relation to where they were on the platform or tracks (χ2 (1, N = 118) = 3.80, p < 0.28).
or tunnel when the train arrived (χ2 (1, N = 52) = 6.56, p < 0.01). Together, these two positions (track or tunnel) proved be the most risky compared to being at one of the other sites of the station. The mortality risk of track and tunnel is no surprise, but it is important to highlight it because it affects our way of thinking about suicide prevention.
4.4. Behavioral patterns
We recorded on average 7 different types of observable behaviors at the station for each person before the suicidal self-directed violence. As many as 785 activities or behaviors were recorded and some of the most frequent are "crossing the security line", "standing up", "getting help" and "sitting on the bench". They were categorized into four categories (Fig. 4) following the categories suggested in Table 1 partially based on Mishara et al. (2016). Note that this categorization is suggestive not confirmatory.
Although we are not able to tell the exact time in minutes spent before attempt, our figures support the hypothesis that most in-dividuals do not jump right away in front of the first train when they enter the platform, they instead spend some time at platform. Several individuals showed signs of ambivalence. This means, for example, that they "walk the entire platform", "go between plat-forms", "leave items on platform", "enter and exit the train" or "start jumping but stop or hesitate to jump". This included "attempting to protect itself from the train’s impact" and "hesitating before entering the station" (Fig. 4 (a)). Some also showed signs of lethargy or detached behavior, careless (Fig. 4(b)). For example, they leave items on the platform or do not pay for the train ticket. Fig. 4(c) indicates out of place behaviors, which appeared to be symptomatic of someone with a mental health problem or intoxicated, such as odd gestures, talking to oneself, or having motions, which appeared to be out of place. Finally, Fig. 4(d) illustrates examples of behavior types most closely linked to suicide attempts, such as sitting on platform edge. In the next section, we further illustrate potential differences in behavior by groups.
4.5. Visualization of suicide related behavior over time
Fig. 5 shows two groups of individuals in our database visualized using VISUAL-TimePAcTS: (1) those showing potential suicidal behavior-‘at risk’, and (2) those classified as ‘suicides, suicide attempts, or interrupted suicide attempt’. In the figure, each individual is represented by a continuous path showing their sequence of performed activities. The length of each activity corresponds to its duration and color shows the activity type (see color legend in Fig. 5). Individuals are sorted along the horizontal axis by group (‘at risk’ to the left and ‘suicides, suicide attempts, or interrupted suicide attempt’ to the right), and within each group individuals are sorted by sex (women, men, unknown) and time period of the attempt (morning, afternoon and evening/night). Time of day is rep-resented on the vertical axis. In addition to performed activities, the figure can alternatively display individuals’ settings/places or socialization during the logged time slots. Moreover, arbitrary sorting of the individuals is possible with respect to any two arbitrary metadata variables available for the data set, such as gender, age, day, month. Drawing the activity sequences of several individuals next to each other on the same screen in this manner allows for a comprehensive overview of the data and enables comparisons to be made between them. For example, note that showing lethargic behavior (purple color) and different suicidal behavior (brown color) were more common among those belonging to the group 2 than among those from group 1. Visualization of behavior over time can be particularly relevant to show how quick rescue services vary in reacting depending on stations and types of calls.
4.6. Modelling situational suicide risk
In order to further investigate the situational conditions that relate to suicidal self-directed violence in the transit, we modelled which behaviors and places run a higher risk for suicide. Table 2 shows results of the conditional logistic regression with a case- crossover design. Being near the edge of the platform (A) and crossing the white line of safety (B1-2) was associated with an increased probability (greater odds) to die by suicide compared to being in all other places on the platform. Being in the middle of the platform was associated with lower probabilities (odds) of dying by suicide compared to being in all other places. The center of the platform (not on the white line or at the edges) is thus a suicide protective place.
Table 3 shows the relationship between showing a particular behavior against all other types of behaviors and the likelihood of suicide, suicide attempt, or interrupted suicide. Some variable combinations is associated with an increased likelihood of suicide while others, reduces it. To show ambivalent behaviors (such as going back and forth to the white security line of the platform) or suicidal behavior (e. g. sitting on the edge of the platform or heading in the direction of the ‘white’ security line) is associated with an increased risk of suicide or suicide attempt.
However, arriving by train to the platform (not entering from the station entrance) proves to be a protective factor against suicide, perhaps because other passengers have had a chance to intervene, to be noticed during the trip that ‘something was wrong with that person’, or just because those who show suicidal behavior and wonder in the transit system are less prone to be suicidal than those who enter from the main entrance with the pure objective to die by suicide.
Table 4 shows the intersection of the most behaviors of committing or attempting suicide and the most common locations where this occurred. Most (15 of 16) instances of walking in the area beyond the security line occurred in a single type of location, at the corners of the platform. In contrast, the instances of sitting on the edge of the platform occurred in the platform middle (3 of 7), in the platform corners (2 of 7), and also in other locations on the platform (2 of 7).
5. Discussion
Based on the information from the surveillance cameras gathered using the template, we were able to classify and visualize the sequence of different types of suicidal behavior. Using the graphs and visualization tool VISUAL-TimePAcTS, we identified the most vulnerable environments, all types of behaviors and times with regard to the frequency of suicide (or suicide attempts) at each section of the platform.
Despite some temporal variation, the occurrence of cases recorded using CCTV cameras does not show a distinct pattern over the hours of the day. Among the suicides, weekends include twice as many cases as weekday. No major differences were found over the months of the year but they were slightly higher in January and June. Our results also show that the time of day did not correspond to where an individual spent time at the station, namely, at platform, tunnel or track. Although we are not able to tell the exact time in minutes spend before attempt, our figures reveal that most individuals do not jump in front of the first train, irrespective if they enter the platform by another train or through the station entrance. The length of time spent by the individual in the platform is valuable for suicide prevention because it allows one to identify the individual at risk, call on support, and rescue services on time. Although there is a large variation in behaviors between individuals observable in CCTV, from 1 behavior to 31 different types of behavior (standard deviation 4.3), these numbers are similar to those compiled by Mishara et al. (2016) if we compare with the frequency of the occurrence of easily observable behaviors in a suicide attempt, which is about 6 (min 3, max 41).
More than half of the individuals were captured by CCTV at the entrance to the station, some of them jumped over the barriers or passed through without paying. As Mishara et al. (2016) indicate, in most cases it appeared that there was nothing unusual about the behavior of these individuals when they entered the stations. However, in some cases they seemed to have difficulties when they bought their ticket or showed signs of ambivalence, they went back and forth. In almost all cases (93%) that were detected at the entrance, the individual was alone. In other words, we were not able to identify any visible companion at the moment of the footage. It is not difficult to connect these findings with suicide attempts because they reflect what we could expect in terms behavior just before a suicide. However, it is more challenging to be able to identify behaviors on CCTV cameras that were not strictly suicidal (e.g., caused by intoxication of any kind) and that, at the end, lead to suicide.
As initially suggested by Dinkel et al. (2011) and Mishara et al. (2016), the immediate previous location of an individual just before a suicide attempt affects the outcome, whether it leads to suicide or not. Both the descriptive statistics and results from modelling corroborate our previous hypotheses, that there are places in the station that are more risky than others are. Half of those who constituted our total sample of individuals were on the platform edge, that is, on the security white line painted on the platform; a fourth went on the track, in the tunnel, and the rest were somewhere else in the platform. However, among those who were classified as “suicide attempters”, half of them were already on the track or inside the tunnel—places that proved to be the most lethal ones compared to being at one of the other locations, either the platform or any other place at station. Our models that take into account the number of times each person spent time on these locations confirmed this pattern.
Modelling results showed the central section of the platform was a protective location against the outcome suicide while spending time at the corner/end of the platform over the security line is the highest risk combination of behavior and location. Note that although some of these behaviors (alone or in combination with location) are important to explain the suicide geography, they may not become significant in these models since they are not simultaneously recurrent. In other words, the combination of behavior and location simultaneously are not frequent enough to become significant in the model most likely because our observations of behavior
Fig. 5. VISUAL-TimePAcTS visualization showing the activity sequences of the individuals in our study. Individuals are sorted along the horizontal
axis by group (‘at risk’ to the left and ‘suicides, suicide attempts, or interrupted suicide attempt’ to the right) and, within each group, by sex (women, men, and unknown) and time period of the attempt (morning, afternoon and evening/night). Time of day is represented on the vertical axis. Color represents the activity type. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)
and locations are relatively small in comparisons with the total possible combinations of behavior and locations at platform.
None of the variables that signify individual risky behaviors were significant in the model (for example, "leaving items on the platform", "not paying for the ticket" or "displaying lethargic/mindless state"). Nor were significant most variables that in the literature called "strange or odd behavior" (such as, for example, shouting, crying, talking to themselves, fighting with someone, being in a lethargic state). These individual behaviors were not so frequent, which means that they are not captured in the model. It may also be because some of these behaviors are difficult to observe remotely with CCTV cameras. It should also be mentioned that it is difficult to determine whether a person intends to die by suicide, based solely on filmed material. Behaviors that may seem "odd/secluded/ strange" do not necessarily need to be signs of suicidal behavior, but they can nevertheless be signs of vulnerability that should therefore be noticed and acted upon from a purely injury prevention perspective.
Finally, suicide in transit environments is gendered. More men than women were among those who performed suicidal self-directed violence acts in Stockholm’ transport networks, confirming previous literature with reference to gender. However, regarding age, this sample is composed of younger adults and fewer elderly males compared to our initial expectations. These individual differences are observable when we visualize the data by gender. The intersection of gender and age may be an important aspect to be considered in suicide prevention.
Table 2
The relationship between being in a certain place versus all other types of place and the odds of committing suicide (or attempting suicide and being saved) among 118 persons.
Types of place % OR 95% CI P-value
Near the platform corner (A) 9.5 8.58 0.82, 90.35 0.073*
Just before suicide: white line (B) 6.7 7.43 0.74, 74.50 0.088*
At beginning of platform 43.8 0.27 0.05, 1.33 0.108
In the middle of the platform 22.9 0.14 0.01, 1.41 0.095*
In the tunnel 10.5 1.41 0.32, 6.34 0.651
In the tracks 6.7 2.90 0.49, 17.09 0.240
*P<0.10 **P<0.05 ***P<0.01, Completed with unadjusted conditional logistic regression models. OR = Odds ratio, CI = Confidence Intervals.
The 118 persons spent time in a total of 19 locations in the stations and on the platform leading up to the time of the suicide or suicide attempt. Together the 6 locations in the table account for 42.3% of the 19 locations. Time near the entrance of the station acounted for an additional 25.8% location instances and other 11 other locations together accounted for a total of an additional 31.9% of instances.
Table 3
The relationship between exhibiting a certain behavior versus all other types of behaviours and the odds of committing suicide (or attempting suicide and being saved) among 118 persons.
Types of behavior % OR 95% CI P-value
Arrives by train to station 11.0 0.15 0.02, 1.17 0.070*
Walks in the white line 16.8 2.54 1.05, 6.15 0.039**
Moves back and forth in the platform white line 9.1 3.97 1.44, 10.94 0.008***
Walks between the platforms 6.5 1.05 0.25, 4.37 0.951
Leaves objects in the platform 4.6 4.74 1.13, 19.88 0.330
Sits on the edges of the platform 5.0 44.30 5.58, 351.44 0.001***
Rises (from sitting or laying down) 14.4 2.58 0.72, 9.21 0.144
Someone holds, hugs, attempt to help 13.2 0.76 0.21, 2.76 0.675
Walks the whole length of the platform 9.1 1.84 0.67, 5.04 0.234
Arrives on the platform 10.3 0.46 0.13, 1.61 0.224
*P<0.10 **P<0.05 ***P<0.01, Completed with unadjusted conditional logistic regression models. OR = Odds ratio, CI = Confidence Intervals.
Together these behaviours accounted for 48.9% of the behaviours the individuals exhibited during their time in the station and on the platform. Of the other behaviours no single behavior accounted for more than 2% of the total behaviours.
Table 4
The relationship between most frequent behavior, most common types of places and the odds of committing suicide (or attempting suicide and being saved)
Type of place
Type of behavior Stations’ entrance Platform middle Platform corners Other platform Total
Pays fare at entrance 41 0 0 0 41
Arrives by train to station 6 17 0 3 26
Walks in the white line 0 0 15 1 16
Sits on the edges of the platform 0 3 2 2 7
6. Conclusions
The main goal of this article was to describe, visualize, and assess the situational conditions in which suicides takes place in the transit system, in particular, on the platforms of metro stations in Stockholm, Sweden. With the help of the recorded material, we were able to see, among other things, that most people who attempt suicidal self-directed violence acts in the metro have specific behavioral patterns that can be analyzed with visualization tools and statistical analyses. Our results attempted to answer our three research questions. They show that the incidence of suicide varies over time (RQ1); somewhat over the day, but with fewer weekly and monthly variations. Regarding RQ2, the individuals’ behavior and heir location in the platform also affected the outcome. For instance, showing ambivalent behavior was associated with suicide risk, such as going back and forward over the security line or sitting at the platform corner. We found also that the individuals’ characteristics related to the outcome, namely whether the self-directed violence act is fatal or not. More men than women attempted suicidal self-directed violence acts. However, they were younger than expected. Overall, this study provided more in-depth knowledge of the risky behaviors and places that may lead to suicide and suicide attempts. The novelty of the study is that it identifies a number of behaviors and locations on the platform as indications of risk of ongoing suicide attempts using footages from CCTV cameras. This approach provides a framework for developing prevention strategies based upon in-depth analyses of suicides data captured using CCTV footage.
This study has a number of limitations. Our results show that relevant but less observable behavior is not captured in videos because not all cameras are close to the person or they recorded the individual from ‘wrong’ angle. The change in CCTV position on the platform or the possibility to move the angle of the cameras can be a solution to better capture individual’s behavior at certain stations. Another limitation is that less frequent types of behavior can be important for explaining suicide, but that they are not always ‘caught’ with conditional logistic regression with case-crossover design. In other words, we expect that there is bias towards the null hypothesis. This happens because for each site we looked at whether it was at the site associated with a higher (or lower) risk of suicide compared to being in all other places on the platform. In the case that some of the other sites were associated with a higher risk of suicide, the model was less likely to find a significant effect, meaning that these results are very conservative. One possible way to deal with this issue in future studies is to create indices of behavior (or indices of locations) that can aggregate some homogeneous behavior together, so they have a chance to more frequently appear in the model, either alone or in combination with location (his might give a chance for less frequent types of behavior and/or location to affect the results). As noted above, our analyses were conducted only in the sample overall rather than separately by sex, age, or other subgroup that may be of interest, given that our sample size was not large enough to perform the stratified or interaction analyses that could have investigated whether the results we found are consistent or vary. We encourage further studies using similar designs to investigate these issues further.
New studies can focus on developing automated computer programs to carry out real-time analysis of the films and then warn drivers or passengers. But they may also be basic qualitative situational analyzes that shed light on the interaction between man and the environment. A survey of the stations’ environment is important. To be able to refine the guidelines for suicide prevention, one must also have better knowledge of the individual’s characteristics in addition to gender and age, which in turn would require an information link between different actors who work at the stations.
This study shows that both behavior and location (alone or together) have independent influence on the outcome of suicide and suicide attempts. What we do not know is whether and how the station environment affects the probability of suicide. Previous studies have already shown a connection between station environment, neighborhood and suicide levels. For example, in Stockholm, if there are walls, they can reduce the visibility and monitoring possibilities and can create a more isolated place to attempt suicide. Moreover, lighting has been internationally linked to fewer suicides but we still do not know the mechanisms that explain this relationship (Matsubayashi et al., 2013). One has linked blue lighting with “a calming effect” in people, but that effect does not seem to be as great as one had first expected (Ichikawa et al., 2014). So far, it is unclear how light (type, intensity and location) affects those individuals who spend time at platforms. This is an important research issue that deserve further research.
This study has presented the first step in a research area that requires deeper knowledge and analysis with a longer period of time than the one available in the current project. The spatiotemporal pattern of behavior, as illustrated in this study is useful information but not yet enough to support a suicide preventive toolkit. In the future, it is essential to collect in-depth qualitative information from each case to better understand behaviors. For example, the subjects’ characteristics can be integrated into the analysis perhaps using a multilevel model framework. There is a reason for integrating this individual information because as shown in the later sections of this article, gender, age and other individual characteristics matter in the suicide behavior in transit environments. Although there is much to do, this analysis is the first step in providing a better understanding of the situational conditions of suicide and what is required to know before detailed interventions can be suggested to prevent suicide situationally using CCTV footages.
Financial disclosure
This research project was financed by the Swedish Public Health Authority.
Authors’ contributions
VC, DW, KV, contributed substantially to the conception, design and execution of the study, in particular analysis and interpre-tation and writing of the paper. UKN initiated the project and participated in the observational stage of the study (CCTVs). AG provided access to the data and systematized the information contained in the videos from Stockholm Transportation Company using the template of analysis.
Declaration of competing interest
The authors whose names are listed in this article certify that they do not have any conflict of interests of any nature.
Acknowledgements
We would like to thank the Swedish Public Health Authority for providing financial support for this project, as well as Region Stockholm which supported the project and data collection. Thanks to all participants of the seminar on “Situational Suicide Pre-vention” in February 2018 at the KTH Royal Institute of Technology, Stockholm, Sweden.
References
Barker, E., Kolves, K., De Leo, D., 2017. Rail-suicide prevention: systematic literature review of evidence-based activities. Asia Pac. Psychiatr. 9 (3), e12246 https:// doi.org/10.1111/appy.12246.
Ceccato, V., Uittenbogaard, A., 2016. Suicides in commuting railway systems: the case of Stockholm county, Sweden. J. Affect. Disord. 198, 206–221. https://doi.org/ 10.1016/j.jad.2016.02.051.
Ceccato, V., Wiebe, D.J., Eshraghi, B., Vrotsou, K., 2017. Women’s mobility and the situational conditions of rape: cases reported to hospitals. J. Interpers Violence. https://doi.org/10.1177/0886260517699950.
Cerezo, A., 2013. CCTV and crime displacement: a quasi-experimental evaluation. Eur. J. Criminol. 10 (2), 222–236. https://doi.org/10.1177/1477370812468379. Clarke, R.V., Poyner, B., 1994. Preventing suicide on the London Underground. Soc Sci Med 38 (3), 443–446.
Crosby, A.E., Ortega, L.A.G., Melanson, C., 2011. Self-directed Violence Surveillance: Uniform Definitions and recommended Data Elements. Retrieved from, Atlanta. De Leo, D., Krysinska, K., 2008. Suicidal Behaviour by Train Collision in Queensland, 1990–2004. Aus. N. Z. J. Psychiatr 42 (9), 772–779. https://doi.org/10.1080/
00048670802277263.
Dinkel, A., Baumert, J., Erazo, N., Ladwig, K.-H., 2011. Jumping, lying, wandering: analysis of suicidal behaviour patterns in 1,004 suicidal acts on the German railway net. J. Psychiatr. Res. 45 (1), 121–125. https://doi.org/10.1016/j.jpsychires.2010.05.005.
Emmerson, B., Cantor, C., 1993. Train suicides in brisbane, Australia, 1980-1986. Crisis 14 (2), 90–94.
Erazo, N., Baumert, J., Ladwig, K.H., 2004. Sex-specific time patterns of suicidal acts on the German railway system. An analysis of 4003 cases. J. Affect. Disord. 83 (1), 1–9. https://doi.org/10.1016/j.jad.2004.04.012.
Goldman, R., Pea, R., Barron, B., Derry, S.J. (Eds.), 2014. Video Research in the Learning Sciences. Routledge, Abingdon. H¨agerstrand, T., 1970. What about people in regional science? Pap. Reg. Sci. Assoc. Reg. Sci. Assoc. Meet. 24, 7–21.
Havˆarneanu, G.M., Burkhardt, J.-M., Paran, F., 2015. A systematic review of the literature on safety measures to prevent railway suicides and trespassing accidents. Accid. Anal. Prev. 81, 30–50. https://doi.org/10.1016/j.aap.2015.04.012.
Haw, K., Hadfield, M., 2011. Video in Social Science research. Routledge, Abingdon.
Ichikawa, M., Inada, H., Kumeji, M., 2014. Reconsidering the effects of blue-light installation for prevention of railway suicides. J. Affect. Disord. 152–154, 183–185. https://doi.org/10.1016/j.jad.2013.09.006.
Ladwig, K.-H., Baumert, J.J., 2004. Patterns of suicidal behaviour in a metro subway system: a study of 306 cases injured by the Munich subway, 1980–1999. Eur. J. Publ. Health 14 (3), 291–295. https://doi.org/10.1093/eurpub/14.3.291.
Lee, D., Arthur, I.T., Morrone, A.S., 2017. Using video surveillance footage to support validity of self-reported classroom data. Int. J. Res. Method Educ. 40 (2), 154–180. https://doi.org/10.1080/1743727X.2015.1075496.
Maclure, M., 1991. The case-crossover design: a method for studying transient effects on the risk of acute events. Am. J. Epidemiol. 133 (2), 144–153. Maclure, M., Mittleman, M.A., 2000. Should we use a case-crossover design? Annu. Rev. Publ. Health 21, 193–221. https://doi.org/10.1146/annurev.
publhealth.21.1.193.
Matsubayashi, T., Sawada, Y., Ueda, M., 2013. Does the installation of blue lights on train platforms prevent suicide? A before-and-after observational study from Japan. J. Affect. Disord. 147 (1), 385–388. https://doi.org/10.1016/j.jad.2012.08.018.
Mishara, B.L., Bardon, C., Dupont, S., 2016. Can CCTV identify people in public transit stations who are at risk of attempting suicide? An analysis of CCTV video recordings of attempters and a comparative investigation. BMC Publ. Health 16 (1), 1245. https://doi.org/10.1186/s12889-016-3888-x.
Posner, K., Brodsky, B., Yershova, K., Buchanan, J., Mann, J., 2014. The classification of suicidal behavior. In: Nock, M. (Ed.), The Oxford Handbook of Suicide and Self-Injury. Oxford University Press, Oxford, pp. 7–22.
Qvistr¨om, M., Fridell, L., K¨arrholm, M., 2020. Differentiating the time-geography of recreational running. Mobilities 14 (4), 575–587.
Radbo, H., Andersson, R., 2012. Patterns of suicide and other trespassing fatalities on state-owned railways in greater Stockholm; implications for prevention. Int. J. Environ. Res. Publ. Health 9 (3), 772–780. https://doi.org/10.3390/ijerph9030772.
Radbo, H., Svedung, I., Andersson, R., 2005. Suicides and other fatalities from train-person collisions on Swedish railroads: a descriptive epidemiologic analysis as a basis for systems-oriented prevention. J. Saf. Res. 36 (5), 423–428. https://doi.org/10.1016/j.jsr.2005.08.003.
Sweden, T.P. H.A.o., 2020. Statistik Om Suicid. Retrieved from. https://www.folkhalsomyndigheten.se/the-public-health-agency-of-sweden/living-conditions-and- lifestyle/suicide-prevention/.
Taylor, E., 2010. I spy with my little eye: the use of CCTV in schools and the impact on privacy. Socio. Rev. 58 (3), 381–405. https://doi.org/10.1111/j.1467- 954X.2010.01930.x.
Too, L.S., Spittal, M.J., Bugeja, L., McClure, R., Milner, A., 2016. Individual and community factors for railway suicide: a matched case–control study in Victoria, Australia. Soc. Psychiatr. Psychiatr. Epidemiol. 1–8. https://doi.org/10.1007/s00127-016-1212-9.
Uittenbogaard, A., Ceccato, V., 2014. Safety in Stockholm’s underground stations: an agenda for action. Eur. J. Crim. Pol. Res. 20 (1), 73–100. https://doi.org/ 10.1007/s10610-013-9212-8.
Uittenbogaard, A., Ceccato, V., 2015. Temporal and spatial patterns of suicides in Stockholm’s subway stations. Accid. Anal. Prev. 81, 96–106. https://doi.org/ 10.1016/j.aap.2015.03.043.
van Houwelingen, C.A.J., Beersma, D.G.M., 2001. Seasonal changes in 24-h patterns of suicide rates: a study on train suicides in The Netherlands. J. Affect. Disord. 66 (2), 215–223. https://doi.org/10.1016/S0165-0327(00)00308-6.
Vrotsou, K., Ellegård, K., Cooper, M., 2009. Exploring time diaries using semi-automated activity pattern extraction. Int. J. Time Use.Res. 6 (1), 1–25.
Vrotsou, K., Ynnerman, A., Cooper, M., 2013. Are we what we do? Exploring group behaviour through user-defined event-sequence similarity. Inf. Visual. https://doi. org/10.1177/1473871613477852.
Wikstr¨om, P.-O., Ceccato, V., Hardie, B., Treiber, K., 2010. Activity fields and the dynamics of crime. J. Quant. Criminol. 26 (1), 55–87. https://doi.org/10.1007/ s10940-009-9083-9.
Zelniker, E.E., Gong, S., Xiang, T., 2008. Global Abnormal Behaviour Detection Using a Network of CCTV Cameras. In: The Eighth International Workshop on Visual Surveillance - VS2008, Marseille, France. Paper presented at the.
