People who are admitted to inpatient psychiatric facilities are routinely assessed for suicide risk (Sullivan et al. Reference Sullivan, Barron, Bezmen, Rivera and Zapata-Vega2005; Bowers et al. Reference Bowers, Banda and Nijman2010; De Leo & Sveticic, Reference De Leo and Sveticic2010; Worchel & Gearing, Reference Worchel and Gearing2010; Simon, Reference Simon2011). In the USA, the 2015 Hospital National Patient Safety Goals set by the Joint Commission into Patient Safety recommended that all patients admitted to psychiatric hospitals should have a risk assessment that ‘identifies specific patient characteristics … that may increase or decrease the risk for suicide’ (2015). The aim of this risk assessment is to identify ‘high-risk patients’ who might warrant closer monitoring or more clinical resources (Sokolov et al. Reference Sokolov, Hilty, Leamon, Hales, Simon and Hales2006; Vine, Reference Vine2009). Despite the widespread acceptance of the importance of suicide risk assessment of inpatients, little is known about the statistical properties of suicide risk assessment in this setting.
A 2011 meta-analysis of 29 controlled studies examining the clinical factors associated with inpatient suicide identified seven studies that used multiple risk factors to define high-risk and lower risk groups. The pooled odds of suicide among high-risk patients compared with lower risk patients was 10.9 with a sensitivity and specificity of 64% and 85%, respectively (Large et al. Reference Large, Smith, Sharma, Nielssen and Singh2011). Since then several large studies have described suicide risk categorization among psychiatric inpatients (Madsen et al. Reference Madsen, Agerbo, Mortensen and Nordentoft2012; Madsen & Nordentoft, Reference Madsen and Nordentoft2012; Hunt et al. Reference Hunt, Bickley, Windfuhr, Shaw, Appleby and Kapur2013; Lieb et al. Reference Lieb, Palm, Meyer, Sarubin, Mokhtari-Nejad and Riedel2014; Levi et al. Reference Levi, Werbeloff, Pugachova, Yoffe, Large and Davidson2016). Moreover, the 2011 meta-analysis lacked some elements of a contemporary data synthesis; it did not assess the strength of reporting in the primary research, did not include results obtained by personal communication and lacked an assessment of risk of bias due to study methods (Stroup et al. Reference Stroup, Berlin, Morton, Olkin, Williamson and Rennie2000; Moher et al. Reference Moher, Liberati, Tetzlaff and Altman2009).
We report a systematic review and meta-analysis of the studies describing high-risk models for suicide among psychiatric inpatients. Our primary aim was to assess the strength of suicide risk assessment to discriminate between high-risk inpatients compared with lower risk inpatients by meta-analytic estimates of the pooled odds of suicide among high-risk patients, the sensitivity and specificity, the positive predictive value and the area under the curve (AUC) statistics. Secondary aims were to explore potential moderators of between-study heterogeneity in the primary research and to systematically review the constituent risk factors used in high-risk models for inpatient suicide.
Methods
We conducted a registered meta-analysis (PROSPERO; CRD42016054207) according to PRISMA guidelines (Moher et al. Reference Moher, Liberati, Tetzlaff and Altman2009). We included peer-reviewed, published, cohort and controlled studies that combined two or more clinical factors to define a stratum of psychiatric patients at high risk of suicide. We defined inpatient suicide according to the convention used in primary research to include the suicide of patients in psychiatric facilities and those on approved or unapproved leave (Large et al. Reference Large, Smith, Sharma, Nielssen and Singh2011).
Search strategy
We searched titles and papers written in English or with English language titles and abstracts indexed in Medline from 1948 to week 2 January 2017, EMBASE from 1974 to week 2 January 2017 and PsycINFO from 1967 to week 2 January 2017 using the terms (suicid* AND hospit* or inpatient* or in-patient*) (Fig. 1). We hand searched the reference lists of included studies and of relevant review articles and meta-analyses (Cassells et al. Reference Cassells, Paterson, Dowding and Morrison2005; Bowers et al. Reference Bowers, Banda and Nijman2010; Large et al. Reference Large, Smith, Sharma, Nielssen and Singh2011; Walsh et al. Reference Walsh, Sara, Ryan and Large2015; Chan et al. Reference Chan, Bhatti, Meader, Stockton, Evans and O'Connor2016; Large et al. Reference Large, Kaneson, Myles, Myles, Gunaratne and Ryan2016; Madsen et al. Reference Madsen, Erlangsen and Nordentoft2017). ML and AC independently winnowed first by examination of title, then abstract, then full text (Fig. 1).
Flow chart of searches for studies reporting high-risk models for inpatient suicide.
Study selection
We included papers that: (i) reported on cohort or controlled studies of current psychiatric inpatients; (ii) compared the clinical factors of groups of surviving patients to those of patients who died by suicide; (iii) used two or more clinical variables to define a high-risk group for suicide either by use of an ad hoc scale with a cut-off point or by using multivariate statistical methods; (iv) reported the numbers of true positives, false positives, false negatives and true negatives or the sensitivity and specificity. Where other effect size data were reported, true and false positives and negatives were back-calculated and then confirmed with the authors of the primary research. Where studies reported multivariate statistics without reporting the diagnostic properties of the high-risk model, we included data proved by correspondence with the authors. We excluded studies that: examined suicide among outpatients or recently discharged patients; reported on suicide attempts but not suicides; examined non-psychiatric patient groups or exclusively examined biological or treatment-related risk factors.
Data extraction
Two authors (ML and HM) independently extracted the data. The effect size data recorded were the number of true positives, false positives, false negatives and true negatives. Most studies dichotomized the patients into high- and lower risk groups. Where multiple cut-off points were reported, the data with the highest odds ratio (OR) were used.
Moderator variables and the assessment of reporting strength
Three continuous moderator variables were collected to potentially explain between-study heterogeneity: the year the study was published (because suicide risk categorization might have improved over time); the number of potential suicide risk variables initially examined (because studies including a larger number of variables are prone to chance associations) and the number of variables in the high-risk model (because risk categories that rely on more detailed clinical information might be more accurate).
The risk of bias in the primary studies was assessed using a five-item strength of reporting scale derived from the Newcastle-Ottawa scale for the assessment of reporting strength of non-randomized studies (Wells et al. Reference Wells, Shea, O'Connnell, Peterson, Welsh and Losos2013). These variables were whether the study: (i) defined suicide using an external mortality database (rather than hospital mortality data because local suicide data are believed to underestimate suicide rates) (Walsh et al. Reference Walsh, Sara, Ryan and Large2015); (ii) recorded the clinical risk factors prospectively (rather than by examining medical records of suicide cases and controls); (iii) used a survival analysis (to control for time at risk in the tests for significance of individual risk factors); (iv) used a statistical model to define the high-risk group (rather than a cut-off point in a scale of equally weighted risk items); (v) reported the results of the risk categorization in the form of sensitivity and specificity or numbers of true and false positives and negatives (as opposed us needing to back-calculate these statistics from effect size data and/or obtaining them by correspondence). Studies were awarded one point for each item.
Data synthesis
Random-effects meta-analysis was used to calculate pooled estimates of the diagnostic OR for suicide among those who were assessed as being at high risk v. lower risk and the sensitivity and the specificity using Comprehensive Meta-Analysis (CMA; Version 3, Biostat, Englewood, New Jersey, USA). A random-effects model was chosen a priori for all analyses because of the differences in study populations and definitions of high-risk strata. Between-study heterogeneity in effect size was examined using the I 2 and Q-value statistics and prediction intervals (PIs). Between-group heterogeneity (sensitivity analysis) was examined without assuming a common within-study variance and the significance of between-group heterogeneity was determined with Q-value statistics.
Random-effects meta-regression (method of moments) was used to examine whether the year of publication, the number of variables initially considered and the number of variables in the high-risk model were associated with between-study heterogeneity. A meta-analytic estimate of the receiver operator curve and AUC was calculated using Meta-DiSc (Zamora et al. Reference Zamora, Abraira, Muriel, Khan and Coomarasamy2006).
The possibility of publication bias was assessed using Egger's regression (Egger et al. Reference Egger, Davey Smith, Schneider and Minder1997) and was quantified using Duval and Tweedie's trim and fill method (Duval & Tweedie, Reference Duval and Tweedie2000). The group of studies that directly reported true and false positives and negatives or sensitivity and specificity was compared with those where those figures were back-calculated from effect size data and/or obtained by correspondence.
Results
Searches
The searches identified 17 relevant studies reporting 18 samples (hereafter referred to as studies) of inpatients that were categorized by suicide risk after the suicides had occurred (Fig. 1, Table 1). There were no studies that prospectively examined the predictive properties of a suicide risk scale or predetermined set of risk factors. The earliest study was published in 1998, the median year of the studies was 2002 and the most recent was published in 2016. Eleven studies reported the effect size data directly and seven studies were included after back-calculation and/or correspondence with the authors. There was disagreement about the two of 72 effect size data points and two of 180 moderator or strength of reporting data points. These disagreements were resolved by re-examination by three authors (ML, HM and NM).
Studies reporting high-risk models for inpatient suicide
Study characteristics
The studies examined 191 944 people [mean per study 10 663, standard deviation (s.d.) = 29 651, median = 256] of whom 1718 died by suicide (mean per study 95.4, s.d. = 91.4, median = 56).
Six studies used a cohort design (Krupinski et al. Reference Krupinski, Fischer, Grohmann, Engel, Hollweg and Moller1998, Reference Krupinski, Fischer, Grohmann, Engel, Hollweg and Moller2000; Spiessl et al. Reference Spiessl, Hubner-Liebermann and Cording2002; Neuner et al. Reference Neuner, Schmid, Wolfersdorf and Spiessl2008; Madsen et al. Reference Madsen, Agerbo, Mortensen and Nordentoft2012; Lieb et al. Reference Lieb, Palm, Meyer, Sarubin, Mokhtari-Nejad and Riedel2014) and 12 studies used a case–control design. The six cohort studies and two studies with a nested case–control design (Lin et al. Reference Lin, Hung, Liao, Lee, Tsai and Chen2014; Levi et al. Reference Levi, Werbeloff, Pugachova, Yoffe, Large and Davidson2016) used independent variables (suicide risk factors) that were systematically recorded at the time of admission to hospital, while 10 studies ascertained risk factor data by a retrospective examination of the medical records of suicides and controls. The mean number of variables initially collected per studies in the primary research was 78.6 (s.d. = 74.9, range 14–272, median = 65), while the mean number of variables included in the high-risk models was 6.1 (s.d. = 3.2, range 2–16, median = 5). Regression models were used to define the high-risk groups in 11 studies, while seven studies used a cut-off score in a bespoke scale of equally weighted risk factors derived from either univariate or multivariate significance tests. The dependent variable of suicide was defined using external mortality register data in 13 studies, while five studies used hospital mortality data. All studies defined inpatient suicide as the suicide of a registered inpatient irrespective of leave status at the time of the death.
Meta-analysis
The pooled odds of suicide in high-risk groups compared with the lower risk groups was 7.1 [95% confidence interval (CI) 4.2–12.2] (Fig. 2) and the AUC was 0.83 [standard error (AUC) = 0.06] (Fig. 3) indicating a strong effect size (Rosenthal, Reference Rosenthal1996). Between-study heterogeneity in ORs was very high, Q-value = 143, df (Q) = 17, p value < 0.001, I 2 = 88.1%, 95% PI = 0.80–63.1 (Fig. 2). Two studies failed to identify any suicides in the high-risk group (Spiessl et al. Reference Spiessl, Hubner-Liebermann and Cording2002; Neuner et al. Reference Neuner, Schmid, Wolfersdorf and Spiessl2008). The first quartile, median and third quartile pooled ORs were, 3.4, 8.8 and 26.1 and the highest was 94.8. An analysis that excluded the two studies with no suicides in the high-risk group had an OR that was 37% higher than the sample of 18 studies (OR 9.7, 95% CI 6.0–15.6, I 2 = 85.3%). The 12 case-controlled studies had a pooled odds of 9.7 (95% CI 6.3–15.0, I 2 = 75.1%) and the six cohort studies had a pooled odds of 1.7 (95% CI 0.38–7.6, I 2 = 93.3%).
Forrest plot of studies reporting high-risk models for inpatient suicide. Footnote to Fig. 2. Diamond represents the pooled odds ratio and 95% confidence interval.
Meta-analytic receiver operator curve (ROC) and 95% confidence intervals of high-risk models for inpatient suicide.
An Egger's regression that included all 18 studies was not significant (intercept = 0.99, t-value = 0.76, df = 16, p value = 0.46), while a Duval and Tweedie's trim and fill method identified four hypothetically missing studies with a weaker association between high-risk status and suicide which, if included, would have returned a lower adjusted OR of 5.1 (95% CI 3.1–8.6). An Egger's test suggested publication bias toward studies with a higher OR after excluding the two studies with no suicides (intercept = 3.2, df = 14, t-value = 2.96, p value = 0.01).
The pooled sensitivity of a high-risk categorization was 53.1% (95% CI 38.2–67.5%, I 2 = 95.9) indicating that just over half of the suicides occurred in the high-risk groups. The pooled specificity of a lower risk categorization was 84.2% (95% CI 71.6–91.9%, I 2 = 99.9). Excluding the two studies with no suicides, an examination of the receiver operator curve plot shows that the sensitivity values ranged from 10% to almost 90% and specificity values ranged from 65% to 100%. The true-positive rate among the high-risk patients among the six cohort studies (positive predictive value) was 0.43% (95% CI 0.014–1.3%, I 2 = 95.9).
Meta-regression
Between-study heterogeneity in ORs was significantly associated with the year in which the study was published (coefficient = −0.079, standard error = 0.030, z-value = −2.67, p value = 0.008) with more recent studies tending to have lower ORs. Between-study heterogeneity was not associated with the number of variables that were initially examined (coefficient = 0.003, standard error = 0.004, Z-value = 0.91, p value = 0.36) or the number of variables included in the high-risk model (coefficient = 0.073, standard error = 0.092, Z-value = 0.80, p value = 0.42).
Strength of reporting
Studies that used external mortality databases to ascertain suicide cases had a higher pooled OR than studies that relied on hospital data, but this difference fell short of statistical significance (Table 2). The eight studies that prospectively collected risk factor data had significantly lower pooled odds of suicide than the 10 studies that used a retrospective examination of medical records (Table 2). The prospective studies included all six cohort studies, and this group of studies also had a much larger mean samples size – 31 387 (s.d. = 47 079) – than the controlled studies, mean sample size – 306 (s.d. = 430). There was no difference in the pooled odds according to whether high-risk groups were defined using a cut-off score on a score of equally weighted risk factors or used a statistically defined high-risk group. Studies that directly reported the diagnostic data in the form of the number of true and false positives and negatives, or by reporting the sensitivity and specificity had a higher pooled odds of suicide than studies where the diagnostic data were back-calculated and/or obtained by personal communication (Table 2). One study used a survival method to examine potential risk factors (Madsen et al. Reference Madsen, Agerbo, Mortensen and Nordentoft2012). No study scored fewer than 2 out of 5 on the strength of reporting scale, 11 samples scored 2, six scored 3 and one study scored 4. The 11 studies with a lower strength of reporting score (⩽2) had a lower pooled OR than seven studies with a higher (⩾3) strength of reporting score, but this difference was not statistically significant (Table 2).
Effect size according Strength of Reporting Items and Strength of Reporting Score
Risk factors included in the high-risk models
The 18 models included 110 variables. Sixteen of the 18 models included at least one risk factor relating to suicidal behavior. Thirteen models had an item for previous suicidal behavior and seven had an item for recent suicidal behavior. More broadly defined suicidality was included in two models. Depressed mood or an affective disorder was included in 11 models. Indicators of more psychiatric hospital care were present in eight models and male sex was included in seven. Current social stress or loss was in six models, schizophrenia was in five and being employed or in training was included in four models. Delusions were included in three models, as was older age. Being previously absent without leave, being married, being aggressive, being involuntarily hospitalized, having a shorter duration of illness, having a family history of suicide, lacking insight, objective signs of depression, previous psychological therapy, and significant medication side effects were included in two models. The 21 factors that were only present in a single model included suicidal ideation, suicide plans, female sex, personality disorder, substance use, treatment resistance and chronic mental illness.
Discussion
Research conducted over the last 20 years suggests a statistically strong association (OR 7.1) between high-risk strata and completed suicide among current inpatients. At first glance, this analysis of the combined research effort suggests that there will be about seven times the proportion of suicides among high-risk patients compared with lower risk patients. The AUC of >0.83 also indicates a large effect size (Dolan & Doyle, Reference Dolan and Doyle2000). It suggests that if a patient who had died by suicide while an inpatient and another patient who had not died by suicide were randomly selected from a sample of inpatients, there would be an 83% chance that the patient who died by suicide would have been deemed at higher risk of suicide than the patient who did not suicide.
The pooled OR was achieved with a sensitivity of 53.1% and a specificity of 82.4%, meaning that about one in six patients will be considered to be high risk, and that just over half of all the suicides can be expected to be in this group. Data from cohort studies suggested that the proportion of all high-risk patients that suicide (the positive predictive value) was 0.43%, meaning that about 1 in 230 high-risk inpatients will suicide before discharge. A positive predictive value of 0.49% can also be calculated using the meta-analytically derived sensitivity and specificity in this study and a base rate of one inpatient suicide per 676 admissions as estimated in a recent meta-analysis of inpatient suicide rates (Walsh et al. Reference Walsh, Sara, Ryan and Large2015). This is similar to the estimated positive predictive values of 0.4% (Levi et al. Reference Levi, Werbeloff, Pugachova, Yoffe, Large and Davidson2016) and 0.23% (Madsen & Nordentoft, Reference Madsen and Nordentoft2012) in the two largest primary studies of inpatient suicide. The estimated OR, the sensitivity and specificity and the positive predictive value can also be compared with those of a recent meta-analysis of suicide risk categorization among longitudinal cohorts of psychiatric patients (OR 4.84, sensitivity = 56%, specificity = 79%, positive predictive value = 5.5%) (Large et al. Reference Large, Kaneson, Myles, Myles, Gunaratne and Ryan2016). Although risk assessment might offer a stronger discrimination between suicide and non-suicide among inpatients than in longitudinal studies, the positive predictive value is much lower in inpatient settings as a result of the lower number of suicides over the shorter period of hospitalization.
Despite the encouraging nature of some these statistics, the extreme between-study heterogeneity means that it cannot be assumed that a future study or a direct application of suicide risk assessment to a population of inpatients would have similar results. The PIs we report suggest that there is 95% probability that a future study will lie somewhere between having no ability to predict suicides (with an OR of about 1) and having a very powerful ability to predict suicides with an OR of 63. Moreover, four factors suggest that our estimate of the strength of all of the diagnostic metrics is also likely to be higher than what might be achieved in practice.
First, all of the primary studies derived their high-risk groups retrospectively, leading to likely inflation of the effect size by overfitting due to chance capitalization in the high-risk models (Cohen, Reference Cohen1986). The primary studies we included were very vulnerable to this effect because they examined an average of almost 80 risk factors, all used p values of 0.05 or higher as a threshold for inclusion in the high-risk model, making some chance associations almost inevitable. Therefore, some factors included in the risk models would not be likely to contribute to the strength of the risk assessment in future studies (Buchanan, Reference Buchanan2008). These risk factors with chance associations with suicide might include some of the risk factors that were only used in a single high-risk model, but might also include some other factors that were used in several models. For example, older age was a risk factor in three high-risk models, despite evidence from other studies that suggests that older age is not a risk factor for inpatient suicide (Large et al. Reference Large, Smith, Sharma, Nielssen and Singh2011).
Second, the effect size data were likely inflated by publication bias – we found some statistical evidence for hypothetically missing studies, and found a higher effect size in studies that reported the diagnostic data directly when compared with those studies that we could only include after back-calculation and correspondence with the authors.
Third, the effect size was likely inflated by the 10 studies that collected risk factor data retrospectively and had a pooled OR that was over four times higher than the eight studies that collected risk data prospectively.
Finally, compounding the uncertainty about the statistical strength of suicide risk categorization is uncertainty about what risk factors might best be used in practice. The most consistently used factors in high-risk models were suicidal behavior and indicators of depression. Contrary to our expectations, we found no evidence that either the number of variables that were initially examined or the number of variables in the high-risk model was associated with predictive strength. This suggests that the number of suicide risk variables that might be meaningfully combined using regression methods or ad hoc risk scales is quite small. It might not be fruitful to use more complex models with many risk factors. The strength of the association between some individual risk factors and suicide also suggest that additional data have a limited effect on risk categorization. The risk factors depressed mood and previous self-harm have each been estimated by meta-analysis to confer a risk of suicide with an odds of about four in inpatient settings (Large et al. Reference Large, Smith, Sharma, Nielssen and Singh2011). Our pooled estimate, based on an average of six factors per model, did not double the diagnostic odds of either of these single factors alone.
Our estimates of the sensitivity and positive predictive value raise questions about how risk assessment should be used in inpatient settings. It is reasonable to assume that a risk assessment that relies on valid risk factors (such as the presences of depression and past history of suicidal behavior) will be associated with increased odds of suicide and therefore will provide some real information about the patient. However, if risk assessments are to be used, two caveats need to be clearly understood.
The first caveat is that validly classified ‘low-risk’ patients are still at real risk of suicide – our study suggests that almost half of all the inpatient suicides were not high risk. In an era when there is quite broad agreement that we should aim for zero suicides in health care (IIMHL, 2016), we need to be cautious about over-reliance on tests for future suicide that only identify about half of all future suicides and are potentially falsely reassuring about patient safety.
The second caveat follows from the low PPV that indicates that a large number of patients who will be classified as high risk will never go on to suicide. This suggests that a high-risk classification should not result in interventions that subject many patients, who will never suicide, to excessive intrusion or coercion.
Together these caveats suggest that although risk assessment does provide some information, it is not useful as a basis of clinical decisions. Information about risk might contribute to the clinicians thinking, and probably ought to inform discussions with the patient and their family about treatment options, but it is difficult to conceive of a suicide preventing intervention that should be provided to all high-risk patients (few of whom will suicide) and yet withheld from low-risk patients (among whom half of all suicides occur).
In practice, it is inevitable that some patients will raise our anxieties about suicide. Often this anxiety will not really be formed on a probabilistic basis; the patient will be making direct suicide threats or repeated suicide attempts. Current suicide threats and repeated self-harm or suicide attempts clearly indicate the patient's distress and need for treatment without recourse to speculation about what they might or might not do in the future.
Knowledge of the severe limitations of risk assessment raises more general issues about the basis of clinical decision making in inpatient settings. A thorough and sympathetic assessment of the patient's individual circumstances and current treatment needs should always form the basis of an inpatient treatment plan. The patient's perceived risk of suicide is likely to be relevant in these considerations but should not be determinative or even especially significant. The treatment plan should always be discussed and negotiated with the patient and his or her family and carers; the latter becoming particularly important if the patient is not competent to make treatment decisions.
The limitations of the naturalistic studies we included further complicate the real-world impact of these results. The meta-analysis was unable to consider the impact of any interventions that might have been provided to people who were categorized as being at high risk of suicide. Successful interventions provided to high-risk patients in the primary studies may have the effect of reducing the odds of suicide in that group. On the other hand, it is possible that highly restrictive interventions imposed upon some patients considered to have been high risk may actually have increased suicides, via a mechanism dubbed nosocomial suicide (Large & Ryan, Reference Large and Ryan2014). The extent of these possible effects cannot be estimated without studies that directly investigate the impact of providing increased resource allocation or closer clinical surveillance to high-risk patients.
Conclusion
Despite widespread enthusiasm for risk assessment in hospital settings, the psychometric properties of suicide risk categorization for suicide among psychiatric inpatients remain uncertain. Despite a strong statistical association, estimates of the sensitivity and positive predictive value are not encouraging. Future studies might test the predictive properties of a scale based on predetermined set of risk factors. Further research might achieve stronger and more reproducible risk categories using novel risk factors, perhaps even biological risk factors (Chang et al. Reference Chang, Franklin, Ribeiro, Fox, Bentley and Kleiman2016) or more sophisticated statistical techniques to combine clinical risk factors (Kessler et al. Reference Kessler, Warner, Ivany, Petukhova, Rose and Bromet2015). However, even significant improvements in the statistical discrimination might not yield a greatly improved positive predictive value because of the low base rate of inpatient suicide (Szmukler et al. Reference Szmukler, Everitt and Leese2012). Finally, it should not be forgotten that the ultimate value of risk categorization depends on its potential for application. Even a strong statistical discrimination between high- and lower risk groups lacks meaning if there is no rational basis for interventions that should be provided to patients categorized as high-risk (the vast majority of whom will not suicide) yet should not be given to lower risk patients (Large & Ryan, Reference Large and Ryan2012), among whom about half of all suicides occur.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291717002537.
Acknowledgements
The authors would like to thank Dr Bettina Hubner-Liebermann, Dr Chian-Jue Kuo, Dr Martin Lieb, Dr Tine Madsen, Tanja Schmidt, Dr Florian Seemuller, Dr Roc Tavcar for correspondence about their published research. The study was not funded and no funding body played any role in study design, data collection and analysis, the decision to publish or the preparation of the manuscript.
Declaration of Interest
Matthew Large: Dr Large provides expert opinion to courts on matters related to suicide. Amy Corderoy: No conflicts of interest to declare. Nicholas Myles: No conflicts of interest to declare. Hannah Myles: No conflicts of interest to declare. Michael Davidson: No conflicts of interest to declare. Mark Wieser: No conflicts of interest to declare. Christopher Ryan: Dr Ryan provides expert opinion to courts on matters related to suicide.
Contributions
Matthew Large: study conception, overall supervision, data acquisition, data analysis; preparation of the manuscript. Amy Corderoy: data acquisition, preparation of the manuscript. Nicholas Myles: data acquisition, data analysis, preparation of the manuscript. Hannah Myles: data acquisition, data analysis, preparation of the manuscript. Michael Davidson: data acquisition, study design, preparation of the manuscript. Mark Weiser: data acquisition, study design, preparation of the manuscript. Christopher Ryan: study design, critical review and preparation of the manuscript.
