Suicidal behavior is one of the leading causes of injury and death worldwide. Presently, suicide accounts for nearly one million deaths each year (WHO, 2012). For every death, there are an estimated 25 non-fatal suicide attempts and millions more who experience suicidal thoughts (Nock et al. Reference Nock, Borges, Bromet, Cha, Kessler and Lee2008; Crosby et al. Reference Crosby, Han, Ortega, Parks and Gfoerer2011). Over the last five decades, there have been marked increases in research focused on the understanding, treatment, and prevention of suicidal behavior. Despite these efforts, suicide rates remain virtually unchanged (CDC, 2015).
Prior self-injurious thoughts and behaviors (SITBs) are often identified as some of the most robust predictors of future SITBs. Among SITBs, suicide attempt history is consistently cited as one of the strongest risk factors for future suicidal behaviors (Fawcett et al. Reference Fawcett, Scheftner, Fogg, Clark, Young, Hedeker and Gibbons1990; Joiner et al. Reference Joiner, Conwell, Fitzpatrick, Witte, Schmidt, Berlim, Fleck and Rudd2005). This claim has become widely accepted and highly influential – for instance, a recent WHO (2014) report stated that ‘by far the strongest risk factor for suicide is a previous suicide attempt’. Supporting this position, a large body of research has consistently linked suicide attempt history to later suicidal ideation (Miranda et al. Reference Miranda, Gallagher, Bauchner, Vaysman and Marroquín2012; Links et al. Reference Links, Nisenbaum, Ambreen, Balderson, Bergmans, Eynan, Harder and Cutcliffe2012), attempts (Borges et al. Reference Borges, Angst, Nock, Ruscio and Kessler2008; O'Connor et al. Reference O'Connor, Smyth, Ferguson, Ryan and Williams2013), and death (Suokas et al. Reference Suokas, Suominen, Isometsä, Ostamo and Lönnqvist2001; Wenzel et al. Reference Wenzel, Berchick, Tenhave, Halberstadt, Brown and Beck2011). Some studies cite over a 70-fold increase in the likelihood of a subsequent attempt (Sanchez-Gistau et al. Reference Sanchez-Gistau, Baeza, Arango, González-Pinto, de la Serna, Parellada, Graell, Paya, Llorente and Castro-Fornieles2013) and close to a 40-fold increase in the likelihood of death (Harris & Barraclough, Reference Harris and Barraclough1997). However, in stark contrast to these findings, several studies have reported substantially smaller (Wenzel et al. Reference Wenzel, Berchick, Tenhave, Halberstadt, Brown and Beck2011; Van Dulmen et al. Reference Van Dulmen, Mata, Claxton, Klipfel, Schinka, Monica and Bossarte2013) or non-significant (Tejedor et al. Reference Tejedor, Diaz, Castillon and Pericay1999; Brådvik & Berglund, Reference Brådvik and Berglund2009) effects. These discrepancies raise doubt about the true effect of prior suicide attempts on future suicidal thoughts and behaviors.
Features of prior attempts [e.g. number, recency (i.e. time since last attempt), intent, lethality, method] have also been proposed to be important indicators of risk (Beck et al. Reference Beck, Kovacs and Weissman1979; Rudd et al. Reference Rudd, Joiner and Rajad1996; Joiner et al. Reference Joiner, Rudd and Rajab1997; Peruzzi & Bongar, Reference Peruzzi and Bongar1999). Yet large discrepancies exist in this literature as well, with some studies reporting moderate-to-strong effects (Nock et al. Reference Nock, Park, Finn, Deliberto, Dour and Banaji2010; Roaldset et al. Reference Roaldset, Linaker and Bjørkly2012; Troister et al. Reference Troister, Davis, Lowndes and Holden2013) and others small or non-significant effects (Sher et al. Reference Sher, Carballo, Grunebaum, Burke, Zalsman, Huang, Mann and Oquendo2006; Keilp et al. Reference Keilp, Oquendo, Stanley, Burke, Cooper, Malone and Mann2010; Chen et al. Reference Chen, Shyu, Lin, Chen, Ho, Lee and Chou2013).
There also is disagreement about the effects of other SITBs – namely, suicidal ideation, suicide plans, and non-suicidal self-injury (NSSI) – on the likelihood of future suicidal thoughts and behaviors. Although it is generally accepted that ideation confers risk for later suicidal thoughts and behaviors, the magnitude of its effects varies widely (Ramchand et al. Reference Ramchand, Griffin, Harris, McCaffrey and Morral2008; Anderson, Reference Anderson2011). The role of planning for suicide on later suicidal thoughts and behaviors also remains in question. Some researchers have proposed that planning for suicide is a necessary precursor of potentially lethal suicidal behavior (Witte et al. Reference Witte, Merrill, Stellrecht, Bernert, Hollar, Schatschneider and Joiner2008; Van Orden et al. Reference Van Orden, Witte, Cukrowicz, Braithwaite, Selby and Joiner2010) whereas others have argued that lack of planning and impulsivity confers greater risk (Maser et al. Reference Maser, Akiskal, Schettler, Mueller, Endicott, Solomon and Clayton2002; Simon et al. Reference Simon, Swann, Powell, Potter, Kresnow and O'Carroll2002). With respect to NSSI, early research focused heavily on distinguishing the behavior from suicidal outcomes (Brausch & Gutierrez, Reference Brausch and Gutierrez2010; Joiner et al. Reference Joiner, Ribeiro and Silva2012); however, emerging research indicates that the longitudinal effects of NSSI on suicidal behaviors may be much stronger than originally anticipated (Asarnow et al. Reference Asarnow, Porta, Spirito, Emslie, Clarke, Wagner, Vitiello, Keller, Birmaher, McCracken, Mayes, Berk and Brent2011; Wilkinson et al. Reference Wilkinson, Kelvin, Roberts, Dubicka and Goodyer2011).
The effects of indirect experiences with suicidal thoughts and behavior, such as a family history of suicidal thoughts and behaviors and exposure to the suicidal thoughts and behaviors of others (e.g. family, peers, etc.), also are unclear. Several studies report that family history of suicidal behavior is associated with a >10-fold increase in the likelihood of future suicidal behavior (Dahlsgaard et al. Reference Dahlsgaard, Beck and Brown1998; Soloff & Chiappetta, Reference Soloff and Chiappetta2012; Chan et al. Reference Chan, Shamsul and Maniam2014); however, many others have failed to find any effect (Valtonen et al. Reference Valtonen, Suominen, Mantere, Leppämäki, Arvilommi and Isometsä2006; Soloff & Fabio, Reference Soloff and Fabio2008; Dennehy et al. Reference Dennehy, Marangell, Allen, Chessick, Wisniewski and Thase2011). Similar discrepancies exist in the literature examining the effects of exposure to others’ SITBs (Lewinsohn et al. Reference Lewinsohn, Rohde, Seeley and Baldwin2001; Swanson & Colman, Reference Swanson and Colman2013), raising questions about suicide contagion and clustering as well as best practices for postvention.
Given all of these inconsistencies in the literature prompts the question: ‘What are the effects of prior SITBs on future suicidal thoughts and behaviors?’ The present meta-analysis addresses this crucial question and includes four primary aims. First, we provide a descriptive summary of the existing longitudinal literature addressing this question. We focus only on longitudinal studies because we are interested in determining whether prior SITBs confer risk for later suicidal thoughts and behaviors. Although cross-sectional and retrospective studies can provide insight about correlates (i.e. associated features) of a phenomenon, they are not useful in determining risk factors because temporal precedence cannot be inferred from these designs (Kraemer et al. Reference Kraemer, Kazdin, Offord, Kessler, Jensen and Kupfer1997). Second, we examine what, if any, effects prior SITBs, features of prior SITBs, family history, and exposure to SITBs have on future suicidal ideation, attempts, and death. Third, we evaluate the potential moderating effects of sample age, sample severity, and study follow-up length. We do so because differences in methodology related to these issues may influence effects or prior SITBs. Fourth, we consider these findings in the context of their clinical utility, quantified as: (1) improvement in diagnostic accuracy above chance and (2) odds ratios considered in terms of absolute risk of suicide ideation, attempts, and death.
Data sources, study selection, inclusion criteria
We conducted a comprehensive search through December 2014 using PubMed, PsycInfo, and Google Scholar. We then searched the reference sections of all papers identified through these sources. Search terms were variants of ‘longitudinal’ and ‘suicide’, which included: ‘longitudinal’, ‘longitudinally’, ‘predicts’, ‘prediction’, ‘prospective’, ‘prospectively’, ‘future’, ‘later’, and ‘self-injury’, ‘suicidality’, ‘self-harm’, ‘suicide’, ‘suicidal behavior’, ‘suicide attempt’, ‘suicide death’, ‘suicide plan’, ‘suicide thoughts’, ‘suicide ideation’, ‘suicide gesture’, ‘suicide threat’, ‘nonsuicidal self-injury’, ‘NSSI’, ‘self-mutilation’, ‘deliberate self-harm’, ‘DSH’, ‘self-cutting’, ‘cutting’, ‘self-burning’, and ‘self-poisoning’.
A total of 2179 unique papers were identified. Based on abstracts, 1578 studies did not meet inclusion criteria. Inclusion required that the paper include at least one longitudinal analysis predicting suicide ideation, attempts or death using any SITB variable in any population, any year, and any geographic location. Published peer-reviewed studies were excluded on the basis of three criteria: (1) analyses that were not longitudinal; (2) analyses that did not examine discrete suicide-relevant outcomes; and (3) analyses that were conducted in the context of a primary treatment study. After reading the remaining articles in full, a total of 172 studies were retained (see Fig. 1 for PRISMA flowchart). Supplementary Table S1 includes details regarding each eligible study; Supplement S2 includes a complete reference list for all eligible studies.
We elected to focus on suicidal ideation, attempts, and death as outcomes for two reasons. First, we were primarily interested in determining the effects of prior SITBs on suicidal outcomes.Footnote 1 Footnote † Accordingly, it was necessary that outcomes only include self-injurious thought or behavior outcomes that involved some level of suicidal intent. In other words, the thought or behavior must involve two key elements: (1) it is self-directed and (2) involves some non-zero level of intent to die. Therefore, variables that were not specific to suicidal SITBs [e.g. deliberate self-harm (i.e. self-directed injury with or without suicidal intent), parasuicide (i.e. self-directed injury with unclear suicidal intent), etc.] were not examined as outcome variables. Second, we were interested in understanding the specificity of effects on discrete suicide outcomes. As such, we required that outcome variables did not combine discrete forms of suicidal thoughts and behaviors. Of note, however, these variables (e.g. combined forms of SITBs, deliberate self-harm, parasuicide, etc.) were examined as predictors of suicide ideation, attempts, and death.
To ensure case independence, duplicate cases were removed (n = 8). This occurred when either the same data were re-analyzed across multiple publications and/or multiple follow-up assessments using the same predictors were included in a single study. In the former case, the most inclusive study was retained; in the latter, only estimates at the final assessment were retained.
All statistical tests were examined for each study. Any statistical test where a SITB variable was used to predict suicide ideation, attempt or death outcome within a study was termed a ‘prediction case’ and retained for analysis. There were a total of 494 unique prediction cases.
Data extracted from each eligible study included: authors, publication year, follow-up length, number of SITB participantsFootnote 2 , sample type (i.e. general population, clinical, self-injurious), sample age group (i.e. adult, adolescent, mixed), predictor variable, outcome type, and any statistics relevant to any longitudinal analyses using SITB to predict suicide ideation, attempts, or death. Regarding sample type, general population samples included participants not selected for a history of SITB or psychopathology; clinical samples included participants selected based on psychopathology; and SITB samples included participants selected based on SITB history. Category and subcategory codes were assigned to each predictor. Agreement was reached on all category assignments for each predictor by lead and co-authors.
We also considered the issue of study quality. Assessments of study quality are indicated when there is a substantial degree of methodological variability that may influence the accuracy of results. In particular, meta-analyses of treatment studies often warrant quality assessments as methodological differences among treatment studies are common. For instance, treatment studies can vary considerably with respect to design type (e.g. case-control, randomization, double-blinding), sample (e.g. severity, representativeness), and treatment provision (e.g. fidelity, adherence, compliance), among many other factors. These common methodological differences across treatment studies can substantially influence the accuracy of results. However, in contrast to many treatment meta-analyses, the present meta-analysis includes studies that are highly uniform. This is because inclusion criteria in the present meta-analysis constrained the study pool by requiring that eligible studies share a common core design (i.e. longitudinal) and outcome (i.e. suicidal ideation, attempts, or death).
Although finer-grained methodological differences still exist (e.g. length of follow-up, population severity, etc.), there are no established objective criteria that could inform a priori hypotheses about how these differences may relate to study quality. For instance, there is no objective precedence for determining how the specific length of follow-up (e.g. shorter or longer) or type of sample (e.g. general community, clinical, or self-injurious) impacts accuracy of prediction. To explore possible influences of these differences on meta-analytic results, moderator analyses were conducted. Heterogeneity across studies was also accounted for by using random-effects models in analyses. Both moderator analyses and random-effects models are described in detail below.
Meta-analyses were performed with Comprehensive Meta-Analysis, version 2 (CMA, 2014); diagnostic accuracy statistics were analyzed with MetaDiSc, version 1.4 (Zamora et al. Reference Zamora, Muriel and Abraira2014). When available, unadjusted estimates were used.Footnote 3 When odds ratios (ORs) were not reported, they were calculated based on correlations, independent group means, risk ratios, and 2 × 2 contingency tables. Hazard ratios (HRs) were analyzed independently.Footnote 4
Between-study heterogeneity was quantified by I 2 tests. As high levels of between-study heterogeneity were expected, a random-effects model was used for all meta-analyses. Fixed-effects models assume that the underlying (i.e. true) effect size is identical across studies; any observed variance is assumed to be a result of chance. In contrast, random-effects models assume a distribution of similar (but not identical) effects across studies. Accordingly, the combined effect estimated by a random-effect model represents the mean of a distribution of true effects rather than a single true effect. Whereas fixed-effects models only estimate within-study variance random-effects models estimate both within- and between-study variance. Given that systematic between-study variance is common and expected across studies (e.g. due to differences in population, methodology, etc.), random-effects models are typically more appropriate than fixed-effects models in meta-analyses. Systematic between-study variance (i.e. heterogeneity) is accounted for in random-effects models in the weighting and calculation of each prediction case.
Moderator analyses were used to test potential moderating effects of sample age groups, sample severity, or study follow-up length. Slightly diverging from moderation analyses in primary studies, moderation analyses in meta-analyses test whether variation in effect size among studies is associated with differences in selected covariates (i.e. moderators). To test the effects of moderators on effect estimates, we employed meta-regression using a random-effects model using unrestricted maximum likelihood. For sake of interpretability, we provide means and confidence intervals for each analysis involving a categorical moderator (i.e. sample age groups, sample severity) and beta coefficients and slopes for analyses using a continuous moderator (i.e. follow-up length).
Publication bias was quantified using multiple indices, including classic fail-safe N, Orwin's fail-safe N, Begg & Mazumdar's rank correlation test, and Egger's regression test, funnel plot symmetry, and Duval & Tweedie's trim-and-fill test. Diagnostic accuracy was evaluated in terms of sensitivity, specificity, likelihood ratios (LR+, LR−), pooled diagnostic OR, and area under the curve (AUC) using receiver-operated characteristic (ROC) curves. Estimates >3 standard deviations (s.d.) from the mean (n = 13; 2.63%) were omitted as outliers, resulting in a total of 481 unique prediction cases included in analyses. Of note, when outliers were included in analyses, results were highly consistent with those when outliers were excluded.
Studies spanned 1965–2014. Suicide attempt was the most common outcome (47.80%), followed by death (40.50%) and ideation (11.60%). Nine cases (1.90%) were classified as ‘protective factors’ rather than risk factors; as such, they were excluded from further analysis. The median number of SITB participants across studies was 79 (mean = 710.53, s.d. = 4441.78, range = 4–48 649); however, 4.65% of studies failed to provide clear information about number of SITB participants. Adolescent-only samples comprised 19.50% of all cases, most of which examined attempt (58.50%) and ideation (35.10%).
The median follow-up length across all studies was 52 months (mean = 82.52, s.d. = 102.29, range = 1 week–648 months). The most frequent follow-up interval was 25–60 months (21.40%). Approximately 20% had follow-up intervals of 121 months or longer. Less than 6% had follow-ups shorter than 6 months and <1% had a follow-up shorter than 1 month. Death outcome cases were associated with the longest follow-up (median = 120, mean = 138.99 months, s.d. = 128.94, range = 1–648), followed by attempt (median = 24, mean = 49.53 months, s.d. = 55.34, range = 1–360) and ideation (median = 24 months, mean = 47.16 months, s.d. = 53.52, range = 1–168).
With respect to sample severity, the majority of cases involved either self-injurious (42.60%) or clinical (38.00%) samples; only 19.30% involved general population samples. Death outcomes most commonly involved self-injurious samples (75.90%), followed by clinical (18.97%) and general population (5.13%) samples. Clinical samples were most common for attempt (56.52%); self-injurious (22.17%) and general population (21.30%) samples were equally represented. For ideation, general population samples were the most common (60.71%) followed by clinical (28.57%) and self-injurious (10.71%) samples.
Overall prediction and publication bias
For overall prediction analyses, the estimate reflects the pooled effect of all prior SITBs, regardless of category type, on the outcome of interest. A total of 54 prediction cases were included, resulting in a weighted mean OR of 2.07 [95% confidence interval (CI) 1.76–2.43]. Between-study heterogeneity was extreme (I 2 = 93.56%) and there was consistent evidence of publication bias. Accounting for publication bias, the effect would be reduced to an overall weighted mean OR of 1.33 (95% CI 1.10–1.61). See Table 1 and Fig. 2 for detailed publication bias statistics and funnel plots. No cases were available for diagnostic accuracy analyses.
OR, Weighted mean odds ratio.
Classic and Orwin's fail-safe N values represent the number of studies required to nullify the observed effects; Begg & Mazumdar's rank correlation test computes the rank order correlation between effect estimates and standard error; Egger's test of the intercept uses precision (i.e. the inverse of the standard error) to predict the standardized effect (i.e. effect size divided by the standard error). The size of the effect is reflected in the slope and bias is reflected in the intercept (B 0); Missing cases under Duval & Tweedie's trim & fill are the number of cases estimated as missing below the mean.
For OR analyses, 189 prediction cases were included, generating a weighted mean OR of 2.16 (95% CI 2.01–2.33). Heterogeneity was high (I 2 = 90.76%). Multiple indices indicated significant publication bias. Accounting for publication bias would reduce the weighted mean OR to 1.68 (95% CI 1.57–1.81). See Table 1 and Fig. 2.
A total of 57 cases included sufficient information for diagnostic accuracy analyses. Overall accuracy was slightly better than chance (AUC = 0.60). Sensitivity was 26% (95% CI 24–27), indicating extremely limited ability to identify true positive cases. Specificity was 87% (95% CI 87–88), suggesting the ability to identify true negative cases was acceptable; however, this is likely a statistical artifact resulting from the low-base rates of the outcome and predictors, which necessarily generate a high number of true negatives and low number of false negatives. The LR+ was 1.77 (95% CI 1.57–2.04), LR- was 0.84 (95% CI 0.80–0.88), and the pooled diagnostic OR was 2.25 (95% CI 1.85–2.73).
OR analyses included 144 cases. Heterogeneity was high (I 2 = 75.49%). Publication bias was indicated across multiple indices. The weighted mean OR was 1.54 (95% CI 1.39–1.71), and decreased to 1.51 (95% CI 1.36–1.68) when publication bias was considered. See Table 1 and Fig. 2.
Diagnostic accuracy analyses included 52 cases. Overall accuracy was poor (AUC = 0.62). This likely was a result of extremely poor sensitivity (10%, 95% CI 10-11). Specificity was acceptable (86%, 95% CI 86-86); however, this again is likely in part a statistical artifact. The LR+ was 1.57 (95% CI 1.37–1.81), LR- was 0.96 (95% CI 0.90–0.99) and pooled diagnostic OR was 1.76 (95% CI 1.45–2.15).
Prediction trends across years of research
The first studies using prior SITBs to predict ideation, attempts, or death were published in 1983 (Adam et al. Reference Adam, Valentine, Scarr and Streiner1983), 1971 (Bagley & Greer, Reference Bagley and Greer1971), and 1965 (Motto, Reference Motto1965), respectively. Meta-regressions based on 1-year intervals indicated no significant change in predictive ability across time for suicidal ideation (b = −0.05, p = 0.08), attempts (b = −0.006, p = 0.50), or death (b = 0.007, p = 0.15).
Risk factor category analyses
The full results of the risk factor category analyses are presented in Table 2. As shown, prior suicide ideation significantly increases the likelihood of all outcomes, with the strongest estimate for subsequent ideation (weighted mean OR 3.12). Effects were weaker in the prediction of attempts (weighted mean OR 1.88) and death (weighted mean OR 1.95).
a Estimates are not reported for analyses involving ⩽3 cases, as the small number of cases compromise the reliability of estimates; risk factor categories with <3 cases across all three outcomes are not listed in the table.
n, number of prediction cases; OR, weighted mean odds ratio; CI, confidence interval; dashes (i.e. ‘–’) indicate that information was not available; suicide attempt, history (yes/no) of suicide attempt; ambivalence, endorsing an equivalent wish to die v. wish to live during a suicide attempt; concerning reaction, regret of not dying as a result of a suicide attempt or expressing continued wishes to die by suicide following a suicide attempt; maximum intent, highest level of intent preceding any suicide attempt; maximum lethality, highest level of lethality associated with any suicide attempt; number, number of past attempts; preparations, evidence of preparatory behaviors preceding attempt; violent attempt, suicide attempt using a highly fatal method (e.g. firearm, jumping from height, hanging); recency, time elapsed since prior attempt; reason, reason attributed for attempting suicide; NSSI, non-suicidal self-injury; SITB, self-injurious thoughts and behaviors; DSH, deliberate self-harm.
A history of suicide plans was associated with significantly increased odds of suicide death (weighted mean OR 1.44). There were an insufficient number of cases to produce reliable estimates for other suicide-relevant outcomes.
A history of suicide attempt (presence or absence) was associated with significantly increased odds of all outcomes, with the strongest effects for attempt (weighted mean OR 3.61) and weakest for suicide ideation (weighted mean OR 1.58).
Suicide attempt features
When considered as a single construct, the weighted mean ORs were significant for attempt (weighted mean OR 1.60) and death (weighted mean OR 1.30) but not ideation (weighted mean OR 0.79).
The effects of specific features also were tested. Only estimates for attempts and death are presented as there were too few ideation cases. In the prediction of suicide attempts, number (weighted mean OR 1.92), recency (weighted mean OR 2.51), maximum lethality (weighted mean OR 1.54) and maximum intent (weighted mean OR 1.23) of past attempts were significant. General (rather than maximum) lethality and intent were not significant. Too few cases were available to provide reliable estimates for the remaining features [i.e. level of preparations (n = 2), onset of first attempt (n = 1), history of a violent attempt (n = 2), and concerning reactions following attempts (n = 3)].
In the prediction of suicide death, only history of a violent attempt (weighted mean OR 1.89) and intent (weighted mean OR 1.20) were significant. Number, lethality, level of preparations, degree of intoxication, and having a concerning reaction following an attempt were not. Reliable estimates could not be provided for attempt recency (n = 1), reason for past attempt (n = 3), and history of interrupted attempt (n = 1).
A prior history of NSSI was associated with significantly increased odds of suicide attempt (weight mean OR 4.27). There were insufficient cases to test associations with suicide ideation or death.
SITB and deliberate self-harm (DSH)
Several studies included more general predictors of SITB or DSH that did not specify what type of thought or behavior was present. A prior history of SITB was associated with significantly increased odds of suicide attempt (weighted mean OR 2.26) and death (weighted mean OR 2.04). DSH was associated with increased odds of suicide death (weighted mean OR 1.51). There were insufficient cases to test the other associations in this domain.
Family history of SITB
Family history of SITB was associated with significantly increased odds of suicide ideation (OR 2.13) and attempt (OR 1.57) but not death. Given the small number of prediction cases, results were not reliable for specific forms of family history (e.g. history of death v. attempt, etc.).
Exposure to SITB of others (e.g. friends, family members, schoolmates, etc.) was associated with significantly increased odds of suicide ideation (weighted mean OR 1.56) and attempt (weighted mean OR 2.05). There were too few cases available to provide reliable estimates for more specific forms of exposure.
Across all outcomes, weighted mean ORs were statistically equivalent across clinical (ideation: OR 2.81, 95% CI 1.98–4.00; attempts: OR 2.45, 95% CI 2.20–2.71; death: OR 2.10, 95% CI 1.49–2.94) and general population samples (ideation: OR 2.17, 95% CI 1.82–2.59; attempts: OR 2.27, 95% CI 1.98–2.60; death: OR 2.82, 95% CI 1.65–4.81). Estimates were substantially weaker among self-injurious samples (ideation: OR 0.84, 95% CI 0.17–4.17; attempts: OR 1.46, 95% CI 1.14–1.86; death: OR 1.42, 95% CI 1.31–1.54).
Predicting ideation, weighted mean ORs for adult (OR 2.38, 95% CI 1.26–4.49) and adolescent (OR 1.94, 95% CI 1.70–2.20) samples were not significantly different; the estimate for mixed samples was not significant (OR 1.67, 95% CI 0.02–102.14). Predicting attempts, adult (OR 2.28, 95% CI 2.05–2.53), mixed (OR 1.96, 95% CI 1.26–3.05), and adolescent (OR 2.08, 95% CI 1.85–2.33) samples were statistically equivalent. For death, the effect was strongest for adolescent samples (OR 2.52, 95% CI 1.74–3.65); adult (OR 1.69, 95% CI 1.42–1.99) and mixed (OR 1.35, 95% CI 1.21–1.50) samples were statistically equivalent.
There were no significant effects of follow-up length on the prediction of ideation (b = −0.001, p = 0.59) or death (b < 0.001, p = 0.79). For attempts, predictive ability significantly worsened as follow-up length increased, though the effect was small (b = −0.002, p < 0.001).
Our findings indicate that prior SITBs are statistically significant risk factors for suicide ideation, attempts, and death; however, effects were considerably weaker than anticipated. Overall weighted mean ORs for ideation and attempts were only slightly above 2.0; for death the estimate was near 1.5. Adjusting for publication bias further reduced estimates. Effects were consistent regardless of sample severity, sample age groups or study follow-up lengths. Diagnostic accuracy analyses were in line with these findings, with prior SITBs being characterized by acceptable specificity (i.e. true negative rate), extremely poor sensitivity (i.e. true positive rate), and AUCs marginally above chance. We emphasize that these findings reflect the effects of prior SITBs when studied within the narrow methodological constraints of the existing literature.
Beyond overall predictive power, we also evaluated the effects of specific risk factor categories. Prior suicide ideation was the strongest predictor of later ideation; NSSI and suicide attempt history conferred the most risk for later suicide attempts; and suicide attempt history and suicide ideation were among the strongest predictors of suicide death. Surprisingly, features of past attempts were relatively weak predictors, with most features producing estimates close to 1.5 and several others generating non-significant effects. Family history of SITBs also conferred relatively little risk for later suicide ideation or attempts and had no significant effect on suicide death. Exposure to SITBs was a stronger predictor of suicide attempts than anticipated with effects comparable to factors like prior suicidal ideation and number of past attempts.
Although some predictors did emerge as stronger than others, it is critical to evaluate these results with respect to their clinical utility. We evaluated clinical utility on two domains. First, we considered the improvement of diagnostic accuracy. As noted above, knowledge of SITBs offered only slight improvement above chance. Second, we considered the magnitude of OR estimates in terms of the absolute risk of suicide ideation, attempts, and death. Suicidal thoughts and behaviors are rare. The likelihood of death by suicide in the United States in a given year is 12.5/1 00 000 (i.e. 0.000125); attempts are estimated to be 25 times more likely (i.e. 0.0031). The strongest predictor in this meta-analysis was NSSI (weighted mean OR 4.27) predicting future suicide attempts. Based on this estimate, the presence of NSSI would still only increase the odds of an attempt to 0.013 – a figure still nearly zero. Moreover, these calculations are based on prediction over a 1 year interval; most clinicians are tasked with determining risk over a period of days or weeks.
The present findings suggest that, in terms of absolute (rather than relative) odds, prior SITBs are weak risk factors for future suicide ideation, attempts, and death at least within the narrow methodological confines in which they have been studied to date. It is possible, however, that prior SITBs are powerful predictors when considered in the context of other risk factors (e.g. psychopathology). Results from sample severity moderation analyses offer some insight into this issue. Studies of self-injurious and clinical samples typically involve more stringent comparison groups, thereby controlling for a host of risk factors not accounted for in general samples. If prior SITB were indeed stronger in the context of other risk factors, we would expect that effects of prior SITB would be substantially stronger in general samples. However, sample severity moderation analyses generated only small effects, suggesting that, even in the context of other risk factors, prior SITBs remain weak risk factors in an absolute sense. Results of these analyses were surprising as we had anticipated more pronounced differences between sample severity groups; however, we note again that these results only reflect the evidence of the existing literature and should therefore be interpreted within the narrow methodological constraints of researched published to date. Studies that assess this issue more directly and extend beyond existing methodological confines are needed. For instance, approaches that involve machine learning algorithms to combine large number of risk factors over a short follow up period may be particularly promising (Kessler et al. Reference Kessler, Warner, Ivany, Petukhova, Rose, Bromet, Brown, Cai, Colpe, Cox, Fullerton, Gilman, Gruber, Heeringa, Lewandowski-Romps, Li, Millikan-Bell, Naifeh, Nock, Rosellini, Sampson, Schoenbaum, Stein, Wessely, Zaslavsky and Ursano2015).
Based on the present meta-analysis, we recommend four primary directions for future research. First, features of prior attempts, family history of SITB, and exposure to SITB have received very little empirical attention. Widespread consensus exists in the field that prior attempt features (e.g. lethality, number, intent, etc.) in particular are extremely important risk factors (Joiner et al. Reference Joiner, Walker, Rudd and Jobes1999; Peruzzi & Bongar, Reference Peruzzi and Bongar1999). Our findings are inconsistent with this belief; however, we caution that results are based on remarkably few prediction cases within each specific feature category. As such, we strongly recommend prioritizing longitudinal studies of prior attempt features.
Second, there is a need for improved assessments, particularly for the assessment of intent and plans. The majority of intent cases used the Suicide Intent Scale (SIS; Beck et al. Reference Beck, Schuyler, Herman, Beck, Resnik and Lettieri1974); the majority of planning cases used the planning subscale of the same scale. The inclusion of planning within the SIS necessarily conflates these constructs, raising questions about the true effects of these features considered independently. Efforts focused on improving the understanding and measurement of specific features of attempts, particularly intent and planning, are critical.
Third, studies focused on acute or short-term prediction are needed. Presently, the average follow-up is nearly 7 years; yet, clinicians are tasked with determining risk over the period of hours or days. Our findings indicated that longer follow-ups did not improve predictive power and, in some cases, significantly weakened it. Moving forward, studies would benefit from incorporating shorter follow-ups with strong potential for producing clinically useful results (Glenn & Nock, Reference Glenn and Nock2014).
Fourth, clarifying the mechanisms underlying the relations between prior SITBs and future suicidal thoughts and behaviors would be informative. Viable mechanisms must account for the present pattern of findings. This includes (1) why prior SITB are generally weak predictors of future suicidal SITB and (2) why suicidal and non-suicidal SITBs confer statistically equivalent risk for future suicidal behaviors.
In sum, prior SITBs do confer risk for future suicidal ideation, attempts, and death; however, effects are substantially weaker than anticipated. When considered in the context of the extremely low prevalence of suicidal thoughts and behaviors and the clinical demands that call for very short prediction timeframes, prior SITBs may not improve prediction much beyond chance levels. Our findings highlight critical gaps in study design, assessment, and the mechanisms that lead prior SITBs to confer risk for future suicidal thoughts and behaviors. Future research must address these gaps in order to make significant progress in the prevention of suicidal thoughts and behaviors.
For supplementary material accompanying this paper visit http://dx.doi.org/10.1017/S0033291715001804.
This work was supported in part by funding from the Military Suicide Research Consortium (MSRC), an effort supported by the Office of the Assistant Secretary of Defense for Health Affairs (Award No. W81XWH-10-2-0181). This work was supported in part through a training grant (T32MH18921) from the National Institute of Mental Health. Opinions, interpretations, conclusions and recommendations are those of the authors and are not necessarily endorsed by the MSRC or the Department of Defense.
Declaration of Interest