Results are limited in four important ways. First, our analysis was based on retrospective reports. Enthusiasm for the strength of the risk index should consequently be reserved until it is cross-validated prospectively. Risk factors that vary over time (e.g., marital status, income) are of special concern because they might have changed after attempts. Retrospective recall of parental psychopathology also might have been influenced by attempts. Second, the single NCS-R question used to distinguish attempts from gestures may have yielded a less accurate classification of intent than one based on clinical assessment. Third, we considered only a restricted set of DSM-IV disorders and assessed neither their severity nor chronicity, which might have downplayed the associations between disorders and attempts. Fourth, the low base rate of attempts led to low statistical power, which caused us to work with coarse predictor classifications and an unweighted summary score to minimize risk of over-fitting the model, possibly resulting in underestimation of the strength of association between the risk index and attempts. Efforts to compensate for low power by including substantively meaningful variables in the risk index that were not statistically significant might have led to over-fitting.
It is noteworthy, in light of these limitations, that most of the significant socio-demographic risk factors – young age, low socio-economic status, previous marriage – are consistent with prior findings (Kessler et al., 1999
; Moscicki, 1997
; Petronis et al., 1990
), as is the finding that past attempts are powerfully related to future attempts (Brown et al., 2000
; Goldstein et al., 1991
; Hulten et al., 2001
; Moscicki, 1997
). The elevated risk of attempts among Non-Hispanic Blacks, in comparison, is not consistent with previous findings, although no previous research has examined race-ethnicity as a correlate of attempts among ideators.
The finding that respondent mental disorders are not reliably associated with suicide attempt over and above reports of parental psychopathology is surprising, given previous evidence that respondent mental disorders mediate the relationship of parental psychopathology with offspring suicide attempts (Brent et al., 1996
; Brent et al., 2002
). Quite a different pattern was found in associations with ideation in the total sample and with plans among ideators, where OR’s of parent disorders were much smaller than those of respondent disorders. This pattern is most plausibly interpreted to mean that the associations of respondent disorders with attempts are mediated by ideation and plans, while respondent perceptions of parent disorders might be markers either of latent causes (e.g., genetic influences) (Brent & Mann, 2005
; Statham et al., 1998
) or of features of respondent disorders that are not assessed in the CIDI (e.g., severity, chronicity). Adjudication among these competing possibilities in future research will require a more rigorous assessment of parent disorders and a more fine-grained dimensional assessment of respondent disorders.
The finding that information about diagnosis is not strongly related to attempts among ideators controlling for plans has practical importance because it is much easier to administer the simple NCS-R questions about suicide ideation and plans than to carry out diagnostic interviews. However, as previous research found disorder severity, which was not considered in our analysis, to be an important predictor of suicide attempts (Fergusson et al., 2005
; Oquendo et al., 2004
; Sokero et al., 2005
), future work is needed to evaluate whether measures of disorder severity would improve the accuracy of the risk index.
The presence of a suicide plan, often considered the preeminent indicator of imminent risk, did not emerge as the strongest correlate of attempts. This does not mean that plans need not be assessed, as they are important both for clinical management (e.g., treatment contracting, removing the means for suicide) and prediction of attempts. Yet 43% of attempters described their attempts as unplanned. This characterization has to be incorrect in a rigorous sense, as all suicide attempts, including impulsive attempts, are “planned” even if the “plan” occurred only a few seconds before the attempt. However, attempters who report not having a plan presumably mean that they lacked a plan conceived prior to the situation in which the attempt occurred. Future research should investigate this issue by debriefing “unplanned” attempters about the sequence of thoughts and decisions that led up to their attempts.
The finding that a simple retrospectively reported risk index is strongly related to 12-month suicide attempts among ideators would be of great clinical value if the index was corroborated prospectively, especially in light of concerns regarding the effectiveness of detecting and intervening with people at high risk of suicide attempts (Gaynes et al., 2004
; Kessler et al., 2005a
). These results are of special clinical value given that the risk factors in the final index can be assessed with relative ease in clinical settings and that most of the factors (e.g., age, ethnicity, income, marital status, history of parental psychopathology) are unaffected by attempts to “fake good.” The results reported here are sufficiently promising to warrant an attempt at large-scale prospective replication, for which a precedent exists (Pokorny, 1983
). To ensure sufficient statistical power, such an investigation should focus on people with suicidal ideation who are assessed with an expanded version of the current risk index and tracked to document subsequent attempts. To maximize generalizability, this research would ideally be conducted with a minimally selected sample, such as patients who screen positive for suicidal ideation in primary care settings. Analyses should evaluate more refined coding of the current risk factors as well as risk factors not studied here (e.g., a broader set of disorders, measures of disorder severity or chronicity). It would be useful to develop the model in one subsample and cross-validate it in another. It would also be useful to consider alternative methods of data collection, such as computer-assisted self-assessment, to improve completeness and honesty of reports (Greist et al., 1973
; Kobak et al., 1996
; Turner et al., 1998