The increased rate of suicide attempts prior to study entry in participants with bipolar disorder appears to largely reflect an earlier age of onset which has been previously noted with bipolar disorder (McMahon et al., 1994
). While a trend was noted in our retrospective analyses of increased prior suicide attempts with bipolar disorder compared with unipolar disorder, the differences observed were statistically significant, after controlling for age of onset, only for severe suicide attempts. These findings differ somewhat with an analysis of the Epidemiologic Catchment Area database which revealed an excess of suicide attempts in individuals with bipolar disorder even after controlling for age of onset (Chen and Dilsaver, 1996
). While frequency data may have been suggestive of increased suicidal behaviors during follow-up in participants with bipolar disorder, polarity did not independently predict suicidal behaviors when modeled in a mixed-effects grouped-time survival analysis. Age at cycle onset, hopelessness at study intake, and active substance abuse did independently predict suicidal behavior during mood cycles. The likelihood ratio tests revealed no significant main effect by polarity interactions and therefore demonstrated no evidence of differential effects of risk factors for suicidal behavior by polarity.
Prior studies have contrasted risk of suicide between those with bipolar and unipolar disorders with mixed results. Several studies have assessed suicide risk factors in unipolar and bipolar disorders (Fawcett et al., 1990
; Krupinski et al., 1998
; Schneider et al., 2001
; Angst et al., 2002
; Oquendo et al., 2007
; Tondo et al., 2007
), and some have conducted separate analyses by polarity (Black et al., 1988
; Kallner et al., 2000
; Angst et al., 2005a
). An assumption of the former studies is that risk factors do not vary by diagnosis and the latter studies do little to validate this assumption. To date, the assumption that established risk factors do not vary by polarity has not been empirically tested in a well-defined prospective cohort. None of these studies have statistically assessed, using an interaction term, differential effects of risk factors by polarity. A community sample of registered admissions did assess for differential effects of risk factors by polarity using an interaction term, but given limitations in available data was not able to test for any traditionally established risk factors apart from age, instead focusing on variables related to time from hospital admission (Hoyer et al., 2004
). Another study assessed interactions though did not assess differential affects by diagnosis or polarity (Young et al., 1994
). Our study uniquely addresses this long-maintained assumption for some established suicide risk factors and supports the now empirically-validated assessment of risk factors independent of polarity in affective disorders.
Of the 4,204 cycles assessed, 40 ended in suicide and 384 contained suicide attempts. To ensure adequate power to assess the differential effect of risk factors for suicidal behavior by polarity, we utilized suicidal behavior, representing suicide attempts or completions, as an outcome. The inclusion of suicide attempts may nonetheless limit the generalizability of our findings to the prediction of risk for completed suicide. Those who complete suicide may represent an overlapping but not identical population from those who attempt suicide (Beautrais, 2001
). Our focus on polarity further did not allow the assessment of a differential effect of risk factors by bipolar subtype. The lower cumulative depression burden in those with bipolar disorder reflects those with bipolar I, but not bipolar II spending less time depressed. Comparison by bipolar subtype would further allow assessment of other risk factors, such as mixed states or cycling. Our analyses controlled for a linear effect of age on suicide, which may not adequately control for a non-linear age effect. A significant linear effect of age was nonetheless revealed and analysis of the data did not suggest a non-linear age effect. The adjustment of our diagnosis for prospective changes in the psychiatric status of participants strengthens the comparison of these groups by reducing the risk of misclassification. Length of follow-up considered, our observed rate of conversion from unipolar to bipolar disorder is generally consistent with the previous literature (Akiskal et al., 1983
; Akiskal et al., 1995
; Goldberg et al., 2001
; Angst et al., 2005b
). Screening of vital statistics and obtaining death certificates reduces our risk of misclassification for suicide; however, it remains possible that those who died of unknown causes may have actually died of suicide and have been misclassified. The Collaborative Depression Study uniquely provides comprehensive demographic, diagnostic, and phenomenological data for a large clinical sample of individuals with affective disorders followed for an extended period of time. Other advantages of this sample include rigorous clinical evaluations and low rates of loss to follow-up.
The mixed-effects grouped-time models used in this analysis pose several advantages for the assessment of suicide risk factors in a longitudinal study of this duration. Many studies in psychiatry using survival analysis have utilized more traditional time-to-event approaches such as the logrank test or Cox proportional hazards models (Leon et al., 1990
). These approaches assume independence among observations and therefore cannot include repeated observations per participant. To accommodate changes in risk factors over time, we changed the unit of analysis to mood cycle, allowing risk factors to be re-assessed each cycle. The mixed-effects grouped-time survival method represents a modern approach to examine correlated observations, such as multiple mood cycles, in a single model (Hedeker et al., 2000
). This allows for correlation in within-subject mood cycles and for the number of mood cycles to vary widely. All prospectively observed mood cycles were able to be analyzed together, allowing us to assess polarity as a predictor of suicidal behavior and a differential effect of risk factors by polarity within a mood cycle. Many of these risk factors vary over time and our statistical modeling was able to account for this.
The lack of any independent or differential effects of polarity on suicide risk poses several clinical and public health implications. Previous research has suggested that suicide risk factors may differ between those with alcoholism and those without (Murphy et al., 1992
). Our data does not support any such differences by polarity in affective disorders. While the presence of an affective disorder shapes suicide risk assessment, polarity may be of limited relevance for suicidal risk or the effect of other suicide risk factors. This information may inform current efforts to develop standardized tools to augment the clinical suicide risk assessment. In the light of growing interest in the use of standardized suicide assessment psychometrics, it will be important to clearly delineate that suicide risk factors do not vary by diagnosis. While the current study suggests no differences in risk factors between the bipolar and unipolar disorders, these findings may not generalize outside of the affective disorders. Suicide risk remains difficult to predict (Goldstein et al., 1991
) and public health initiatives should continue to target the identification and appropriate treatment of affective disorders as a general preventative approach (Rihmer, 1996
; Mann et al., 2005