All data for this exploratory analysis were drawn from the STEP-BD study; study characteristics have been described in detail elsewhere (Sachs et al. 2003). Briefly, the 7-year study followed 4360 outpatient participants, ages 15 years or older, with BD of every subtype. The study was conducted across 22 academic medical centers in the United States and was approved by the institutional review board of the participating institutions. All participants provided written informed consent.

The primary study measure was the Clinical Monitoring Form (CMF) (see below). After excluding participants who had no CMF data, our initial dataset included 242 individuals who attempted suicide and eight who died by suicide. Two transgender individuals, both of whom were in the group exhibiting no suicidal behavior, were removed from the analysis due to small sample size. The data were restricted to only include individuals who completed at least three study assessments and who remained in the study for more than 60 days (113 individuals with a suicidal behavior event (SB group) and 3081 individuals without an event (No SB group). The data were further restricted to include only individuals who reported any SI during the study period (which varied by participant). This final dataset comprised 92 participants (64.1% female) in the SB group and 1863 participants (58.9% female) in the No SB group.

Data on suicidal behavior, defined as either a suicide attempt or a death by suicide, were collected from the Severe Adverse Events (SAE) and Care Utilization (CU) forms as previously described (Ballard et al. 2020).

The CMF, which was the primary outcome measure in the STEP-BD study, was used to evaluate participants at each outpatient visit with their treating psychiatrist. The CMF measures the severity of DSM-IV mood symptoms (including SI) and other pertinent clinical characteristics within the prior 2-week period. The form has been well-validated and correlates strongly with other mood rating scales such as the Hamilton Depression Rating Scale and the Montgomery-Åsberg Depression Rating Scale (Sachs et al. 2002). Data from the CMF were merged with data from the Affective Disorder Examination (ADE), a semi-structured interview conducted at study entry, which contained the same clinician-administered rating scales found on the CMF. Additionally, information on substance use was drawn from the MINI International Neuropsychiatric Interview (MINI), a brief structured interview for the major Axis I psychiatric disorders also conducted at study entry. Additional details about these forms can be found in the Additional file 1: Methods.

Three explanatory variables were used to examine the relationship between SI characteristics and odds of a future suicidal behavior event: SI variability, SI intensity, and SI persistence. All three were derived from the SI scale of the CMF. Given the ordinal quality of the scale, a statistic for ordinal dispersion was used to assess SI variability over the study period (Blair and Lacy 2000). Briefly, the statistic defines an index of ordinal concentration by taking the sum of the squared difference of the cumulative relative frequency of each level of the scale and the state of maximal dispersion (which is equal to 0.5), and then normalizing by the value for maximal dispersion (given the number of levels of the scale). The difference of this number and one then provides a measure of dispersion of the ordinal variable on a scale from 0 (no dispersion) to 1 (maximal dispersion), where the produced values can be interpreted as the percent of the maximal possible dispersion for the given variable. This approach makes no continuous assumptions about the data and has been shown to produce little bias in estimates with large sample sizes, as is the case with our data (Blair and Lacy 2000). SI intensity was measured as the median SI score per participant over the study period, and SI persistence was measured as the proportion of visits with any reported SI out of the total number of visits over the study period for each participant. Using the ordinal dispersion statistic, measures of affective instability were derived from the DSM-IV mood symptom scales separately for mania and depression (excluding SI), in line with previous approaches (Stange et al. 2016a, b). Severity of depressive symptoms over the study period was measured by summing the absolute value of depressive symptom scores at each visit, again excluding SI, and then averaging these total scores over the study period.

Because past studies have noted a relationship between Personality Disorders Questionnaire (PDQ) score and suicidal behavior in the STEP-BD study (Antypa et al. 2013), this measure was also included (see Additional file 1: Methods).

Summary statistics were calculated for demographics and SI variability, SI intensity, and SI persistence for the SB and No SB groups. Logistic regression was used with suicidal behavior as the outcome variable, and each SI measure of interest (SI variability, SI intensity, or SI persistence) was used as the explanatory variable in separate models to examine the relationship between SI characteristics and likelihood of a future suicidal behavior event.

Covariate selection was conducted to minimize the potential for spurious relationships between predictors of interest and outcome, and Pearson correlation coefficients were calculated with variables of interest and covariates to estimate potential sources of both confounding and collinearity. Based on literature supporting their relationship with either suicidal behavior, SI, or both, the following covariates were included in all models: age at study entry, gender, history of suicide attempt, PDQ score, severity of depressive symptoms, affective instability in depressive and manic symptoms, and alcohol or other substance abuse at study entry (Antypa et al. 2013). Number of total study visits was also included in the models for SI variability, SI intensity, and SI persistence to address possible confounding. To test the hypothesis that affective instability would modify the relationship between SI variability and suicidal behavior, each affective instability measure was interacted with SI variability. As a sensitivity analysis to protect against spurious relationships due to time in study, all of the models were run with only participants who exited the study within two years of study entry; results were similar (see Additional file 1: Table S1).

Multiple imputation was used to account for missingness in the data, which included PDQ score (SB group: 29.8%; No SB group: 21.7%), history of suicide attempt (SB group: 3.2%; No SB group: 1.1%), alcohol abuse (SB group: 6.3%; No SB group: 7.6%), and substance abuse (SB group: 6.4%; No SB group: 7.6%). Regression results from both complete case analyses and multiple imputation analyses were found to be similar. Pooled multiple imputation results are reported below; see Additional file 1: Methods for additional information and Additional file 1: Table S2 for complete case results. All analyses were run using Stata version 16 (StataCorp, College Station, TX, USA).