The methods and procedures for the Stanley Foundation Bipolar Treatment Outcome Network (SFBN) study (1995–2002) have been previously described in detail [24, 25]. All participants provided written informed consent, approved by the respective local Institutional Review Boards (IRBs), prior to study enrollment. Study details are described in brief below.

A total of 935 adults were recruited from four U.S. and three European sites for this naturalistic, longitudinal study, which involved the prospective assessment of clinical mood states. All participants were diagnosed using the Structured Clinical Interview from the DSM-4 [10]. Patients were evaluated at least monthly, with treatment adjustments made as clinically indicated. Patients met DSM-4 criteria for bipolar I disorder (BD-I), bipolar II disorder (BD-II), bipolar disorder not otherwise specified (BD-NOS), or schizoaffective disorder–bipolar type. Patient characteristics stratified by bipolar disorder subtype are presented in Table 1. An interrater reliability was maintained between clinical sites (κ values: YMRS, 0.7 and IDS-C, 0.85) [4].

Individuals who participated in defined, open-label, or double-blind clinical trials were excluded from the cohort [24, 26]. The current analyses focuses on patients who completed at least one visit with concurrent assessments of manic and depressive symptoms using the Young Mania Rating Scale (YMRS) [27] and the Inventory of Depressive Symptomatology–Clinician-Rated (IDS-C) [28, 29].

Mood states were classified using scores from the YMRS and the IDS-C, both of which are clinician-administered and have demonstrated strong reliability and validity [27,28,29]. These measures were completed during study visits, with symptoms assessed on the day of visit and for the preceding 3 and 7 days for YMRS and IDS-C, respectively.

To identify mood states, we applied symptom-based cut-offs derived from prior work which used YMRS and IDS-C thresholds in a naturalistic study of SFBN data (see Table 2) [4, 30]. These mixed state definitions are not formal diagnostic criteria, but rather dimensional, symptom-based constructs operationalized for the purposes of this study. Depressive symptoms were defined as an IDS-C R score of 15 or higher, indicating at least mild depression. Depression with mixed features was defined as the co-occurrence of depressive symptoms (IDS-C R ≥ 15) alongside mild hypo/manic symptoms, as indicated by YMRS scores between 3 and 11. Hypo/mania with mixed features was defined as the presence of depressive symptoms (IDS-C R ≥ 15) with moderate hypo/manic symptoms, characterized by YMRS scores of 12 or greater.

For depression symptom categorization, we used the full IDS-C score; however, for statistical analyses, we utilized a modified version of the IDS-C (IDS-C R) in which item 18 (suicidal ideation) was removed to avoid circularity, ensuring that the predictor (depression severity) does not include variance from the outcome variable. The IDS-C and IDS-C R scores were highly correlated (r = 0.99, p < 0.0001).

The YMRS and IDS-C share several items that assess overlapping symptoms, including sleep disturbances, irritability, and psychomotor agitation. To determine whether the observed associations were driven by these overlapping symptoms rather than distinct mood state effects, we conducted a sensitivity analysis in which these items were removed. Specifically, we excluded items 4 (Sleep) and 5 (Irritability) from the YMRS, and items 1–3 (Insomnia), 6 (Irritability), and 24 (Psychomotor agitation) from the IDS-C in subsequent analyses. This allowed us to isolate the relationship of interest from potential artifact.

SI was assessed using item 18 of the IDS-C, which evaluates the presence and severity of suicidal thoughts. Item 18 is scored as follows: 0 indicates no thoughts of suicide or death; 1 indicates a sense of emptiness or the belief that life is not worth living; 2 indicates suicidal thoughts occurring several times per week for several minutes; and 3 indicates suicidal thoughts occurring daily, with or without a plan or past suicide attempt. For our analyses, we dichotomized suicidal ideation to IDS-C item 18 = 0 and IDS-C item 18 ≥ 1.

Statistical analyses were conducted using SPSS version 28 (IBM Corp., Armonk, NY, USA). Generalized estimating equations (GEE) were used to examine the association between mood symptoms—specifically depressive, hypo/manic, and their co-occurrence—and SI assessing both the strength and direction of these relationships. The primary analyses utilized categorical models, with mood symptoms categorized based on predefined thresholds that capture distinct mood presentations.

The GEE models were specified using binary logistic regression with a robust covariance estimator and an independent working correlation matrix to account for repeated observations within individuals. The dependent variable in all models was the IDS-C item 18 score, which reflects the presence and severity of SI. The primary predictor variables were the categorical indicators of depressive, moderate, and mild hypo/manic symptoms. An interaction term was included to capture the potential synergistic effect of co-occurring depressive and hypo/manic symptoms.

The interpretation of the GEE model coefficients followed conventional guidelines for binary logistic regression. A positive β-value indicates an increased likelihood of suicidal ideation as symptom severity increases, whereas a negative β-value suggests a decreased likelihood. Odds ratios (OR) were derived from the β-coefficients using the standard transformation OR = e(β) providing a measure of odds for SI based on mood symptom severity. A statistically significant interaction term indicates that the relationship between mood symptoms SI differs from the additive contributions of depressive and hypo/manic symptoms. Age and study week were included as linear covariates in all models to control for potential time- and age-related effects. Additional covariates included gender, both as a main effect and in interaction with mood symptoms, and bipolar subtype.

To assess the combined effect of depressive and hypo/manic symptoms, we calculated combined ORs by exponentiating the sum of the relevant β-coefficients (e.g., ORcombined = exp(β_depression + β_hypomania + β_interaction)). This allowed us to quantify how co-occurring depressive and hypo/manic symptoms modify SI beyond their individual contributions. For combined OR calculations, only statistically significant β-values were included to ensure that the resulting estimate reflected meaningful associations. If a β-coefficient was not significant (p > 0.05), its corresponding OR was not included in the combined calculation, as its contribution to the overall effect was uncertain. Similarly, if the revised scales (e.g., excluding overlapping items) altered statistical significance, the corresponding OR was excluded, ensuring that the final combined OR was based only on robust, validated effects.

In addition to the primary categorical models, secondary analyses were conducted using continuous YMRS and IDS-C R scores as predictor variables to replicate methodology from prior research [21,22,23, 31]. The continuous models followed the same GEE specifications as the categorical models; however, instead of dichotomizing symptoms, mood symptoms were modeled as continuous variables to capture variability in symptom severity without the use of predefined clinical categories. The interaction between continuous depressive and hypo/manic symptom scores was included to identify potential nonlinear relationships and mixed symptom presentations.