Person-specific dynamics between negative emotions and suicidal thoughts

Abstract

Introduction:

Recent technology has enabled researchers to collect ecological momentary assessments (EMA) to examine within-person correlates of suicidal thoughts. Prior studies examined generalized temporal dynamics of emotions and suicidal thinking over brief periods, but it is not yet known how variable these processes are across people.

Method:

We use data EMA data delivered over two weeks with youth/young adults (N = 60) who reported past year self-injurious thoughts/behaviors. We used group iterative multiple model estimation (GIMME) to model group- and person-specific associations of negative emotions (i.e., fear, sadness, shame, guilt, and anger) and suicidal thoughts.

Results:

29 participants (48.33%) reported at least one instance of a suicidal thought and were included in GIMME models. In group level models, we consistently observed autoregressive effects for suicidal thoughts (e.g., earlier thoughts predicting later thoughts), although the magnitude and direction of this link varied from person-to-person. Among emotions, sadness was most frequently associated with contemporaneous suicidal thoughts, but this was evident for less than half of the sample, while other emotional correlates of suicidal thoughts broadly differed across people. No emotion variable was linked to future suicidal thoughts in more than 14% of the sample.

Conclusions:

Emotion-based correlates of suicidal thoughts are heterogeneous across people. Better understanding of the individual-level pathways maintaining suicidal thoughts/behaviors may lead to more effective, personalized interventions.

1. Introduction

Suicide remains a leading cause of death worldwide (CDC, 2020; Word Health Organization, 2021). Despite decades of research, mental health professionals’ ability to predict imminent risk of suicidal thoughts and behaviors remains weak (Franklin et al., 2017). One of the challenges to prediction in suicide is that suicide risk factors are idiosyncratic; with no single variable accounting for suicide risk equally across individuals (Kaurin et al., 2022; Kuehn et al., 2022a; Leenaars, 2002; Sewall & Wright, 2021; Franklin et al., 2017). A focus on averaged risk effects, estimated across people, may occlude idiographic within-person processes which could aid in the precise detection of individual-level correlates of suicide. The present study focuses on new statistical models applied to intensive longitudinal data to quantify heterogeneity in within-person processes.

Nomothetic approaches, such as multi-level models, are the most conventional method for identifying group-level suicide risk factors. These methods assume that effects apply equally across all individuals. For example, Al-Dajani et al. (2022) collected daily diaries for 28 days with 78 adolescents recently discharged from psychiatric hospitalization for suicide risk. Using multi-level models, Al-Dajani et al. (2022) found that certain coping strategies predicted next day suicidal thoughts. Although multi-level models allow for the simultaneous analysis of between- and within-person variance, coefficients from multi-level models reflect averaged effects across all participants. In other words, these estimates are constructed “top-down” at the group-level, and individual level heterogeneity is simply estimated as variance. Idiographic methods (n = 1), on the other hand, can directly model person-level heterogeneity while also estimating the degree to which associations are shared across people. Idiographic techniques can account for person-level heterogeneity through a “bottom-up” approach. They do so by first estimating person-specific, individual-level models and then generalizing to group-level inferences based on the person-specific models only if there is sufficient homogeneity in the individual-level models.

Findings from nomothetic and idiographic based studies have been discrepant. Specifically, in the Al-Dajani study discussed above, nomothetic analyses found specific coping behaviors (e.g., seeking professional help) were associated with next day suicidal thinking However, other recent studies have observed between-subject heterogeneity in within-person coping responses to suicidal thoughts (Kuehn et al., 2022a). This study, also based on daily diaries in the month following discharge from psychiatric hospitalization due to suicide risk, used a case series design of idiographic models of three participants. Findings suggested that the associations between coping behaviors and suicidal thoughts differed across the three individuals. Despite some evidence of heterogeneity, most research on suicidal thoughts and behaviors remains focused on identifying “common causes” for SITBs across people. If there is a large degree of heterogeneity in person-specific effects, the search for common causes is unlikely to be fruitful and may impede progress in identifying person-specific risk factors for SITBs.

The affect regulation hypothesis (Kuehn et al., 2022b) provides an ideal framework for estimating both group and person-specific effects. On the group-level, negative emotions are theorized to be a key driver of self-injurious thoughts and behaviors (SITBs) through a negative reinforcement pathway (Kleiman et al., 2018a). Negative emotions such as fear, shame, sadness, guilt, and anger have also been found to be a proximal predictor of SITBs (Armey, Crowther, & Miller, 2011; Mou et al., 2018; Victor et al., 2019). Indeed, a recent individual participant data meta-analysis of intensive longitudinal studies found support for this model (Kuehn et al., 2022b). However, negative emotions have often been treated as a composite measure of distress precluding the examination of specific affective states. This conflicts with clinical interventions which are often designed to target difficulties in specific affective states (e.g., behavioral activation for sadness/depression (Jacobson, Martell & Dimidjian, 2001) or exposure for fear/anxiety disorders (Abramowitz, Deacon, & Whiteside, 2019). Although at least one study examined specific negative emotions (i.e., anxiety/agitation and shame/self-hatred) as proximal predictors of suicidal thoughts and behaviors (Bentley et al., 2021), recent person-centered research has found that the measurement structure of affect, often conceptualized in a bivalent negative versus positive fashion, differs across individuals ranging from two to eight factors (Foster & Beltz, 2022). This implies that individuals may vary in their daily experience of emotions and potentially their relationship to suicidal thoughts as a result.

One “bottom-up” analytic technique that models both group and individual effects is Group Iterative Multiple Model Estimation (GIMME; Gates & Molenaar, 2012). Prior work using GIMME has shown within- and between-person differences in domains of affect and behavior in borderline personality disorder (Dotterer et al., 2020; Beltz & Gates, 2017; Wright et al., 2015), alcohol use problems (Foster, 2019), and in general models of physical and mental health (Kelly et al., 2020). Recent work by Kaurin et al. (2022) used GIMME to demonstrate significant heterogeneity in links between suicidal ideation and composite negative affect (among other variables) within individuals diagnosed with borderline personality disorder. In this study, Kaurin et al. (2022) identified two distinct subgroups of participants, who differed in their relations between hostility, negative affect, and positive affect. There was no evidence for any group-level associations with suicidal thoughts.

Results from studies that use this blend of nomothetic and idiographic modeling can provide a more accurate understanding of individual-level effects within the group. For example, how relevant and potent is each risk for each individual sampled? Are there common patterns in risk relevance and potency across people as the affect regulation hypothesis posits? In this way, results from this line of work can inform group-level theories, as well as group-level prevention efforts. The present study used GIMME models in ecological momentary assessment (EMA) data to model both group- and individual-level associations between specific negative emotions (i.e., fear, shame, sadness, guilt, and anger) and suicidal thoughts among a high-risk for suicide sample. Given evidence demonstrating considerable individual-level heterogeneity (Kuehn et al., 2022a; Kaurin et al., 2022), we hypothesized that our group-level model would be sparse (i.e., few risks would explain SITBs uniformly across high-risk individuals), but that our individual-level models would be relatively denser, reflecting significant individual-level heterogeneity (i.e., links between specific components of negative affect would have different relevance and potency in explaining variation in SITBs across people).

2. Materials and methods

2.1. Participants

The present study used data from a project that recruited youth/young adults at risk for suicide through a combination of sources. Most of the 60 original participants (76.27%) were recruited from online sources (e.g, Facebook, Instagram, etc.) while others were recruited through flyers in the community and local treatment providers. Participants expressed interest in the study through an online screening survey where they reported on suicidal thoughts and behaviors in the past month. Eligible participants were contacted to attend an in-person baseline session in which they provided informed assent/consent, completed self-report and semi-structured assessments, and received training in the EMA protocol. Data was collected from 2019 to 2020 and was included in the Kuehn et al (2022b) meta-analysis.

Participants were recruited based on suicide risk to increase the probability of observing frequent suicidal thoughts. Eligibility criteria for the study included access to an internet enabled smartphone, familiarity with the English language, ability to come to an in-person meeting for consent, identity/age verification, and a baseline interview^b. Participants were also recruited if they were 1) between the ages of 16 and 20 years old, 2) endorsed either one episode of non-suicidal self-injury in past two weeks with past-month suicidal ideation (≥ 31 on the Suicide Ideation Questionnaire) or a had a suicide attempt in the past year; 3. Parent/primary caregiver consent (for those 16-17) with adolescent assent to study (16-17) OR adult (18-20) participant consent.

2.2. Procedure

All procedures were in accordance with the Declaration of Helsinki and were approved by the Institutional Review Board at the University of Washington. Following the baseline session, described above, participants completed both signal- and event-contingent surveys via their smartphones over the course of the next 14 days. Signal-contingent surveys were delivered on a pseudo-random schedule five times per day between the hours of 8:00AM and 10:00PM with a minimum of one-hour between signals (average of 5.68 hours between completed responses including overnight; 4.32 hours between completed EMA responses excluding overnight). A reminder text message was sent if the participant did not click on the survey link within 30 minutes. Survey links were deactivated one hour after delivery. Participants saved a direct link in their phone to complete an event-contingent survey when they experienced a SITB that was not reported in signal-contingent surveys. To increase compliance, study personnel called participants to check-in when they missed three surveys in a row. Participants were given feedback on their performance halfway through the study and were encouraged to increase their engagement if compliance was low. At the end of the 14 days, participants were compensated $1 for each completed survey and an additional $30 if they completed more than 80% of signal-contingent surveys. Participants were not compensated for event-contingent surveys.

Of 4,200 possible signal-contingent responses (60 subjects, 70 signals per person), participants completed 3,413 valid signal-contingent surveys (81.26%) and they submitted an additional 19 event-contingent surveys. 31 of the 60 participants did not report any suicidal thoughts during the study period to analyze (despite having a recent history of suicidal thoughts). 29 participants (48.33% of the sample) reported at least one suicidal thought during the 14-day window. Among these 29 participants, there were 2,030 possible responses (48.33% of 4,200) with 372 missing observations (81.78% complete observations; range = 45.71% - 100%; median = 85.71%). All 19 event-contingent surveys were included in the present study. Participants were more likely to miss surveys in the morning (b = −0.33; OR = 0.72; 95% CI OR = 0.67 – 0.76; p < .001) and later in the EMA period (e.g., on day 13 or 14; b = .05; OR = 1.05; 95% CI OR = 1.03 – 1.08; p < .001). Participants reporting higher levels of suicidal ideation at baseline were also less likely to miss surveys (b = −0.01; OR = 0.99, 95% CI OR = 0.97 – 0.99; p = .01). Missing data was handled in GIMME through full information maximum likelihood.

2.3. Measures

2.3.1. Momentary emotions:

Negative emotions were assessed with a prompt worded, “In the past 10 minutes, how much have you felt the following emotions.” Response items were adapted from the negative affect subscale of the Positive-Affect Negative-Affect Scale (PANAS; Watson et al. 1988) and included anger, shame, sadness, fear, and guilt. Participants rated their emotions on a 100-point visual analogue scale ranging from 1 (low) to 100 (high). Participants were presented with a check box to distinguish between the absence of an emotion and a missing response. Multilevel reliability statistics were high and there was high within-person variability of emotions across time (R_KF (fixed time effects/reliability of all items across all timepoints) = .99; R_1R (random time effects/generalizability of a single time point across all five emotions) = .33; R_KR (random time effects/generalizability of all emotions across all time points) = .98).

2.3.2. Suicidal thoughts:

Self-injurious thoughts and behaviors were first queried with an item worded, “Since the last assessment, have you thought about harming yourself?” A binary, “Yes” or “No”, response option was provided. Participants were instructed to answer ‘yes’ to this item if they experienced any suicidal or non-suicidal self-injurious thoughts or behaviors. To differentiate between non-suicidal and suicidal thoughts, those who answered ‘yes’ were then asked their intention to kill themselves within the past 30 minutes (ICC = .20; response options ranged from 0 [‘none], 1 [‘briefly thought about it but no plan to act’], 2 [‘thought about it some with no intention to act’], 3 [‘thought about it a lot and started to act but unsure’], and 4 [‘specific plan with intention to die’]). Participants reporting a ‘3’ or ‘4’ triggered an emergency response plan in which an automatic email was sent to study staff who contacted participants within 24 hours to conduct a risk assessment and provide emergency referrals. We created a binary variable indicating the presence/absence of suicidal thoughts. These items were adapted from Nock et al. (2009) and were comparable to other EMA studies of SITBs.

2.4. Data analytic plan

We first generated descriptive statistics for the n =29 participants who experienced frequent suicidal thoughts over the 14-day follow-up period and compared them to the excluded n = 31 sample to determine any group level differences that may explain imminent risk of suicidal thoughts over the 14-day observation period. Participants included in the present study reported higher baseline levels of suicidal ideation (t (df = 57.70) = −3.40, p = .001), PROMIS anxiety (t (df = 57.98) = −2.30, p = .03), and PROMIS depression (t (df = 56.57) = −2.30, p = .03) than excluded participants. There were no differences in terms of age or sexual orientation.

We then used the GIMME package (Lane et al., 2023) in the R 4.3.2 environment (R Core Team, 2013) to test group-level and person-specific correlates between negative emotions and suicidal thoughts. GIMME (Gates & Molenaar, 2012) uses threshold cutoffs to model the frequency of a shared effect within the sample (e.g., with a .75 threshold a group level path would be modeled if a bivariate relationship was significant, p < .05, in 75% of the person-specific models). We tested four GIMME models, increasing the threshold for group-level effects starting with 25% and increasing to 75% (37.5% and 50% as interim models). This allowed us to determine the frequency of group-level effects for increasingly larger subsets of the sample (i.e., to determine the level of heterogeneity in within-person associations). Indices of model fit were extracted and compared between models. Next, we examined person-specific plots from the 75% threshold model. We describe themes present in the individual-level models and present exemplar plots from four individuals in the results section (see Supplementary Materials for individual models of all 29 participants). As is common practice in idiographic research (Kaurin et al., 2022; Kuehn et al., 2022a), these four individuals were chosen to highlight the range of effects described below.

3. Results

3.1. Descriptive

Demographics of the sample are reported in Table 1. Participants (full sample, N = 60) were youth and young adults 16-20 years old (M_age = 18.58; 76.67% assigned female-at-birth; 61.67% lesbian, gay, or bisexual). Bivariate correlations between negative emotions and suicidal thoughts are presented in Table 2. Of the subset of participants reporting at least one instance of a suicidal thought during the two-week study, participants reported an average of 6.97 suicidal thoughts (SD = 8.53, median = 4, range = 1 – 36).

Table 1.

Sample demographics

Baseline Characteristic
Age (mean, SD)	18.78	1.15
	n	%
Male Sex	4	13.79
Heterosexual Males	2	6.90
Gay or bisexual Males	2	6.90
Other orientation	0	0
Female Sex	25	86.21
Heterosexual Females	8	27.59
Lesbian or bisexual females	14	48.28
Other Females	3	10.35
Race/Ethnicity
White	15	51.72
Asian	4	13.79
Black	1	3.45
Native American	0	0
More than one identity	4	13.79
Hispanic	4	13.79
Middle Eastern	1	3.45

Note. Limited to the n=29 sample analyzed in person-specific models.

Table 2.

Group-level bivariate correlations between suicidal thoughts and negative emotions

Variable	1.	2.	3.	4.	5.
1. Sadness
2. Shame	.53***
3. Fear	.43***	.44***
4. Guilt	.47***	.67***	.43***
5. Anger	.38***	.29***	.28***	.26***
6. SI	.37***	.29***	.18***	.18***	.17***

Note. Correlations are for n = 29 sample only.

3.2. Group Iterative Multiple Model Estimation (GIMME)

3.2.1. Group-level paths

For group models with a 25% threshold (see Figure 1a for the group-level paths), on average, fit indices suggested these individual models fit the data well (χ² = 38.81, df = 36.14, CFI = .98, NNFI = .98, RMSEA = .03, SRMR = .09); all 29 person-specific models had excellent fit. At the group-level, we found an autoregressive effect of suicidal thoughts such that suicidal thoughts at t-1 predicted an increased odds of suicidal thoughts at t. In addition, we observed a contemporaneous association between sadness and the intensity of suicidal thoughts, indicating that at least 25% of the participants experienced higher levels of sadness during a suicidal thought.

Figure 1. Group-level paths with a threshold of .25, .375, .5, and .75.

Note. 25% threshold (A.), 37.5% threshold (B.), 50% threshold (C.) and 75% threshold (D.). Gray lines represent all paths from person-specific models. Dotted lines depict lagged effects while contemporaneous paths are shown in solid lines. Black lines signify group-level effects. Looped lines suggest autoregressive paths.

We increased our threshold to 37.5% (see Figure 1b for group-level paths) and individual models again suggested excellent fit (χ² = 40.27, df = 36.24, CFI = .98, NNFI = .97, RMSEA = .04, SRMR = .09). All 29 models demonstrated excellent fit. Group-level model results remained similar to results from the group-level model using a 25% threshold. We increased our threshold to 50% (Figure 1c) and found similar results to earlier models in terms of fit (χ² = 41.96, df = 37.72, CFI = .98, NNFI = .98, RMSEA = .04, SRMR = .09); however, the contemporaneous association between sadness and suicidal thoughts was no longer present. This suggests that less than 50 percent of the sample reported a positive association between sadness and suicidal thinking.

Finally, we increased the group-level threshold to 75% (Figure 1d) to determine which of the two shared effects were present in a clear majority of the sample. Model fit indices suggested excellent fit (χ² = 43.07, df = 38.76, CFI = .98, NNFI = .97, RMSEA = .04, SRMR = .10) and all individual models also had excellent fit. Once again, in the group-level model, only an autoregressive effect of suicidal thoughts remained. There were no contemporaneous or lagged associations with suicidal thoughts. Thus, for most of the sample, reporting suicidal thoughts at the previous timepoint was associated with suicidal thinking at the next timepoint, although the direction and magnitude varied between individuals.

3.2.2. Person-specific models

We then examined person-specific models with the most stringent group-level threshold of 75%. We highlight models from four participants in Figure 2; all individual models are presented in the Supplementary Materials. The most common pathways were a positive autoregressive effect with ideation predicting ideation (present in 15 of the 29 models; 51.72%) and a contemporaneous association between sadness and SI (present in nine of the 29 models; 31.03%; see Table 3). Person-specific models ranged in the number of associations with suicidal thinking from zero to five. Across individual models, there were 17 unique paths between specific emotions and suicidal thoughts. Of these 17, ten paths (58.82%) represented positive correlations, suggesting an increase in a negative affective state was associated with increases in the severity of suicidal thinking, while seven (41.18%) were negative. Eight of the paths depicted a contemporaneous association, while nine were lagged. The average person-specific model was relatively sparse, demonstrating only 1.03 contemporaneous path between specific negative emotions and SITBs (range = 0 – 4) and only 0.55 lagged paths between negative emotions at t-1 and SITBs at t (range = 0 – 4). Themes from these individual models are described below.

Figure 2. Four exemplary person-specific models with group-level threshold set to 75%.

Note. Person A, Person B, Person C., and Person D models depicted. Positive associations are represented in red while negative correlations are shown in blue. Dotted lines are lagged effects while solid lines represent contemporaneous effects.

Table 3.

Unique paths between specific emotions and suicidal thoughts in person-specific models (n=29)

	N	%
# of paths (N, mean)	46	1.59
Positive Contemporaneous
Sad	9	31.03
Anger	7	24.14
Fear	1	3.45
Shame	8	27.59
Guilt	1	3.45
Positive Lagged (emotion predicting SI)
Shame	2	6.90
Anger	3	10.35
Guilt	2	6.90
Sad	4	13.79
Negative Contemporaneous
Anger	2	6.90
Sad	1	3.45
Shame	1	3.45
Negative Lagged
Fear	1	3.45
Shame	1	3.45
Guilt	1	3.45
Sad	2	6.90

^*Notes: Paths from person-specific models with 75% group threshold.

3.3. Variation in the magnitude and direction of the auto-regressive effect.

Although an auto-regressive effect was present for all 29 individuals, the direction (e.g., positive versus negative) varied between these person-specific models. In 15 of the 29 individual models (51.72%), there was a positive auto-regressive effect, which indicated that a suicidal thought at the previous time point predicted an increased odds of reporting a suicidal thought at the next time point. However, for 14 of the 29 models (48.28%), there was a negative auto-regressive effect, suggesting that for almost half the sample a suicidal thought at the previous time point predicted a decreased odds of a suicidal thoughts at the next survey. The average auto-regressive path was b = 0.00 (OR = 1.00, OR range = 0.67 – 1.42), meaning that, on average, suicidal thinking at t-1 was not associated with an increased odds of a suicidal thought at the next time point.

3.4. Univariate associations with single emotions.

11 of the 29 person-specific models indicated an autoregressive effect and a single association with one specific emotion. The most common path, present in 31 percent of person-specific models, was a contemporaneous association between sadness and suicidal thoughts. One example of such person-specific model is presented in Figure 2a. For the individual highlighted, they had a positive autoregressive effect in which suicidal thinking at t-1 predicted an increased odds of suicidal thinking at t, and a positive contemporaneous association with sadness (b = 0.33, OR = 1.39), meaning that a one standard deviation increase in sadness was associated with a 1.39 increase in the odds that this individual thought about suicide at the same timepoint. Another individual had a negative autoregressive effect and a lagged association between sadness and suicidal thoughts (b = 0.43, OR = 1.54). After sadness, the next most prevalent univariate association was a contemporaneous path between shame and suicidal thoughts (present in less than 28 percent of person-specific models).

3.5. Multi-variate, complicated, patterns with various emotions.

While the most common linkage was a positive contemporaneous association between a specific negative emotion and suicidal thoughts, ten of the 29 (34.48%) person-specific models included multiple emotions that were associated with suicidal thoughts (see Figure 2b/2c for two examples). Some of these models included both positive and negative associations as well as contemporaneous and lagged paths, suggesting about a third of the individuals reported a complex interplay of emotions surrounding suicidal thinking. For example, in the individual model presented in Figure 2b, they had a negative autoregressive effect and positive contemporaneous associations between suicidal thoughts and between fear and suicidal thoughts and sadness. They also had a negative contemporaneous association with anger. This all suggests that this persons’ suicidal thoughts tended to be brief, and that they often felt both afraid and sad, but not angry, when thinking about suicide. Additionally, they had a negative lagged association between fear and suicidal thoughts, in which they reported lower fear prior to thinking about suicide.

4. Discussion

The present study provides evidence for between-person heterogeneity of within-person associations of specific emotions and SITBs. Most researchers use a uniform construction of negative affect thus precluding insights into the relationships between specific affective states and SITBs (Foster & Beltz, 2022). Additionally, most suicide research uses nomothetic methods that demonstrate the average-level associations between risk factors and SITBs, even though clinical work and intervention call for the use of a person-specific framework to address heterogeneity within risk factors and the strength of a person-specific intervention. This study tests these assumptions by directly estimating models at the individual level and evaluating between-person consistency (i.e., prevalence) in the relevance and potency of links between discrete affective states and suicidal thinking.

4.1. Theoretical Implications

On the group-level, only sadness was correlated with suicidal thoughts; however, this effect was present only in a subset of individuals (< 50%). The present analysis adds nuance to affect regulation models of SITBs (Armey et al., 2011; Kleiman et al., 2018a. Kuehn et al., 2022b). This affect regulation hypothesis, in which a reduction in negative affect is thought to negatively reinforce SITBs, assumes similar mechanisms across all people who experience suicidal thoughts and each time an individual person is thinking about suicide. The present study questions some of these assumptions by highlighting significant individual-level heterogeneity.

Given the robust findings across several datasets, it is likely true that the affect regulation hypothesis applies to most of the individuals experiencing SITBs; however, the present analyses suggest that SITBs arise as a function of different affective experiences across individuals. Specifically, different affective states were variably related to SITBs across individuals. For one individual, fear, anger, and sadness were all closely related to suicidal thoughts, while for another person, only shame was correlated with SITBs. Consequently, SITBs should be understood, studied, and tracked with attention to the specific and unique affective experiences that elevate SITB risk, rather than continuing to estimate and develop treatment approaches that are intended as “one size fits all” strategies for reducing risk uniformly across patients reporting SITBs.

One important avenue for future research is to compare nomothetic versus idiographic predictive models. The present dataset is not ideal for this purpose as there were a relatively small number of observations per person and a limited number of suicidal thoughts observed. However, other studies have developed idiographic predictive models for suicide risk (Wang et al., 2023; Reeves, 2021). To our knowledge, it is not yet known whether these idiographic models improve predictive performance.

4.2. Treatment Implications

About 50 percent of individuals who die by suicide do not have any involvement with mental health care (Stene-Larsen & Reneflot, 2019). These individuals are, however, often exposed to healthcare providers through routine screening and assessments. The findings presented in this study have implications for both just-in-time adaptive interventions (Coppersmith et al., 2022) and the development of adjunctive tools to complement psychotherapy. For example, there has been great interest in developing digital health interventions to reduce the burden of care and reach individuals not involved in the mental health care system. Ecological momentary interventions (EMI) are one such application and are generally defined as mobile phone applications which deliver therapeutic content during high-risk states (Versluis et al., 2016). GIMME models could help to personalize the content delivered through EMIs by identifying personally relevant SITB risk factors to then inform the development of real-time content in helping an individual reduce the factors most closely related to suicidal thinking. GIMME may also be used as a completement to traditional therapy to help inform participants of their patterns over time, guide therapists' case conceptualizations, and to inform therapists which coping strategies to teach to individual clients (Yin, Hughes, & Rizvi, 2023).

These findings could be also used to streamline more intensive interventions. Dialectical Behavior Therapy (Linehan, 1993) is a comprehensive evidence-based treatment for individuals at high-risk for suicide and teaches mindfulness, emotion regulation, distress tolerance, and interpersonal effectiveness skills. When a DBT therapist learns a client has experienced a SITB, they often complete a behavioral chain analysis assessing the cognitive, emotional, and behavioral sequence of events leading up to a SITB. A DBT therapist could use these idiographic person-specific models to inform their chain analyses, as a contemporaneous and data-driven way to precisely target and reinforce specific DBT-based strategies (Cheavens et al., 2022; Webb et al., 2022).

Personalized interventions, based on idiographic data from EMA, have been developed for mood and anxiety disorders (Fisher et al., 2019) as well as eating disorders (Levinson et al., 2023) with promising results from open trials. Given the seemingly large degree of heterogeneity in within-person associations of suicidal thoughts, personalized interventions also seem warranted for individuals at risk for suicide. However, much more research is needed to inform the development of such interventions. Specifically, researchers need a better understanding of the minimum sampling duration and rate, as well as the frequency of positive suicidal thoughts, to create informative and replicable results. Additionally, given that affective variables were not associated with suicidal thoughts for most of the individuals in the present study, more idiographic research understanding other key suicide specific variables is needed for informing interventions aimed at reducing risk for individuals without affective dysregulation (e.g., perceived burdensomeness, lack of connectedness). Finally, obtaining feedback from therapists and potential patients regarding the utility of idiographic models as well as identifying how to best communicate these models to patients are all needed to guide treatment decisions and maximize the effectiveness of personalized interventions.

4.3. Limitations

There are a few limitations to the present study. First, these models are based on single-item self-reports in which participants were asked to reflect on their affective states. Individuals at high-risk for suicide likely are less accurate in naming their emotions (Iskric et al., 2020). Additionally, including contextual information would improve our understanding of the environmental circumstances and cognitive appraisals associated with these affective states (e.g., if someone was feeling fear in context of an objectively threatening situation). Theoretical and empirical findings highlight the role of appraisals, or interpretations, in emotional reactivity (Cole et al., 2019). It is thus unclear the extent to which these reports reflect deficits in emotion differentiation, inaccurate or unhelpful interpretations, or specific responses to environmental circumstances, which would indicate the use of different psychosocial intervention strategies.

The replicability of specific pathways in network models is likely another limitation to the present study (Borsboom et al., 2018; Robinaugh et al., 2020). Due to the multivariate nature of network models, the decision to include or exclude specific variables may change paths between nodes. However, use of a temporal network model – which draws information from many repeated observations within each person – somewhat mitigates concerns about replicability compared to other network estimation approaches (e.g., partial correlation networks estimating links between variables by pooling single observations across individuals). This concern is also attenuated by the fact that one the major conclusions of this study, the large degree of heterogeneity in person-specific models of high-risk individuals, is consistent with prior research (Kaurin et al., 2022; King et al., 2020; Kuehn et al., 2022a; Kleiman et al., 2018b). The present study does not address disparate levels of severity of suicidal thoughts. It is possible that there are differences in heterogeneity around varying severities or sub-constructs of suicidal thoughts. Finally, more than half of the original sample size did not report a suicidal thought. As such, we made assumptions about how self-reports might generalize across time (e.g., if someone gives a single report of a suicidal thought, we assumed it was accurate rather than interrogating for the presence of temporal biases in how often these events occur or how much a person affiliates with a frequency in these experiences and/or feels influenced by them. As such, these findings should be interpreted with caution regarding the extent of between-subject heterogeneity and the replicability of the three subgroups.

Nonetheless, the present study extends the current understanding of relations between negative emotions and suicidal thinking. Negative emotions and suicidal thoughts were linked in most of the person-specific models; however, these associations varied in substantial ways across participants that signal the need for researchers and clinicians to track and consider heterogeneity in the determinants of SITBs at the individual level going forward. Given that treatments developed on nomothetic assumptions over the past five decades have been ineffective at reducing suicide deaths at a population level (Fox et al., 2020; Harris et al., 2022), perhaps a better understanding of the personalized set of factors corresponding to suicidal thinking may lead to more effective interventions.