Emotion Differentiation and Dynamics Predicting Suicide Outcomes in High‐Risk Adults

ABSTRACT

Objectives

Negative emotions have been well studied in relation to suicidal thoughts and behaviors (STBs); however, little research in these areas have been informed by affective science frameworks that consider the complex and time varying nature of STBs. Emotion differentiation is well‐validated as one such core construct in affective science. The present study is the first to examine negative emotion differentiation in relation to STBs and in a sample at heightened risk for suicide. In addition, to investigate how emotion differentiation relates to real time measures of emotion regulation, we examined the relationship between negative emotion differentiation and facets of affective dynamics (i.e., affective inertia, intensity, and variability).

Methods

Adults leaving psychiatric inpatient care (n = 68) with recent STBs completed semi‐structured interviews, self‐report measures, and a 4‐week ecological momentary assessment (EMA) protocol in the period from 2021 to 2023.

Results

Results indicated no significant relationship between negative emotion differentiation and STBs (suicidal ideation duration, frequency, controllability, severity, nor suicide attempt history) when measured retrospectively or concurrently. However, negative emotion differentiation was significantly related to only some negative affective dynamics (i.e., inertia and variability).

Conclusions

These findings have important clinical implications regarding type and timing of treatment for STBs.

1. Introduction

Suicide remains a leading cause of death in the United States (Centers for Disease Control ²⁰²¹). Suicidal ideation (SI) is even more common, with 12.3 million American adults reporting serious thoughts of ending their own life each year (SAMHSA ²⁰²²). SI is also one of the most common reasons for psychiatric hospitalization (Bowers 2005). Further understanding of SI may yield insights into the process leading to suicide itself and help ameliorate the suffering associated with SI independently by incorporating those insights into SI interventions (Jobes and Joiner 2019).

1.1. Negative Emotions and STBs

Unsurprisingly, suicide risk is elevated among individuals with psychopathologies characterized by emotional difficulties (e.g., Dome et al. ²⁰¹⁹; Lucht et al. ²⁰²²; Rizk et al. ²⁰¹⁹). More specifically, emotional difficulties are well documented as correlates of recent (e.g., past month) and current SI, including greater emotional reactivity (Law et al. 2015; Palmier‐Claus et al. ²⁰¹²), lability (Arria et al. 2009), and intensity (Lynch et al. 2004). Although research examining affect in relation to suicidal thoughts and behaviors (STBs) is largely based on cross‐sectional, self‐report data, within‐person changes in negative affect (NA) are implicated as a robust correlate of SI (Armey et al. 2020; Schatten et al. ²⁰²¹). As one example, findings of a recent meta‐analysis indicated an increase in NA before, and a decrease in NA after, SI occurring in daily life (Kuehn et al. 2022).

Affective processes are clearly implicated in STBs, but more work is needed to understand the way that dynamic emotional processes over time relate to STBs. Affective dynamics, the patterns of fluctuations in emotion and its components over time (Kuppens et al. 2012), include dimensions such as affective variability, inertia, and intensity. Affect variability reflects the range or amplitude of an individual's emotional states across time (Houben et al. 2015), with some evidence suggesting elevated affective variability is relevant to risk of STBs among individuals at high risk for psychosis (Palmier‐Claus et al. ²⁰¹²), young adult women with a history of nonsuicidal self‐injury and/or STBs (Victor et al. 2021), people with borderline personality disorder (Links et al. 2008; Nisenbaum et al. ²⁰¹⁰), and in a general community sample (Zhu et al. 2024). Affect inertia refers to the tendency of emotions to carry over from one moment to the next, and therefore resistance of affect to changes (Kuppens et al. 2010; Suls et al. ¹⁹⁹⁸). Only one study has observed affective inertia as a risk factor for SI in high‐risk young adult women (Victor et al. 2021). Affect intensity is defined as the strength of emotional states (American Psychological Association ²⁰²⁴). Of all affective dynamics, elevated NA intensity has been most robustly linked to STBs—for example, more frequent suicidal behaviors (Osman et al. 1999). Elevated NA is also robustly associated with the presence of psychopathology (i.e., across internalizing disorders; Scott et al. ²⁰²⁰).

Intensive longitudinal methods are particularly well suited to study affect and STBs given the dynamic nature of both over minutes, hours, and days. Ecological momentary assessment (EMA) is one such method that has demonstrated convergence with cross‐sectional findings while providing a more detailed portrait of SI in the short‐term (Armey et al. 2020; Gee et al. ²⁰²⁰). EMA data allows for the use of robust statistical methods to estimate within‐person affective dynamics while accounting for missing and non‐normal data, such as through dynamic structural equation modeling (Asparouhov et al. 2018; Hamaker et al. ²⁰²¹).

1.2. Emotion Differentiation as an Element of Emotion (Dys) Regulation

The presence of negative emotionality may confer risk for STBs in the context of emotion dysregulation, which is itself associated with STBs (e.g., Law et al. ²⁰¹⁵; Rajappa et al. ²⁰¹²); in contrast, adaptive emotion regulation (i.e., cognitive reappraisal) has been identified as protective against STBs (Franz et al. 2021). STBs themselves may function as maladaptive strategies used to regulate intense NA (Coppersmith et al. 2023) insofar as SI may improve one's affective state temporarily (Kleiman et al. 2018) but contribute to long‐term distress. The inability to employ effective emotion regulation may confer greater vulnerability for STBs than NA alone. Prior work addressing this topic has relied on self‐report measures of trait‐like tendencies to use specific emotion regulation strategies or on overall perceived efficacy of those strategies (e.g., Caprara et al. ²⁰⁰⁸; Gratz and Roemer ²⁰⁰⁴).

Emotion differentiation, the ability to recognize and categorize related but distinct emotions (e.g., anger versus fear; Barrett et al. ²⁰⁰¹), involves labeling emotions as distinct affective states versus global emotional experiences (i.e., all bad/negative or good/positive). Emotion differentiation may influence emotion regulation (Carpenter and Trull 2013), as an awareness of distinct emotions can facilitate the selection and use of effective and targeted emotion regulation strategies (Barrett et al. 2001). Thus, it is unsurprising that those with difficulties in emotion differentiation are more likely to choose harmful coping strategies, such as substance use (Anand et al. 2017), aggressive behaviors (Edwards and Wupperman 2017), disordered eating (Mikhail et al. 2020; Selby et al. ²⁰¹⁴), and transdiagnostic psychopathology (Ozomaro et al. 2025).

Emotion differentiation is typically examined separately for negative emotions (i.e., negative emotion differentiation, NED) and for positive emotions (i.e., positive emotion differentiation, PED) with most psychopathology research focusing on NED. High NED has been linked to adaptive outcomes such as successful decision making in risk‐taking conditions (Li and Ashkanasy 2019) and empathic accuracy (Erbas et al. 2016). Low NED has been linked to maladaptive outcomes including greater reactivity to negative events (Starr et al. 2017), engagement in nonsuicidal self‐injury (Zaki et al. 2013), and higher emotion‐related impulsivity in borderline personality disorder (Tomko et al. 2015). Interestingly, although NED has not yet been directly examined in relation to affective inertia or variability, one recent, large EMA study of adolescents demonstrated a relationship of greater NED with elevated NA and decreased PA (Lo et al. 2025), which may indicate a tradeoff between the mental effort of emotion naming rather than engagement in other affective regulatory efforts (Nook 2021). Although one prior study found no significant relationship between NED and SI frequency among adolescents (Al‐Dajani et al. ²⁰²⁴), no studies to date have investigated the relationship between STBs and NED in adults despite the importance of negative emotions broadly in understanding STBs across the lifespan (i.e., Armey et al. ²⁰²⁰; Núñez et al. ²⁰²⁰; Selby et al. ²⁰¹³).

1.3. The Present Investigation

Research on affective experiences and STBs is limited by cross‐sectional methods and a reliance on samples with low severity STBs. To fill this gap, we examined NED and NA dynamics captured during a 4‐week EMA period. Research has also been limited by reliance on community samples with low severity STBs. To address the issue of low base rates of STBs, we enrolled adults with recent STBs who were leaving psychiatric inpatient care, a period of increased risk for suicide (Chung et al. 2017).

Aim 1: Evaluate the relationship between concurrent NA dynamics and NED. We hypothesized that lower NED would be associated with higher NA intensity, inertia, and variability.

Aim 2: Examine the association between NED and recent history of STBs. We hypothesized that lower NED would be associated with higher self‐rated past week SI severity and the presence of self‐rated lifetime suicide attempt at baseline.

Aim 3: Examine the association between NED and concurrently reported STB during EMA. We expected that lower NED would be associated with greater aggregate SI frequency reported during EMA.

2. Method

2.1. Participants

Data were gathered as part of a larger study focused on affect, cognition, sleep, and suicide risk (parent study preregistration here: https://osf.io/4txz5). Participants were adults (age 18+) recruited between 2021 and 2023 while admitted to one of two psychiatric inpatient hospital units in the southern United States based on past week SI or suicidal behavior (i.e., suicide attempt, preparatory behavior, aborted/interrupted attempt). The study involved interviews, questionnaires, actigraphy, and a 4‐week EMA protocol. Exclusion criteria included inability to complete study procedures due to low English fluency, impaired cognition (orientation score < 5 on the MoCA), or acute psychosis or mania that interfered with study participation. The baseline interview and questionnaires were completed while participants were in the hospital, and a 4‐week EMA protocol was started after discharge using a personal smartphone or loaner phone. A total of 130 participants were enrolled in the parent study based on power analyses for primary study aims; however, only a subset of participants (n = 68) met the necessary requirements for the present analyses (see data analytic strategy). The analytic sample (60.3% cisgender women) primarily identified as non‐Hispanic/Latinx (76.5%) and White (80.9%). Participants' mean age was 35.82 years (SD = 12.7, range = 18–67). Most (85.3%, n = 58) reported a lifetime history of suicide attempts. Other demographic and clinical characteristics are presented in Table 1.

TABLE 1.

Descriptive demographic and clinical characteristics.

Variable	n (%)
Ethnicity
Non‐Hispanic	52 (76.5%)
Hispanic/Latinx	16 (23.5%)
Race
White	55 (80.9%)
Black/African American	8 (11.8%)
Asian/Asian‐American	0 (0%)
Hawaiian/Pacific Islander	0 (0%)
Native American/Alaskan Native	3 (4.4%)
Biracial/Multiracial/Other	6 (8.8%)
Biological sex
Male	26 (38.2%)
Female	41 (60.3%)
Missing	1 (1.5%)
Education level
Grade 8 (no high school)	1 (1.5%)
Some high school	6 (8.8%)
High school or GED	21 (30.9%)
Some college (no degree)	19 (27.9%)
Two‐year college degree	10 (14.7%)
Four‐year college degree	5 (7.4%)
Some graduate or professional degree	4 (5.9%)
Graduate or professional degree	2 (2.9%)
EMA STBs*
Proportion of participants with SI	60 (88.3%)
Proportion of surveys with SI	2434 (33.7%)

Abbreviations: EMA, ecological momentary assessment; SI, suicidal ideation; STBs, suicidal thoughts and behaviors.

^*Between all included participants (n = 68), 7222 EMA surveys were completed.

2.2. Recruitment and Enrollment Procedures at Baseline

All study procedures were approved by the Texas Tech University Institutional Review Board as well as the respective research review entities at the recruiting hospitals. Members of the hospitals' treatment teams identified potentially eligible patients and provided a brief description of the study to patients to ascertain potential interest. Research staff then approached participants who had provided verbal consent for contact on the inpatient unit, completed written informed consent procedures, and then completed initial eligibility screening questions to confirm past week STBs. Enrolled participants then completed semi‐structured interview measures not used for the present analyses, in addition to self‐report measures and computer tasks. Participants received a $30 gift card for this session.

2.3. EMA Procedures

Researchers trained participants in a 4‐week EMA protocol using detailed instructions and practice items. Surveys were scheduled for pseudo‐random distribution seven times a day for 28 days, starting the day after hospital discharge. Participants provided the research team with “black out” periods (e.g., times when participants would be unavailable to respond to surveys) and sleep and wake times to set bounds around which surveys could be sent. Survey notifications were sent via text message and contained a unique Qualtrics link only available within 30 min of the notification. A reminder message was sent if the link was not opened within 10 min of the first notification. Survey notifications were scheduled using proprietary software developed by the lead investigator's lab. At the end of the EMA period, participants received a $50 gift card, regardless of surveys completed, in addition to bonus payments: $1 for each survey completed (up to $196) and $25 for completing at least 75% of the surveys.

2.4. Participant Safety Protocols

Graduate student and post‐baccalaureate assessors were trained in suicide risk assessment and safety planning procedures. Participants were notified during EMA training, on the EMA survey landing page, and at the bottom of each survey page that responses were not monitored in real time, and they were directed to contact a crisis hotline if they were in need of immediate support. If a participant indicated suicidal behavior on an EMA survey, a report was sent to the research team via email for safety monitoring and follow up with the participant within 24 h. Safety response procedures involved a relatively less intrusive approach compared to many existing practices (Bentley et al. 2021), which may have reduced reactivity from participants in response to research team follow up, relative to other, more active crisis response strategies.

2.5. Measures and Calculation of Relevant Constructs

2.5.1. Demographic Characteristics

We assessed whether self‐reported demographic information, including age, race, ethnicity (non‐Hispanic White individuals and participants from other racial/ethnic groups), biological sex, and education attainment (attended at least some college vs. high school education or less), were confounds in emotion related abilities (Mankus et al. 2016). NED was not significantly associated with age (r = −0.05; p = 0.71), gender (t[65] = 1.35, p = 0.18), ethnicity/race (t[66] = −1.65, p = 0.10), or education (t[66] = 1.57, p = 0.12).

2.5.2. Baseline STBs

Lifetime, past year, past month, and past week STBs were assessed using an adapted version of the Columbia‐Suicide Severity Rating Scale (C‐SSRS; Posner et al. ²⁰¹¹). The C‐SSRS has demonstrated strong validity among psychiatric inpatients (Madan et al. 2016). The current study examines several constructs including history of suicidal behavior and SI (severity, frequency, duration, and controllability). Suicidal behavior, defined as a suicide attempt for the present study, was rated as either present (“1”) or absent (“0”) for each period (e.g., lifetime, past week). SI features were each rated on a scale with a higher rating indicating more severe SI (see Table 2). Graduate and post‐baccalaureate assessors that administered the C‐SSRS were trained to reliability. Interrater reliability review was conducted by trained graduate students and post‐baccalaureates, with disagreements in ratings discussed among the research team and resolved by the senior investigator on the study.

TABLE 2.

Descriptive characteristics of baseline‐reported past week SI.

Past week SI	N (%)	M (SD)
Controllability		3.86 (1.17)
Easily able to control thoughts (1)	3 (4.4)
Can control thoughts with little difficulty (2)	5 (7.4)
Can control thoughts with some difficulty (3)	15 (22.1)
Can control thoughts with a lot of difficulty (4)	16 (23.5)
Unable to control thoughts (5)	25 (36.8)
Frequency		3.77 (1.20)
Less than once a week (1)	1 (1.5%)
Once a week (2)	10 (14.7%)
2–5 times a week (3)	20 (29.4%)
Daily or almost daily (4)	7 (10.3%)
Many times each day (5)	28 (41.2%)
Duration		3.24 (1.44)
Fleeting – a few seconds to minutes (1)	11 (16.2%)
Less than 1 h/some of the time (2)	9 (13.2%)
1–4 h/a lot of the time (3)	18 (26.5)
4–8 h/most of the day (4)	9 (13.2%)
More than 8 h/persistent or continuous (5)	19 (27.9%)
Severity		4.04 (1.35)
Wish to be dead (1)	7 (10.3%)
Non‐specific active suicidal thoughts (2)	5 (7.4%)
Active suicidal ideation with methods, no intent (3)	3 (4.4%)
Active suicidal ideation with some intent to 4	16 (23.5%)
Active suicidal ideation with intent and plan (5)	37 (54.4%)

Note: Each row may not equal the total sample size (N = 68) as some participants were missing response data for specific items. Variables were dichotomized, with lower severity (“0”) denoted by unbolded font and higher severity (“1”) denoted by bolded font.

Abbreviation: SI, suicidal ideation.

Consistent with the overall focus on participants receiving acute inpatient psychiatric care for recent SITBIs, and our inclusion criteria of at least passive SI (e.g., wish to be dead) within the past week, baseline SI characteristics were negatively skewed (see Table 2). As an ordered categorical approach was not appropriate for our data given the cell sizes of our categories (Long 1997), we dichotomized each baseline SI characteristic so that the lower severity group was coded as “0”, and the higher severity group was coded as “1”. Cut points for categorizing dichotomous variables were based on observed distributions, informed by decisions regarding interpretability and conceptually reasonable divisions to ensure that cell sizes of each group were adequate for analyses (see Table 2).

2.5.3. Momentary SI

SI at the momentary level was assessed using two items inquiring about passive SI “at this moment” in each EMA survey: “I feel that life is not worth living” and “I want to die”. Items measuring active SI were not included in the present study due to ethical concerns (i.e., that the research team was not available to follow up with high‐risk responses in real time). Each item was rated from 1 (not at all) to 5 (very much). To provide a between‐person estimate of EMA‐reported SI, we calculated the proportion of completed EMA surveys in which either SI item was rated as a 2 or higher (i.e., any SI endorsed). Higher scores indicated a greater proportion of EMA surveys with SI.

2.5.4. Momentary NA

At each EMA prompt, participants were asked to rate seven negative emotions following the stem “At this moment, I feel…” on a scale from 1 (not at all) to 5 (very much). Six items (afraid, nervous, angry, irritable, sad, and blue) were drawn from the Positive and Negative Affective Schedule, extended version (Clark and Watson 1994), which has robust evidence of validity and sensitivity to change (Clark and Watson 1994; Watson et al. ¹⁹⁸⁸). An item was added assessing hopelessness (“hopeless”) in line with previous research on hopelessness and suicide risk (Beck et al. 1985). Ratings for two positive emotion items (cheerful and happy) were not used for the present analyses.

Person‐level NED was determined by calculating intraclass correlation coefficient (ICC) scores for the seven NA items aggregated across the EMA period for participants with at least 20 completed surveys (n = 68), though there is a relative lack of guidance regarding what constitutes an adequate number of observations needed for NED indices (Thompson et al. 2021). Although ICC scores should range from 0 to 1, negative ICC scores are possible due to measurement error (Shrout and Fleiss 1979). We re‐scored negative ICC values as zero, defining them as high differentiators (Cohen et al. 2003). ICC scores were Fisher z‐transformed and subtracted from 1 so that higher scores reflect greater NED (Thompson et al. 2021). Thus, although calculated from dynamic data, NED was estimated as a single value per participant which reflects the extent to which the NA items hang together within each assessment over the EMA period.

For aim 1, mean NA intensity, NA variability, and NA inertia were calculated using dynamic structural equation modeling (DSEM), an approach that is robust to missing data, zero‐inflated categorical variables, and variable time between observations (Hamaker et al. 2021) and appropriate for estimating affective dynamics (Hamaker et al. 2018). In contrast to other techniques, such as estimation of mean squared successive differences (MSSD; Wang et al. ²⁰¹²), DSEM has two advantages: (a) the ability to effectively parse variability observed at the within‐ versus between‐person level without conflation with auto‐regression, and (b) the ability to select meaningful lag intervals that differ from observed lag. Survey‐level NA was calculated as an observed average of the seven NA items per survey; these scores were used to model group‐mean‐centered NA intensity, NA variability, and NA inertia in the EMA period (see Statistical Analytic Plan below).

2.6. Data Analytic Strategy and Preparation

Although the larger parent study aims and hypotheses were preregistered, analyses for the present study were not preregistered. All data for these analyses, as well as syntax used, have been made publicly available on OSF and can be accessed at https://osf.io/4txz5.

2.6.1. Missing Data

Computation of EMA parameters (NED, NA intensity) allows for variability in the number of completed surveys given an adequate number of observations per participant. For regression analyses (aims 2 and 3), participants with missing data on outcome variables were excluded per analysis via listwise deletion; therefore, sample size ranged from 68 to 64. Handling of missing data for DSEM models is discussed below.

2.6.2. Statistical Analytic Plan

MPlus statistical software, version 8.6 (Muthén and Muthén 2017) was used for DSEM analyses and IBM SPSS version 29.0.1.0 (IBM Corp ²⁰²³) was used for regressions, bivariate associations, and descriptive analyses. For visualization of DSEM modeling, see Figure 1.

FIGURE 1.

Dynamic structural equation model overview. Using a two‐level dynamic structural equation model, we calculated person‐specific estimates of mean NA (${\mathit{NA}}^{\left(b\right)}$), NA auto‐regression ($\varnothing \mathit{NA}$), and NA variability ($\mathit{\text{logψNA}}$). At the within‐person level, NA for a particular individual (w) at a specific time (t; ${\mathit{NA}}_t^{\left(w\right)}$) was predicted by NA at the prior time interval (${\mathit{NA}}_{t-1}^{\left(w\right)}$) and NA residual variance ($\mathit{\psi NA}$). Random effects are depicted as circles in lines. Arrows at the between‐person level reflect the allowed covariances between all parameters, including negative emotion differentiation (NED) and the covariate age.

To examine the between‐person effects of NA intensity, variability, and inertia in association with NED (aim 1), we specified a two‐level random effects model using DSEM to calculate person‐specific estimates of overall NA intensity (mean), variability, and inertia. These models specified a two‐hour time lag using the TINTERVAL option. TINTERVAL is a data processing option that handles missing data by dealing with subject level observation timing by allowing for unequally spaced time observations through timeline discretization and estimation of missing data for time slots without observations (Muthén et al. 2024). Within‐person NA variance and NA auto‐regression were calculated as random effects to allow examination of latent mean centered NA variability and NA inertia alongside latent mean centered NA intensity at the between‐person level. At the between‐person level, NA variability, NA inertia, NA intensity, and NED were allowed to freely covary with each other.

Preliminary models examined, separately, the between‐person covariances between NA dynamics and potential covariates (age, race/ethnicity, education, and biological sex; see Table 3); only age exhibited a significant association with NA inertia. As a result, age was included in the full model testing NED in association with NA dynamics for aim 1.

TABLE 3.

Dynamic structural equation model tested covariates for aim 1.

Variable	β[95% CI]
	NA intensity	NA variability	NA inertia
Age	0.20 [−0.06, 0.44]	0.25 [−0.48, 0.01]	0.33 [0.06, 0.56]*
Biological sex	0.09 [−0.22, 0.38]	0.26 [−0.04, 0.51]	0.19 [−0.13, 0.46]
Race/ethnicity	–0.09 [−0.39, 0.22]	0.16 [−0.15, 0.43]	–0.08 [−0.36, 0.23]
Education level	0.17 [−0.14, 0.46]	−0.17 [−0.45, 0.14]	–0.02 [−0.31, 0.29]

Note: Each covariate was tested in a separate model. Biological sex was coded as two groups (male, female); race/ethnicity was coded as a binary variable where 0 = Non‐Hispanic/Latinx White and 1 = all other participants; education was coded as a binary variable where 0 = high school education or less and 1 = some college education or greater.

Abbreviation: NA, negative affect.

^* p < 0.05.

Models were estimated using Bayesian estimation with a minimum of 5000 iterations. Model convergence was determined using the Potential Scale Reduction, with values under 1.1 indicating stable fit. Once models exhibited convergence, they were repeated with double the iterations required for convergence, to ensure stability of model parameters. Statistically significant parameters for the relationship between each parameter (i.e., intensity, inertia, variability) and NED were determined by estimating a 95% credibility interval (CI). CI intervals that do not span over zero indicate statistical significance at α = 0.05.

For aim 2, NED was entered into four separate logistic regression models as a predictor, each with one binary SI severity outcome variable. A further binary logistic regression was used to investigate the relationship between NED and lifetime suicidal behavior. For aim 3, a linear regression model was used to test the relationship between NED and concurrently reported proportion of surveys with SI during EMA. The assumptions of a linear regression were tested (e.g., normality, homoscedasticity, linearity, independence of observations) and no significant deviations were observed; further, no significant outliers were identified via Cook's Distance.

2.6.3. Power Analyses

As the above represents secondary data analyses, a priori power analyses were not conducted. However, data from this study are within range for good estimation of random means, autoregressions, and variances in DSEM (i.e., t = 100 and N > 50), based on prior simulation findings (Schultzberg and Muthén 2018). Power analyses for aims 2 and 3 were conducted in G*Power (Faul et al. 2007) using post hoc sensitivity analyses, given that sample size was already determined. For a binary logistic regression (aim 2), the present sample size of 68 yields 15% power to detect small effects (OR = 1.437), 49% power to detect medium effects (OR = 2.477), and 86% power to detect large effects (OR = 4.268). For a linear model (aim 3), a sample size of 68 yields 16% power to detect small effects (Cohen's f = 0.02), 80% power to detect medium effects (Cohen's f = 0.15), and 99% power to detect large effects (Cohen's f = 0.35).

3. Results

Descriptive characteristics of SI at baseline are provided in Table 2. The majority of participants (88.2%) reported at least some SI during EMA. In total, non‐zero SI was endorsed on 33.7% of surveys; aggregated NA intensity during the EMA period ranged from 1.03 to 4.38 across participants (M = 1.83, SD = 0.71). Participants completed an average of 105.68 surveys (SD = 27.26) out of 196, indicating an average completion rate of 53.92%, which was not correlated with NED (r = −0.152, p = 0.22). Participants completed most surveys in the afternoon (42.4%, 12 pm to 6 pm), followed by evenings (35.7%, 6 pm to 12 am), morning (19.5%, 6 am to 12 pm), and night (2.39%, 12 am to 6 am).

3.1. Tests of Primary Hypotheses

First, consistent with hypotheses, higher NA variability (β = −0.59, 95% CI [−0.74, −0.38]) and higher NA inertia (β = −0.41, 95% CI [−0.63, −0.14]) were related to lower NED when adjusting for participant age. In contrast, NA intensity was not related to NED (β = −0.12, 95% CI [−0.38, 0.17]).

Second, across binary logistic regressions of the between‐person relationship between NED and recent SI features, there was no significant association between NED and past‐week SI severity (OR = 2.05, 95% CI [0.63, 6.68]); controllability (OR = 1.44, 95% CI [0.44, 4.69]); duration (OR = 3.34, 95% CI [0.94, 12.17]); or frequency (OR = 1.35, 95% CI [0.43, 4.13]; see Table 4). Results for the fifth binary logistic regression analyzing the relationship between NED and lifetime suicidal attempt similarly demonstrated no significant association (OR = 5.70, 95% CI [0.97, 33.72]).

TABLE 4.

Logistic regression analyses of NED as a predictor of past week baseline SI and suicide attempt.

Model	β [95% CI]	SE β	Wald's χ2	df	p	Odds ratio
SI controllability
Constant	0.53	0.26	34.08	1	0.04
NED	0.366 [0.44, 4.69]	0.60	0.37	1	0.54	1.44
SI duration
Constant	−0.23	0.26	0.78	1	0.38
NED	1.22 [0.94, 12.17]	0.65	3.50	1	0.06	3.34
SI frequency
Constant	0.15	0.25	0.33	1	0.56
NED	0.30 [0.43, 4.13]	0.58	0.26	1	0.61	1.35
SI severity
Constant	0.24	0.25	0.87	1	0.35
NED	0.72 [0.63, 6.68]	0.60	1.42	1	0.23	2.05
Lifetime SA
Constant	2.06	0.42	23.79	1	< 0.001
NED	1.74 [0.97, 33.72]	0.91	3.70	1	0.05	5.70

Note: Each outcome variable was entered individually into separate logistic regressions with NED as the predictor; the referent group for the outcome variable was “No” for SA history.

Abbreviations: NED, negative emotion differentiation; SA, suicide attempt; SI, suicidal ideation.

Last, we examined the association between NED and concurrently reported STBs during the EMA period. Results of the linear regression testing the between‐person relationship between NED and EMA‐reported SI indicated no significant relationship between these constructs (β = 0.08, SE = 0.09, 95% CI = [−0.13, 0.24], p = 0.54).

4. Discussion

To our knowledge, this study was the first to examine the relationship between NED and STBs, and to examine NED in relation to other NA dynamics. A strength of the current work was our focus on individuals who were recently discharged from a psychiatric inpatient hospital reporting STBs in the past week. A further strength was our use of EMA, which diminishes retrospective recall biases, and allowed us to form a sophisticated measure of NED based on the covariance of emotion state ratings aggregated across a four‐week period.

Our first aim was to examine whether NED was related to other NA dynamics (inertia, intensity, and variability) using multilevel modeling, conceptualizing affective dynamics as an implicit measure of emotion regulation in daily life. Indeed, most prior research using cross‐sectional measures focuses on STBs in relation to only one aspect of affective dynamics, such as NA intensity (e.g., Armey et al. ²⁰²⁰). These results also have the added utility of providing an alternative and naturalistic measurement of emotion regulation beyond traditional retrospective self‐report, which often measures an individual's perceived efficacy of emotion regulation strategies (e.g., Caprara et al. ²⁰⁰⁸; Gratz and Roemer ²⁰⁰⁴). Our hypotheses were partially supported, in that lower NED was significantly related to greater NA variability and NA inertia. These findings may have important implications for how negative emotions unfold over time among those with recent STBs. As NA variability and inertia may be risk factors for psychopathology (Links et al. 2008; Nisenbaum et al. ²⁰¹⁰), these findings suggest that low NED may indicate emotion regulation disturbances among those with recent STBs. Namely, in the context of low NED, individuals at high risk for suicide experience greater NA variability and inertia, in line with prior findings (Victor et al. 2021). Notably, we did not find a significant association between NA intensity and NED in the present study, which may be driven by high levels of NA in our sample (e.g., ceiling effects).

Our latter two hypotheses examined how NED related to SI severity, assessed in relation to the past week (at baseline) and prospectively across the EMA period. Unexpectedly, we found no significant relationships between past week SI characteristics at baseline or aggregated EMA frequency of SI and NED. Our sample size left us underpowered to detect medium and small effects, which may explain these null findings, as a recent meta‐analysis observed small to medium effects between NED and other maladaptive behavioral outcomes (Seah and Coifman 2022). Further, the present sample reported baseline SI severity at baseline that mirrored a bimodal distribution; the resulting need to dichotomize these variables also left us underpowered to detect smaller effects. Our findings are inconsistent with previous research linking poor NED with nonsuicidal self‐injury (Zaki et al. 2013), a known risk factor for STBs (Klonsky et al. 2013). We further observed null findings with respect to NED and suicide attempt history, which may also be interpreted with consideration for our low power to detect small effects and the high‐risk nature of the sample. Overall, findings suggest that other components of the emotion regulation process (e.g., emotion distress tolerance), rather than NED, may be more critical to understanding STBs.

4.1. Limitations

Results from the present study should be interpreted in light of factors that impact generalizability of our findings. First, our EMA items captured only passive SI, limiting our ability to assess more acute momentary SI. One possible impact of including items assessing active SI could be greater variability in SI severity (i.e., death ideation versus SI with a method in mind), which would allow for the examination of NED alongside a greater continuum of momentary SI risk. This is an important distinction given that NED may be related to active, but not passive, SI, which cannot be compared based on the wording for the present EMA items. Inclusion of a greater range of captured SI would also have the potential to increase statistical power. These issues also extend to the baseline SI measures which were dichotomized due to small cell sizes, therefore decreasing statistical power.

Additionally, our safety procedures, such as contacting participants following reported suicidal behavior, while necessary for participant well‐being, may have impacted study results. Researcher observation and response to EMA data may influence participant behavior (Bentley et al. 2024; Bentley et al. ²⁰²¹) and can contribute to concealment on subsequent surveys (Knorr and Ammerman 2025); however, the risk response approach used in the present study is less intrusive than much of the existing literature (Bentley et al. 2021). Further, participation in an EMA study may itself lead to changes in the phenomena being studied, for example, by creating reactivity to study safety protocols in willingness to report SITBs (Bentley et al. 2024).

With respect to statistical approaches, we used ICC to operationalize NED, a common technique within the field (Thompson et al. 2021); however, this approach cannot capture unique differences for specific emotion items. Idiographic or network approaches to examining specific emotions may be helpful for understanding the covariance of emotions at the momentary level within individuals. Previous work has also noted issues in using ICC to estimate emotion differentiation (i.e., inability to account for third variables), which warrant further attention in improving our understanding of NED (Nook et al. ^2021b).

Lastly, characteristics of our sample are relevant to interpreting study weaknesses. For example, the relatively small sample size of the present study limits our ability to detect significant effects. Future studies may consider increasing sample sizes for similar between person analyses (e.g., by combining EMA samples, when appropriate). Finally, the current sample is unique in several ways that may limit generalizability. Specifically, the sample was recruited from a region with documented shortages of mental health care providers (HPSA ²⁰²³), impacting availability of post‐discharge psychiatric care. This sample was also relatively homogenous in racial and ethnic identity.

5. Conclusions

Our findings provided a novel, methodologically sophisticated test of NED, STBs, and dynamic affective processes in a high‐risk sample by capturing NED across a 4‐week period, rather than using retrospective self‐report from a single point in time. Results indicate that NED may not be associated with STBs in this population. However, we did observe that NED was associated with facets of NA dynamics (i.e., affect inertia and variability). Thus, NED may be important in understanding emotion regulation processes in high‐risk adults, even if not directly associated with past or concurrent STBs. Future research may also benefit from examining whether NED shows within‐person variability at the momentary level and how these fluctuations may confer risk for suicide; such an approach would also allow for examining the temporal ordering of NED and STBs. Broadly, these findings suggest there is continued value in probing theoretically derived research questions regarding affect and STBs using more granular approaches. Namely, person‐centered multi‐level modeling may offer a richer understanding of how affect changes over time and how these affective dynamics may coincide with or predict STBs. These methods may inform context‐specific and individualized targets for intervention to reduce SI and ultimately prevent suicidal behavior.

With respect to clinical implications, our results suggest, that, for more severe or chronic STBs present in high‐acuity clinical samples, interventions geared towards improving emotion differentiation (i.e., emotion naming and emotion labeling) may not be most effective over the short term. However, it is inappropriate to interpret our findings to mean emotion‐focused modalities are unhelpful for STBs in general—instead, we suggest that further work is needed to clarify the underlying mechanisms of emotion focused therapies (e.g., changes to processes other than NED) and which mechanisms are most useful for specific populations facing STBs.

Author Contributions

Elizabeth C. Hoelscher: conceptualization (lead), methodology (lead), data curation (lead), formal analysis (lead), writing – original draft (lead). Sarah E. Victor: conceptualization (equal), software (lead); supervision (lead), funding acquisition (lead), resources (lead); writing‐review and editing (lead). Sheri L. Johnson: conceptualization (equal), Writing‐review and editing (equal). Jason Van Allen: conceptualization (equal), Writing‐review and editing (equal). Leslie Brick: conceptualization (equal), formal analyses (supporting), Writing‐review and editing (equal).

Funding

Research reported in this manuscript was supported by the National Institute of Mental Health, within the National Institutes of Health, under grant number R21MH124794 (PI: S.E. Victor). Additional support for this work was provided by APAGS/Psi Chi Junior Scientist Fellowship (PI: E.C. Hoelscher). While co‐author L.A.D. Brick was affiliated with the Warren Alpert Medical School of Brown University at the time this work was completed, her affiliation at the time of publication is with the University of New Mexico. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Ethics Statement

This study received approval from the Texas Tech University Institutional Review Board.

Consent

All participants provided informed consent prior to participating in this study.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are openly available in Open Science Framework at https://osf.io/t64r8/, reference number DOI 10.17605/OSF.IO/T64R8.