Investigating Predictors of Differential Trajectories of Suicidal Desire Across an Intensive Longitudinal Design

ABSTRACT

Objective

To understand the evolution of suicide risk, repeated assessments of suicide‐related symptoms are necessary. Though previous studies have examined sample‐level effects, there is a dearth of research examining possible differential risk trajectories, including vulnerable subgroups, and how subgroup differences may produce distinct pathways of suicide risk.

Methods

Community‐based adults (n = 237) with severe suicidal ideation completed six daily prompts across a 14‐day period. Growth mixture models identified subgroups (i.e., classes) of trajectories of SI. One‐way ANOVAs examined differential predictors of these classes' average intensity, variability, stability, and baseline levels of suicide risk factors.

Results

A three‐class linear trajectory was best‐fitting. The largest class (Class 3, n = 146) exhibited decreases in suicidal desire; the second class (Class 2, n = 54) displayed increasing suicidal desire; and the smallest group (Class 1, n = 37) maintained moderate, stable levels of suicidal desire. Class 2 generally exhibited higher mean intensity/baseline levels, lower mean variability, and mixed stability of suicide risk factors.

Conclusions

Although our study design does not imply causality, a subset of participants may be emotionally reactive to repeated assessments. Future research should ethically investigate and address why some individuals may experience heightened suicide risk during repeated assessments and analytically accommodate class‐specific trajectories of suicidality.

Assessing suicide‐related symptoms, including thoughts, desires, intent, and behaviors, is an integral part of conducting research that furthers our understanding, assessment, and prevention of suicide. In particular, understanding trajectories in suicide risk over time—whether over the course of hours, days, weeks, months, or years—necessitates the repeated assessment of suicide‐related symptoms. Despite the importance of this research, concerns have been raised regarding the iatrogenic nature of repeatedly assessing suicide‐related symptoms and risk, namely whether individuals would experience higher levels of suicidal ideation when prompted with, and in reaction to, questions and stimuli about suicide (Lakeman and FitzGerald ²⁰⁰⁹). This question has been thoroughly tested empirically, with remarkably consistent findings: assessing suicidal ideation and behavior has not been found to increase suicidal ideation nor the likelihood of engaging in suicidal behavior (Cha et al. 2016; Coppersmith et al. ²⁰²²; Harris and Goh ²⁰¹⁷; Hom et al. ²⁰¹⁸; Husky et al. ²⁰¹⁴; Kivelä et al. ²⁰²³; Law et al. ²⁰¹⁵; Rogers ²⁰²¹). These findings are underscored by a meta‐analysis of 13 studies, where no evidence of iatrogenic effects was found (DeCou and Schumann ²⁰¹⁸). Notably, this holds true across studies that take place over a long period of time—up to a year—with frequent inquiries about participants' experiences with suicidal thoughts and suicide attempt history (Coppersmith et al. 2022; Hom et al. ²⁰¹⁸), and across at‐risk populations, including adolescents (Gould et al. 2005), high‐risk community‐based adults (Rogers 2021), Veterans (Bryan et al. 2016; Poindexter et al. ²⁰¹⁹), patients with borderline personality disorder (Law et al. 2015), adult psychiatric inpatients (Schatten et al. 2022), and undergraduate students (Bender et al. 2019; Hom et al. ²⁰¹⁸).

However, most studies have observed trajectories of suicidal ideation and/or behavior across entire samples, without consideration of potential differential trajectories across distinct subgroups. In notable exceptions, phenotypes of real‐time patterns of suicidal thoughts have been explored, largely in samples of adolescent and adult psychiatric inpatients (both in the hospital and immediately following discharge; Czyz et al. ²⁰²²; Kivelä et al. ²⁰²⁵; Kleiman et al. ²⁰¹⁸; Spangenberg et al. ²⁰²³). Each study found a varying number (i.e., 3 to 5) of profiles that differed based on the mean intensity and variability of suicidal ideation. For instance, Kleiman et al. (2018) found a 5‐profile solution with two classes at relatively low mean intensity of suicidal ideation (with low and moderate variability), one class at moderate mean intensity of suicidal ideation (with high variability), and two classes at relatively high mean intensity of suicidal ideation (with low and high variability). Spangenberg et al. (2023) examined profiles in two samples and found differing subgroups among them: one sample with four classes varying in mean and variability (i.e., low‐stable, high‐unstable, low‐moderately stable, and moderate‐unstable) and one sample with three classes (i.e., high‐stable, low‐stable, and low‐unstable). Kivelä et al. (2025) found three profiles consisting of varying frequencies/intensities/variabilities of suicidal ideation (i.e., high frequency/intensity‐moderate variability, moderate frequency/intensity‐high variability, moderate frequency‐low intensity/variability). Finally, Czyz et al. (2022) examined trajectories of suicidal thoughts across the course of their studies—both during and following psychiatric hospitalization—and found subgroups characterized by low, declining, and persistently high suicidal ideation. Phenotypes characterized by higher intensity and lower variability tended to have more severe depression symptoms and a greater likelihood of engaging in suicidal behavior (Czyz et al. 2022; Kivelä et al. ²⁰²⁵; Kleiman et al. ²⁰¹⁸; Spangenberg et al. ²⁰²³).

On the other hand, individuals who have more reactivity and volatility in affective states and suicide risk factors may be more likely to be sensitive to the effects of repeated assessments of suicide‐related symptoms. One study of individuals at “ultra‐high risk” for suicide found that instability in negative affect predicted frequency of suicidal thoughts (Palmier‐Claus et al. ²⁰¹²); the authors theorized that instability may contribute to suicidal thoughts via repeated reactivation and elaboration of latent suicidal cognitions. Moreover, as found in a study of individuals with a previous psychiatric illness/diagnosis, small shifts in negative affect may be capable of reactivating suicidal thoughts and ideas (Lau et al. 2004). Considering the nature of suicide risk assessment, sensitivity to small shifts may be related to some participants reporting drops in mood and elevated thoughts of suicide. Sensitivity to small shifts can be seen in individuals with psychiatric disorders that demonstrate difficulties with maintenance, reactivity, and/or avoidance of affective states, such as posttraumatic stress disorder and/or borderline personality disorder (Kivelä et al. 2023; Maraz et al. ²⁰²²).

Nonetheless, these nuanced possibilities have rarely been empirically tested among subgroups of suicidal participants. Examination of potential subgroups who may experience reactivity (e.g., worsened mood and/or increased suicidal desire within the short term in response to specific prompts or surveys) to designs with repeated measurements of suicide‐related symptoms would allow researchers and clinicians to tailor study safety protocols and/or interventions. Notably, although not well studied, reactivity could occur either immediately or cumulatively over time; it could also occur within specific contexts (e.g., poor sleep, stressful life events). Prior studies examining digital phenotypes based on repeated assessments/intensive longitudinal designs focused predominantly on the intensity and variability of suicidal thoughts themselves, as opposed to the patterns of multiple risk factors that may impact trajectories of risk. Examining potential predictors of different risk trajectories—including a variety of suicide risk factors, such as suicidal intent, interpersonal theory of suicide constructs (Chu et al. 2017), general and suicide‐specific rumination (Rogers and Joiner 2017, 2018), hopelessness (Beck et al. 1985), agitation (Busch et al. 2003; Rogers et al. ²⁰¹⁶), and perceived stressfulness of life events (Liu and Miller 2014)—will aid identification of individuals who may need further resources. Furthermore, prior studies also typically examined patterns of risk among individuals either undergoing or recently discharged from psychiatric hospitalization, limiting the generalizability of these findings to real‐world settings among those who are not necessarily engaged in mental health treatment. Extending this work to a high‐risk community sample will yield further insights regarding trajectories of suicide‐related outcomes during intensive longitudinal designs, since there may be not only population‐level differences (e.g., adolescent vs. adult, inpatient vs. community), but also different levels of statistical noise (e.g., research conducted in naturalistic vs. hospital settings) and non‐psychiatric drivers of suicidality that may be more generalizable to everyday interpretations of suicide risk.

Thus, the present study aimed to (1) identify patterns/profiles of change in suicidal desire across the course of a two‐week ecological momentary assessment (EMA) study through growth mixture modeling; and (2) examine profile group differences in the intensity (i.e., mean levels across the assessment period), variability (i.e., within‐person standard deviations, reflective of average within‐person dispersion), stability (root mean squared successive differences [RMSSD], reflective of the average degree of change across successive prompts), and baseline levels (i.e., levels during the initial prompt) of various suicide risk factors to see if there are identifiable/potentially malleable predictors of these profiles. We focused on suicidal desire as an outcome, as opposed to suicidal intent or another proximal suicide‐related endpoint (e.g., suicidal behaviors), given that suicidal desire reflects a relatively more common and fluid marker of human suffering (“ideational morbidity”) than suicidal intent. Although suicidal intent represents a behavioral marker that may predict subsequent suicide attempts, suicidal desire is worth understanding as a clinical outcome in its own right (Jobes and Joiner 2019), especially as it pertains to understanding responses to repeated assessments. Based on prior research on trajectories, we hypothesized that up to three risk trajectories would emerge, with most participants exhibiting either decreases or no change in suicidal desire across the course of the study (given that past research has largely found small decreases over time in aggregate), and a minority of participants exhibiting small increases in suicidal desire. However, we made no a priori hypotheses regarding which risk factors' intensity, variability, or stability would predict each trajectory, given that many of these factors have not been previously examined in relation to risk trajectories.

1. Methods

1.1. Procedures

Participants were recruited from online targeted social media advertisements that provided information about the study. Prospective participants were directed from these advertisements to an online screening to assess current suicidal ideation severity, as assessed by the Beck Scale for Suicide Ideation (BSS; Beck and Steer ¹⁹⁹¹). Those with a score of 11 or higher on the BSS, and who were aged 18 years or older, English‐speaking, and who lived in the United States, were eligible to participate. All eligible participants were then invited to download the Ethica Data mobile application via a link or QR code. Once downloaded, participants completed a baseline survey in which informed consent was obtained, alongside several sociodemographic characteristics, baseline psychological symptoms, and current and lifetime suicide‐related histories, through a series of self‐report questionnaires.

Participants responded to six daily EMA prompts across 14 consecutive days, with prompts occurring randomly within stratified equivalent intervals (e.g., 9:00 a.m. to 11:00 a.m., 11:00 a.m. to 1:00 p.m.) between 9:00 a.m. and 9:00 p.m. in participants' local time zones. Prompts were scheduled to not occur within 30 min of completion/expiration of another prompt. Although participants were encouraged to complete prompts right away, they had up to 1 h to complete each prompt before its expiration to ensure that data on current symptoms were being collected. Considering the high‐risk nature of the sample, a list of national mental health resources and instructions to create a coping card/safety plan (Stanley and Brown 2012) were provided at the end of each EMA assessment. After completion of the EMA portion of the study, participants were compensated with a $20.00 Amazon gift card. To encourage adherence to the protocol, a bonus $20.00 gift card was awarded to participants who attained an 80%+ response rate. Data collection took place between February and May 2019. All procedures were approved by the relevant university's Institutional Review Board.

1.2. Participants

The final sample included individuals who downloaded the EMA application, completed the baseline survey, and remained enrolled for at least 3 days of EMA monitoring. Specifically, 237 community‐based adults (61.6% cisgender women, 86.9% White/European American, 38.4% bisexual/pansexual, 58.2% single/never married, and 43.9% who completed at least some college), aged 18 to 55 years (M = 27.12, SD = 8.60), were included in the study. A more thorough breakdown of sociodemographic characteristics can be found in Table 1. All participants reported current significant suicidal ideation at baseline, most had a history of suicide plans (85.2%), and more than half reported at least one lifetime suicide attempt (67.5%).

TABLE 1.

Sociodemographic characteristics of full sample.

	N	%
Gender
Cisgender man	16	6.8%
Cisgender woman	146	61.6%
Transgender man	21	8.9%
Transgender woman	4	1.7%
Non‐binary/non‐conforming	38	16.0%
Other	2	0.8%
Age	M = 27.12, SD = 8.60
Race/Ethnicity a
White/European American	206	86.9%
Black/African American	9	3.8%
Hispanic or Latino/a	17	7.2%
Asian	9	3.8%
Pacific Islander	3	1.3%
American Indian/Native American	4	1.7%
Other	1	0.4%
Sexual orientation
Heterosexual/straight	79	33.3%
Gay/Lesbian/homosexual	40	16.9%
Bisexual/pansexual	91	38.4%
Asexual	16	6.8%
Did not specify
Marital status
Single/never married	138	58.2%
Cohabiting	41	17.3%
Married	37	15.6%
Separated	10	4.2%
Divorced	7	3.0%
Widowed	0	0.0%
Education
Less than high school	9	3.8%
High school diploma/GED	36	15.2%
Some college	104	43.9%
Associates degree	22	9.3%
Bachelor's degree	47	19.8%
Master's degree	16	6.8%
Doctoral/professional degree	3	1.3%
Suicide‐related histories
Lifetime suicide plan	202	85.2%
Lifetime suicide attempt	160	67.5%

^a Categories were not mutually exclusive.

Participants completed, on average, 12.29 days of EMA monitoring (SD = 2.92) and 4.13 prompts per day (SD = 1.82), reflecting a response rate of 87.8% for days completed and 68.8% of assessments completed each day. In total, 12,781 prompts were completed (69.1%), with 63.7% of the 237 participants completing a minimum of 80% of the EMA prompts and consequently receiving the compensation bonus. Differences in response rates across sociodemographic and clinical characteristics are reported in Rogers (2021).

1.3. Measures

1.3.1. Suicide‐Related Outcomes

During each prompt, participants rated their current suicidal desire (“How strong is your current desire to make a suicide attempt?”) and intent (“How strong is your current intent to make a suicide attempt?”) on visual analog scales ranging from 0 (No desire/intent) to 100 (Definite desire/intent). The suicidal desire variable was used to identify latent trajectories of suicidal desire, whereas characteristics of suicidal intent were used as a predictor of the resultant trajectories.

1.3.2. Affective States

Participants rated their current affective states, including depression, anxiety, happiness, psychological pain, hopelessness, and agitation on a visual analog scale ranging from 0 (low) to 100 (high).

1.3.3. Thwarted Interpersonal Needs

Four items from the Interpersonal Needs Questionnaire (Van Orden et al. 2012) were adapted to assess current perceived burdensomeness (“Since the last assessment, I have felt like a burden on the people in my life” and “Since the last assessment, I have felt useless”) and thwarted belongingness (“Since the last assessment, I have felt like I belong” and “Since the last assessment, I have felt lonely”). Participants rated each item on a 7‐point scale ranging from 1 (Not at all) to 7 (Extremely), with select items reverse‐coded such that higher scores reflect more severe perceived burdensomeness and thwarted belongingness. Between‐person reliability was excellent (R _kF = 0.99 and 0.95), and within‐person reliability of change was good (R_C = 0.79 and 0.72) for perceived burdensomeness and thwarted belongingness, respectively.

1.3.4. Rumination

Two items (i.e., “I have been dwelling on my mistakes, failures, or losses” and “I keep thinking about something negative that has happened”) were drawn from adaptations made by Ruscio et al. (2015) to items from the Ruminative Responses Scale (Nolen‐Hoeksema and Morrow ¹⁹⁹¹) to assess general rumination. Likewise, two items (i.e., “When I have thoughts of suicide, I cannot escape these thoughts” and “When I have thoughts of suicide, I imagine the process of how I want to kill myself”) were adapted from the Suicide Rumination Scale (Rogers et al. 2022) to assess suicide‐specific rumination. During each prompt, participants rated each item on a 5‐point scale ranging from 0 (Not at all) to 4 (Almost always). Between‐person reliability was excellent (R _kF = 0.99) and within‐person reliability of change was good (R _C = 0.78) (Shrout and Lane 2012).

1.3.5. Perceived Stress

During each prompt, participants responded to an open‐ended item assessing the most significant positive or negative event that had occurred since the prior assessment. They were then asked a follow‐up question, which was administered regardless of whether or not participants provided an open‐ended response, that probed how stressful the event was perceived to be on a 0 (Not at all stressful) to 100 (Extremely stressful) visual analog scale; this rating was used to assess perceived stress in analyses.

1.4. Data Analytic Strategy

Growth mixture models, using maximum likelihood estimation with the expectation–maximization algorithm, were estimated to identify subgroups (i.e., classes/profiles) of trajectories of suicidal desire across the study period at the day level (i.e., suicidal desire averaged across days for each participant). Estimation was carried out sequentially on one‐, two‐, three‐, and four‐class solutions, examining both linear and quadratic patterns of change using the lcmm package in R (Proust‐Lima et al. ²⁰¹⁷). Goodness‐of‐fit was evaluated using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), with lower values reflecting better model fit, and entropy, with values closer to 1 indicating good classification. Interpretability of the latent trajectories and the relative size of classes (i.e., avoiding classes representing less than 5% of the sample) were also considered in determining the ideal number of latent trajectories. Posterior class probabilities were then derived for the best‐fitting model to assign individuals to specific classes. Of note, although we focus specifically on suicidal desire in this paper, we report sensitivity analyses focused on trajectories of suicidal intent as an outcome for interested readers in the supplemental materials (Table S1).

A series of general linear models (i.e., one‐way ANOVAs) were then conducted to examine differential predictors of these classes. Specifically, the average intensity, variability, stability, and baseline level of each risk factor were compared across the latent trajectories. To assess average intensity, mean scores of each risk factor were calculated for each participant across the 2‐week assessment period. To assess average variability, within‐person standard deviations were calculated for each risk factor across the assessment period. To assess stability, the RMSSD (von Neumann et al. ¹⁹⁴¹), which is reflective of the average stability or magnitude of change of scores across successive assessments (such that larger relative values equal less stability across consecutive prompts), was calculated for each risk factor. Finally, to assess baseline levels, the first EMA prompt was used. In each model, Welch‐corrected F‐tests were used when the assumption of homogeneity of variance was violated, and Tukey's HSD was used as a post hoc correction. Additionally, due to the large number of statistical tests run (61 group difference analyses in total, in addition to pairwise comparisons), we adjusted our Tukey‐corrected post hoc alpha further to p = 0.01.

2. Results

2.1. Trajectories of Suicidal Desire Across the Study Period

Aggregated across the full sample, suicidal desire significantly decreased across the study period (B = −0.14, SE = 0.04, p < 0.001, 95% CI [−0.21, −0.06]). A summary of model statistics for all growth mixture models is presented in Table 2. Based on goodness‐of‐fit indices and interpretation of the class sizes and models, the three‐class linear trajectory model was selected as the optimal set of trajectories for these data (visualized in Figure 1). Class 1 consisted of 37 participants who did not significantly change in suicidal desire, on average, across the study period (B = −0.002, SE = 0.02, p = 0.921, 95% CI [−0.05, 0.04]). Class 2 consisted of 54 participants who significantly increased in suicidal desire across the study period (B = 0.13, SE = 0.02, p < 0.001, 95% CI [0.09, 0.16]) and who had significantly higher starting points on suicidal desire than Class 1 (B = 1.94, SE = 0.41, p < 0.001). Finally, Class 3 consisted of 146 participants who significantly decreased in suicidal desire across the study period (B = −0.05, SE = 0.01, p < 0.001, 95% CI [−0.07, −0.03]) and who had significantly lower starting points on suicidal desire than Class 1 (B = −1.93, SE = 0.36, p < 0.001).

TABLE 2.

Summary statistics from growth mixture models.

Model	AIC	BIC	Entropy	Class 1 (%)	Class 2 (%)	Class 3 (%)	Class 4 (%)
Linear: 1 class	23700.94	23721.75	1	100
Linear: 2 classes	23689.95	23721.16	0.89	4.22	95.78
Linear: 3 classes	23588.27	23629.89	0.84	15.61	22.78	61.60
Linear: 4 classes	23594.27	23646.30	0.59	19.83	22.78	57.38	0
Quadratic: 1 class	23649.18	23683.86	1	100
Quadratic: 2 classes	23657.18	23705.73	0.001	46.41	53.59
Quadratic: 3 classes	23602.24	23664.67	0.90	5.49	90.71	3.80
Quadratic: 4 classes	23572.51	23648.80	0.59	7.59	25.32	67.09	0

Abbreviations: AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion. Entropy is used to evaluate the precision of class assignments for individual participants; higher entropy indicates clear class separation, whereas low entropy suggests more ambiguous class membership.

FIGURE 1.

Trajectories of suicidal desire across the three identified classes.

There were significant differences in response rates across these classes (F [2, 226] = 4.04, p = 0.019), with lower response rates in Class 2 (M = 64.4%) than in Class 3 (M = 72.9%), though this was non‐significant (at a 0.01 alpha level) after correcting for Type I error (p = 0.038). Moreover, there were significant age differences across classes (F [2, 216] = 5.52, p = 0.005), such that participants in Class 3 (M = 25.89, SD = 7.03) were significantly younger than those in Class 2 (M = 30.54, SD = 10.72, p = 0.003). However, there were no differences in class across gender (χ ² [10] = 15.88, p = 0.103), any race (χ ² [2] = 1.03 to 4.52, ps = 0.104 to 0.596), sexual orientation (χ ² [8] = 5.61, p = 0.691), marital status (χ ² [10] = 6.39, p = 0.782), or suicide attempt history (χ ² [2] = 0.56, p = 0.755).

2.2. Predictors of Change Profiles

2.2.1. Mean Intensity

There were significant differences in the mean intensity of every suicide risk factor across the three trajectory classes (see Table 3 for means, standard deviations, and F‐statistics). Specifically, Class 2 had significantly higher levels of suicidal intent, depression, anxiety, psychological pain, hopelessness, perceived burdensomeness, rumination, suicide‐specific rumination, agitation, and perceived stress, and lower levels of happiness than both Class 1 (ps < 0.001 to =0.008) and Class 3 (ps < 0.001). Class 2 also had significantly higher levels of thwarted belongingness than Class 3 (p < 0.001), but not Class 1 (p = 0.208). Moreover, Class 1 had significantly higher levels of suicidal intent (p < 0.001), depression (p = 0.002), psychological pain (p = 0.008), hopelessness (p < 0.001), perceived burdensomeness (p = 0.006), and rumination (p < 0.001) than Class 3, but no differences in anxiety (p = 0.378), happiness (p = 0.399), thwarted belongingness (p = 0.029), suicide‐specific rumination (p = 0.025), agitation (p = 0.092), or perceived stress (p = 0.359).

TABLE 3.

Means (standard deviations) of average intensity of suicide risk factors, stratified by trajectory class.

Risk factor	Class 1	Class 2	Class 3	F
Suicidal intent	29.11 (16.73)	48.73 (29.11)	9.29 (11.61)	95.48***
Depression	62.03 (20.15)	77.07 (23.32)	47.27 (23.46)	32.80***
Anxiety	55.39 (24.51)	73.06 (20.28)	49.65 (23.54)	19.31***
Happiness	35.83 (17.42)	19.23 (16.06)	40.11 (18.25)	26.19***
Psychological pain	54.08 (24.48)	74.87 (23.81)	40.03 (25.61)	36.63***
Hopelessness	54.05 (22.69)	78.20 (18.50)	36.75 (22.60)	68.92***
Perceived burdensomeness	3.30 (1.54)	4.62 (1.24)	2.48 (1.44)	43.28***
Thwarted belongingness	4.43 (1.20)	4.84 (0.89)	3.89 (1.17)	14.44***
Rumination	2.20 (0.92)	2.96 (0.90)	1.58 (0.86)	47.47***
Suicide‐specific rumination	2.19 (0.89)	2.85 (0.80)	1.69 (1.11)	25.06***
Agitation	41.47 (23.60)	59.32 (26.18)	32.29 (22.22)	25.03***
Perceived stress	41.34 (20.31)	56.36 (22.04)	36.22 (19.12)	18.94***

Note: Each cell contains: mean (standard deviation).

^*** p < 0.001.

2.2.2. Variability

Significant differences in the within‐person variability of every suicide risk factor, except for suicide‐specific rumination, agitation, and perceived stress, were found across the three trajectory classes (see Table 4 for means, standard deviations, and F‐statistics). Namely, Class 2 had significantly less variability in depression and happiness than Class 1 (ps < 0.001 to = 0.0098) and Class 3 (ps < 0.001). Furthermore, Class 2 had significantly less variability than Class 3 (ps < 0.001 to = 0.003), but not Class 1 (ps = 0.016 to 0.728), in anxiety, psychological pain, hopelessness, perceived burdensomeness, and rumination. Class 3 had significantly less variability in suicidal intent than Class 1 (p < 0.001) and Class 2 (p < 0.001). Finally, there were no pairwise differences in thwarted belongingness after adjusting our Type I error rate (ps = 0.014 to 0.854).

TABLE 4.

Means (standard deviations) of average variability of suicide risk factors, stratified by trajectory class.

Risk factor	Class 1	Class 2	Class 3	F
Suicidal intent	11.83 (6.16)	12.99 (7.85)	6.66 (6.79)	18.89***
Depression	18.42 (8.59)	10.54 (8.48)	19.14 (9.81)	15.74***
Anxiety	18.81 (8.86)	14.97 (8.91)	21.24 (9.87)	7.97***
Happiness	17.46 (7.54)	12.18 (8.67)	18.26 (8.11)	10.21***
Psychological pain	16.58 (8.71)	12.55 (9.48)	18.29 (10.99)	5.57**
Hopelessness	18.08 (9.95)	12.73 (8.72)	19.38 (10.29)	8.17***
Perceived burdensomeness	1.05 (0.49)	0.71 (0.45)	1.20 (0.61)	14.01***
Thwarted belongingness	0.92 (0.56)	0.86 (0.43)	1.10 (0.52)	4.81**
Rumination	0.81 (0.43)	0.62 (0.37)	0.88 (0.40)	7.81***
Suicide‐specific rumination	0.59 (0.47)	0.53 (0.34)	0.51 (0.36)	0.57
Agitation	19.77 (10.62)	16.80 (8.69)	20.41 (11.28)	2.09
Perceived stress	29.54 (12.72)	27.94 (12.45)	29.83 (12.52)	0.41

Note: Each cell contains: mean (standard deviation).

^** p < 0.01.

^*** p < 0.001.

2.2.3. Stability

There were significant differences in the stability of suicidal intent, depression, anxiety, happiness, hopelessness, perceived burdensomeness, and rumination across the three trajectory classes (see Table 5 for means, standard deviations, and F‐statistics). Specifically, Class 2 had more stable depression than Class 1 (p = 0.003) and Class 3 (p < 0.001), Class 2 had more stable anxiety and happiness than Class 3 (ps < 0.001 and = 0.001), but not Class 1 (ps = 0.015 and 0.131), and Class 3 had more stable suicidal intent than Class 1 and Class 2 (ps < 0.001). However, after adjusting our alpha to 0.01, there were no significant pairwise differences in hopelessness (ps = 0.012 to 0.928), perceived burdensomeness (ps = 0.014 to 0.943), or rumination (ps = 0.022 to 0.816).

TABLE 5.

Means and standard deviations of average stability of suicide risk factors, stratified by trajectory class.

Risk factor	Class 1	Class 2	Class 3	F
Suicidal intent	13.10 (6.88)	14.39 (8.13)	6.77 (5.76)	32.20***
Depression	19.79 (10.09)	12.88 (9.61)	20.28 (9.54)	11.54***
Anxiety	21.66 (11.25)	17.29 (9.99)	23.71 (10.35)	7.25***
Happiness	22.26 (9.43)	16.48 (11.09)	21.98 (8.93)	6.84**
Psychological pain	18.98 (9.84)	16.90 (11.82)	19.71 (10.30)	1.34
Hopelessness	22.46 (12.38)	16.85 (10.96)	21.74 (9.79)	4.70*
Perceived burdensomeness	1.24 (0.60)	0.99 (0.69)	1.27 (0.60)	4.04*
Thwarted belongingness	1.17 (0.61)	1.10 (0.57)	1.28 (0.53)	2.32
Rumination	0.93 (0.40)	0.80 (0.50)	0.98 (0.39)	3.55*
Suicide‐specific rumination	0.59 (0.39)	0.58 (0.39)	0.57 (0.37)	0.06
Agitation	22.35 (10.17)	20.49 (10.50)	24.30 (11.63)	2.32
Perceived stress	37.89 (15.79)	36.53 (15.53)	38.08 (12.71)	0.24

Note: Each cell contains: mean (standard deviation).

^* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

2.2.4. Baseline Levels

Significant differences were found across the three trajectory classes (see Table 6) in suicidal intent, depression, happiness, psychological pain, hopelessness, perceived burdensomeness, suicide‐specific rumination, and agitation. Class 2 had higher baseline levels of suicidal intent than Classes 1 and 3 (p < 0.001), and Class 1 had higher baseline levels of suicidal intent than Class 3 (p = 0.001). Furthermore, Class 2 had significantly higher baseline levels of depression, happiness, psychological pain, hopelessness, perceived burdensomeness, suicide‐specific rumination, and agitation than Class 3 (ps < 0.001 to 0.004), as well as higher levels of hopelessness than Class 1 (p = 0.001). Class 1 also had higher baseline suicide‐specific rumination than Class 3 (p = 0.008).

TABLE 6.

Means (standard deviations) of baseline levels of suicide risk factors, stratified by trajectory class.

Risk factor	Class 1	Class 2	Class 3	F
Suicidal intent	33.04 (24.66)	54.69 (29.50)	15.91 (17.08)	42.73***
Depression	66.78 (27.35)	76.43 (27.11)	55.29 (27.91)	8.08***
Anxiety	59.70 (34.23)	69.49 (28.54)	59.38 (28.35)	1.59
Happiness	35.59 (26.82)	18.17 (22.97)	42.99 (24.07)	13.41***
Psychological pain	52.93 (29.23)	72.74 (30.91)	50.44 (30.28)	7.16**
Hopelessness	55.26 (29.99)	81.26 (26.08)	47.72 (27.71)	18.89***
Perceived burdensomeness	3.56 (1.80)	4.69 (1.59)	3.17 (1.80)	9.64***
Thwarted belongingness	4.59 (1.64)	5.09 (1.03)	4.44 (1.38)	2.91
Rumination	2.60 (1.10)	2.96 (1.18)	2.29 (1.25)	3.99*
Suicide‐specific rumination	2.69 (0.86)	2.77 (1.01)	2.03 (1.03)	9.60***
Agitation	57.96 (32.25)	62.49 (31.02)	42.66 (30.43)	6.52**
Perceived stress	45.12 (36.57)	64.97 (33.36)	46.43 (37.12)	3.60*

Note: Each cell contains: mean (standard deviation).

^* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

3. Discussion

The aims of the present study were twofold. First, we sought to identify various presentations (i.e., classes/profiles) and trajectories (i.e., linear versus parabolic) of suicidal desire across a 14‐day study period among those at high risk for suicide. Once classes were established, we then explored class‐specific predictability patterns; that is, we compared each group's intensity, variability, stability, and baseline levels regarding a number of suicide‐relevant risk factors.

A three‐class linear trajectory model best fit the data on suicidal desire. Class 1 (n = 37; 15.7%) showed no change across a 14‐day study period, Class 2 (n = 54; 22.8%) showed increasing suicidal desire, and Class 3 (n = 146; 61.6%) showed decreasing desire. The majority demonstrated reduced suicidal desire, consistent with overall sample trends, but a notable subset showed increased risk, underscoring important heterogeneity in trajectories. These findings align with previous research, which also found a small subset of participants with increasing suicidal desire (Salagre et al. 2021; Zhuojun et al. ²⁰²⁵). This may reflect a regression towards participants' natural baseline levels of suicidal ideation. Thus, future research would benefit from longer durations of data collection (e.g., months). Alternatively, these participants may have reacted negatively towards repeated suicide risk assessment. Thus, it is recommended to include opportunities for participant insight (i.e., awareness of their own fluctuations in suicidal desire) and reflection (i.e., participant‐generated alternative explanations for such changes) which may be qualitatively analyzed. If study‐related distress is, in part, driving the changes in suicidal desire, we would recommend that future research is responsive to any specific ethical concerns and considerations that emerge. Thus, a crucial first step towards resolution is to understand why this group presented with elevated suicide risk.

One avenue towards answering such a question requires an understanding of the differential predictability patterns across the three groups. Our study did just this, comparing each group's mean intensity, variability, stability, and baseline levels regarding a number of suicide‐relevant risk factors. Consistent with prior work showing that higher mean symptom levels are strongly associated with elevated suicide risk (Franklin et al. 2017; Kleiman et al. ²⁰¹⁸), participants in Class 2 (i.e., those with higher initial suicidal desire that increased over time) displayed significantly greater mean levels of suicidal intent, depression, anxiety, psychological pain, hopelessness, perceived burdensomeness, thwarted belongingness (only compared to Class 3), rumination (general and suicide‐specific), agitation, and perceived stress, and lower levels of happiness. Class 1 (i.e., stable, intermediate desire) also exhibited elevated suicidal intent, depression, psychological pain, hopelessness, perceived burdensomeness, and rumination as compared to Class 3. Though speculative, future research should examine if the differences between Class 2 and Class 3—higher mean levels of anxiety, suicide‐specific rumination, agitation, and perceived stress—may represent differentiating markers of short‐term escalation.

Class 2 was also significantly more homogenous (i.e., exhibited less variability) and stable across several constructs, showing less within‐person variability in symptoms of anxiety, psychological pain, hopelessness, perceived burdensomeness, and rumination compared to Class 3. One interpretation may be that Class 2 is characterized by chronic presentations of low mood (i.e., higher depression and lower happiness). Thus, future research may benefit from incorporating clinical assessments to determine if the severity or duration of a co‐occurring mood disorder may partially mediate subsequent increases in suicidal desire throughout repeated assessment (e.g., persistent depressive disorder, major depressive disorder—recurrent; see American Psychiatric Association ²⁰²²). These individuals (Class 2) also exhibited more severe baseline levels of symptoms for most risk factors, highlighting potential factors that may help predict who may be at risk for increasing symptoms, reactivity to study procedures, and/or adverse events (and who may benefit from more thorough monitoring and intervention during study participation). Importantly, we do not suggest that higher‐risk individuals should be excluded from study participation, as it is important to understand and prevent suicide among those with the greater acuity and severity of symptoms. However, further consideration of safety procedures during study participation is warranted based on these results. On the other hand, Class 3 was more homogenous in suicidal intent (i.e., consistently lower ratings of suicidal intent) than both classes, and more stable in suicidal intent than Class 1. Given that repeated assessments have been shown to have therapeutic effects by enhancing self‐awareness and emotional processing, future research should examine if this subgroup may particularly benefit from continued structured monitoring and feedback interventions (Finn and Tonsager 1992; Hom et al. ²⁰¹⁸).

Finally, it is worth considering how the nature of the present sample (a community‐based sample of adults experiencing significant suicidal ideation at baseline) may have contributed to these findings. Community‐based samples may capture a broader range of variability in both risk and protective processes relative to highly selected clinical samples (e.g., inpatients hospitalized for suicidality), which are often characterized by acute severity and more homogenous elevations in risk. At the same time, we acknowledge that the present study was not designed to directly test differences between community and psychiatric samples, and thus, we were unable to empirically determine the specific mechanisms driving these possible differences. One plausible interpretation, however, is that individuals in community samples may differ meaningfully in their exposure to and engagement with mental health services. For example, compared to individuals recently discharged from psychiatric hospitalization, community participants may be less likely to have received formal interventions (e.g., psychotherapy, psychiatric care), may lack established support systems (e.g., ongoing relationships with mental health providers), and may have had limited exposure to skills‐based approaches for managing suicide risk (e.g., coping with interpersonal distress or regulating acute affect). Relatedly, community samples may include individuals who are disclosing suicidal ideation for the first time. Thus, future research should examine if such factors may, in part, explain why one group reported increases in ideation over the course of the study period.

3.1. Limitations and Future Directions

As with all research, it is important to note the limitations within these findings. First and foremost, this study was observational in nature and did not experimentally manipulate the number or frequency of assessments. Thus, we are unable to infer that different trajectories of suicidal desire were attributable to study design and repeated assessments, as opposed to regression to the mean or the influence of any other external factors. Likewise, without a control condition, we cannot determine if those in Class 2 would have increased in suicidal desire despite participation in an ecological momentary assessment study design. We also did not explicitly ask debriefing questions on reactivity, though this was occasionally mentioned in open‐ended feedback (reported in Rogers ²⁰²¹). Second, this sample was predominantly White, women, and single; thus, this work may not generalize to other sociodemographic groups. As such, we encourage ethical replication of this research (i.e., combative against the potential for iatrogenic effects), which we further outline below. Third, EMA prompts were only delivered between 9:00 a.m. and 9:00 p.m. in participants' local time‐zones; it would be preferable in future research to optimize the timing of these prompts and tailor them to participants' typical wake hours. Fourth, due to the average intensity differences between Class 2 versus Class 1 and Class 3, there is a need to examine other potential factors that were not collected in this study. Such examples may include emotion dysregulation, distress tolerance, and emotional reactivity. It would also be useful for future research to investigate the specific patterns of change that are associated with reactivity to prompts. For instance, do increases in suicidal desire happen immediately? Is there a cumulative effect (sensitization versus habituation) to suicide‐related cues over time for some individuals? Are there certain contexts (e.g., poor sleep, stressful life events) that may predict instances of reactivity?

Moreover, as this study did not involve direct researcher intervention throughout the study period, future research may consider inclusion of clinicians more generally to combat against iatrogenic effects. One example may include the addition of a clinical interview or diagnostic evaluations before beginning the assessments. Such inclusions may aid in conceptualization as to why Class 2 elevated in suicidal desire throughout the duration of the study. Alternatively, another interpretation of the lack of researcher intervention may be that participants were more comfortable disclosing actual suicidal desire (i.e., leading to more accurate measurement), which is consistent with research indicating that suicidal ideation is more frequently reported in EMA designs rather than end‐of‐study assessments (Czyz et al. 2018; Gratch et al. ²⁰²⁰). Although this may benefit understanding of the course of suicidal desire, there may also be benefits in including clinical intervention throughout study participation (e.g., comprehensive risk assessments and risk management, opportunities for qualitative data collection on fluctuations in suicide risk, etc.). Thus, future research should consider how and when to involve clinicians into the data collection process (for comprehensive risk management protocol, see Bentley et al. ²⁰²⁴). However, it is also important to see how participants and individuals with lived experience perceive these protocols, as well as how it may impact study participation and honest disclosure/responding.

Nevertheless, it is equally notable to acknowledge this study's strengths. This is the first study, to our knowledge, to identify three different groups (i.e., classes/profiles) trajectories of suicidal desire over a 14‐day study period. Moreover, such differences were identified among those already presenting with significant suicide risk. This study also contradicts previous research (see Husky et al. ²⁰¹⁴; Kivelä et al. ²⁰²²) and identifies that there may be a group that experiences iatrogenic effects of repeated assessment of suicidal desire (i.e., elevations in suicidal desire). To our knowledge, this was also the first investigation to examine variability in suicide risk factors, in addition to variability of suicidal ideation and intent themselves. Presented in conjunction with class‐specific patterns on suicide‐relevant risk factors, this work lays critical foundations for the future of suicide risk assessment, management, and intervention efforts.

4. Conclusions

Examining contemporary methodologies like ecological momentary assessments within clinical psychology is essential to understand individual trajectories of suicide risk. Findings from this investigation reveal three distinct classes/profiles of respondents who engaged in repeated assessment spanning a 14 day study duration: The largest group exhibited initially lower levels of suicidal ideation, which decreased over the assessment period; the second largest group displayed higher initial levels of suicidal ideation, which escalated throughout the assessments; and the smallest group maintained moderate levels of suicidal ideation between the first two groups, remaining stable throughout the study. Subsequent studies should continue to ethically explore and effectively address why these groups differ. Hence, future research may elucidate unique risk by incorporating analytical frameworks to accommodate class‐specific trajectories of suicidal ideation.

Author Contributions

Dea Mitaj: conceptualization, writing – original draft, writing – review and editing. Megan L. Rogers: conceptualization, investigation, funding acquisition, writing – original draft, methodology, validation, visualization, writing – review and editing, formal analysis, project administration, data curation, supervision, resources. Morgan Robison: conceptualization, writing – original draft, writing – review and editing. Justin H. Tran: conceptualization, writing – original draft, writing – review and editing.

Funding

This work was supported by funding awarded to the Military Suicide Research Consortium (W81XWH‐16‐20003; W81XWH‐16‐2‐0004), an effort supported by the Office of the Assistant Secretary of Defense for Health Affairs, as well as a grant awarded to the first author by the American Psychological Association. The opinions, interpretations, conclusions, and recommendations expressed do not necessarily represent those of the Military Suicide Research Consortium, Department of Defense, or American Psychological Association.

Ethics Statement

All procedures were approved by the Florida State University Human Subjects Committee (HSC #2018.25295).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Summary statistics from growth mixture models examining suicidal intent.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.