A Change Would Do You Good…or Would It? The Role of Emotion Variability in Female Adolescents’ Depressive Symptoms

One emotion dynamics construct that has been leveraged in research on the development of depression in adolescence is emotion variability, defined as change over time in one’s emotion level across a range of contexts. Higher levels of emotion variability are theorized to reflect broad failures of reactivity and regulation, which may consolidate into clinically altered emotional function, such as is observed in depression (Houben et al., 2016). Laboratory studies of dynamic emotion in adolescence suggest that altered emotion dynamics (at the timescale of seconds to minutes) are implicated in the development and maintenance of depressive symptoms in adolescence (Koval et al., 2012; Kuppens et al., 2010, 2012). Studies of emotion variability have also been conducted using ecologically sensitive methods, (e.g., ecological momentary assessment [EMA], daily diary) to capture adolescent emotional experience in daily life (e.g., Neumann et al., 2011; Silk et al., 2003). Although laboratory methods facilitate the examination of adolescent emotion dynamics in a more rigorously controlled manner, ecologically valid methods like EMA present a valuable opportunity to advance understanding of dynamic emotion processes as they unfold in adolescents’ real-world contexts (i.e. in the presence of peers, parents, and realistic emotion-inducing situations).

Several studies have found that higher levels of NE variability are associated with elevated depressive symptoms (e.g., low/irritable mood, changes in appetite/sleep, excessive guilt) and diagnosis (Maciejewski et al., 2014, 2019; Neumann et al., 2011; Silk et al., 2003, 2011). Beyond concurrent associations, there is evidence that NE variability may play a mechanistic role in depressive symptom maintenance in adolescence (Neumann et al., 2011; Silk et al., 2011). However, several studies have shown that NE variability is no longer a significant predictor of depressive symptoms above and beyond mean levels of NE, likely due to high levels of confounding between the mean and variability (Ringwald & Wright, 2022). Fewer studies have examined associations between depressive symptoms and PE variability, with some studies demonstrating associations between internalizing symptoms and PE variability (Morgan et al., 2017; van Roekel et al., 2016) and other studies not (Forbes et al., 2012). However, research suggests that PE alterations may be relatively unique to depression amongst internalizing disorders, compared to alterations in NE, which are more broadly shared across internalizing psychopathology (Forbes et al., 2004; although see Sequeira et al., 2022), pointing to a potential role for PE variability in the development of anxiety symptoms.

Despite these initial indications of a positive relationship between emotion variability and depressive symptoms, variation in study methodology underlines the need for further investigation. Studies have differed in timescale (hours vs. days; Morgan et al., 2017; Silk et al., 2003, 2011; van Roekel et al., 2016; Maciejewski et al., 2014, 2019; Neumann et al., 2011), and whether they examine NE and PE variability separately (e.g., Silk et al., 2011; Morgan et al., 2017) or in conjunction (e.g., “mood” variability; e.g., Neumann et al., 2011). Finally, studies have differed in the extent to which they adjusted for measures of central tendency, with some studies including no adjustment (e.g., Silk et al., 2003) and some adjusting for average levels of PE/NE (e.g., Maciejewski et al., 2014). Some scholars argue that emotion dynamics measures add limited predictive benefit beyond measures mean-level emotion and advocate for the mean as the strongest predictor of symptoms/disorder (Dejonckheere et al., 2019). However, others have more recently argued that the high degree of dependence between the mean and variabiltiy measures is a confound, and advocate for the use of the mode as a more clincially relevant “set-point” of affect (Ringwald & Wright, 2022). However, to date, no studies have investigated both methods (i.e., adjustment for mean vs. mode) in the same study in adolescence.

Participants were drawn from the first wave of the GIRLS Brain Study which was approved by the Institutional Review Board at the University of Pittsburgh. Advertisements and community postings were used to recruit early adolescents assigned female at birth ages 11–13 for a longitudinal study of neural and social factors in the development of depression and anxiety in adolescents assigned female at birth. Five hundred fifty-two families completed an initial phone/web screening. Two hundred and thirty-five adolescent females met criteria and 197 participants completed the initial screening visit. As the study was designed to examine risk for the development of anxiety and depression, exclusion criteria included current diagnosis or history of DSM-5 anxiety disorder (except specific phobia, due to the narrower scope and decreased functional impairment) and/or major depressive disorder, to ensure participants did not already have a history of anxiety or depression and facilitate the interpretation of results in the context of development of these disorders. The study also excluded for severe psychopathology (diagnoses of bipolar disorder, schizophrenia, or acute suicidality) due to the extreme alterations in emotional experience often characteristic of these disorders. Participants with diagnoses of the hyperactive-impulsive and combined types of attention-deficit/hyperactivity disorder were also excluded due to concerns about movement during an fMRI scanning protocol not pertinent to the present analysis. Additional exclusion criteria included: a diagnosis of autism spectrum disorder; IQ < 70 (assessed with the two-scale Wechsler Abbreviated Scale of Intelligence; Wechsler, 1999); current or past neurological or serious medical condition; fMRI contraindications; uncorrected impaired vision; past head injury or neurological anomalies; and current use of medication that impacts the central nervous system. All enrolled participants provided informed consent (from parents/guardians) and informed assent (from adolescents). The overarching study was designed to have adequate power (β = 0.80, α < 0.05) to detect medium-sized effects. A separate power analysis for the current study was not conducted.

The final sample for the overarching study comprised 129 female adolescents. Female sex assigned at birth was parent-reported. Approximately two-thirds of participants were at elevated risk for social anxiety and depression based on elevated dispositional shyness/fearfulness (i.e., scored 0.75 SD or higher than the sample mean on the Fear and/or Shyness subscales of the Early Adolescent Temperament Questionnaire-Revised; EATQ-R; Ellis & Rothbart, 2001). As noted below, 12 participants were excluded from analyses based on missing the EMA or not meeting minimum response thresholds during the EMA period and some participants had missing data on depression outcomes, resulting in a final analysis sample of n = 117 early adolescents.

At the end of the second lab visit, adolescents were provided with a study smartphone and instructed on completing EMA on a range of emotional, social, and behavioral topics, most of which are not relevant to the current report. Participants were prompted three times on weekdays (early morning prior to school, after school, and evening) and four times on weekend days (morning, early afternoon, late afternoon, evening) for 16 consecutive days (six weekend days, ten weekdays) for a maximum of 54 observations. Prompts were randomly delivered within three-hour blocks in the timeframes described above, with the exception of the morning prompt which was delivered in a morning window set by the participant based on their waketime. Surveys took seven minutes to complete on average; participants received reminders every 15 min for up to an hour, after which time the survey expired.

At the beginning of each EMA survey, participants were asked to report their current mood on eight emotion indicators: happy, sad, joyful, worried, stressed, interested, mad, excited. Each emotion was rated on a 0–100 sliding scale where 0 = Not at all and 100 = Extremely, with the slider initially anchored at 0. To create PE and NE variability composites, the highest of the four indicators of each broad emotion measure (PE = happy, joyful, interested, excited; NE = sad, worried, stressed, mad) was selected within each observation to create an overall PE/NE value for that observation.² Reliability was adequate for both NE (α_within = 0.67, α_between = 0.94; ω_within = 0.68, ω_between = 0.95) and PE (α_within = 0.80, α_between = 0.94; ω_within = 0.80, ω_between = 0.94). Within-person SDs of momentary PE/NE (i.e., each observation) were calculated as a measure of PE/NE variability. Further, mean and modal PE/NE, respectively, were included to represent individuals’ average/set-points of PE/NE (Ringwald & Wright, 2022).

At the baseline visit and 6-, 12-, and 18-month follow-up, female adolescents completed the Mood and Feelings Questionnaire (Messer et al., 1995), a 33-item measure designed to assess symptoms of depression, via an online Qualtrics survey. The MFQ was not designed to have a strict clinical cutoff; however, psychometric studies have found that a score of 29 effectively discriminated clinical cases of depression. Reliabilities ranged from ⍺ = 0.89–0.95.

Age and socioeconomic status were evaluated as covariates in light of established associations with depression (e.g., Abitante et al., 2022; Griffith et al., 2021; Hankin et al., 2015; Maciejewski et al., 2014; Peverill et al., 2021). Age at the baseline visit was calculated in years by subtracting the participant’s birth date from their visit date. Socioeconomic status was approximated by including parent-reported annual gross income in dollars on a 0–10 scale, where 0 = 0–10,000 and 10 = 100,000+. Both age and socioeconomic status were group-mean centered to facilitate interpretation.

Analyses were conducted in R version 4.2.1 (R Core Team, 2022) using the nlme package (Pinheiro et al., 2025) and Mplus Version 7.31 (Muthen & Muthen, 2023).

All models were tested using maximum likelihood estimation to address missing data. Depressive symptom sum scores were used as indicators of a longitudinal growth curve model to capture depressive symptom trajectories, with the intercept fixed at baseline. Iterative model testing was conducted to ensure models were correctly specified. Model improvement was tested (where possible) with a −2 log likelihood test. A means-only model with a random intercept of participant was run first, to examine the decomposition of the variance at the between- and within-person levels, followed by the addition of fixed and random effects of linear time. We evaluated effect sizes by use of pseudo-R², following recommendations by Cohen (2013) for effect sizes in multiple regression of small (R² = 0.02), medium (R² = 0.15), or large (R² = 0.35).

Finally, to test the contributions of each predictor of interest (e.g., PE/NE variability) on depressive symptoms, a series of models were fit. The longitudinal growth model of depressive symptoms served as the baseline model. Once the fit of the longitudinal growth model was established, covariates were added to the model. Covariates were tested sequentially in the baseline model; if the covariate was not a significant predictor of depressive symptoms, the covariate was not retained. (Although see below regarding missingness related to socioeconomic status.) For models testing PE/NE variability, modal (or mean) PE/NE was added first, to test incremental improvement in model fit with the addition of emotion variability beyond the adjustment for central tendency.

We evaluated patterns of missingness on depressive symptom assessments according to the recommendations outlined by Hayes and colleagues (Hayes et al., 2024). We examined associations between possible auxiliary variables (all predictors and covariates, plus shyness, fearfulness, and baseline parent depressive symptoms) and missingness in depressive symptom reports at any wave of assessment. We first conducted t-tests of all possible auxiliary variables, which indicated significant associations between socioeconomic status and modal NE and depressive symptom missingness, such that participants with lower socioeconomic status and lower modal NE were more likely to be missing on one of the follow-up depression assessments. When all possible auxiliary variables were included in a logistic regression, no variables were significant predictors of missingness (all ps > 0.05). We also ran logistic regression models testing all linear interaction and quadratic terms to test for non-linear sources of missingness; no predictors were significant (all ps > 0.05). Finally, we ran a model including only socioeconomic status and modal NE; in that model, socioeconomic status was marginally significantly predictive of missingness (p =.058) and modal negative emotion was no longer significant (p =.994). As a result, we assumed data were missing at random and included socioeconomic status as a covariate in all models.

Iterative model testing was performed to identify the best-fitting, most parsimonious model. Model fit statistics are in Table 3. First, a means-only model of depressive symptoms was assessed, including a random intercept for participant. The intraclass correlation coefficient (ICC) indicated just over half of the variance was at the between-person level (ICC = 0.53) and a little less than half was at the within-person level. Second, a fixed effect of time was added, and was not a significant predictor of depressive symptoms, indicating there was not a significant linear increase or decrease on average across the 18-month span (B = 0.02, p =.637). Third, a random linear effect of time was added. A −2 log likelihood test indicated the model significantly improved; however, the intraclass correlation coefficient indicated there was relatively little variance accounted for by the random slope of time (ICC = 0.002).

Covariates were tested individually, both as direct effects and in interaction with time (i.e., testing differences in slope dependent on age). Age was not significantly associated depressive symptoms either independently (p =.512) or in interaction with time (p =.185). Socioeconomic status was not significantly associated with depressive symptoms intercept (p =.116) but significantly interacted with depressive symptom slope (B = 0.03, p =.018), indicating that participants with lower socioeconomic status had more negative slopes.

A representative sample of participant means, modes and change in emotion over the course of the 16-day EMA period is depicted in Fig. 1. First, NE mode was entered into the model as a direct effect and in interaction with time. Modal NE significantly predicted depressive symptoms at a medium effect size, but did not interact with depressive symptom slope (see Table 4). Second, NE variability was added as a direct effect and in interaction with time. NE variability significantly positively predicted depressive symptoms at a small effect size, but did not interact with depressive symptom slope, indicating that higher levels of change in NE over the EMA period were associated with higher baseline levels of depressive symptoms but not with change in depressive symptoms over time.

Next, we entered PE mode into the baseline longitudinal growth curve model as a direct effect and in interaction with depressive symptom slope. PE mode was not significantly associated with depressive symptoms at baseline or in interaction with depressive symptom change over time (Table 5). We then added PE variability and its interaction with depressive symptom slope. PE variability was significantly positively associated with depressive symptoms at baseline at a small to medium effect size, but was not associated with changes in depressive symptom slope (i.e., no significant interaction with time), indicating that higher levels of change in positive emotion during the EMA was associated with higher initial levels of depressive symptoms.

Models for emotional variability were also run using mean affect as an adjustment for central tendency. All models fit adequately (Table 3). Socioeconomic status remained a significant predictor of depressive symptom slope in both models (Tables 4 and 5).

Mean NE significantly predicted depressive symptoms at baseline at a medium effect size, but did not interact with change in depressive symptoms over time (Table 4). When mean NE was included as the adjustment for central tendency, NE variability was not significantly associated with depressive symptoms at baseline or over time. Mean PE was not significantly associated with depressive symptoms at baseline or over time (Table 5). By contrast, PE variability was significantly positively associated with depressive symptoms at baseline at a small to medium effect size, indicating that participants who reported higher levels of change in PE during EMA had higher baseline depressive symptoms.

In addition to adjusting for mean-level emotion, the current study adjusted for an emotional “set point” in conjunction with variability (Ringwald & Wright, 2022), which may aid in understanding variability in emotion in a more ecologically valid and clinically relevant manner. Emotional deviations from one’s set-point may better capture the full “swing” of change in emotion experienced from moment to moment in someone’s life, which may have implications for the perceived severity of change and falicitate the development of more targeted intervention. Further, leveraging modal emotion reinforces the importance of the contextual changes that prompt changes in emotion. Collecting and examining contextual correlates of an individual’s modal emotional state and deviation from that modal state is an important future direction for emotion variability research (Houben et al., 2015; McKone & Silk, 2022), in order to disentangle context-specific and individual difference-mediated emotion change. For example, perhaps female adolescents report higher modal NE but more stability with family and lower modal NE but more variability with peers; such insights would promote more sensitive questions regarding the relation between emotion variability and psychopathology that would more directly inform clinical intervention, such as “What contextual changes are associated with larger changes in emotion and for whom?”

Similarly, collecting data on female adolescents’ attributions in concert with their emotion ratings would facilitate research into cognitive processes that may be especially potent influences and potential prime targets for intervention. Identifying contextual or attributional factors associated with larger deviations from set-points may facilitate the design of strategic interventions; for example, if many female adolescents experience more variability at school, in the evening hours, or when they are with particular people, such contexts may be effective targets for delivering personalized, just-in-time interventions, such as emotion regulation strategies or grounding exercises, via mobile phone.

Socioeconomic status may be one broader context that is important to consider with regard to the development of depressive symptoms. Notably, female adolescents with higher socioeconomic status exhibited more positive slopes of depressive symptoms over time. The sample studied here was of relatively high socioeconomic status, so these findings may be more relevant in the context of middle- to upper-income adolescents, and the findings likely would not generalize to populations with higher rates of participants in poverty, where higher socioeconomic status would likely have a protective effect (Peverill et al., 2021).

There are several limitations to note. First, these findings are likely inflated by shared informant bias, as adolescents self-reported both depressive symptoms and emotion. Repeating these analyses with clinician-rated symptoms, ideally based on both adolescent and parent report, is an important next step. Further, findings should be considered in the context of possible heterogeneity in depression and/or nonlinear associations (Santee & Starr, 2021); it is plausible that very low levels of emotion variability may be indicative of one phenotype of depression (Bylsma,  2021; Koval et al., 2012; Kuppens et al., 2012), and that maintaining moderate levels of emotion variability is most adaptive, indicating appropriate response to contextual changes. Research testing these questions is required. This study was also likely underpowered for the prediction of change in depressive symptoms over time, as symptom levels were low at baseline, and remained relatively stable throughout the course of the study. In other words, there was limited variability in growth in depressive symptoms over time, which was unexpected given the study design (i.e., risk for development of depression). Finally, although the sample collected here is broadly representative of female adolescents and families in the midsize Midwestern city in which it was collected, it is approximately two-thirds white, so findings are not generalizable to populations with higher proportions of racial and ethnic minority groups or to adolescent males.