The dynamic influence of exercise on emotional state among people with a history of mood disorder

Abstract

Background:

Planned or motivated physical activity (exercise) is amenable to direct modification and has been inversely linked to major depression (MDD), suggesting its potential to support recovery in bipolar disorder (BD). However, it is unclear how quickly exercise may impact emotion. We characterized the association of exercise with subsequent emotional state within hours in persons with a history of mood disorders.

Method:

Euthymic participants aged 11–85 years, diagnosed with BD (n=110), MDD (n=157), anxiety without mood disorder (n=96), or no disorder (n=96), self-reported their emotional state (sad, anxious, activated, energetic) scaled 1–7 and whether they were exercising four times per day for two weeks via ecological momentary assessment. We characterized the implicit impact of exercise on subsequent emotion within-person and at the group level using dynamic structural equation modeling.

Results:

Among those with a major mood disorder (BD or MDD), exercise was associated with a strong reduction in subsequent anxiety (−1.706 (95% Credible Interval −2.296, −1.048)), overall and in males and females. However, exercise was not associated with subsequent changes in sad, activated, or energetic feelings. Exercise-emotion dynamics did not differ significantly between individuals with BD compared to MDD, Anxiety without Mood, or no disorder. However, there was variation at the individual within-person level in BD subtypes.

Conclusion:

Exercise appears to drive beneficial impact on emotional state within hours in persons with BD or MDD—particularly by lowering anxiety. Independent replication and clinical trials are needed to draw conclusions about specificity by mood disorder subtype.

1. Introduction

Bipolar disorder (BD) is characterized by dysregulated, fluctuating affect across the domains of emotion, cognition, and activity, with observational work suggesting dysregulated activity is central (Alloy & Abramson, 2010; Cheniaux et al., 2014; Dennison et al., 2021; Johnson et al., 2012; Merikangas et al., 2019; Scott et al., 2017; Shou et al., 2017; Stapp et al., 2023). Gold standard treatments for BD, such as lithium, focus on mood stabilization through the prevention or treatment of acute mood episodes, which is important given the impairment associated with such episodes (Grof, 2010; Kessler et al., 2007). However, such treatments do not work for everyone, take months or longer to take effect in those who do respond, and do not fully address non-acute consequences such as interepisode lability (Cochran et al., 2016; Henry et al., 2008; Prisciandaro et al., 2019; Sperry et al., 2020; Taylor et al., 2021; Wichers et al., 2010). Attention to interepisode stabilization (Crowe et al., 2020) may lessen the impairment that exists during the interepisode period, extend euthymia and prevent severe episodes, and may additionally benefit long-term outcomes such as improved vascular health. There is a critical need for treatments that address interepisode emotional regulation in BD to promote recovery.

Planned or motivated physical activity (exercise) is amenable to direct modification, is an established health indicator to promote brain health (Firth et al., 2020; Goodrich & Kilbourne, 2010; Piercy et al., 2018; Reynolds et al., 2022; USDHHS, 2023; Vancampfort, Rosenbaum, et al., 2016), and has a demonstrated inverse relationship with depressive symptoms and major depressive disorder (MDD) (Bailey et al., 2018; Cooney et al., 2013; da Cunha et al., 2023; Danielsson et al., 2013; Pearce et al., 2022; Schuch et al., 2016; Wang et al., 2022) with numerous potential mechanisms for benefits in BD (Fellendorf et al., 2017; Hearing et al., 2016; Kucyi et al., 2010; Ng et al., 2007; Sa Filho et al., 2020; Sylvia et al., 2010; Thomson et al., 2015). In clinical samples of persons with BD, self-reported physical activity has been inversely associated with mood episode recurrence, hospitalization, time depressed, and depressive level, while not exercising regularly has been associated with lower quality of life and functioning (Melo et al., 2019; Sylvia, Friedman, et al., 2013). This suggests that changes to physical activity, appropriate to a person’s life stage (Vancampfort et al., 2013; Vancampfort, Firth, et al., 2016; Vancampfort et al., 2017), may be an effective adjunctive treatment in BD.

However, gaps remain in the development of exercise interventions. While development, testing, and uptake of physical activity interventions in BD and MDD appear feasible (Adams et al., 2015; Daumit et al., 2020; Daumit et al., 2013; Khoubaeva, 2022; Leone et al., 2018; Sylvia, Janos, et al., 2019; Sylvia et al., 2011; Sylvia, Pegg, et al., 2019; Sylvia, Salcedo, et al., 2013), and benefits may happen over time, we do not know when or how quickly they occur. Additionally, while reduction of depressive symptoms is one potential benefit of physical activity for persons with BD, it is not the only important feature of emotional health, and recommendations derived from experience with MDD may not be directly generalizable given the hallmark impairing elevation of BD. Given the effects of aging on the body (DiPietro, 2001; Varma et al., 2017) as well as the greater prevalence of mood disorders in females (Kessler et al., 2007) and severe outcomes such as vascular diseases in women with BD (Fiedorowicz et al., 2011; Laursen et al., 2013; Ortiz et al., 2022) it is also important to examine potential age and sex differences in the association between exercise and emotional state in mood disorders. Thus, deeper understanding of the dynamic relationship of exercise on several emotional states within hours during euthymia, and how this may vary by age, sex, and mood disorder subtype, is vital to inform intervention targets.

We previously used dynamic structural equation modeling (DSEM), to differentiate emotional dynamics across mood disorder subtypes in the National Institute of Mental Health Family Study of Affective Spectrum Disorders (NIMHFS) (Stapp et al., 2023). That work demonstrated the feasibility and utility of DSEM for modeling implicit, intraindividual processes and real-time dynamics produced by ecological momentary assessment (EMA) in a community-based sample enriched for mood disorders; here, we build on that work to incorporate exercise bouts reported in real-time and their relationship with subsequent emotional state. Specifically, we aimed to 1) characterize exercise-mood dynamics in those with major mood disorders to test whether exercise drives subsequent emotional state within hours, examining potential differences by age and sex, 2) test whether there are group differences in exercise-emotion dynamics in BD compared to MDD and those without mood disorders, and 3) explore exercise-emotion dynamics at the individual-level in BD, disaggregated by bipolar subtype.

2. Methods

2.1. Data source and sample eligibility

The NIMHFS (Merikangas et al., 2014; Merikangas et al., 2019) began enrollment in 2004 and concluded enrollment in 2020. Probands were recruited from a community screening of the greater Washington, DC area, local health newsletters and announcements, the National Institutes of Health (NIH) Clinical Center general volunteer core, and the NIMH Mood and Anxiety Program. The original study was approved by the NIH Combined Neuroscience Institutional Review Board (IRB). Adult participants provided written informed consent and minors provided assent in addition to guardian consent. The sample subset for this study includes participants of the NIMHFS with directly ascertained diagnostic interviews and EMA data, including individuals with lifetime major mood disorder, including BD (including both Bipolar-I and Bipolar-II), or MDD, and those without, including individuals with anxiety without mood disorder or those with no psychiatric disorders. The current analysis uses de-identified data and was determined to be exempt from IRB review by George Washington University.

2.2. Procedures

2.2.1. Psychiatric diagnosis

The NIMHFS Diagnostic Interview for Affective Spectrum Disorders ascertains diagnostic criteria and subthreshold phenomenology for current and lifetime Diagnostic and Statistical Manual of Mental Disorders (DSM) disorders (American Psychiatric Association, 2000, 2013). The interview was developed based on earlier diagnostic interviews for genetic epidemiologic studies (e.g.,(Nurnberger et al., 1994) and does not use skip-outs. Experienced clinicians interviewed probands directly, in addition to conducting blinded, direct interviews with relatives. Inter-rater reliability of all diagnostic categories was high, with intraclass correlations of 0.87 and above for all diagnostic categories. The best-estimate DSM-IV diagnoses used herein were based on interviews, family history, and clinician consensus ratings.

2.2.2. Ecological momentary assessment

For 14 consecutive days, participants used a pre-programmed mobile device that administered 4 brief electronic assessments per day, which occurred approximately every 4 waking hours. Several options of schedules were available to participants to accommodate typical sleep/wake cycles, with up to a few minutes of random variation in delivery around the set times. Research staff confirmed participants were euthymic, and did not proceed with EMA if participants with mood disorders were in-episode.

2.3. Measures

2.3.1. Exposure: Exercise (physical activity)

All participants were queried at each timepoint about current participation in exercise, physical leisure, and/or sports. These assessments of motivated/planned physical activity were combined within each individual assessment period as a binary exposure of exercising (yes/no) at that timepoint.

2.3.2. Outcome: Momentary emotional state (mood)

At each timepoint, participants rated from 1 to 7 on a Likert-type scale to indicate their present momentary level of emotions, with the higher number corresponding to the italicized variables for which the emotional states herein are named: very happy/cheerful (1) through very sad/depressed/unhappy (7), very relaxed/calm (1) though very anxious/nervous (7), very inactive/quiet (1) through very active/aroused (7; activated), and very tired/sluggish (1) through very energetic/lively/excited (7). These four dimensions correspond with Larsen and Diener’s mood circumplex, which we have previously used to uncover multivariate affective dynamics of mood disorders (Larsen, 1992; Stapp et al., 2023).

Continuous variables are latent person-mean centered automatically during the DSEM process, such that the value at each assessment represents deviation from the individual person’s mean, irrespective of the group mean, with latent variables for the between-level components to account for measurement error, eliminate Nickell’s bias and Ludtke’s bias, and accommodate missing data (Asparouhov et al., 2017).

2.3.3. Key covariates

2.3.3.1. Age.

A unique feature of the NIMHFS is that these deeply-phenotyped participants with EMA data range from adolescence through the cusp of “oldest old.” Here, we included age at EMA collection. When exploring effect modification of exercise-mood dynamics by age among those with mood disorders, we stratified at age 45 years, which was both the sample median and a useful cutpoint marking midlife and older relative to younger participants (objective 1). When conducting conditional DSEM (focused on between-group differences in dynamics by diagnostic groups), we included age measured continuously as a covariate (objective 2).

2.3.3.2. Sex.

The NIMHFS collected data on sex and did not query gender at enrollment. Here, we explored effect modification of exercise-mood dynamics among those with mood disorders by stratifying by sex (objective 1), and when conducting conditional DSEM we included sex as a covariate (objective 2).

2.4. Statistical Analysis

Sample statistics were obtained in Stata 18. All latent modeling (DSEM) was conducted in Mplus 8.11. DSEM combines time series, multilevel, and structural equation modeling to analyze intensive longitudinal data. It is estimated with Bayesian methods, based on the Markov chain Monte Carlo algorithm via Gibbs sampler, accommodates up to 80% missing data (Asparouhov et al., 2017) and in the present study makes use of diffuse priors, 2 chains, and at least 2000 iterations in initial runs, 5,000 iterations in replication runs for every model, and confirmation runs of at least 10,000 iterations in any model that shifted between initial and replication runs.

2.4.1. Unconditional DSEM (Objective 1)

The unconditional model of exercise-emotion dynamics is illustrated in Figure 1, adapted from Hamaker and colleagues (Hamaker et al., 2018). It is a lag-1 multilevel vector autoregressive (VAR(1)) model with average levels (μ) of emotional state (sad, anxious, activated, energetic) and endorsement of exercise, autocorrelation (φ) regression coefficients for inertia of emotional state and exercising, cross-lagged (φ) regression coefficient for the effect of exercise at the prior timepoint on emotional state at the subsequent timepoint, and log of the variance of the innovations (log(π)) capturing residual variance and within-person intensity of change from timepoint to timepoint for emotion and exercising, all of which are random (varying across individuals).

Figure 1. Dynamic structural equation model of interrelationship of exercise and emotion within hours.

Note: Panel A decomposes intensive longitudinal data into dynamic within-person and stable between-person components. Individual means are the within-person mean level of a given indicator observed across all timepoints, modeled at the between-person level. In Panel B, autocorrelation is the within-person relationship from one timepoint to the next, via the autoregression parameters (inertia), which uses past values to predict the current value. Cross-lags at the within-person level are cross-regressions measuring how a given indicator is influenced by a different indicator at a prior timepoint. Innovation variances include residual variance and reflect within-person intensity of change from timepoint to timepoint. Panel C shows the between-person level model, which includes the between-person observed individual means and the random effects (black dots) in the within-person level model. Style adapted from Hamaker et al. 2018.

The means shown in the results represent the fixed or averaged effects in a given subgroup/stratum (i.e., group-level average across all individual averages) for each of the random effects. The 95% credible intervals (CrI) can be interpreted as there being a 95% probability that the true (unknown) estimate would lie within the interval, given the evidence provided by the observed data. There still may be important individual differences even if group means do not ‘significantly’ differ from zero as represented by a 95% CrI that excludes zero. For the two states (emotion and exercising), the numbers themselves are assumed to be a reflection of the unobserved, underlying continuous normal process that gives rise to observed responses (McNeish et al., 2024). Because all forms of correlation are measured simultaneously in the VAR(1) model, the effect estimates produced account for all inertias, innovations, and group means of individual mean levels of a given emotional state and exercise, their cross-lag, and their covariances.

2.4.1.1. Post hoc sensitivity analyses of age groups within major mood disorders.

Beyond stratifying those with mood disorders at the age midpoint (under 45 versus ages 45 and older), we examined subsets of age, truncating the youngest and oldest 5% and 10% of each primary age stratum. For the younger age group, where the primary age stratum was 11 to 44 years, the sensitivity analyses examined ages 16 to 44 and 18 to 44, representing the truncation of the youngest 5% and 10%, respectively. For the older age group, where the primary age stratum was 45 to 80 years, the sensitivity analyses examined ages 45 to 67 and 45 to 64, representing the truncation of the oldest 5% and 10%, respectively.

2.4.2. Conditional DSEM (Objective 2)

For ‘predictor’ modeling, at the between-level we tested the association of individual mood disorder classes (BD compared to MDD, BD compared to Anxiety without Mood, BD compared to no disorder), adjusted for age and sex, with the random effects described above in the unconditional models (i.e., means, inertias, cross-lag, innovation variances) of exercise-emotion dynamics.

2.4.3. Exploration and visualization of exercise-emotion dynamics at the individual-level in bipolar disorder, disaggregated by bipolar subtype (Objective 3)

We assumed that testing differences in exercise-emotion dynamics between bipolar subtypes (predictor regression modeling) would not yield meaningful results because of the relatively small number of each subtype (n=56 for Bipolar-I, n=54 for Bipolar-II), yet we wanted to attempt to understand the dynamics in these subtypes. During DSEM, a dataset of plausible values can be generated for each participant. Thus, to understand the within-person effect of exercise on emotion in bipolar subtypes, we generated a distribution of plausible values of this effect for each participant, consistent with Bayesian philosophy (McNeish et al., 2024). Specifically, we generated a distribution of 200 plausible values for each participant with Bipolar-I and Bipolar-II, balancing precision and computational time. Mplus further generates the mean and lower and upper limits of the 95% CrI, which we extracted and used to generate caterpillar plots in R, where it is visible whether the estimated exercise-emotion effect was non-null for each individual participant and emotional state.

3. Results

3.1. Sample characteristics

Four hundred fifty-nine participants with Bipolar-I (n=56), Bipolar-II (n=54), MDD (n=157), Anxiety without Mood Disorder (n=96), or no psychiatric disorder (n=96) had EMA data (Table 1) and contributed a total of 24,516 observations. The average age was 41.4 years (standard deviation, 19.3; range, 11–85). Sixty-three percent of participants were female. We have previously reported that the majority of the NIMHFS report White race, but racial and ethnic composition does not differ across diagnostic groups and is not included in models (Stapp et al., 2023). Anxiety disorders are common (78% prevalence) among participants with mood disorder. Those who reported exercising at least once (N=208; 437 exercise observations) were not significantly different from those who did not in terms of age or sex. There was good compliance with EMA reports, with participants completing over 76% of 25,704 possible total observations.

Table 1.

Sample Characteristics of Participant Subset from the NIMH Family Study of Affective Spectrum Disorders with Exercise-Emotion Dynamics Measured by Ecological Momentary Assessment

	Total	Major Mood Disorder	No Threshold-Level Mood Disorder
	N=459	N=267	N=192
Age in years, mean (SD), Range [min, max]	41.4 (19.3) [11, 85]	42.9 (16.9) [11, 80]	39.4 (22.2) [11, 85]
Sex
Male	170 (37%)	86 (32%)	84 (44%)
Female	289 (63%)	181 (68%)	108 (56%)
Diagnosis Group, N (%)
Bipolar Disorder, Type I	56 (12.2%)	56 (21%)	-
Bipolar Disorder, Type II	54 (11.8%)	54 (20.2%)	-
Major Depressive Disorder	157 (34.2%)	157 (58.8%)	-
Anxiety without Mood	96 (20.9%)	-	96 (50%)
No Psychiatric Disorder	96 (20.9%)	-	96 (50%)
Comorbidity, N (%)
Anxiety Disorder	306 (66.7%)	210 (78.7%)	96 (50%)
Substance Use Disorder	84 (18.3%)	76 (28.6%)	8 (4.2%)
Eating Disorder	33 (7.2%)	30 (11.2%)	3 (1.6%)

Note: SD, Standard Deviation. Those with Mood Disorders meet full diagnostic criteria for Bipolar I, Bipolar II, or Major Depressive Disorder, which are mutually exclusive. Non-mood comorbid disorders are not mutually exclusive and do not round to 100%; anxiety disorders include agoraphobia, obsessive-compulsive disorder, generalized anxiety disorder, panic disorder, post-traumatic stress disorder, separation anxiety, social anxiety, and specific phobia.

3.2. Exercise-emotion dynamics in major mood disorders, broadly and by age and sex strata

In participants with a major mood disorder (BD or MDD), self-reported exercise was significantly associated with subsequently lower anxiety (−1.706 (95% CrI −2.296, −1.048)), but was not associated with subsequent activation, energy, or sadness (Table 2, top row). In the initial runs, exercise drove subsequently lower activation and energy, as well, but these did not hold up to replication.

Table 2.

Group-level Estimates of the Cross-Lagged Effect of Exercise to Subsequent Emotional State Within Hours in a Community Sample Aged 11 to 85 years

	Exercise → Activated, φVX,i	Exercise → Anxious, φAX,i	Exercise → Energetic, φEX,i	Exercise → Sad, φSX,i
	Est. Mean (95% CrI)	Est. Mean (95% CrI)	Est. Mean (95% CrI)	Est. Mean (95% CrI)
Major Mood Disorders (MMD: BD|MDD)
MMD overall (n=267, obsv=14194)	−0.551 (−1.204, 0.271)	−1.706 (−2.296, −1.048)	−0.435 (−1.066, 0.173)	0.258 (−0.295, 0.63)
MMD male (n=86, obsv=4543)	−0.51 (−2.077, 0.929)	−1.623 (−2.728, −0.709)	−0.589 (−1.45, 0.478)	0.198 (−0.324, 0.676)
MMD female (n=181, obsv=9651)	−0.553 (−1.307, 0.113)	−0.829 (−1.822, −0.076)	−0.265 (−1.57, 0.653)	0.426 (−0.18, 1.272)
MMD <45 years (n=132, obsv=6891)	−0.323 (−1.544, 0.856)	2.159 (0.961, 3.585)	−0.948 (−2.382, 0.13)	1.596 (0.573, 2.805)
MMD 45+ years (n=135, obsv=7303)	−0.47 (−1.184, 0.075)	−0.315 (−0.903, 0.365)	−0.226 (−1.068, 0.828)	0.089 (−0.484, 0.584)
Mood Disorder Classes and Comparison Groups
BD (n=110, obsv=5749)	−0.801 (−1.93, 0.422)	−0.5 (−1.467, 0.611)	0.1 (−0.707, 1.091)	0.279 (−0.477, 0.96)
MDD (n=157, obsv=8445) a	−0.294 (−1.123, 0.611)	0.567 (−0.718, 1.453)	−0.585 (−1.243, 0.388)	0.36 (−0.729, 1.261)
Anxiety w/o Mood (n=96, obsv=5184)b	0.939 (−0.382, 2.179)	−0.465 (−1.234, 0.472)	0.74 (−0.445, 1.896)	−0.216 (−0.803, 0.317)
No psychiatric disorder (n=96, obsv=5138)c	−1.201 (−2.664, 0.36)	0.221 (−0.991, 1.59)	0.518 (−1.195, 1.777)	0.214 (−2.553, 2.627)
Bipolar Disorder Subtypes
Bipolar-I (n=56, obsv=2980)	−1.018 (−3.158, 1.188)	−0.504 (−1.695, 0.843)	−0.864 (−2.618, 0.912)	0.137 (−1.165, 1.439)
Bipolar-II (n=54, obsv=2769)	−1.21 (−2.817, 0.601)	−0.328 (−4.482, 3.13)	0.151 (−2.251, 2.998)	0.479 (−1.506, 2.565)

Note: BD, Bipolar Disorder; CrI, Credible Interval; MDD, Major Depressive Disorder; MMD, Major mood disorder (BD or MDD).

Bold cells do not contain zero, analogous to ‘statistical significance’ at the two-sided p<0.05 threshold. Bipolar Disorder is a combined variable of individuals with either Bipolar-I or Bipolar-II. Point estimates (posterior means) and 95% credible intervals for mean dynamics at the between-person level, stratified by diagnostic group. Each means with its corresponding credible interval is from a unique stratified model, where estimates are derived from a single multivariate model that simultaneously adjusts for all dynamics including individual mean levels, inertias, cross-regressions, and innovation variances.

^{a, b, c}Conditional models in which the effect of BD compared to another disorder (MDD, Anxiety without Mood Disorder, No psychiatric disorder), adjusted for age and sex, was regressed on exercise-emotion dynamics were not significantly different (i.e., the cross-lag of exercise to subsequent emotion was not significantly associated with--different by--diagnostic group).

Stratified analysis (Table 2, rows two through five) demonstrated the exercise to lower anxiety effect was present for both males (−1.623 (95% CrI −2.728, −0.709) and females (−0.829 (95% CrI −1.822, −0.076)). When stratifying by age, among those younger than middle age exercise was unexpectedly associated with subsequently higher anxiety (2.159 (95% CrI 0.961, 3.585)) and sadness (1.596 (95% CrI 0.573, 2.805)), in contrast to adults middle-aged and older who experienced a non-significant drop in anxiety after exercising.

All DSEM effect estimates (mean levels, inertia, innovation variances, and cross-lags) for all diagnostic groups and models are reported fully in the Supplement, Table S1. From initial to replication analyses, there was some fluctuation in raw values (rare, minor shifts from positive to negative or vice versa, or by ~0.1 unit of change higher or lower), especially across the non-significant estimates, and therefore contrary to accepted norms of presenting ‘marginal’ associations we have expressly chosen not to interpret non-significant estimates. We used the replicated results in presenting findings, which were mostly similar to or more conservative than the initial runs; in all cases demanding a confirmatory run with at least 10,000 iterations, any shifts observed when moving from the initial to the replication was confirmed and thus substantiated by the 10,000-iterations run.

3.2.1. Post hoc sensitivity analyses

The post hoc analyses (Table 3) were motivated by the observation of unexpected anxiety and sadness responses to exercise among the younger half of the sample with mood disorders, which were qualitatively different than overall in major mood disorders and among other strata (older, male, female; described above). Among younger participants with mood disorders, truncating the youngest 5% (ages 11 to 15, i.e., excluding tweens and young teens) yielded effects of exercise lowering anxiety (−1.839 (95% CrI −2.966, −0.378)) in line with all other groups, which was maintained, albeit non-significantly, in the analysis truncating the youngest 10%. Only when capping the youngest 10% did the association of exercise on subsequent sadness become null, as observed in the rest of the sample. Among those middle-aged and older, results from all sensitivity analyses truncating the oldest 5% and 10% were fully consistent with analyses in all participants aged 45 years and above. All sensitivity analyses made use of the same sequence of replicating iterations in main analyses, and results presented in Table 3 are from replication runs of at least 5,000 iterations.

Table 3.

Sensitivity analyses of group-level estimates of the cross-lagged effect of exercise on subsequent emotional state within hours in persons with mood disorders, comparing main age strata to truncation of 5% and 10% youngest and oldest ages

	Exercise → Activated, φVX,i	Exercise → Anxious, φAX,i	Exercise → Energetic, φEX,i	Exercise → Sad, φSX,i
	Est. Mean (95% CrI)	Est. Mean (95% CrI)	Est. Mean (95% CrI)	Est. Mean (95% CrI)
Main analysis: MMD <45 years (n=132, obsv=6891)	−0.323 (−1.544, 0.856)	2.159 (0.961, 3.585)	−0.948 (−2.382, 0.13)	1.596 (0.573, 2.805)
5% truncation: MMD age >15 to <45, n=117, obsv=6133	0.305 (−1.444, 1.536)	−1.839 (−2.966, −0.378)	−0.859 (−2.127, 0.703)	1.506 (0.322, 2.931)
10% truncation: MMD age >17 to <45, n=103, obsv=5441	−0.325 (−2.337, 1.101)	−1.91 (−3.604, 0.125)	−0.715 (−2.594, 0.597)	0.892 (−0.866, 2.084)
Main analysis: MMD 45+ years (n=135, obsv=7303)	−0.47 (−1.184, 0.075)	−0.315 (−0.903, 0.365)	−0.226 (−1.068, 0.828)	0.089 (−0.484, 0.584)
5% truncation: MMD age 45 to <68, n=120, obsv=6463	−0.533 (−1.368, 0.35)	−0.507 (−1.451, 0.307)	−0.415 (−1.343, 0.507)	0.002 (−0.615, 0.635)
10% truncation: MMD age 45 to <65, n=108, obsv=5823	−0.609 (−1.496, 0.135)	−0.34 (−1.228, 0.38)	−0.276 (−1.29, 0.55)	0.056 (−0.662, 0.684)

Note: CrI, Credible Interval; MMD, Major mood disorder (Bipolar Disorder [either Bipolar-I or Bipolar-II] or Major Depressive Disorder).

Bold cells do not contain zero, analogous to ‘statistical significance’ at the two-sided p<0.05 threshold. Point estimates (posterior means) and 95% credible intervals for mean dynamics at the between-person level, stratified by age group. Each means with its corresponding credible interval is from a unique stratified model, where estimates are derived from a single multivariate model that simultaneously adjusts for all dynamics including individual mean levels, inertias, cross-regressions, and innovation variances.

3.3. Exercise-emotion dynamics in mood disorder classes and comparison groups, comparing bipolar disorder to major depressive disorder, anxiety without mood, and no disorder

Within each mood disorder class and comparison group (Table 2, middle rows), participants with BD (Bipolar-I or Bipolar-II), MDD, anxiety without mood disorder, and no psychiatric disorder, exercise was not associated with a significant change in self-reported emotional state at the subsequent timepoint (see Supplement, Table S1, for all DSEM estimates across all models).

Conditional models regressed BD relative to another disorder group (MDD, Anxiety without Mood Disorder, or no psychiatric disorder) on exercise-emotion dynamics, adjusted for age and sex. In all conditional models, the cross-lag of exercise to subsequent emotion was not significantly associated with—i.e., different by—diagnostic group (Supplement, Table S2), reflecting what was observed in unconditional models.

3.4. Effect of exercise on emotion in bipolar disorder subtypes, group-level and within-person

Perhaps not unexpectedly, when drilling down to further levels of specificity, and thus smaller groups, exercise was not associated with subsequent changes in emotion for persons with Bipolar-I or Bipolar-II at the group level (Table 2, bottom two rows; see Supplement, Table S1 for all DSEM estimates).

Although differences in cross-lags between exercise and subsequent emotions were not significantly different between diagnostic groups (i.e., Table 2), variation at the individual within-person level was apparent (Figure 2; Figure S1a–h for large versions). For over half of persons with Bipolar-I exercise blunted subsequent feelings of being activated, anxious, and energetic; interestingly, exercise had the weakest link with sadness, and a stronger link with reducing anxiety (Figure 2a). In contrast, most persons with Bipolar-II had a null relationship between exercise and all emotional states at the following timepoint (Figure 2b).

Figure 2. Within-person Effect of Exercise on Emotion in Bipolar Disorder, types I and II.

Figure note: Within-person Effect of Exercise on Emotion in Bipolar Disorder, types I and II. Numeric values in caterpillar plots are raw, not standardized, and should not be interpreted unto themselves. Each row represents one participant, with distributions of individual, within-person effects estimated with 200 plausible values. Solid markers indicate that the 0 was outside the person-specific 95% credible interval (open markers indicate that 0 was inside the person-specific credible interval, analogous to ‘non-significance’). Each Figure panel shown in full-size in the Supplement, Figures S1a–S1h.

4. Discussion

Among those with a major mood disorder (BD or MDD), self-reported exercise was associated with a strong reduction in subsequent anxiety within hours, a relationship observed in both males and females, and those aged 16 years and older. Self-reported exercise-emotion dynamics did not differ significantly between individuals with BD compared to MDD, Anxiety without Mood, or no disorder. However, within BD subtypes there was variation at the individual within-person level in which most persons with Bipolar-I were estimated as experiencing lower feelings of activation, anxiety, and energy after exercising but emotional states were unchanged for most persons with Bipolar-II.

The most prominent, consistent mood change following exercise was reduction in anxiety—broadly across mood disorders, in both males and females. Our earlier work on this sample showed greater variability of anxiety than of mood in people with Bipolar-I and greater reactivity of anxiety across all mood disorders subgroups (Lamers et al., 2018), as well as higher levels of fragmentation of both mood and anxiety among those with Bipolar-II but not Bipolar-I (Johns et al., 2019). In light of pervasive comorbidity between anxiety disorders and both BD (Merikangas et al., 2011) and MDD (Kessler et al., 2003) and greater fluctuation in anxiety symptoms in daily life, consideration of anxiety as a therapeutic target makes sense (Goldstein, 2023), and it would appear that exercise may be a reasonable method of engaging that target. A large body of work demonstrates a large magnitude of effect for exercise on reducing anxiety in people with depression and/or anxiety disorders (Banyard et al., 2025). Relatedly, meta-analytic work demonstrates exercise has a large magnitude of effect on depression over time among those with MDD and/or anxiety disorders, with an estimated number needed to treat between 2 and 3 (Banyard et al., 2025; Heissel et al., 2023). There is less work in BD specifically, however a small body of work suggests an observed within-day reduction in depressive symptoms following exercise (Walsh et al., 2023) and a similar trend in the context of multi-week activity interventions (Lafer et al., 2023; Weinstock et al., 2016). Although we observed that exercise was not associated with subsequent changes in sadness among most participants, this was over a short timeframe, and a single item, in contrast to the broader construct of depression that invokes not only sadness but also somatic symptoms and anhedonia. It is possible that exercise’s route to decreasing depression is not via sadness, but rather through other components of depression. Such specificity of emotional-somatic mechanisms across mood disorder subtypes would be worth testing directly in future trials, particularly given our prior observations of unique dynamics of active and energetic feelings in Bipolar-I and sadness and anxiety in MDD (Merikangas et al., 2019; Stapp et al., 2023) and our observation in the current study of a potentially tighter link between exercise and emotional response in more persons with Bipolar-I than with Bipolar II. Finally, our findings suggest that benefits of exercise on anxiety manifest quickly enough to be observed within hours, although this observational study was not designed to capture the exact moment of change. Others have shown that objectively measured physical activity may have a stronger effect on mood over one hour than over a longer period of three hours (Hollands et al., 2020). Future research should examine immediate-term impacts, such as pre-post exercise bout, in naturalistic environments in which most persons’ therapeutic exercise would take place. This could potentially be accomplished using EMA triggered by wearables that detect, for example, increased heart rate for a sustained period or other sensor-informed context-sensitive EMA (Dunton, 2017).

The qualitatively different emotional reactions between younger and older strata of persons with mood disorders was unexpected, though merits consideration. Specifically, when stratifying by age, contrary to the blunting of anxiety observed in mood disorders overall and in other subgroups, in the younger half of the sample exercise was followed by increases in anxiety and sadness. Our post hoc sensitivity analyses showed that when truncating the youngest 5% of the sample, representing ages 11 through 15, that anxiety no longer increased following exercise and instead decreased among the remainder of the younger stratum (i.e., ages 16 to 44 years). Truncating the youngest 10% also nullified the link with sadness, as observed overall and in all other subgroups. That excluding the youngest participants brought the analyses in line with all other groups suggests that there is greater variation in and potentially qualitatively different exercise-emotion dynamics among those tweens and young teens; however, there were too few participants in our sample to test that hypothesis directly, and therefore this would be an important question for investigation in a larger sample of adolescents with mood disorders. One consideration may be the role of structured activities, including both organized sports as well as the school day itself. While physical activity is known widely to be protective of mental health (He et al., 2018), organized sports may generate stress (e.g., through games, challenging practices, interpersonal dynamics, environmental exposure (Council on Sports et al., 2011)) that could yield a qualitatively different impact on emotion relative to, for example, a middle-aged adult participating in self-directed exercise at the gym. Relatedly, the intensity or duration of exercise may vary by age group, which could influence the emotional response (Reed & Buck, 2009). It is also possible that the impact of being interrupted in organized sports or the school day (e.g., physical education class) produces stress or emotional impacts that are not directly related to the physical activity but appear as such in the context of analysis. The NIMHFS provided numerous options of EMA ‘schedules’ to participants, however investigators including youth or others with similarly low control over their schedule should consider during study design and interpretation how structural or contextual factors (e.g., social pressure, school-mandated vs. voluntary activity) may influence measurement of their constructs of interest. An additional consideration is that tween and young teens receiving mood disorder diagnoses could suggest a severity of illness or underlying vulnerabilities that could, in turn, influence exercise-emotion dynamics, though all participants completing EMA were euthymic.

We did not find group-level statistical differences in exercise-emotion dynamics between diagnostic subtypes, including comparing those with BD to MDD, BD to anxiety without mood disorder, BD relative to individuals with no mental disorder, and Bipolar-I to Bipolar-II. This may imply that for those inclined to self-initiate exercise, the benefits of and guidelines in physical activity may be generalizable across mood disorders. Conversely, this could be due to sample size. Although the overall sample size was adequate to conduct DSEM, the greater number of significant findings in models with higher N may suggest that null findings in individual disorder classes and bipolar subtype may be due to reduced sample size rather than representing true negative findings, which points to the need for replication and caution before generalizing findings from intervention studies of depression and MDD to BD. Moreover, as data sources, access, and harmonization grow and improve in the coming years, so too will the ability to develop statistical innovations around personalized digital health profiles and responsive feedback.

4.1. Limitations and Implications

First, given the focus on ecological validity in EMA, the questionnaires did not query activities/feelings between assessment timepoints, and thus there may have been additional exercise bouts—and subsequent impact on emotion—that are not accounted for in our models. Adding questions about inter-assessment health behaviors such as exercise would be useful for future data collection, with the potential trade-off of some ecological validity. The NIMHFS EMA procedure assessments occurred at approximately the same time each day for a given person (with differences across participants depending on their chosen schedule), whereas some investigators prefer less predictable assessments. A participant might schedule their exercise so that it occurs before or after the expected time range of a particular EMA timepoint and/or choose to report their ‘current’ activity as one that had just completed or was about to begin. It is important for research studies to balance frequency and (un)predictability of data collection with potential participant burden. Second, our approach for the current study involved a time grid based on wakeful hours to preserve interpretability, however future studies contrasting within-day versus between-day influence of exercise on mood (Walsh et al., 2023) would be a promising future direction, particularly to understand how long the benefits may be observed (in terms of hours of days), under what conditions, and the role of sleep. Third, our data do not include dynamics of acute episodes; however, this was by design, as our goal was to characterize interepisode dynamics. Therefore, these findings tend to apply to those who are outside of acute episodes and may not generalize to within-episode dynamics. Exercise is already promoted as an intervention to reduce or potentially prevent depression in the general population, and testing the impact of exercise within an acute episode is the domain of clinical trials and health systems quality improvement with active monitoring to support patients. That said, persons with mood disorders, on average, spend most of their lives in euthymia and therefore findings are still important to understanding potential mechanisms and promoters of recovery. Fourth, some research suggests self-reported physical activity may over-estimate activity compared to actigraphy (Vancampfort, Firth, et al., 2016). However, this is less of a concern in EMA, and there is practical utility in self-reported physical activity given that this is what providers rely on in community care and most longitudinal clinical research in BD (Vancampfort, Rosenbaum, et al., 2016). Fifth, we characterized exercise categorically, without information on dose (e.g., intensity, duration); however, intensity levels (e.g., moderate, vigorous) and duration may influence the magnitude of the emotional response and should be investigated further. Future work will incorporate dose, including by employing objectively assessed physical activity that can be analyzed both dimensionally as well as controlling for individual's average level over the time of observation (Husky et al., 2025).

4.2. Conclusion

We examined the association of exercise on subsequent emotional state across the full mood circumplex in persons with mood disorders, finding that the most salient emotional change following exercise was reduction in anxiety, broadly and in males and females. Although research generally points to an inverse relationship between physical activity and depression, it could be that at a more granular level those changes occur through reductions in feelings of anxiousness rather than sadness, with fluctuations in anxiety as the more common real-time manifestation of all subtypes of mood disorders. Therefore, testing such specific emotional pathways and their somatic impacts could be a promising future direction for clinical investigation—in particular, anxiety reduction may be a promising primary outcome for lifestyle-based adjunctive treatments in mood disorders. More work is needed in young adolescents, both to directly study exercise-emotion dynamics in younger persons experiencing or at risk for mood disorders and to understand and address potential structural issues around their ability to participate in research and its potential impact on measurement. While the lack of difference across mood subtypes in exercise-emotion dynamics could suggest generalizability across diagnoses for those inclined to self-initiate exercise, independent replication and direct clinical trials are needed before drawing strong conclusions regarding specificity by mood disorder subtype or implications of lifestyle psychiatry across age groups.

Highlights.

• The prominent emotional change following exercise was reduced anxiety, not sad mood

• Reduced anxiety following exercise occurred broadly across mood disorders

• Exercise was associated with lower subsequent anxiety in both males and females

• Changes in post-exercise anxiety and sadness initially differed across age strata

• Tween and young teen exercise-emotion dynamics require further investigation

Data statement:

Data from the NIMH Family Study are available upon request; qualified investigators may contact the NIMH Family Study Principal Investigator, Kathleen R. Merikangas. To protect research participant identities, only deidentified/coded or summarized data will be made available for sharing