Specificity of affective dynamics of bipolar and major depressive disorder

Abstract Objective Here, we examine whether the dynamics of the four dimensions of the circumplex model of affect assessed by ecological momentary assessment (EMA) differ among those with bipolar disorder (BD) and major depressive disorder (MDD). Methods Participants aged 11–85 years (n = 362) reported momentary sad, anxious, active, and energetic dimensional states four times per day for 2 weeks. Individuals with lifetime mood disorder subtypes of bipolar‐I, bipolar‐II, and MDD derived from a semistructured clinical interview were compared to each other and to controls without a lifetime history of psychiatric disorders. Random effects from individual means, inertias, innovation (residual) variances, and cross‐lags across the four affective dimensions simultaneously were derived from multivariate dynamic structural equation models. Results All mood disorder subtypes were associated with higher levels of sad and anxious mood and lower energy than controls. Those with bipolar‐I had lower average activation, and lower energy that was independent of activation, compared to MDD or controls. However, increases in activation were more likely to perpetuate in those with bipolar‐I. Bipolar‐II was characterized by higher lability of sad and anxious mood compared to bipolar‐I and controls but not MDD. Compared to BD and controls, those with MDD exhibited cross‐augmentation of sadness and anxiety, and sadness blunted energy. Conclusion Bipolar‐I is more strongly characterized by activation and energy than sad and anxious mood. This distinction has potential implications for both specificity of intervention targets and differential pathways underlying these dynamic affective systems. Confirmation of the longer term stability and generalizability of these findings in future studies is necessary.


INTRODUCTION
Hallmark features of bipolar disorder (BD) involve fluctuation of mood, cognition, and activity, whereas major depressive disorder (MDD) is characterized by episodes of lower mood, cognition, and activity (American Psychiatric Association, 2000. This fluctuation not only defines acute episodes, but also may influence interepisode function (Cochran et al., 2016;Henry et al., 2008;Prisciandaro et al., 2019;Sperry et al., 2020;Taylor et al., 2021;Wichers et al., 2010). This highlights the importance of better understanding affective fluctuation in mood disorder subtypes, with its potential for illuminating underlying differences between disorders and targets for intervention.
There are two major challenges to measuring affective fluctuation.
The first is that it may not be well-captured by reports that rely on retrospective assessments across a long reporting window-especially in the context of active fluctuation or occasional problems with insight.
And current affective states can influence participants' appraisal of their trait affect, underscoring the value of repeat measures (Brose et al., 2013). There has been increasing use of ecological momentary assessment (EMA), which involves many repeated measures within and between days that facilitate collection of dynamics of mood states in real time (Johns et al., 2019;Lamers et al., 2018;Merikangas et al., 2019). By design, EMA is less vulnerable to recall bias and has better ecological validity than retrospective assessments; it is also particularly well-suited to capturing implicit intraindividual processes and real-time dynamics (Burchert et al., 2021;Myin-Germeys et al., 2009;Schneider et al., 2020;Trull & Ebner-Priemer, 2020). The second challenge is that affect is a complex, multifaceted construct. A circumplex model of affect posits that much of emotional experience can be captured by dimensions of pleasant to unpleasant valence and high to low activation, which are orthogonal, and their product (Feldman Barrett & Russell, 1998;Larsen & Diener, 1992;Posner et al., 2005;Russell, 1980;Yik et al., 1999). Larsen and Diener's (1992) circumplex involves a symmetric and fully dimensional approach, labeling subjective feeling tone (valence) and activation (arousal) and their combinations. Using four bipolar scores to represent the four major dimensions thereby represents the full circumplex structure of affective states.
Despite the relevance of theory-informed design of EMA studies of affective dynamics (Hall et al., 2021) and comprehensiveness of the full circumplex, prior studies have tended to rely on either one or two indicators of mood or scales of positive and negative affect composed of a wide variety of indicators. This neglects the multifaceted nature of affect, and may obscure the driving components and reduce comparability across studies with different composite indicators. Evidence from such studies using summary scales or broad meta-analytic categories suggests associations between MDD with higher average levels, range (variability), and lability (temporal instability) of negative affect (Aan Het Rot et al., 2012;Houben et al., 2015;Scott et al., 2020;Thompson et al., 2012) and lower positive affect (Aan Het Rot et al., 2012;Thompson et al., 2012). These same summary approaches link BD or (hypo)mania with variability of positive emotions and variability and instability of negative affect (Gruber et al., 2013;Houben et al., 2015;Knowles et al., 2007;Sperry & Kwapil, 2022). These prior find-ings on euthymic affective dynamics in BD are more mixed than those in depression research, which may be due in part to the need to disambiguate individual components of the circumplex. Moreover, earlier reports, including our own work (Lamers et al., 2018), modeled inertia, variability, and instability separately and without cross-lags. That prevented concurrent examination of interrelationships of affective states themselves and their forms of variability.
Substantial evidence from community and clinical samples has documented distinctions in the familiality and genetics, onset and course, consequences, patterns of comorbidity, and treatment response between the BD subtypes of bipolar disorder type I (BPI) and bipolar disorder type II (BPII) as well as between BD and MDD (Mcintyre et al., 2022;Merikangas et al., 2014;Song et al., 2018;Tondo et al., 2022). Our earlier work distinguished different patterns in people with a history of BPI, BPII, MDD, and controls, including greater changes in sad mood and energy in response to changes in motor activity among persons with BPI . However, that work considered sad mood and energy separately without including their dynamic interrelationships.
We therefore employed dynamic structural equation modeling (DSEM) (Asparouhov et al., 2017) to examine the full emotional circumplex and the interrelationships in the underlying dimensions in persons with each of the major mood disorder subtypes. This approach combines time-series, multilevel, and structural equation modeling to allow for individual differences in parameters that define an individual's time-series process, thereby providing myriad possibilities without the strict assumptions of prior approaches to modeling EMA data. In a euthymic community-based sample enriched for mood disorders, our specific aims were to (1) characterize unconditional real-time affective dynamics stratified by mood disorder subtype among all four dimensions from the mood circumplex and (2) identify significant differences in those affective dynamics conditional on mood disorder subtypes.
Based on phenotypic and dynamic differences across mood disorders described above, we hypothesized that there would be differences in activation between BD and MDD and in valence between BPI and BPII, as well as unique patterns of emotional cross-reactivity among mood disorder subtypes.

Sample
The sample for the present study is a subset (N = 362) of participants  Larsen and Diener (1992), which was adapted from Russell (1980) and Watson and Tellegen (1985). Dimensions within the circumplex are as described by Larsen and Diener, and italicized words outside the circumplex are the individual emotions presented in our ecological momentary assessment (EMA) questionnaire. Words additionally bolded reflect our naming conventions for the variables/dimensions, with the scores from the Likert-type scale.
tant information is obtained from both interval-and signal-contingent recording of dynamic constructs (Hall et al., 2021).
The set of EMA questions regarding momentary affect used in this analysis were the same across both devices, representing the four major emotional dimensions with eight corresponding octants of the mood circumplex (Larsen & Diener, 1992; Figure 1).  (Schultzberg & Muthén, 2017). The analyses in this study made use of at least 5000 iterations, two chains, with a thinning of 10 (i.e., one in 10 iterations was saved). Additionally, we conducted a sensitivity analysis with at least 10,000 iterations of each model; these produced nearly identical results and we present them side-byside with our main results in Tables S1 and S2. At each iteration of the Markov chain Monte Carlo algorithm, missing data are sampled from their conditional posterior, which takes into account an individual's neighboring observations and autocorrelation as well as residual variance. DSEM does not use listwise deletion and can accommodate data missing at random. For this study, timepoints were treated as being equally spaced given their representation of approximately four waking hours each. In simulations, more densely spaced time grids could be inserted with reasonable estimation with as much as 80% missing data (Asparouhov et al., 2017), but our approach for the current study involved a naturalistic grid based on wakeful hours, where missing data would only exist for actual missed timepoints to preserve interpretability. Certain data are available on request; they are not publicly available due to privacy. Figure 2 demonstrates the DSEM model of affective dynamics (Hamaker et al., 2018). This is a lag-1 multilevel vector autoregressive model (VAR(1)) with random individual means (μ), autoregressive and cross-lagged (φ) regression coefficients, and log of the variance of the innovations (log(π)) for momentary emotional dimensions (sad, anxious, active, energetic). Random refers to the effects being allowed to vary across individuals. Individual means are the within-person mean level of a given emotional dimension across all timepoints, such as sadness (μ S ), which are modeled at the between-person level (in essence, group means of individual mean levels). Autocorrelation is the within-person relationship from one timepoint to the next, via the autoregression parameters, which uses past values to predict the current value. It can also be referred to as inertia or carryover, for example, the extent to which an individual's mood state such as sadness carries over from one timepoint to the next (φ SS ). Higher positive inertia suggests it takes longer to return to equilibrium after an increase at the prior timepoint. At the within-person level, we also measure the way a given indicator of momentary mood is influenced by a different indicator of momentary mood at a prior timepoint; this is referred to as the cross-regression, representing the relationships across mood states and time, or cross-lags. For example, we measure to what extent prior anxiety spills over into subsequent sadness (φ SA ).

Model decomposition and interpretation
Innovation variances on the within-person level include everything that was not measured explicitly but affects the course of the observed variables, such as individual variability of exposures and reactivity to those influences or external factors (Hamaker et al., 2018), and reflect within-person intensity of change from timepoint to timepoint, or lability. Innovation variances at the within-person level are modeled at the between-person level using their log to ensure they are positive for all individuals.
As presented in the results, means are the fixed or averaged effects in the sample or stratum for each of the random effects; variances of those means (shown in Tables S1 and S2) reflect the variability between the participants for each of the random effects. Tables include log inno-vation variances, with select exponentiated numbers presented here in the text. All means and variances in unconditional stratified models and regression coefficients in predictor modeling include 95% credible intervals (CIs), which 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, or, stated another way, given the observed data the effect has 95% probability of falling within the interval. Credible intervals that do not contain zero provide evidence that the mean differs from zero; however, when group means do not differ from zero it does not imply that the parameter is unimportant, because there may be meaningful individual differences.
Because all forms of correlation are measured simultaneously in the VAR(1) model, the effect estimates produced account for all cross-lags, inertias, innovations, and group means of individual mean levels of all four emotional dimensions and their covariances at once.
Unconditional models present all forms of emotional variability and cross-reactivity stratified by BPI, BPII, BD (a composite variable of BPI and BPII), MDD, and controls. For the multilevel regression analyses ("predictor modeling"), we included an observed predictor for the random effects at the between level, including individual mood disorder subtypes compared to controls and to each other, adjusted for age and sex (binary self-reported). Herein, we present unstandardized results, which allow for an interpretation close to the original scales that were used in measuring the variables; this is salient given the practicality of interpreting a scale that ranges from 1 to 7 with defined anchors.

Affective dynamics in mood disorder subtypes
To synthesize findings and aid interpretation, Figure S1 visualizes directionality and significance of within-stratum effect estimates (Table 2), and directionality and significance of pairwise differences between diagnostic groups from predictor models (bipolar subtypes compared to each other and controls [

Inertia
We found positive inertia for almost all affective states within strata, indicating a tendency for increases in affective states to carry over TA B L E 1 Sample characteristics comparing participants with a major mood disorder or no disorder who have wave 1 EMA data in the NIMH Family Study.  (100) Note: Anxiety disorders included agoraphobia, obsessive-compulsive disorder, generalized anxiety disorder, panic disorder, posttraumatic stress disorder, separation anxiety, social anxiety, and specific phobia. Anxiety and substance use disorders are not mutually exclusive, and therefore could be comorbid with a mood disorder. a Self-reported racial or ethnic identities other than White were combined here to protect participant privacy and prevent re-identification due to small cell sizes. Numbers of participants who did not report a racial or ethnic identity are also not presented by mood group for privacy reasons.
from one timepoint to the next, with the exception of active among controls (  Note: Bold means are those that are estimated to differ from 0 based on the bounds of the 95% credible interval. Bipolar disorder is a combined variable of individuals with either bipolar-I or bipolar-II. Each column of means with its corresponding credible interval is from its unique multivariate stratified model, with estimates that simultaneously adjust for all dynamics including individual mean levels, inertias, cross-regressions, and innovation variances. Abbreviations: L2.5, lower 2.5% of credible interval; U2.5, upper 2.5% of credible interval.

TA B L E 3
Regression coefficients and 95% credible intervals for predictor modeling testing differences in affective dynamics between bipolar disorder subtypes and controls.

DISCUSSION
We used DSEM to develop finer grained characterization of affective dynamics in mood disorders. We found that people with mood disorders displayed higher levels of sad and anxious mood and lower levels TA B L E 4 Regression coefficients and 95% credible intervals for predictor modeling testing differences in affective dynamics between major mood disorder categories and controls.

0.096
Note: Bold effect estimates are those that are estimated to differ from 0 based on the bounds of the 95% credible interval. Each column of effect estimates (regression coefficients) is from a unique, multivariate pairwise model that simultaneously adjusts for all dynamics including individual mean levels, inertias, cross-regressions, and innovation variances. All models adjust for age and sex. Abbreviations: Est., effect estimate; L2.5, lower 2.5% of credible interval; U2.5, upper 2.5% of credible interval.
of energy as compared to controls. People with BPI had lower average activation, and lower energy that was independent of activation, compared to MDD or controls. However, increases in activation were significantly more likely to perpetuate in those with BPI. Those with BPII were characterized by higher lability of sad and anxious mood compared to BPI and controls but not MDD. Compared to BD and controls, those with MDD exhibited cross-augmentation of sadness and anxiety, and sadness blunted energy.
Although the aforementioned average levels of higher sadness and anxiousness and lower energy were consistent across mood subtypes compared to controls, lower active levels were likely specific to BPI.
Prior evidence is mixed with respect to BD and average levels of "positive affect" (Aan Het Rot et al., 2012;Gruber et al., 2013;Knowles et al., 2007). This finding may be due to our assessing pure activation in its own dimension rather than crossed with valence. Specifically, nomenclature and measures of "positive" and "negative" affect build on Watson and Tellegen's (1985) work. That nomenclature is somewhat misleading because the constructs of negative and positive affect cross octants-that is, they involve both valence and arousal/activation (Larsen & Diener, 1992). Doing so may introduce measurement error and inhibit understanding of specificity. Previous findings suggest the centrality of energy and activity dysregulation in BD (Cheniaux et al., 2014), the salience of lower mean activity in euthymic BD (Cheniaux et al., 2014;Merikangas et al., 2019;Scott et al., 2017), and the importance of disentangling valence from activation. Previous findings also showed remitted BD and MDD groups aligned in reporting greater negative affect as well as sad mood measured individually compared to controls. The present study is larger than prior work, which may have allowed us to identify some differences between mood disorder subtypes not previously observed (Gruber et al., 2013;Lamers et al., 2018;Knowles et al., 2007;Scott et al., 2020). Moreover, by using DSEM, we are the first to have included all affective dynamics and circumplex dimensions in the same model, which may attenuate all but the most robust differences between mood disorder subtypes. Replication is needed.
Affect broadly tended to carry over from one timepoint to the next but the tendency for sadness inertia was significantly stronger in BPI and MDD. Elevation in activation was significantly more likely to perpetuate in BD, especially BPI, for whom anxiety was fleeting. In contrast, energetic feelings were less likely to perpetuate for those with MDD compared to controls. Our findings of greater sadness inertia and lower anxiety inertia in BPI differ from Lamers et al. (2018), who did not find significant differences between mood or anxiety groups in sad and anxious inertia examined univariately. Some participants from that earlier report (Lamers et al., 2018) are analyzed herein. However, our modeling approaches are quite different (univariate vs. multivariate)-reflecting the recent statistical advancements of this field, our sample has grown meaningfully larger, we did not include persons with anxiety disorders but no mood disorders, and we expanded our scope to the full emotional circumplex.
Our finding of the tendency for increased activation to perpetuate over time in BD is complementary with decades of work on the behavioral activation/approach system as a potential characteristic or prognosticator of BD (Alloy & Abramson, 2010;Johnson et al., 2012).
One of the only other studies to use DSEM to examine affective dynamics did not find evidence of cross-lags, that is, augmenting or blunting effects, between positive and negative affect in internalizing disorders (Scott et al., 2020); their sample was smaller and young, and the use of summary scales may have obscured more subtle fluctuations.
We found that those with MDD experienced a feedback loop of crossreactivity between sad and anxious levels, as well as a drop in energy after increased sadness. This is important when considering the significant tendency for sadness to persist, which may then amplify anxiety and, in turn, precede additional increased sadness. These findings may provide real-time evidence of negative attentional biases and cognitive styles, even in euthymia, that have been considered important in understanding mechanisms of recurrence in MDD and anxiety disorders (Peckham et al., 2010). Increased anxiety blunted subsequent energy in controls relative to MDD and BPII, and augmented sadness relative to BPI. Anxiety demonstrated starkly different cross-reactivity in BPII, augmenting subsequent energy and activation. Another qualitative difference in cross-reactivity of emotions was between energy and activation in BPI. Compared to BD, increased activation blunted subsequent energy in MDD and controls and energy augmented activation for controls. These patterns were likely driven by BPI, for whom activation, though self-perpetuating, did not spill over into energy, and energy did not spill over into activation. More research is therefore needed on mechanistic drivers of energy and activation, disambiguated, in BPI.
Higher innovation variances of sad and anxious across mood groups-likely driven by BPII and MDD-suggest greater withinperson variation from timepoint to timepoint, or more affective lability.
These findings were mostly consistent with Lamers et al. (2018), although we tested differences between BD subtypes and were therefore able to detect lower sad and anxious lability in BPI relative to BPII. The higher instability of "negative emotions" confirms prior studies of symptom patterns of people with BD (Houben et al., 2015;Scott et al., 2020). Within-person lability in energy was significantly larger in BD, likely driven by BPI, compared to both MDD and controls. Previous meta-analysis has demonstrated an association between BD and variability of "positive emotions" (Houben et al., 2015), so it may be that energy, a product of pleasant valence and higher activation, is one specific example of such.
An important caveat is the subtlety of mood fluctuations; overall, emotional ratings trended toward somewhat happy, mostly calm, somewhat inactivated/quiet, and neutral or balanced with respect to energy. Additionally, higher sadness might alternatively be labeled lower happiness. Despite these subtleties, a distinct picture emerged, such as higher unpleasant valence across mood groups that was more persistent within and among affective states, and also displayed higher intensity of changes (lability). The findings of a recent meta-analysis of adolescents and young adults showed that high instability and variability of negative affect were associated with more depressive symptoms among individuals with low average negative affect and fewer depressive symptoms among those with high average negative affect (Maciejewski et al., 2022). In combination with our findings, this raises the question of whether variation and cross-reactivity in affect can at times serve as a potential protective "mechanism" or perhaps simply as an indicator/trait of reduced dysregulation-both a "lifting" from more challenging emotions and a drive to return to homeostasis of balanced emotions.

Limitations with offsetting strengths and future directions
As participants were not in-episode for this study, we did not take into account the recency or severity of episodes on affective dynamics, focusing instead on lifetime diagnoses for this cross-sectional analysis.
Additionally, over time, some participants without a diagnosis at the time of EMA (e.g., adolescents or young adults) may eventually develop a psychiatric disorder. However, our findings are focused on the relationship between current affective dynamics and history of a disorder up until the time of EMA, and we do not attempt to predict future trajectories. Such features and trajectories of disorder status may be a promising future direction for investigations of mood transitions that incorporate intensive longitudinal data in longitudinal cohorts. Two weeks is not long enough to cover full hormone cycles in females, which are known to be linked to changes in energy and mood among other physiology (Slyepchenko et al., 2021). Investigators and clinicians interested in uncovering sex differences in fundamental affective dynamics of mood disorders as well as their interrelationships with biologic influences should sample experiences over longer periods.
Timepoints were treated equally based on wakeful hours for interpretability, which may affect strength of association (Asparouhov et al., 2017). Additionally, interpretation of findings from interval-contingent schedules of roughly equal time periods may be different from fully random signal-dependent schedules (Hall et al., 2021). We also did not examine within-day versus between-day fluctuation in affect, which may differ for positive affect in individuals with high hypomanic scores (Sperry & Kwapil, 2022); future investigations will quantify the impact of sleep on continuation versus interruption of affective dynamics.
Although theoretically informed, there is not an objective truth to which these findings can be validated, and words for emotions are not the same thing as emotions (Larsen & Diener, 1992). However, a circumplex framework for understanding affect and emotional dynamics has been well-studied in clinical and nonclinical samples. Our findings build upon that tradition and make use of state-of-the-science modeling to understand real-time variability of individual emotions and their cross-reactivity, simultaneously adjusted for each other, in a richly phenotyped community-based sample, which is a key advancement in understanding affective dynamics of mood disorders. Although we statistically adjusted for age in our models, future work should directly interrogate within-person change in depth of emotional expression longitudinally in persons with mood disorders as well as in larger samples of youth to better understand how development may impact measurement of emotion using the circumplex (Posner et al., 2005).
There are other approaches to understanding affective dynamics-positive and negative affect being predominant. However, for this initial novel application of DSEM characterizing dynamics in mood disorder subtypes, we did not believe the solution to understanding the complexity of the emotional circumplex was to subsume individual emotions under scales that cross octants. Future research should examine the utility of individual octants or dimensions in contrast to data-reducing approaches such as scales or common latent factors in understanding differences across mood disorders (Bos et al., 2022;Taylor et al., 2021) and general affective disturbance and risk in nonclinical samples (Sperry & Kwapil, 2022;Trampe et al., 2015). Finally, we did not examine what is driving these fluctuations and thus cannot determine whether they are stochastic, innate, or sparked by health behaviors or comorbidities, for example (Alloy & Abramson, 2010;Johnson et al., 2012;Lewis et al., 2022;Pemberton & Fuller Tyszkiewicz, 2016;Quick et al., 2022;Stapp et al., 2022). Our perspective is that there is inherent worth in understanding these characteristics and dynamics as they are, irrespective of putative cause. Future research on multiple health and social behaviors and contexts simultaneously with multivariate affective dynamics in real time would enrich understanding of antecedents and intervention points.

CONCLUSION
These findings demonstrate differences in the affective dynamics of BD compared to MDD between episodes. Future studies should examine the stability of the patterns across episodes that may then serve as early warning signals of upcoming mood transitions (Bos et al., 2022;Van De Leemput et al., 2014;Yee et al., 2021). Given the exploratory nature of this approach, future work should seek to independently replicate these findings and assess their readiness for clinical applications (Bosley et al., 2019) such as digital feedback, therapy and psychoeducation, or pharmacology. Whether said dynamics can be fully operationalized, predicted, and (self)regulated will have significant clinical import. Herein, we have demonstrated DSEM's utility in multivariate characterization of dynamic affect, finding that mood disorder subtypes are differentiated by distinctive patterns of affective dynamics, which may deepen our understanding of mood disorders.

ACKNOWLEDGMENTS
We thank J. Swendsen for his significant contributions to the development of this project and dedicate this article to him. We also wish to thank our study participants. This work was supported by the Intramural Research Program of the National Institute of Mental Health (ZIA MH002804; clinical protocol NCT00071786).

CONFLICT OF INTEREST STATEMENT
The authors declare no conflicts of interest.

DATA AVAILABILITY STATEMENT
Certain data are available on request; data are not publicly available due to privacy.