Affective Dynamics Across Internalizing and Externalizing Dimensions of Psychopathology

Affective Dynamics and Psychopathology

A range of affective dynamic constructs have been described using diverse terminology and methods of assessment (Krone, Albers, Kuppens, & Timmerman, 2018; Kuppens & Verduyn, 2015). Although not technically a temporal dynamic, mean intensity (i.e., the within-person average level of an emotion across repeated measures) is an important aspect of an individual’s emotional profile that is robustly associated with psychological well-being (Krone et al., 2018; Spindler, Stopsack, Aldinger, Grabe, & Barnow, 2016). Intensive longitudinal studies (e.g., using EMA) have demonstrated that high negative affect (NA) and low positive affect (PA) mean intensity characterize individuals with social anxiety disorder (Kashdan & Farmer, 2014), borderline personality disorder (BPD; Ebner-Priemer et al., 2015), depression (Erbas, Ceulemans, Lee Pe, Koval, & Kuppens, 2014), and high levels of neuroticism (Erbas et al., 2014; D. J. Miller, Vachon, & Lynam, 2009), which itself is associated with internalizing problems (Kotov, Gamez, Schmidt, & Watson, 2010). Evidence of associations between NA and PA mean intensity in daily life and externalizing problems is sparse. However, disinhibition-related traits (i.e., low conscientiousness and agreeableness) are associated with both low PA (Carstensen, Pasupathi, Mayr, & Nesselroade, 2000; Komulainen et al., 2014) and externalizing problems (Ruiz, Pincus, & Schinka, 2008), which suggests that externalizing disorders may be associated with low PA but not necessarily high NA. Notably, internalizing and externalizing disorders are rarely examined concurrently in existing EMA research, which limits firm conclusions about the extent to which differences in NA and PA mean intensity are transdiagnostic as opposed to disorder-specific.

Although research examining affect intensity and externalizing pathology using intensive longitudinal methods is limited, there is research using retrospective self-report of NA and PA; however, this work has yielded conflicting results. For instance, pathological gambling (MacLaren, Fugelsang, Harrigan, & Dixon, 2011) and substance use disorders (Cheetham, Allen, Yücel, & Lubman, 2010; Elkins, King, McGue, & Iacono, 2006) have been associated with elevated NA and mixed results for PA. Aggressive behavior, however, may be associated with low PA, particularly with respect to social relationships (Fanning, Berman, & Guillot, 2012), and psychopathy has been tied to low PA and elevated NA (Garcia, Adrianson, Archer, & Rosenberg, 2015). Other work has failed to find evidence of a link between PA and externalizing pathology (Watson, Stasik, Ellickson-Larew, & Stanton, 2015).

Inertia and variability are commonly studied features characterizing temporal persistence and fluctuation of emotions, respectively (Krone et al., 2018; Kuppens & Verduyn, 2015). Inertia (e.g., autoregression, carry-over) refers to the degree to which an emotion persists from one occasion to the next (Kuppens, Allen, & Sheeber, 2010); higher inertia indicates less flexibility to adapt to changing circumstances. Variability describes the amplitude of deviations from an individual’s mean level over time (Krone et al., 2018), usually quantified as a within-person standard deviation. Variability metrics that take inertia into account, such as the mean squared successive difference between consecutive emotion measurements (Trull et al., 2008), capture moment-to-moment fluctuations; higher levels indicate a more unstable emotional profile (i.e., affective instability). Greater variability per se is generally thought to reflect heightened affective reactivity to internal or external events. Some affective dynamic models, such as the Flex3 model (Hollenstein, 2015), highlight the importance of dynamic (real-time) flexibility within a particular context as distinct from reactive flexibility, which occurs in response to shifting contexts, environments, or situations.

Although some degree of inertia and variability of emotions is considered normative (Cunningham, Dunfield, & Stillman, 2013; Kuppens, Oravecz, & Tuerlinckx, 2010) and the role of emotional inertia may vary on the basis of emotional context (Koval, Sütterlin, & Kuppens, 2016), higher levels of both were linked to a wide range of psychopathology. Specifically, a recent meta-analysis of 79 short-term affective dynamics studies (Houben et al., 2015) suggested that higher NA and PA inertia and variability were associated with multiple types of symptoms, although effect sizes were generally smaller for PA dynamics. However, PA variability was not uniformly associated with maladaptive constructs. For example, PA variability was positively associated with extraversion, negatively associated with depression, and marginally positively associated with positive emotionality (Houben et al., 2015). Another recent study found that higher PA instability predicted subsequent increases in life satisfaction, which suggests that PA instability might reflect adaptive flexibility and reactivity to environmental input, including positive or rewarding events (Spindler et al., 2016). These findings are consistent with research suggesting that depressed individuals demonstrate less contextually appropriate sensitivity to positive as well as negative stimuli (Bylsma, Morris, & Rottenberg, 2008). Overall, greater NA variability appears to be largely transdiagnostic across the internalizing and externalizing spectra, but the nature of relations between PA variability and psychopathology may be more complex. Furthermore, no studies have examined associations between inertia and externalizing symptoms. Thus, it remains unclear whether NA and PA inertia characterize affective dynamics across both internalizing and externalizing disorders.

Augmentation and blunting describe how changes in one emotion relate to near-term changes in another emotion (Pe & Kuppens, 2012). Positive cross-lagged coefficients indicate that increased experience of one emotion predicts an increase in another emotion at a successive time point (i.e., augmentation), whereas negative coefficients indicate increased experience of one emotion precedes decreased experience of another (i.e., blunting). Elevated levels of augmentation and blunting (e.g., stronger connections between distinct emotions over time) are thought to reflect a densely connected emotional network (Pe et al., 2014), which may reflect an emotional system that is more resistant to change and impermeable to external influences (Kuppens & Verduyn, 2015). These dynamics have not been well studied in relation to psychopathology yet hold great potential for improving our understanding of emotion regulation processes.

Covariation describes the extent to which different emotional experiences occur together and is commonly calculated with correlations between multiple emotions assessed at the same time (Krone et al., 2018; Kuppens & Verduyn, 2015). Negative correlations between positive and negative emotions (i.e., when stronger increases in PA are accompanied by stronger concurrent decreases in NA and vice versa) indicate greater affective polarization (Rafaeli, Rogers, & Revelle, 2007). Greater affective polarization in daily life (i.e., experiencing emotional states as all good or all bad) is associated with depression and anxiety (Coifman, Bonanno, & Rafaeli, 2007; Hamaker, Asparouhov, Brose, Schmiedek, & Muthén, 2018) as well as BPD and impulsive behaviors such as risky sexual behavior, substance use, and self-injury (Coifman, Berenson, Rafaeli, & Downey, 2012). These data suggest that polarized affective experiences may be transdiagnostic indicators of pathology across internalizing and externalizing symptoms. These results are consistent with the dynamic-affect model (Zautra, Potter, & Reich, 1997), which suggests that increased emotional intensity (consistent with many types of psychopathology) will contribute to within-person increases in affective polarization.

Limitations of Prior Studies

Although most studies examined affective dynamics in a single diagnostic group relative to healthy control subjects, there is evidence of common affective patterns across various disorders on the internalizing spectrum (Ebner-Priemer et al., 2015; Santangelo et al., 2016; Trull et al., 2015). However, the affective dynamics that uniquely characterize externalizing disorders such as antisocial personality disorder are largely unknown. To date, no studies have examined how internalizing and externalizing symptoms are associated with temporal dynamics of emotions in a single transdiagnostic sample. Such research could aid in delineating pathogenic affective dynamics that are common across all psychopathology as opposed to those that are specific to either high-order dimensions (e.g., internalizing, externalizing) or specific diagnostic categories (e.g., depression, social anxiety, BPD).

Another limitation of previous studies is that affective dynamic constructs were usually examined in isolation without accounting for their covariation with other dynamics (e.g., studying variability without considering inertia or mean affective intensity), which may produce biased or inaccurate estimates of dynamic components (Krone et al., 2018). In addition, the majority of previous studies overlooked patterns of cross-lagged associations or covariance among two or more emotions in the context of other dynamic features such as inertia and variability. Most prior studies also used statistical methods that assume equal spacing between repeated measures; however, unequally spaced observations characterize most experience sampling data. Ignoring differences in the intervals between assessments produces inaccurate estimates of true serial autocorrelation (Oravecz, Tuerlinckx, & Vandekerckhove, 2011).

Recent innovations in dynamic multilevel modeling (e.g., dynamic structural equation models [DSEM]) can address many of these methodological limitations. It is now possible to model (as random effects) individual differences in overall means, autoregressive coefficients, cross-lagged effects, residual variances, and their covariance while also specifying a specific time interval of interest to account for unequal spacing between assessments (Asparouhov, Hamaker, & Muthén, 2018; Hamaker et al., 2018; Jongerling, Laurenceau, & Hamaker, 2015). In this context, the random residual variance (i.e., innovation variance) represents the portion of within-person variability that cannot be predicted by the previous observations or other variables in the model. The innovation variance is of substantive interest as a measure of individual differences in reactivity or exposure to internal or external factors that influence emotional states, such as stressful interactions, substance use, positive events, or sleep quality.

Moreover, with methods such as DSEM, it is now possible to examine multiple affective processes simultaneously to understand cross-lagged effects (i.e., affective augmentation or blunting) and covariance of innovations (i.e., affective differentiation or polarity) in addition to mean intensity, inertia, and innovation variance. For instance, Hamaker et al. (2018) used a DSEM framework to examine the dynamics of both PA and NA assessed via daily self-reports. Their results suggested that in young adults, becoming “stuck” in negative affectivity (higher NA inertia) and greater polarization of PA and NA (more negative covariance of innovations) are characteristic of depressive symptoms and appear to perpetuate them.

Goals and Hypotheses of the Current Study

The overarching goal of our study was to examine relations between empirically identified higher-order dimensions of psychopathology and multiple affective dynamics in a diagnostically diverse urban sample of young women. We used bivariate multilevel cross-lagged models of EMA data within a DSEM framework (Asparouhov et al., 2018; Hamaker et al., 2018) to address many of the limitations in prior studies. Specifically, we estimated multiple temporal dynamic features of PA and NA simultaneously, including mean intensity, inertia, variability (i.e., innovation variance), augmentation and blunting (i.e., cross-lags), and covariance (i.e., polarization) at the hourly level. We then examined transdiagnostic dimensions of psychopathology (derived from confirmatory factor analysis [CFA] of diagnostic categories obtained via clinician-administered semistructured diagnostic interviews) as predictors of all affective dynamics in a multivariate DSEM. On the basis of previous studies examining the structure of psychopathology (Eaton et al., 2011; Kotov et al., 2017; Krueger, McGue, & Iacono, 2001), we expected that two factors, representing internalizing and externalizing dimensions, would emerge from CFA of the diagnostic categories. Given previous findings (for reviews, see Houben et al., 2015; Krone et al., 2018), we also hypothesized that greater levels of internalizing psychopathology would be associated with (a) higher NA mean intensity, inertia, and variability; (b) lower PA mean intensity; and (c) more NA and PA blunting and polarization. Although evidence for affective dynamics and externalizing disorders is sparse, we tentatively hypothesized that externalizing disorders would be associated with lower PA mean intensity, higher NA variability, and affective polarization.

Method

Assessment procedures and measures

Participants completed semistructured diagnostic interviews (described below) in the laboratory followed by a 21-day EMA protocol. All study procedures were approved by the University of Pittsburgh Institutional Review Board (Protocol Number PRO13050549). This project was conducted in accordance with the Declaration of Helsinki. Participants were compensated for their participation.

Diagnostic interviews

Psychiatric disorders were assessed using the Structured Clinical Interview for DSM–IV Axis I Disorders (SCID-I; First, Gibbon, Spitzer, & Williams, 1997) and the Structured Clinical Interview for DSM–IV Personality Disorders (SIDP-IV; Pfohl, Blum, & Zimmerman, 1997). ¹ Interviewers were clinicians at the postbaccalaureate, master’s, and doctoral levels who were trained by the first author (L. N. Scott). The present analyses were restricted to current (i.e., not lifetime) disorders that affected 5% or more of our sample. Prevalence rates for each diagnosis are presented in Table 1. A randomly selected set of interviews (n = 25; 16%) were videotaped and rated by an independent clinical judge for calculation of interrater reliability. Kappas (κs) were calculated for diagnoses with frequencies of 5% or more in the reliability sample. The average κ across diagnostic categories was .81 (SD = .15; range = .62–1.0).

OPEN IN VIEWER

Table 1.

Prevalence Rates for Current Clinical Diagnoses That Were Most Common (Affecting > 5%) in the Study Sample (N = 156)

Diagnosis	n	%
Alcohol use disorder	8	5.1
Antisocial personality disorder	14	9.0
Borderline personality disorder	23	14.7
Drug use disorder	31	19.9
Dysthymia	8	5.1
Generalized anxiety disorder	13	8.3
Major depressive disorder	25	16.0
Panic disorder	8	5.1
Posttraumatic stress disorder	21	13.5
Social phobia	10	6.4
Specific phobia	11	7.1

Ecological momentary assessment of positive and negative affect

After completion of diagnostic interviews, participants were trained to use a touch-screen smartphone on which they completed a 21-day web-based EMA protocol consisting of a daily morning assessment and six daily random prompts to assess their current mood (for additional description of the EMA protocol, see Scott et al., 2017). At each assessment, participants were asked to rate on 5-point Likert scales (1 = not at all, 5 = extremely) the extent to which they experienced different emotions “in the past 15 minutes.” This wording was used to reduce zero inflation of affect ratings commonly seen in momentary assessments in which participants respond regarding their affect as they are filling out the survey rather than how they were feeling before the survey prompt. Items assessing general positive and negative affective states were drawn from the Positive and Negative Affect Schedule–Extended version (Watson & Clark, 1999). PA and NA scales were created by calculating the mean of positive items and the mean of negative items for each person at each measurement occasion. For each scale, between- and within-person reliability estimates were calculated using methods described by Cranford et al. (2006). The scales used, their corresponding items, and between- and within-person reliability estimates for each scale were as follows: NA (ashamed, guilty, hostile, irritable, scared, sad, lonely; .98 and .74) and PA (excited, proud, happy, cheerful, joyful, delighted; .99 and .90). Intraclass correlation coefficients, which represent the proportion of between-person variability in repeated measures from the EMA protocol, were .38 for NA and .42 for PA. Thus, more than half of the variability in these measures was within individuals over time.

Participants completed an average of 112 EMA reports (SD = 31.56), which corresponds to 76% of possible reports. A total of 17,465 observations were included for analysis. Prior simulation studies (Schultzberg & Muthén, 2018) showed that DSEM performs well when the number of subjects and observations are both greater than 100, as is the case in our study. Given the semirandom assessment schedule, the distribution of time intervals between assessments varied widely but was positively skewed; the modal value was 1.6 hr, and the median was 2.2 hr. Thus, the majority of assessments were less than 2 hr apart, which informed our decision to model the data at the hourly interval as further described below.

Analytic procedures

Analyses were conducted using Mplus (Version 8.2; Muthén & Muthén, 2017). We first used CFA to identify higher-order constructs underlying disorders that affected 5% or more of our sample (see Table 1). On the basis of disorders that were common in our sample as well as prior literature on structural models of psychopathology (Eaton et al., 2011; Kotov et al., 2017; Krueger et al., 2001), we constructed our initial model with two factors: (a) Internalizing, which comprised major depressive disorder (MDD), dysthymic disorder, panic disorder, social phobia, specific phobia, posttraumatic stress disorder (PTSD), generalized anxiety disorder, and BPD; and (b) Externalizing, which comprised BPD, antisocial personality disorder, alcohol use disorder, and drug use disorder (collapsed across individual drug abuse/dependence diagnoses).

All diagnostic categories were entered as categorical variables, and the weighted least square mean and variance adjusted (WLSMV) estimator was used. Standard indices were examined to evaluate model fit, including χ², comparative fit index (CFI), Tucker-Lewis index (TLI), and root mean square error of approximation (RMSEA). Excellent model fit is indicated by a nonsignificant χ² test, CFI and TLI values > .95, and RMSEA values < 0.05 (Hu & Bentler, 1999). For well-fitting models, we examined the pattern of factor loadings to identify improper solutions (e.g., standardized factor loadings > 1) and determine whether any diagnoses should be removed from the model because of weak or negative factor loadings (i.e., standardized loadings < .3). We also examined modification indices (MI) to determine whether any theoretically defensible cross-loadings should be allowed (using MI > 5 as the basis). Nested models were compared using the DIFFTEST option in Mplus, which calculates the χ² difference test for WLSMV estimation. After the best fitting CFA model was identified, we extracted the factor scores to be used as predictors in subsequent analyses.

Next, we modeled the hour-to-hour temporal dynamics of NA and PA using bivariate multilevel cross-lagged models estimated using DSEM in Mplus Version 8.2 (Muthén & Muthén, 2017). Specifically, we estimated a first-order vector autoregressive—VAR(1)—model with Bayesian Markov chain Monte Carlo (MCMC) estimation and default (i.e., noninformative) priors. The DSEM framework has several advantages for intensive longitudinal data analysis: (a) Repeated measures of NA and PA are automatically decomposed into within- and between-person variance using latent mean centering, which avoids the downward bias in autoregressive parameters associated with sample mean centering; (b) multiple random effects and their covariances can be modeled simultaneously; (c) unequal intervals between observations due to random sampling or missed assessments can be accommodated by specifying the time interval of interest for time-lagged associations (e.g., TINTERVAL command); and (d) missing data are imputed automatically by sampling from the conditional posterior distribution at each iteration of the MCMC algorithm. Because we were interested in short-term temporal dynamics and most observations were less than 2 hr apart, we specified 1-hr increments as the time interval of interest by setting the TINTERVAL function to a value of 1, scaled to the unit of hours. With this specification, covariance coverage was > 23%, which is well above the minimum recommended coverage for DSEM according to prior simulations (Asparouhov et al., 2018). Our model specifications mirrored those used in a similar study of the temporal dynamics of NA and PA at the daily level (for technical details, see Hamaker et al., 2018). We allowed all random effects to covary with each other at the between-persons level. All models were estimated using default (i.e., noninformative) priors, 50,000 iterations, and two MCMC chains. Model convergence was checked using the proportional scale reduction (PSR) indices for each parameter and examining trace plots for irregularities (e.g., trends and cycles). PSR values close to 1 suggest model convergence.

We first estimated an unconditional bivariate multilevel VAR(1) model with no between-person covariates, which included nine random effects (depicted in Fig. 1):

Individual differences in mean intensity of PA and NA, as indicated by random means µ_PA and µ_NA;

Inertia of PA and NA from one hour to the next, as reflected in higher autoregressive coefficients Φ_PP and Φ_NN;

Augmentation or blunting, as reflected in cross-lagged regressions Φ_PN and Φ_NP;

Variability of PA and NA, as indicated by random logs of the unique parts of the innovations (i.e., innovation variances), log(π_PA) and log(π_NA); and

Affective polarization, as indicated by associations between the unique parts of the innovations of PA and NA within each hour, captured in log of negative covariance between the innovations log(ψ).

As in Hamaker et al. (2018), we specified a common factor for the covariance of the innovations with factor loadings of 1 and −1 for PA and NA to aid in model convergence. The resulting parameter is modeled as the log of the negative covariance at the between-persons level. The predicted covariance of the innovations can thus be calculated by taking the exponentiated predicted log of the negative covariance and multiplying by −1 (for details and formulas, see Hamaker et al., 2018). Generally, a more strongly negative predicted covariance of innovations suggests more polarization such that within-hour increases in NA are accompanied by same-time decreases in PA and vice versa.

OPEN IN VIEWER

Fig. 1.

Path diagram for the unconditional bivariate multilevel VAR(1) model for positive affect (PA) and negative affect (NA) with nine random effects: individual differences in mean intensity of PA and NA (random means µ_PA and µ_NA); inertia of PA and NA from one hour to the next (autoregressive coefficients Φ_PP and Φ_NN); augmentation or blunting (cross-lagged regressions Φ_PN and Φ_NP); individual differences in innovation variances as captured by the random log of the unique parts of the residuals (or innovations) of PA and NA (innovation variances, log(π_PA) and log(π_NA)); and affective polarization, as indicated by associations between the unique parts of the innovations of PA and NA within each hour captured in the log of negative covariance between the innovations, log(ψ). The final unconditional and conditional models included two additional random effects of time regressed on PA and NA (not shown). Yellow ovals indicate within-level variables, and blue ovals represent random means. W = within-individual variables; η _t = a common factor for the covariance of the innovations with factor loadings of 1 and −1 for PA and NA. This figure was adapted with permission from Hamaker et al. (2018).

After establishing convergence of the unconditional model, we estimated a conditional DSEM model in which all random effects at the between-person level were regressed on the factors extracted using CFA, which allowed us to test hypotheses regarding associations between higher-order dimensions of psychopathology and affective dynamics. Given that nearly three quarters of the sample identified as minority race/ethnicity and more than half were receiving public assistance linked with household income below the national poverty level, we decided a priori to include minority race or ethnicity (0 = White and non-Hispanic; 1 = African American, multiracial, or Hispanic/Latinx) and low socioeconomic status (SES; 0 = not receiving public assistance, 1 = receiving public assistance) as covariates (i.e., predictors of all random effects at the between-persons level) in the final model. We also explored several other factors (e.g., number of EMA entries completed, participant age, education) as potential between-person covariates, but they did not emerge as significant predictors and were omitted from subsequent analyses.

Results

Confirmatory factor analysis

The initial two-factor CFA model fit the data well according to all fit indices, χ²(42) = 40.30, p = .55; CFI and TLI ≥ 1; RMSEA < 0.001, but examination of the pattern of factor loadings indicated model misspecification with the cross-loading of BPD on both factors. Specifically, BPD had a standardized factor loading greater than 1 on the internalizing factor and a negative loading on the externalizing factor. Therefore, we respecified the model with BPD loading only on the internalizing factor. The revised, more constrained model demonstrated excellent fit to the data according to all fit indices, χ²(43) = 41.14, p = .55; CFI and TLI ≥ 1; RMSEA < 0.001; there was no significant decrement in model fit compared with the initial model, Δχ²(1) = 0.55, p = .45. All factor loadings were within the expected range and statistically significant. However, an MI greater than 5 suggested that PTSD should be allowed to cross-load on the externalizing factor. This is consistent with evidence of high rates of comorbidity of PTSD with both internalizing and externalizing disorders (Kessler, Sonnega, Bromet, Hughes, & Nelson, 1995; M. W. Miller & Resick, 2007). Hence, the model was again respecified to allow PTSD to load on both internalizing and externalizing factors. This second revised model fit the data well, χ²(42) = 35.24, p = .76; CFI and TLI ≥ 1; RMSEA < 0.001, and resulted in a significant improvement in model fit in comparison with the model with no cross-loadings, Δχ²(1) = 5.82, p = .02. Although the loading of PTSD on internalizing disorders was no longer statistically significant (p = .13), the magnitude of this loading (.35) and the improvement in model fit led us to retain this version of the two-factor model.

Alternative models were explored (e.g., one-factor, three-factor, second-order hierarchical, and bifactor models), but these resulted in either worse fit or nonconvergence. Therefore, we accepted the revised two-factor model (depicted in Fig. 2) as the best fitting model for the diagnostic data. We extracted factor scores for each participant from this model to be used as predictors in subsequent analyses to examine the relations between these two transdiagnostic dimensions of psychopathology and affective dynamics. Additional details on the alternative CFA models tested and evidence for the suitability of the extracted factors for further analysis are provided in the Supplemental Analyses section of the Supplemental Material available online.

OPEN IN VIEWER

Fig. 2.

Results of the final model from the confirmatory factor analysis showing the standardized loadings of individual disorders on latent externalizing and internalizing psychopathology factors. Values on single-headed arrows are standardized factor loadings, and the value on the double-headed arrow is a correlation. Ovals indicate latent variables; rectangles indicate observed variables. All parameters are statistically significant (p < .05) except for the loading of posttraumatic stress disorder on internalizing (p = .13).

Conditional DSEM

In the conditional DSEM, the 11 random effects described above (i.e., two random trends and nine random affective dynamics parameters) were regressed on minority race, low SES, and the extracted internalizing and externalizing factors. Covariances of predictors and outcomes were also estimated. Before testing the full conditional model, we examined collinearity diagnostics, including variance inflation factor (VIF) and tolerance statistics, which revealed no evidence for multicollinearity among the predictors (VIFs > 1 and < 2; tolerance ≥ .62). Unstandardized fixed effects (means), regression parameters, and effect sizes (Cohen’s f²) for internalizing and externalizing factors on each of the nine random affective dynamics from the final conditional model are presented in Table 2. The highest PSR value in the final conditional model was 1.008, which suggests good model convergence. Full unstandardized and standardized results from the final conditional DSEM are provided in the Supplemental Material. Note that the results reported here focus primarily on unique associations that emerge when controlling for both internalizing and externalizing problems and their covariation in the same model. To aid in the interpretation of the full conditional model results, we also provide results from analyses with only internalizing or externalizing as predictors of affective dynamics (after controlling for demographic covariates) in the Supplemental Material (see Supplemental Analyses and Final Conditional Model Results). ²

OPEN IN VIEWER

Table 2.

Unstandardized Estimates (Posterior Means), 95% Credibility Intervals (CIs), and Effect Sizes (f²) From the Conditional Multilevel Model

Variable	Estimated mean	Between-person predictors of random effects
		Externalizing factor	f 2	Internalizing factor	f 2
PA mean intensity (mean; µPA)	1.85 [1.66, 2.05]	−0.04 [−0.42, 0.34]	.01	−0.05 [−0.29, 0.18]	.01
NA mean intensity (mean; µNA)	1.25 [1.18, 1.33]	−0.03 [−0.18, 0.12]	.01	0.17 [0.07, 0.27]	.10
PA inertia (autoregression; ΦPP)	0.47 [0.38, 0.55]	−0.24 [−0.40, −0.08]	.08	0.05 [−0.06, 0.14]	.01
NA inertia (autoregression; ΦNN)	0.45 [0.35, 0.54]	−0.04 [−0.23, 0.14]	.01	0.06 [-0.06, 0.17]	.02
NAt-1→PA t augmentation (cross-lag; ΦPN)	−0.18 [−0.30, −0.08]	−0.02 [−0.21, 0.17]	.01	0.01 [−0.10, 0.12]	.01
PAt-1→NA t augmentation (cross-lag; ΦNP)	−0.04 [−0.06, −0.02]	0.002 [−0.04, 0.04]	.01	−0.02 [−0.04, 0.01]	.04
PA variability (innovation variance; log(πPA))	−1.84 [−2.19, −1.48]	0.71 [0.01, 1.42]	.04	−0.47 [−0.91, −0.04]	.04
NA variability (innovation variance; log(πNA))	−4.22 [−4.74, −3.72]	0.06 [−0.93, 1.06]	.01	0.97 [0.37, 1.58]	.08
Polarization (negative covariance of innovations; log(ψ))	−3.91 [−4.36, −3.46]	0.66 [−0.25, 1.56]	.03	0.18 [−0.40, 0.76]	.02

Note: N = 156 (17,465 observations). CIs are presented in brackets. Boldface type indicates effects that are statistically significant (i.e., the CIs do not include zero). Effect sizes (f²) were calculated for effects of internalizing and externalizing on the basis of model R² values from models with and without each of these variables (for all conditional model results and R² values, see the Supplemental Material). Effects of covariates (i.e., time at the within-person level; race/ethnicity and low socioeconomic status at the between-person level) and covariances among predictors and dependent variables (random effects) were also estimated (for these parameter estimates and fully standardized results for all conditional models, see Supplemental Material). PA = positive affect; NA = negative affect.

Partially consistent with our hypotheses, both with and without externalizing problems in the model, internalizing disorders were associated with greater NA mean intensity (µ_NA) and NA variability (i.e., NA innovation variance, log(π_NA)). In addition, the internalizing factor was associated with lower PA variability (PA innovation variance, log[π_PA]), but this emerged only after we controlled for externalizing disorders. Thus, internalizing disorders were uniquely associated with more intense and variable NA and less variable PA. Contrary to hypotheses, internalizing problems were not significantly associated with greater NA inertia (Φ_NN), PA mean intensity (µ_PA), augmentation or blunting (Φ_PN,i and Φ_NP,i), or polarization—log(ψ _i )—in any model. Unexpectedly, in models with and without internalizing disorders, externalizing psychopathology was associated with lower PA inertia (i.e., Φ_PP). Externalizing disorders were also associated with greater PA variability—PA innovation variance, log(π_PA)—only after we controlled for internalizing disorders. These findings suggest that externalizing disorders are uniquely associated with more fleeting and variable PA. It should be noted, however, that all effect sizes were quite small according to Cohen’s (1988) criteria (i.e., f² = .02 and .15, which corresponds to small and medium, respectively). The most robust effect emerged for the association between internalizing disorders and NA mean intensity (f² = .10).

Discussion

This is the first study to examine how superordinate dimensions of psychopathology relate to multiple dynamic features of NA and PA at the hourly level in daily life. As hypothesized, two factors representing internalizing and externalizing dimensions emerged from CFA of categorical diagnoses. Partially consistent with our expectations, greater internalizing disorders were associated with higher NA mean intensity and variability (i.e., innovation variance) as well as lower PA variability after controlling for externalizing disorders. These findings suggest that internalizing disorders are uniquely associated with greater overall negative affectivity and reactivity to internal or external events, which may reflect greater stress sensitivity combined with less reward sensitivity. Results did not support our hypotheses regarding associations between internalizing disorders and NA inertia, PA mean intensity, cross-lagged associations (i.e., augmentation/blunting), or polarization of NA and PA. Also contrary to our expectations, externalizing disorders were associated with lower PA inertia and higher PA variability after controlling for internalizing disorders and were not associated with lower PA mean intensity, higher NA variability, or affective polarization.

Our results suggest that disorders along the internalizing spectrum reflect a greater tendency to experience negative emotion but not necessarily more persistent emotion, extremes of emotion (as with polarization), or affective blunting. Our finding that internalizing disorders are linked to more intense and variable NA as well as less variable PA is consistent with the character of these disorders and their underlying personality constructs. In particular, evidence from prior studies suggested that internalizing disorders are characterized by elevated neuroticism (Kotov et al., 2010), which is linked to higher negative emotionality, lower positive emotionality, and more negative affective variability and reactivity to stressors (Komulainen et al., 2014). Although we did not find an association between internalizing disorders and low PA mean intensity, the negative association between internalizing disorders and PA variability may reflect less flexibility and reactivity to positive or rewarding input (Spindler et al., 2016). These results are consistent with prior findings that demonstrated that depressed individuals display low PA variability (Houben et al., 2015) and less sensitivity to positive stimuli (Bylsma et al., 2008). Our results expand on prior work by demonstrating that these dynamics characterize internalizing disorders more broadly rather than depression specifically.

Although we anticipated, on the basis of prior work (Coifman et al., 2007; Hamaker et al., 2018), that internalizing pathology would be associated with affective polarization, this was not the case in the present study. This may be due to the concurrent modeling of NA mean intensity, which was associated with internalizing symptoms and was also significantly correlated with NA-PA polarization (see Table S2 in Supplemental Material). This finding is consistent with the dynamic-affect model (Zautra et al., 1997), which suggests that high emotional intensity contributes to within-person increases in affective polarization. This result further underscores the importance of modeling different affective dynamics in the multivariate context to account for the dependencies between different dynamic features.

Our results for externalizing disorders are more novel and distinct than those for internalizing symptoms. Our study is the first to demonstrate a negative association between externalizing disorders and PA inertia, which suggests that increases in positive emotions are more fleeting in young women with more externalizing pathology. Furthermore, when both internalizing and externalizing disorders were included in the same model, externalizing disorders were associated with greater PA variability, whereas internalizing disorders were associated with less PA variability. Although higher PA variability was associated with several maladaptive constructs in previous studies (Houben et al., 2015), these findings were mixed, and some studies suggested that PA variability is adaptive (Spindler et al., 2016). Our results suggest that heightened PA variability is uniquely associated with externalizing problems in a multivariate model that accounts for NA and various affective dynamics, but only after controlling for internalizing disorders, which are associated with the opposite pattern (i.e., less PA variability).

Greater variability in positive affect among individuals with externalizing disorders could potentially reflect hypersensitivity to reward cues. This pattern is consistent with behavioral studies (Murray, Waller, & Hyde, 2018) and Gray’s (1982, 1987) reinforcement-sensitivity theory, which claims externalizing disorders are characterized by heighted activation of the behavioral-activity system combined with hypoactivation of the behavioral-inhibition system. Given our observation in the current study of more fleeting PA among people with externalizing disorders, it is plausible to speculate that the transience of positive emotional experiences in these individuals may lead to increased sensation-seeking behaviors to alleviate boredom or boost PA (Cloninger, 1987; Zuckerman, 1996). Thus, externalizing disorders may be characterized by a sensitive behavioral-activity system that is prone to activation but once activated, may quickly become deactivated, which leads to the repetitive reward-seeking behaviors characteristic of externalizing disorders, such as substance use or impulsive behavior. These interpretations are speculative, however, and should be further explored in future work.

Some noteworthy patterns of covariation emerged from the CFA of individual disorders. First, BPD did not load on the externalizing factor, which diverges from several prior studies (Eaton et al., 2011; Kotov et al., 2017; Krueger et al., 2001). However, some studies found that BPD was more strongly associated with externalizing disorders in men than in women (James & Taylor, 2008; Johnson et al., 2003). Hence, the exclusive loading of BPD on the internalizing factor may be due in part to the all-female sample. Furthermore, even when BPD does cross-load on both dimensions, internalizing problems typically account for more BPD variance than externalizing problems (Eaton et al., 2011). Nonetheless, the factor loading of BPD on the internalizing factor in the current study was unusually high (.96), which indicates that internalizing disorders accounted for 92% of the variance in BPD. In addition, we did not expect PTSD to load on both internalizing and externalizing disorders, although this result is consistent with prior findings of internalizing and externalizing posttraumatic stress disorder subtypes (M. W. Miller & Resick, 2007; Wolf, Miller, Harrington, & Reardon, 2012).

Strengths of this study include the intensive repeated assessment of affect (i.e., seven times daily for 21 days) and the use of advanced statistical tools to model several affective dynamics simultaneously, which allowed us to control for covariation among different dynamic features of affect. Another important improvement from prior literature was our ability to model innovation variance. This approach takes inertia and other affective dynamics into account to assess variability of affect from one occasion to the next. In addition, whereas most previous work ignored differences in time lags between repeated assessments, we addressed unequally spaced observations explicitly in the model and can thus draw conclusions about a specific time interval of 1 hr.

Another strength and innovative aspect of the current study is our examination of how higher-order dimensions of psychopathology in a transdiagnostic at-risk sample relate to multiple affective dynamics. Comorbidity in psychiatric diagnoses is the norm, rather than the exception (Kessler, Chiu, Demler, Merikangas, & Walters, 2005), and recent innovations in the field support the validity of transdiagnostic approaches to psychiatric concerns (cf. Hierarchical Taxonomy of Psychopathology, Kotov et al., 2017, and NIMH Research Domain Criteria, Krueger & DeYoung, 2016). We revealed new findings on how higher-order dimensions of psychopathology relate to affective dynamics, and these novel results represent a significant extension of the prior disorder-specific dynamics literature. For instance, this is the first study to suggest that internalizing and externalizing disorders are characterized by distinct affective dynamics and may therefore benefit from distinct interventions. This project is further consistent with the goals of the Research Domain Criteria framework in its use of assessment techniques at multiple levels of assessment (e.g., self-report, behavior). A better understanding of how emotional signatures differ or remain consistent across diagnostic boundaries is essential for developing more effective interventions that target transdiagnostic affective processes. However, it should be noted that the effects observed here were small in magnitude, and hence, their clinical utility is not yet clear. An important direction for future research is to evaluate how changes in affective dynamics relate to longitudinal changes in clinical outcomes. We comment on this further below.

Although this study had several notable strengths, several limitations should be discussed. First, most prior studies examining the structure of psychopathology have done so in large, mixed-gender epidemiological samples, which differ markedly from the current at-risk community sample of 156 women in a restricted age range (18–24 years). All participants were selected for the current study on the basis of clinically significant symptoms (i.e., aggressive or self-harm-related urges or behaviors), and hence, the sample included few people in the healthy range of functioning. A large proportion of the sample was also socioeconomically disadvantaged, identified as minority race, and reported substantial exposure to violence in their homes and communities, all factors that are social determinants of poorer health, both psychological and medical. Thus, although the sample was drawn from an understudied population at high risk for difficulties with emotion regulation and both internalizing and externalizing disorders, results should be interpreted in the context of these characteristics of the study sample and may not necessarily generalize to men, older adults, healthier community participants, or individuals in clinical settings who experience less social adversity.

It is likely that our findings regarding the structure of psychopathology in this sample (which are based on CFA models) are affected by the selection criteria used in the study; it is also likely that our findings regarding affective dynamics (which are based on DSEM models) are affected by known gender- and age-related differences in emotional reactivity and affective experiences. However, there was meaningful variability in other relevant demographic characteristics (e.g., educational attainment, SES, employment status) that allow us to draw inferences across racial and SES groups. The CFA results are also limited by the use of dichotomous (rather than dimensional) scores for each diagnosis; although dimensional scores would yield greater variability in diagnostic scores, the study did not include dimensional measures for all diagnoses. This limitation notwithstanding, our application of appropriate CFA methods for categorical data and use of the dimensional scores produced by CFA for the primary DSEM analyses enabled us to investigate affective dynamics in relation to transdiagnostic dimensions of psychopathology in a way that has not been done in previous studies.

Some aspects of our study design and analytic methods also limited our ability to examine all effects of interest. For instance, given the sampling structure of the study itself, we used an interval of 1 hr to examine affective dynamics; however, other durations could be chosen, and our results may not generalize to different time intervals (e.g., several hours or days). This is an important consideration when evaluating our results against existing affective dynamics literature because we would expect affective processes to vary as a function of the time scale of assessment (Hollenstein, 2015). In addition, affect adjectives used in the study were selected to capture both negative and positive valence but were not chosen to assess differing levels of arousal. Thus, our results cannot address dynamics of affective arousal, which should be considered in future work.

We also chose not to examine dynamics of specific emotions such as anger, shame, and fear because these emotional states may be more disorder-specific than general NA and PA dynamics. For instance, a recent study comparing participants with BPD, MDD, and bipolar disorder found that inertia of shame and reactivity of shame and guilt in response to interpersonal challenges were especially characteristic of BPD (Mneimne, Fleeson, Arnold, & Furr, 2018). In addition, although we controlled for linear temporal trends (i.e., changes in NA and PA over time), the complexity of our model precluded examination of other types of cycles, nonstationarity, or contextual moderators of effects. For example, NA inertia may vary as a function of specific situational factors or affective arousal level (De Haan-Rietdijk, Gottman, Bergeman, & Hamaker, 2016), and variability may be influenced by numerous internal or external factors (e.g., drug use, interpersonal stress, sleep quality, exposure to challenging environments), which could not feasibly be modeled because of the complexity of our analyses without these additional parameters. Recent work on dynamic systems models of personality suggested that traits such as neuroticism also exhibit within-person changes over time (Sosnowska, Kuppens, De Fruyt, & Hofmans, 2019), which were not explicitly examined in the current study.

Finally, it remains unknown whether common underlying processes explain transdiagnostic similarities in affective dynamics or whether similar “emotional dynamic outcomes” may emerge from distinct multifinal pathways that vary by diagnostic presentation. Longitudinal research is needed to explicate causal relationships between psychopathology and affective dynamics and the potential mechanisms underlying these relationships. Some preliminary work (Hamaker et al., 2018) suggested that NA inertia and affective polarization prospectively predict the maintenance of depression over time, at least in young adults. Thus, more studies are needed to determine whether and how affective dynamics shape the trajectory of different forms of psychopathology over time and in response to treatment. Further research is also needed to link specific affective dynamics to near-term outcomes of high clinical importance such as suicidality, substance use, or violence. Such data would inform questions about the clinical utility of assessing affective dynamics and might aid in the development of novel real-time interventions to modify the dynamics of affect and prevent such consequences.

Findings from the present study suggest that internalizing and externalizing dimensions of psychopathology are characterized by unique affective dynamics and therefore may benefit from different interventions. For example, interventions to modulate and stabilize negative emotions and promote accessibility of positive emotions may be indicated among individuals with more prominent internalizing symptoms, whereas externalizing symptoms may be more effectively targeted by stabilizing positive affect and increasing adaptive strategies for achieving and sustaining positive emotional experiences. Nevertheless, comorbidity among internalizing and externalizing symptoms is common. Future studies with larger samples could explore potential interactions across these higher-order dimensions. Another important future direction is to identify person-specific affective dynamics that serve as proximal risk or protective factors for harmful behavior. Tracking an individual’s unique affective patterns using real-time monitoring would enhance just-in-time interventions among individuals who are at greatest risk.

Supplemental Material