Introduction
Daily fluctuations in affect are common for everyone, but individuals with recurrent mood disorders (i.e. recurrent major depressive disorder [MDD] and bipolar disorder [BD]) often experience more intense variations (e.g. Bailen & Thompson, Reference Bailen, Thompson, Dozois and Dobson2023). As individuals with recurrent mood disorders typically experience episodic mood swings, there are indications that these mood swings also happen on a moment-to-moment level. Affect dynamics (i.e. how emotions change over time) are linked to psychological well-being and psychopathology development (Bailen & Thompson, Reference Bailen, Thompson, Dozois and Dobson2023; Houben, Van Den Noortgate, & Kuppens, Reference Houben, Van Den Noortgate and Kuppens2015; Kuranova et al., Reference Kuranova, Wigman, Menne-Lothmann, Decoster, van Winkel, Delespaul, Drukker, de Hert, Derom, Thiery, Rutten, Jacobs, van Os, Oldehinkel, Booij and Wichers2021; Maciejewski et al., Reference Maciejewski, Keijsers, van Lier, Branje, Meeus and Koot2019; Sperry, Walsh, & Kwapil, Reference Sperry, Walsh and Kwapil2020; Wichers, Wigman, & Myin-Germeys, Reference Wichers, Wigman and Myin-Germeys2015). However, the bidirectional association between affect and daily life behavior, such as physical activity and smartphone use, remains underexplored. Understanding this relationship is essential for developing effective interventions and promoting resilience in individuals at risk for or currently experiencing mood disorders.
Affect dynamics are typically measured using experience sampling methods (ESM), capturing fluctuations in negative affect (NA) and positive affect (PA) across time and context (Hektner, Schmidt, & Csikszentmihalyi, Reference Hektner, Schmidt and Csikszentmihalyi2006; Larson & Csikszentmihalyi, Reference Larson and Csikszentmihalyi1983). Affect dynamics measures: (1) the magnitude of the ups and downs of affect (i.e. variability), (2) the instability of affect during a day (i.e. instability), and (3) the intrinsic resistance of an emotion to change, where one’s present affect can be anticipated by their affect at the prior moment (i.e. inertia). Research shows that individuals with MDD exhibit higher average levels, variability, instability, and inertia of NA, and lower average levels and higher variability of PA (e.g. aan het Rot, Hogenelst, & Schoevers, Reference aan het Rot, Hogenelst and Schoevers2012; Crowe et al., Reference Crowe, Daly, Delaney, Carroll and Malone2019; Hawes & Klein, Reference Hawes and Klein2024; Houben, Van Den Noortgate, & Kuppens, Reference Houben, Van Den Noortgate and Kuppens2015; Panaite, Rottenberg, & Bylsma, Reference Panaite, Rottenberg and Bylsma2020; Schoevers et al., Reference Schoevers, van Borkulo, Lamers, Servaas, Bastiaansen, Beekman, van Hemert, Smit, Penninx and Riese2021; Stapp et al., Reference Stapp, Zipunnikov, Leroux, Cui, Husky, Dey and Merikangas2023; Thompson, Bailen, & English, Reference Thompson, Bailen and English2021). These differences in affect dynamics may also be present in individuals in remission from depression (Hawes & Klein, Reference Hawes and Klein2024; Schoevers et al., Reference Schoevers, van Borkulo, Lamers, Servaas, Bastiaansen, Beekman, van Hemert, Smit, Penninx and Riese2021; Thompson, Bailen, & English, Reference Thompson, Bailen and English2021). Findings in BD are more heterogeneous (aan het Rot, Hogenelst, & Schoevers, Reference aan het Rot, Hogenelst and Schoevers2012; Faurholt-Jepsen et al., Reference Faurholt-Jepsen, Busk, Bardram, Stanislaus, Frost, Christensen, Vinberg and Kessing2023; Houben, Van Den Noortgate, & Kuppens, Reference Houben, Van Den Noortgate and Kuppens2015; Sperry & Kwapil, Reference Sperry and Kwapil2019; Sperry & Kwapil, Reference Sperry and Kwapil2022; Sperry, Walsh, & Kwapil, Reference Sperry, Walsh and Kwapil2020; Stapp et al., Reference Stapp, Zipunnikov, Leroux, Cui, Husky, Dey and Merikangas2023), likely due to differences across manic, depressive, or euthymic episodes (aan het Rot, Hogenelst, & Schoevers, Reference aan het Rot, Hogenelst and Schoevers2012; Bos et al., Reference Bos, Schreuder, George, Doornbos, Bruggeman, van der Krieke, Haarman, Wichers and Snippe2022; Scott et al., Reference Scott, Victor, Kaufman, Beeney, Byrd, Vine, Pilkonis and Stepp2020; Sperry, Walsh, & Kwapil, Reference Sperry, Walsh and Kwapil2020; Stapp et al., Reference Stapp, Zipunnikov, Leroux, Cui, Husky, Dey and Merikangas2023). Additionally, it is often challenging to determine whether elevated PA represents adaptive, functional high PA or a marker of mania, as heightened PA is a defining feature of manic episodes (American Psychiatric Association, Reference Muskin, Dickerman, Drysdale and Holderness2022).
It is essential to explore how variations in affect dynamics might influence, or be influenced by, daily life behaviors that either promote or hinder psychological well-being. Two prevalent daily life behaviors that have been essential for cultivating mental health habits are physical activity and smartphone use (Marsch, Reference Marsch2021). Research shows a link between physical activity and better mental health consistently and reduced depressive symptoms (Biddle, Reference Biddle2016; Buecker et al., Reference Buecker, Simacek, Ingwersen, Terwiel and Simonsmeier2021; Chekroud et al., Reference Chekroud, Gueorguieva, Zheutlin, Paulus, Krumholz, Krystal and Chekroud2018; Faurholt-Jepsen et al., Reference Faurholt-Jepsen, Frost, Vinberg, Christensen, Bardram and Kessing2014; Mammen & Faulkner, Reference Mammen and Faulkner2013; Ohrnberger, Fichera, & Sutton, Reference Ohrnberger, Fichera and Sutton2017; Pedrelli et al., Reference Pedrelli, Fedor, Ghandeharioun, Howe, Ionescu, Bhathena, Fisher, Cusin, Nyer, Yeung, Sangermano, Mischoulon, Alpert and Picard2020; Rohani, Faurholt-Jepsen, Kessing, & Bardram, Reference Rohani, Faurholt-Jepsen, Kessing and Bardram2018; Wang, Li, & Fan, Reference Wang, Li and Fan2019; Wu et al., Reference Wu, Jhao, Liu, Chen, Hsieh, Wang, Wu and Chien2024). A recent systematic review by Timm, Giurgiu, Ebner-Priemer, and Reichert (Reference Timm, Giurgiu, Ebner-Priemer and Reichert2024) included 66 studies that examined within-subject associations of affective well-being (e.g. higher PA and lower NA) and physical activity using methods such as ESM and accelerometry in everyday life. They showed that even brief periods of physical activity enhance affective well-being, though effect sizes vary.
Similar to physical activity, excessive smartphone use has been associated with MDD and BD depression (Faurholt-Jepsen et al., Reference Faurholt-Jepsen, Vinberg, Frost, Debel, Margrethe Christensen, Bardram and Kessing2016; Pedrelli et al., Reference Pedrelli, Fedor, Ghandeharioun, Howe, Ionescu, Bhathena, Fisher, Cusin, Nyer, Yeung, Sangermano, Mischoulon, Alpert and Picard2020; Rozgonjuk, Elhai, & Hall, Reference Rozgonjuk, Elhai, Hall, Montag and Baumeister2023; Saeb et al., Reference Saeb, Zhang, Karr, Schueller, Corden, Kording and Mohr2015; Wang, Li, & Fan, Reference Wang, Li and Fan2019). However, studies examining the causal bidirectional relationships between affect and smartphone use are scarce and show mixed results (Bradley & Howard, Reference Bradley and Howard2023; Kushlev & Leitao, Reference Kushlev and Leitao2020; Rodman et al., Reference Rodman, Burns, Cotter, Ohashi, Rich and McLaughlin2024; Xu, Bai, Wu, & Liu, Reference Xu, Bai, Wu and Liu2024).
Studying individuals at familial risk for mood disorders offers insight into early affective and behavioral patterns in relation to the onset of mood disorders. Thus far, the two studies that examined familial risk in this context showed differences in affect dynamics for those at familial risk, but these differences were only present for those with current mood symptoms (Abitante et al., Reference Abitante, Cole, Bean, Politte-Corn, Liu, Dao, Dickey, Pegg and Kujawa2024; Pavlickova, Turnbull, Myin-Germeys, & Bentall, Reference Pavlickova, Turnbull, Myin-Germeys and Bentall2015). Unaffected relatives of individuals with BD showed decreased physical activity when compared to individuals without familial risk, indicating that impaired daily life behaviors are not solely the result of BD psychopathology (Stanislaus et al., Reference Stanislaus, Vinberg, Melbye, Frost, Busk, Bardram, Kessing and Faurholt-Jepsen2020). To date, no other studies have examined the effect of familial risk status on affect and objectively measured daily life behaviors. This gap limits our understanding of whether alterations in daily affect–behavior dynamics precede the onset of mood disorders or emerge as a consequence of psychopathology. Examining these processes in individuals at familial risk may potentially provide insight into crucial mechanisms underlying vulnerability to mood disorders.
This study uses data from a longitudinal cohort of BD offspring (Helmink, Mesman, & Hillegers, Reference Helmink, Mesman and Hillegers2024) and uses ESM and Dynamic Structural Equation Modeling (DSEM) (Asparouhov, Hamaker, & Muthén, Reference Asparouhov, Hamaker and Muthén2018) to examine how familial risk and lifetime mood disorders affect the temporal dynamics between affect and behavior. We hypothesized that NA would predict increased screen-on time, as a proxy for smartphone use, and reduced step count, as a proxy for physical activity, while PA would predict the opposite. We also hypothesized that there would be individual differences in the extent to which affect and behavior predicted themselves over time (inertia) and had variability in levels after taking into account other temporal dynamics (innovation variances). We expected familial risk status and recurrent mood disorders to amplify these associations. Combining the rich longitudinal data on lifetime psychopathology in a familial high-risk sample with the moment-to-moment affect and behavior dynamics will help improve our understanding of the underlying mechanisms of intergenerational transmission of psychopathology. Understanding these effects is crucial for identifying potential risk factors and developing targeted interventions.
Methods
Participants and procedure
The current study is an observational experience sampling study (ESM) with passive sensing, conducted over a 14-day period in daily life. This study is part of the 22-year follow-up of the Dutch Bipolar Offspring Study (DBOS; Helmink, Mesman, & Hillegers, Reference Helmink, Mesman and Hillegers2024), a longitudinal cohort of offspring of parents with BD followed from mean age 16 to 38 years (range 11–43). The DBOS was established in 1997 and contains five assessments: baseline (Wals et al., Reference Wals, Hillegers, Reichart, Ormel, Nolen and Verhulst2001), 1-year follow-up (Reichart et al., Reference Reichart, Wals, Hillegers, Ormel, Nolen and Verhulst2004), 5-year follow-up (Hillegers et al., Reference Hillegers, Reichart, Wals, Verhulst, Ormel and Nolen2005), 12-year follow-up (Mesman et al., Reference Mesman, Nolen, Reichart, Wals and Hillegers2013), and 22-year follow-up (Helmink, Mesman, & Hillegers, Reference Helmink, Mesman and Hillegers2024).
A control group was recruited to match the age and background of the cohort, but excluded individuals with a personal or parental history of major psychopathology. Controls in the age range between 30 and 45 years were recruited through online advertisements, and friends and acquaintances of the DBOS participants, in order to represent a similar background. Exclusion criteria were as follows: (1) ever experienced any self-reported psychopathology (BD, MDD, psychosis, anxiety disorder, or personality disorder), or (2) a parent with BD, MDD, or psychosis.
Participants in the cohort study were paid 50 EUR for the three-part study (i.e. interview, online questionnaires, and smartphone study). The control group was paid 25 EUR for participation in the online questionnaires and the smartphone study. Ethical approval was obtained from the Erasmus Medical Centre in Rotterdam and the University Medical Centre Utrecht, the Netherlands. Written informed consent was obtained.
Materials
Lifetime psychopathology
Lifetime psychopathology of offspring was based on DSM-IV Axis I disorders derived from face-to-face semi-structured psychiatric interviews during all five time points. We used the Schedule for Affective Disorders and Schizophrenia for School-Age Children-Present and Lifetime Version (K-SADS-PL) (Kaufman et al., Reference Kaufman, Birmaher, Brent, Rao, Flynn, Moreci, Williamson and Ryan1997) when offspring were below 18 years, and the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) (First, Reference First1997) for those aged 18 and above. K-SADS-PL has high inter-rater agreement with good to excellent test–retest reliability kappa coefficients (Kaufman et al., Reference Kaufman, Birmaher, Brent, Rao, Flynn, Moreci, Williamson and Ryan1997), and the SCID-I has moderate to excellent inter-rater agreement for psychopathology (Lobbestael, Leurgans, & Arntz, Reference Lobbestael, Leurgans and Arntz2011). Diagnostic reliability was supported by the use of trained interviewers (postgraduate students, postdoctoral-level psychologists, or those with an equivalent background) and clinician review of final DSM-IV Axis I diagnoses in consensus meetings (EM, MH). Lifetime psychopathology was not assessed in the control group (exclusion criterion). Participants were grouped into: (1) recurrent mood disorders (≥2 episodes), (2) single mood episode or anxiety disorder, and (3) other/no psychopathology, based on the hierarchical staging models of BD in high-risk offspring (Duffy, Goodday, Keown-Stoneman, & Grof, Reference Duffy, Goodday, Keown-Stoneman and Grof2019).
Current symptoms of depression
Current symptoms of depression were assessed using the Inventory of Depressive Symptomatology – Self-Rated (IDS-SR) at day 0 of mobile assessment (Rush et al., Reference Rush, Giles, Schlesser, Fulton, Weissenburger and Burns1986). The IDS-SR is a self-report questionnaire assessing depressive symptom severity and has shown strong psychometric properties in previous research, including high internal consistency and good convergent validity with other depression measures (Rush et al., Reference Rush, Giles, Schlesser, Fulton, Weissenburger and Burns1986). In the current study, internal consistency was good, with Cronbach’s α = 0.86.
COVID-19 stringency
We used the COVID-19 Stringency Index to account for pandemic-related behavioral constraints (Mathieu et al., Reference Mathieu, Ritchie, Rodés-Guirao, Appel, Giattino, Hasell, Macdonald, Dattani, Beltekian, Ortiz-Ospina and Roser2020).
Mobile assessment: Experience sampling and passive sensing
We used the Avicenna Research app (previously called Ethica Data, www.avicennaresearch.com) for 14 days of ESM and passive sensing (compatible with iOS and Android). Data were collected between September 2021 and December 2021. The daily design and data processing approach of our mobile assessment can be found in Figure 1.
Daily design of the mobile assessment. Note: Green circles represent the ESM questionnaires where negative affect (NA) and positive affect (PA) were assessed. The light blue horizontal rectangle represents the continuous collection of sensor data in the background (step count and screen-on time). In our design, we binned the sensor data (blue vertical rectangles) to be able to examine temporal relationships between affect and behavior. For the creation of the DSEM models, we excluded the first and last sensor data bins and the first ESM questionnaire (as shown in lighter colors).
Experience sampling method
We used ESM to measure daily affect dynamics. During the 14 days of the mobile assessment, participants received five short surveys (±2 minutes) daily, assessing PA (i.e. content, relaxed, calm, excited, energetic, and enthusiastic) and NA (i.e. upset, irritable, nervous, anxious, guilty, sad, bored, ruminating, sluggish, tired, and insecure) on a 1- to 7-point Likert scale. The notifications were set quasi-randomly, within time blocks that were normalized toward the median time of the time block (i.e. 07:30–09:30, 10:30–12:30, 13:30–15:30, 16:30–18:30, 19:30–21:30). We calculated within and between person omega to decide whether to use a circumplex model of affect (i.e. integrating valence and arousal) (Posner, Russell, & Peterson, Reference Posner, Russell and Peterson2005). The total within-person omegas for high arousal NA and PA and low arousal NA and PA ranged between g = 0.65 and g = 0.79, and the between-person omegas ranged between g = 0.93 and g = 0.95. Given that creating scores across the circumplex did not add much information, we elected to create indices for NA and PA without separating valence and arousal.
Passive sensing
To measure daily life behavior dynamics, we used smartphone sensors to assess step count as a proxy for physical activity and screen time as a proxy for smartphone use.
Step counter: The step counter used the standard iOS and Android algorithms to sense steps taken by the participant over a certain duration. For iOS, this was every 5 minutes. For Android, as soon as step count data were available, they were sent to the Avicenna Research app. We calculated the total number of steps per time block. However, based on the Android data, it was impossible to determine with certainty when the steps had been taken during the day. Therefore, we decided to exclude all Android data from the analyses on step count (n = 61), and only include those with iOS devices (n = 67). As a result, all analyses involving step count were restricted to iOS users and should be interpreted as step count analyses within this subgroup rather than in the full sample. To examine whether this restriction may have introduced selection bias, we compared iOS and Android users on demographic, clinical, affective, and behavioral characteristics.
Screen time: Screen time was based on the state of the screen. Screen state was sensed as screen on or screen off with a start time and an end time. We calculated the total time that the screen was turned on. The content of the screen was not recorded.
Data processing
All scripts used for data processing can be found at https://osf.io/r5ycn/. To examine temporal relationships between affect and daily life behavior, we aggregated our behavior data into bins corresponding to the subsequent ESM questionnaire to avoid a double lag that the behavioral data naturally brings (e.g. we calculated the sum of steps for the time between the ESM questionnaire at time 3 and the ESM questionnaire at time 4 to line up with the emotion data from the ESM questionnaire at time 4). Because of the binning approach, it was important to know when the smartphone was actively sensing to account for missingness in sensor data. Sensor activity was calculated as the proportion of each time block for which valid screen state information was available. The raw screen state data consisted of timestamped intervals during which screen state was recorded as either on or off. Periods without recorded screen state information were treated as missing sensor data. Sensor activity, therefore, reflected sensor data coverage rather than screen on time itself. This variable was included as a covariate in all DSEM models (including the unconditional models) to adjust for differences in sensor coverage across time blocks.
Statistical analyses
All scripts used for data analysis can be found at https://osf.io/r5ycn/. Within- and between-person correlations (Spearman’s rho) between affect and behavior were estimated using the ‘psych’ package in R version 4.1.1. DSEM (Asparouhov, Hamaker, & Muthén, Reference Asparouhov, Hamaker and Muthén2018; Hamaker et al., Reference Hamaker, Asparouhov, Brose, Schmiedek and Muthén2018) was used to fit several bivariate multilevel first-order vector autoregressive (VAR [1]) models in Mplus version 8.8. DSEM uses latent mean centering to separate within- and between-person variance and allows for modeling multiple random effects and their covariances simultaneously. We used Bayesian Markov chain Monte Carlo (MCMC) estimation to estimate the DSEM models and handle missing data within each model. Model convergence was assessed using the potential scale reduction (PSR) value, with values below 1.1 considered acceptable, and by inspecting trace and autocorrelation plots (Gelman & Rubin, Reference Gelman and Rubin1992). Models were initially estimated using two chains and a minimum of 1,000 samples, with the first half of the sample counting as burn-in and thinning set to 10. To evaluate the stability of the estimates, all models were subsequently rerun with a larger number of iterations, using 10,000 iterations across two chains with thinning set to 10, resulting in 2,000 MCMC samples, of which the first 1,000 were discarded as burn-in (Hamaker, Asparouhov, & Muthén, Reference Hamaker, Asparouhov, Muthén and Hoyle2021). Final models were retained only when convergence criteria were met, and parameters showed stable trace and autocorrelation plots. When parameter instability remained, we followed recommended best practice procedures for model building and selected the most reliable final model specification (Asparouhov & Muthen, Reference Asparouhov and Muthen2022). Unconditional models were fit for each within-person variable of interest with the sensor activity variable as a covariate. This resulted in eight random effects (four per variable of interest): mean intensity, autoregressive coefficients (i.e. inertia), cross-lagged regressions, and the random logs of unique innovation variances (i.e. leftover variance in the score after considering the means and autocorrelations). To examine whether familial risk status and lifetime psychopathology predicted any of the eight random effects, we ran conditional models including covariates (age, biological sex, current symptoms of depression, COVID-19 stringency, and sensor activity). To compare groups by lifetime psychopathology, we used dummy coding with the other/no psychopathology group as the reference group. A separate conditional model was run to examine the effect of high-risk status on the random effects. We evaluated the 95% credibility intervals (CrI), which indicate the range within which the true parameter value is likely given the data. In other words, there is a 95% probability that the parameter falls within this interval. Intervals not containing zero were considered to have statistically credible differences or effects.
In sum, we tested whether familial risk status (yes or no) and a history of psychopathology (0 = other/no psychopathology, 1 = single mood episode or anxiety disorder, 2 = recurrent mood disorder) moderated the within-person cross-lagged associations. All conditional models were adjusted for age, biological sex, current symptoms of depression, COVID-19 stringency, and sensor activity. While examining the effect of familial risk status, we also included lifetime psychopathology as a covariate.
Results
Participants
Of the original DBOS cohort (n = 140), 100 individuals participated in the 22-year follow-up (71% retention), of whom 82 completed the smartphone assessment. In total, 128 participants were included in the present study (familial high risk: n = 82 [64%], control: n = 46 [36%]). The mean age of the total sample was 37.80 (SD = 3.38), high-risk offspring were slightly older, t(69) = −3.69, p < .001. A total of 73 (57%) females participated in the study. Significant group differences were found in NA (t(5133) = −4.10, p < .001), PA (t(4740) = 9.17, p < .001), and screen-on time (t(5012) = 5.22, p < .001).
A total of 30 offspring participants experienced recurrent mood episodes, 29 individuals reported a single mood episode (n = 25) or an anxiety disorder (n = 4), and 69 participants fell within the other/no psychopathology group. Participants without psychopathology were younger than individuals with recurrent mood episodes (p < .05, 95% CI = [0.16, 3.57]) and had higher step counts than individuals with mood disorders (single mood: p < .05, 95% CI = [−291.97, −13.46], recurrent mood: p < .05, 95% CI = [−375.90, −34.32]). Participants with recurrent mood disorders showed more NA and current symptoms of depression than the other two groups. All three groups differed significantly in PA (recurrent mood < single mood < other/no psychopathology), COVID-19 stringency, and sensor activity. No significant group differences were found by biological sex or screen-on time. All demographic characteristics and test results can be found in Table 1. For in-depth demographic characteristics of our offspring group, we refer to Helmink, Mesman, and Hillegers (Reference Helmink, Mesman and Hillegers2024).
Demographics
Table 1. Long description
The table is divided into three main sections: Demographics, Smartphone measures, and Covariates.
1. Demographics:
- Biological sex (female): Total 73 (57 percent). No significant difference between risk groups (p = 0.11) or psychopathology groups (p = 0.84).
- Age: Total mean 37.80. Familial risk ‘Yes’ (38.66) is significantly higher than ‘No’ (36.26) with p < .001. Recurrent mood disorder group (38.87) is significantly older than the ‘Other’ group (37.00).
2. Smartphone measures:
- Positive affect: Total mean 3.76. Familial risk ‘No’ (3.91) is higher than ‘Yes’ (3.66). Psychopathology groups show a significant decline from ‘Other’ (3.95) to ‘Single’ (3.67) to ‘Recurrent’ (3.33), all p < .001.
- Negative affect: Total mean 0.69. Recurrent mood disorder group (1.10) is significantly higher than ‘Other’ (0.59) and ‘Single’ (0.54) groups.
- Smartphone use (screen time): Total mean 36.45. Familial risk ‘No’ (39.10) is higher than ‘Yes’ (34.95), p < .001. No significant difference between psychopathology groups.
- Physical activity (steps): Total mean 1183.03. The ‘Other’ psychopathology group (1251.83) has significantly more steps than ‘Single’ (1099.12) and ‘Recurrent’ (1046.72) groups.
3. Covariates:
- Current symptoms of depression: Total mean 8.48. The Recurrent mood disorder group (14.17) is significantly higher than ‘Other’ (7.20) and ‘Single’ (5.66) groups.
- COVID-19 stringency: Total mean 40.33. Significant differences exist across all risk and psychopathology categories (p < .001).
- Sensor activity: Total mean 94.66 percent. Significant differences exist between psychopathology groups (p < .001).
Note: Comparisons were made between high-risk offspring and controls with χ2 tests for dichotomous variables, and t tests for continuous variables. Comparisons between lifetime psychopathology groups were made with ANOVA. Groups with different superscripts (a, b, c) differ significantly according to Tukey’s post hoc test (p < .05). †χ2 test.
As step count analyses were restricted to iOS users, comparability analyses between iOS and Android users are presented in Supplementary Material S1. At the group level, iOS users reported higher PA and lower NA than Android users, but did not differ on any of the other characteristics.
Sensor activity coverage was highly skewed, with 82.91% of time blocks showing complete sensor coverage. Time blocks with very low sensor coverage were uncommon, with 4.4% of bins showing sensor activity below 50%.
Correlations
We calculated within- and between-person zero-order correlations for all variables in this study (see Supplementary Material S2). At the between-person level, step count was negatively correlated with NA (moderate effect size), but all other correlations between affect and behavior were non-significant. Within-person correlations showed a positive association between step count and PA. All other within-person correlations between affect and behavior were not significant. All model code and outputs can be found online (https://osf.io/r5ycn/).
Temporal relationships between affect and behavior (unconditional models)
Step count
When examining affect in relation to step count, there was a bidirectional temporal relationship between step count and PA (see Figure 2). If someone took more steps than were usual for them, it was more likely that they experienced more PA than was usual for them at the next time point, and vice versa. No temporal relationship between step count and NA was found. Predictors accounted for 18% of the variance in both PA and NA at the next time point, whereas predictors in the PA model explained 15% and predictors in the NA model accounted for 17% of the variance of step count at the next time point.
Unconditional DSEM models investigating the relationship between affect and step count and screen-on time. Note: Credible random effects are presented as solid black lines. Non-credible random effects are presented as a dashed and gray lines. Standardized estimates are shown with the upper and lower bounds of the 95% credibility interval. NA, ‘negative affect’; PA, ‘positive affect’; t, ‘time point’.
Figure 2. Long description
A multi-panel figure with four quadrants. The top row is labeled Physical activity and the bottom row is labeled Smartphone use. The left column focuses on N A (negative affect) and the right column focuses on P A (positive affect).
* Top-Left Panel: Physical activity and N A. A solid arrow from activity sub t minus 1 to activity sub t has an estimate of .10 [.06, .13]. A solid arrow from N A sub t minus 1 to N A sub t is .33 [.28, .37]. Cross-lagged dashed gray lines show non-credible effects of .00 and .01.
* Top-Right Panel: Physical activity and P A. Solid arrows show auto-regressive effects of .10 for activity and .31 for P A. Solid cross-lagged arrows show a positive relationship from P A sub t minus 1 to activity sub t (.07 [.03, .11]) and activity sub t minus 1 to P A sub t (.06 [.02, .09]).
* Bottom-Left Panel: Smartphone use and N A. Auto-regressive effects are .12 for smartphone use and .36 for N A. A solid cross-lagged arrow from smartphone use sub t minus 1 to N A sub t is .03 [.01, .06]. The reverse path is dashed and non-credible at .01.
* Bottom-Right Panel: Smartphone use and P A. Auto-regressive effects are .12 for smartphone use and .32 for P A. A solid cross-lagged arrow from P A sub t minus 1 to smartphone use sub t is negative .05 [-.08, -.02]. The reverse path is dashed and non-credible at .00.
Each panel includes a third column representing latent variables (activity, smartphone use, N A, or P A) with curved arrows indicating variance estimates ranging from .79 to .90.
Screen-on time
The unconditional DSEM models showed a temporal dynamic between affect and screen-on time (see Figure 2). Experiencing more PA than was usual was associated with decreased screen-on time at the next time point, but not vice versa. This relationship was the opposite for NA: If someone spent more time on their phone than was usual for them, it was more likely that they would experience more NA than was usual for them at the next time point, but not vice versa. These results indicate that affect and behavior influence each other in the expected direction, although the relationship was not bidirectional. Predictors accounted for 10% of the variance in both PA and NA at the next time point, whereas predictors in the PA model explained 18% and predictors in the NA model explained 21% of the variance in screen-on time at the next time point.
Inertia and innovation variance
All variables showed credible inertia paths. This indicates that when someone spent more time on their phone, walked more, or experienced more NA or PA than was usual for them, they were more likely to show similarly elevated behavior and affect at the next time point. In other words, behavior and affect tended to persist over time rather than changing completely from one moment to the next. In addition, there was substantial unexplained variability from one moment to the next, indicating that affect and behavior also fluctuated over time beyond what could be predicted from the previous measurement alone (i.e. credible innovation variance).
Effects of lifetime psychopathology and familial high-risk status (conditional models)
Familial risk status
For the main effects, being at familial risk affected the innovation variance of screen time (see Table 2). This means that those at familial risk showed more stable patterns of phone use over time, with fewer unexpected fluctuations from one measurement to the next. While credible, the 95% credibility interval approached zero, highlighting that this effect should be interpreted with caution and needs replication. None of the other main effects differed credibly between individuals at high familial risk and individuals at low familial risk. For the interaction effects, we did not find any credible differences in affect–behavior relationships between individuals with high and low familial risk.
Results of between-subject predictors of random effects for the model, including familial high-risk status
Table 2. Long description
The table is divided into four primary sections based on affect and activity.
1. Negative affect (N A) and Step count:
- N A mean intensity (mu N A): Estimate mean 2.37 [0.81, 4.14]; High-risk -0.10 [-0.28, 0.10].
- Steps mean intensity (mu S T E P S): Estimate mean 3.99 [0.90, 6.73]; High-risk 0.03 [-0.20, 0.24].
- N A autoregression (phi N A N A): Estimate mean 0.89 [0.12, 1.64]; High-risk 0.04 [-0.27, 0.34].
- Steps innovation variance (log(pi S T E P S)): Estimate mean 2.57 [0.78, 4.37]; High-risk 0.02 [-0.20, 0.23].
2. Negative affect (N A) and Screen-on time:
- N A mean intensity (mu N A): Estimate mean 1.53 [0.51, 2.59]; High-risk 0.02 [-0.09, 0.14].
- Screen time mean intensity (mu S C R E E N): Estimate mean 32.14 [4.56, 59.55]; High-risk -0.13 [-0.27, 0.01].
- N A autoregression (phi N A N A): Estimate mean 0.81 [0.28, 1.31]; High-risk 0.17 [-0.01, 0.33].
- Screen time autoregression (phi S C S C): Estimate mean 0.52 [0.08, 0.93]; High-risk 0.03 [-0.17, 0.22].
- Screen time innovation variance (log(pi S C R E E N)): Estimate mean 6.67 [5.23, 8.10]; High-risk -0.08 [-0.22, 0.06].
3. Positive affect (P A) and Step count:
- Steps mean intensity (mu S T E P S): Estimate mean 3.93 [0.92, 6.76]; High-risk 0.02 [-0.19, 0.24].
- Steps innovation variance (log(pi S T E P S)): Estimate mean 2.61 [0.86, 4.42]; High-risk 0.03 [-0.19, 0.23].
4. Positive affect (P A) and Screen-on time:
- P A mean intensity (mu P A): Estimate mean 2.38 [1.01, 3.81]; High-risk -0.04 [-0.20, 0.11].
- Screen time mean intensity (mu S C R E E N): Estimate mean 31.03 [2.87, 58.76]; High-risk -0.17 [-0.33, 0.00].
- Screen time autoregression (phi S C S C): Estimate mean 0.55 [0.12, 0.97]; High-risk -0.08 [-0.30, 0.18].
- Screen time innovation variance (log(pi S C R E E N)): Estimate mean 6.67 [5.32, 8.04]; High-risk -0.17 [-0.32, -0.00].
Note: Bolded values in the table indicate credible estimates where the 95% C I does not cross zero.
Note: Bold estimates are credible estimates.
Lifetime psychopathology
Regarding main effects, when compared to the other/no psychopathology group, both the single mood and recurrent mood groups showed more inertia of NA within the screen-on time model (i.e. NA tended to last longer over time) and lower mean PA (see Table 3). Moreover, the single mood group showed higher average NA and NA leftover fluctuations than the other/no psychopathology group (i.e. more moment-to-moment changes in NA not explained by previous levels), but no difference was found when comparing the recurrent mood group with the other/no psychopathology group. None of the cross-lagged associations was affected by lifetime psychopathology.
Results of between-subject predictors of random effects for the model, including a history of mood disorders
Table 3. Long description
The table is divided into four main sections.
1. Negative affect and Step count:
- N A mean intensity (mu N A): Estimate mean 2.64 [1.09, 4.33]; Single mood 0.16 [0.01, 0.30]; Recurrent mood 0.07 [-0.10, 0.24].
- Steps mean intensity (mu S T E P S): Estimate mean 3.90 [0.94, 6.58]; Single mood -0.07 [-0.24, 0.10]; Recurrent mood -0.11 [-0.30, 0.08].
- N A autoregression (phi N A N A): Estimate mean 0.89 [0.12, 1.61]; Single mood 0.25 [-0.00, 0.46].
- Steps innovation variance (log(pi S T E P S)): Estimate mean 2.55 [0.80, 4.27].
2. Negative affect and Screen-on time:
- N A mean intensity (mu N A): Estimate mean 1.55 [0.47, 2.62].
- Screen time mean intensity (mu S C R E E N): Estimate mean 37.48 [9.08, 65.91].
- N A autoregression (phi N A N A): Estimate mean 0.78 [0.29, 1.32]; Single mood 0.18 [0.00, 0.33]; Recurrent mood 0.23 [0.04, 0.40].
- Screen time innovation variance (log(pi S C R E E N)): Estimate mean 7.01 [5.67, 8.39].
3. Positive affect and Step count:
- P A mean intensity (mu P A): Estimate mean 1.81 [-0.37, 3.87]; Single mood -0.26 [-0.40, -0.11]; Recurrent mood -0.21 [-0.37, -0.03].
- Steps mean intensity (mu S T E P S): Estimate mean 3.86 [0.95, 6.63].
- Steps innovation variance (log(pi S T E P S)): Estimate mean 2.58 [0.84, 4.37].
4. Positive affect and Screen-on time:
- P A mean intensity (mu P A): Estimate mean 2.40 [1.07, 3.79]; Single mood -0.17 [-0.29, -0.05]; Recurrent mood -0.14 [-0.27, -0.01].
- Screen time mean intensity (mu S C R E E N): Estimate mean 37.05 [8.82, 64.88].
- Screen time autoregression (phi S C S C): Estimate mean 0.58 [0.16, 0.99].
- Screen time innovation variance (log(pi S C R E E N)): Estimate mean 7.04 [5.65, 8.39].
Note: Bold estimates are credible estimates.
Discussion
This was the first study, to our knowledge, that examined the directionality between affect and daily life behavior and the effect of familial risk status and (recurrent) mood disorders on these temporal relationships. Regardless of risk status and psychopathology, individuals showed credible innovation variance and inertia in their affect and daily life behavior. In line with our hypotheses, step count and PA showed a positive bidirectional temporal relationship, but no bidirectional relationship emerged between step count and NA. Moreover, more screen-on time predicted later increased NA, and more PA predicted less screen-on time at the next time point. Contrary to our hypotheses, familial risk status and recurrent mood disorders did not amplify the bidirectional relationships, but some main effects were affected. These results emphasize the unique combination of self-reported affect with objective daily life behavior data and their directionality.
Our results showed credible affect and behavior fluctuations, and they influence each other and themselves throughout the day. However, these effects can vary in direction and strength, and they tend to be small. These results corroborate already existing research on the bidirectional effect of physical activity and PA (Schultchen et al., Reference Schultchen, Reichenberger, Mittl, Weh, Smyth, Blechert and Pollatos2019; Timm, Giurgiu, Ebner-Priemer, & Reichert, Reference Timm, Giurgiu, Ebner-Priemer and Reichert2024), the effect of PA on subsequent physical activity (Cameron, Bertenshaw, & Sheeran, Reference Cameron, Bertenshaw and Sheeran2018; Forster et al., Reference Forster, Richards, Foli, McGowan, Hass, Becker and Miller2021; Niermann et al., Reference Niermann, Herrmann, von Haaren, van Kann and Woll2016; Pannicke et al., Reference Pannicke, Reichenberger, Schultchen, Pollatos and Blechert2020; Peterson et al., Reference Peterson, Charlson, Hoffman, Wells, Wong, Hollenberg, Jobe, Boschert, Isen and Allegrante2012), and the effect of physical activity on subsequent PA (Dunton et al., Reference Dunton, Huh, Leventhal, Riggs, Hedeker, Spruijt-Metz and Pentz2014; Lathia, Sandstrom, Mascolo, & Rentfrow, Reference Lathia, Sandstrom, Mascolo and Rentfrow2017; Wichers et al., Reference Wichers, Peeters, Rutten, Jacobs, Derom, Thiery, Delespaul and van Os2012). A potential biological pathway connecting physical activity and PA could resemble the effect of exercising, where neurotransmitter level changes occur, including increases in endorphins, dopamine, and serotonin (e.g. Meeusen & De Meirleir, Reference Meeusen and De Meirleir1995; Ruscheweyh et al., Reference Ruscheweyh, Willemer, Krüger, Duning, Warnecke, Sommer, Völker, Ho, Mooren, Knecht and Flöel2011; Sutoo & Akiyama, Reference Sutoo and Akiyama2003), which in turn increase PA. Psychologically, experiencing more PA might encourage exploratory behaviors and active coping strategies (Kemner et al., Reference Kemner, Mesman, Nolen, Eijckemans and Hillegers2015), such as engaging in physical activities, which can subsequently enhance future emotional states. This would be in line with the broaden-and-build theory, which states that positive emotions broaden one’s momentary thought-action repertoire (Fredrickson, Reference Fredrickson2004).
Our findings also suggest that experiencing PA correlates with a subsequent reduction in screen-on time, while increased screen-on time is associated with a subsequent rise in NA. This pattern suggests the potential for PA to mitigate engagement with digital devices, possibly by fulfilling emotional needs that reduce the reliance on phone interactions (Blumler, Reference Blumler1979; Kardefelt-Winther, Reference Kardefelt-Winther2014; Unanue, Martela, Vignoles, & Dittmar, Reference Unanue, Martela, Vignoles and Dittmar2024; Van de Casteele et al., Reference Van de Casteele, Soenens, Ponnet, Perneel, Flamant and Vansteenkiste2024; Yang et al., Reference Yang, Ma, Zhang, Xue and Hu2023). Conversely, excessive smartphone use may contribute to NA through several causal pathways, such as social comparison (Festinger, Reference Festinger1954; Midgley et al., Reference Midgley, Thai, Lockwood, Kovacheff and Page-Gould2021; Steers, Wickham, & Acitelli, Reference Steers, Wickham and Acitelli2014); information overload (Arnold, Goldschmitt, & Rigotti, Reference Arnold, Goldschmitt and Rigotti2023; Matthes, Karsay, Schmuck, & Stevic, Reference Matthes, Karsay, Schmuck and Stevic2020); and the disruption of real-world social interactions (Karakose, Yıldırım, Tülübaş, & Kardas, Reference Karakose, Yıldırım, Tülübaş and Kardas2023; Peper & Harvey, Reference Peper and Harvey2018; Yang et al., Reference Yang, Guo, Tao, Meng, Wang, Li, Zhang, Zeng, Tang, Wang, Deng, Zhao, Ma, Li, Xu, Li, Liu, Tang, du and Li2022).
We did not observe an effect of familial risk status or recurrent mood disorders on any of the bidirectional relationships between affect and daily life behavior. A potential explanation for this could be our grouping of those with recurrent MDD and BD into one group due to sample size limitations. There are indications that those with MDD show a different profile in daily affect dynamics than those with BD, for example, high PA is typically seen as healthy and adaptive for those with MDD, whereas for those with BD, it may overlap with (hypo)mania (American Psychiatric Association, Reference Muskin, Dickerman, Drysdale and Holderness2022), blurring the lines between well-being and early signs of a manic mood episode. However, the group of individuals with a history of mood disorders showed higher levels of NA and lower levels of PA compared to individuals without mood psychopathology. Additionally, our results might have been affected by the COVID-19 pandemic. COVID-19 safety measures could constrain individuals’ behavior to be more similar to that of others. Moreover, due to the pandemic, everyone in our study became at risk for psychopathology, not just those with familial risk (e.g. Murphy et al., Reference Murphy, Svob, Van Dijk, Gameroff, Skipper, Abraham, Yangchen, Posner, Shankman, Wickramaratne, Weissman and Talati2023; Pierce et al., Reference Pierce, McManus, Hope, Hotopf, Ford, Hatch, John, Kontopantelis, Webb, Wessely and Abel2021; Witteveen et al., Reference Witteveen, Young, Cuijpers, Ayuso-Mateos, Barbui, Bertolini, Cabello, Cadorin, Downes, Franzoi, Gasior, Gray, Melchior, van Ommeren, Palantza, Purgato, van der Waerden, Wang and Sijbrandij2023).
The findings of this study carry significant lifestyle and clinical implications. Understanding the temporal dynamics and directional influences between self-reported affect and objective daily life behavior can inform targeted interventions. Our results demonstrate that increased screen-on time subsequently elevates NA, while heightened PA reduces screen-on time. Interventions enhancing PA, such as Positive Affect Treatment (Craske et al., Reference Craske, Meuret, Ritz, Treanor and Dour2016), CBT (Whelen & Strunk, Reference Whelen and Strunk2021), or mindfulness (van Agteren et al., Reference van Agteren, Iasiello, Lo, Bartholomaeus, Kopsaftis, Carey and Kyrios2021), may be more effective than those focused solely on reducing screen time. Future studies should explore how we can actively influence the relationship between affect and behavior for treatment purposes, to promote emotional and behavioral resilience.
These results should be interpreted with the following limitations in mind. First, participants used their own smartphones, including multiple platforms. This could cause heterogeneity in the data that was not accounted for. Relatedly, all step count analyses were restricted to iOS users, because step timing could not be determined reliably for Android users. Although iOS and Android users were broadly comparable, they differed in PA and NA at the group level. Therefore, findings involving step count should not be generalized to the full sample without caution. Second, using step counts from a phone to infer physical activity limits our understanding. Participants could carry their phones in different places (e.g. their hands, pocket, bag, or not at all) and are not likely to carry them during physical activities such as swimming or team sports. Third, the screen state was also activated when people received a notification that turned the screen on. For people who received many notifications, the screen state might be inflated. Fourth, even though we had enough power to run the analyses, we were unable to examine the effect of more informative psychopathology groups (e.g. solely individuals with BD, rather than grouping BD with recurrent MDD) due to a small sample size. Fifth, our sample might be less representative, as it predominantly consists of White individuals with a high socioeconomic status. Moreover, our control group is not representative of the community, as none of the individuals have ever experienced any form of psychopathology. Lastly, although DSEM allowed us to examine temporal ordering, the associations may still reflect unmeasured confounding, such as sleep, stress, social interactions, medication use, or even weather. Findings should be interpreted as within-person predictive associations rather than causal effects.
In conclusion, this study is among the first to investigate the directionality between affect and daily life behaviors, as well as the influence of familial risk and recurrent mood disorders on these relationships. Consistent with prior research, a positive bidirectional link emerged between step count, as a proxy for physical activity, and PA. Moreover, increased PA predicted reduced screen-on time, while excessive screen-on time led to a subsequent increase in NA. Although familial risk and psychopathology did not influence cross-lagged associations, our results highlight a key role of PA in affect–behavior dynamics. These insights could pave the way for interventions that target affect and behavior patterns, ultimately promoting resilience and well-being in those at risk for and experiencing recurrent mood disorders.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726105297.
Data availability statement
All code for data processing and analyses can be found at https://osf.io/r5ycn/. The data are not publicly available due to privacy and ethical restrictions, but can be shared upon reasonable request to the corresponding author.
Acknowledgements
We thank all of our participants, field workers, and students over the past 22 years for their help in collecting these data. Illustrations in our figures were designed by Freepik.
Author contribution
Conceptualization: F.G.L.H., M.H.J.H., E.M., and S.H.S.; Formal analysis: F.G.L.H. and S.H.S.; Investigation: F.G.L.H. and E.M.; Methodology: F.G.L.H., E.v.R., M.H.J.H., E.M., and S.H.S.; Writing – original draft preparation: F.G.L.H.; Writing – review and editing: F.G.L.H., E.v.R., M.H.J.H., E.M., and S.H.S.
Funding statement
This work was funded by The Netherlands Organization for Health Research and Development (ZonMw) with project number 60-63600-98-774 (financed by the Dutch Research Council [NWO]), grant 024.005.010 ‘Stress in Action’ by NWO gravitation, grant 22963 from the 7FP of the European Commission, grant 9120818 from the NWO, the Stanley Medical Research Institute, the Fulbright Commission in the Netherlands, and the Prins Bernhard Cultuurfonds. Open access funding provided by Erasmus MC - Sophia Children’s Hospital.
Competing interests
The authors declare that there are no conflicts of interest with respect to the authorship or publication of this article.
Ethical standard
The 22-year follow-up was approved by the Medical Ethics Committee of the Erasmus Medical Centre in Rotterdam, the Netherlands. Former study waves were approved by the University Medical Centre Utrecht, the Netherlands. Written informed consent was obtained. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
