Disentangling the Role of Affect in the Evolution of Depressive Complaints Using Complex Dynamical Networks

The data were obtained through the Boston College daily sleep and well-being survey (Cunningham et al., 2021). The study was set-up during the first wave of COVID-19 (March 20^th 2020 until August 5^th 2020) and participants were recruited online. All English-speaking individuals older than 18 were eligible to participate in the study, resulting in N=1,518 enrolled participants (mean±SD age 35.2±15.1 years old, range 18-90 years old). The participants provided informed consent, and the study received ethical approval from the Institutional Review Board at Boston College. More details on the study and recruitment can be found in Cunningham et al., 2021.

The study started with a demographic survey, and upon completion, participants received daily surveys on their sleep and well-being. The daily surveys were divided into a short and a full version, where the full version included additional questions containing validated assessments of mood (Positive and Negative Affect Schedule [PANAS]; Watson et al., 1988) and depressive complaints (Patient Health Questionnaire-9 [PHQ-9]; Kroenke et al., 2001). The full version was sent on the first three days of the study enrollment. After enrollment, the long survey was sent on two randomly selected days of the week, and the short version was sent on the five remaining days. For a more detailed description of the assessments, we refer to Cunningham et al., 2021.

For the current study, we include the demographics survey, PANAS items, and PHQ-9 assessments from the full version of the questionnaire.

The PANAS is a 20-item questionnaire on the experience of positive (e.g., enthusiastic) and negative (e.g., scared) affect rated on a five-point Likert scale ranging from 1 (“very slightly/ not at all”) to 5 (“extremely”) (Watson et al., 1988). To reduce the number of variables for more power within the conducted statistical analyses (i.e., the estimated network models described in the “Statistical Analyses” section of the current paper), we selected the ten items of the PANAS that have been validated in the short-form (Mackinnon et al., 1999). The reliability of both PA and NA short-form scales is high, with a Cronbach alpha of .78 for the PA scale and .87 for the NA scale. For the assessment of PA, items include: inspired, alert, excited, enthusiastic, and determined, and for NA, items include: afraid, upset, nervous, scared and distressed. In the survey, participants were explicitly asked to rate how they felt in the current moment for each of the attributes (“For each of the following attributes, indicate which description best describes how you currently feel, right now in the moment”).

The PHQ-9 is a 9-item questionnaire to measure depression severity by assessing each of the 9 DSM-IV criteria for depression on a four-point Likert scale ranging from 0 (“not at all”) to 3 (“nearly every day”). The internal reliability of the PHQ-9 is high, with a Cronbach’s alpha ranging between .86-.89 (Kroenke et al., 2001). In the current implementation, the item on suicidal thoughts was omitted. Participants were asked to rate the severity of complaints over the last several days (“In the last several days, how often have you been bothered by any of the following problems: not at all, some of the time, more than half of the time, almost all of the time”).

We selected the study period from March 20^th 2020 until June 26^th 2020, ending three days after the final full survey was sent out. Of the total number of participants enrolled in this study, N=1,355 (89.3%) completed at least one assessment during the selected study period.

The full version of the questionnaire was sent out twice a week on random selected days. Sometimes participants completed the full survey multiple times on a single day: on 79 occasions the survey was completed twice, and on two occasions the survey was completed three times. For these 81 assessments we chose the survey that was completed first. Inspecting the response rate of the full survey over the study period, a clear three-day interval pattern can be seen (see Supplement A, Figure A1-A2). Therefore, we chose to group the days into ‘measurement occasions’, defined by a three-day window (e.g., March 20^th-22^nd 2020 is measurement occasion 1). In this way, the entire study period is grouped into 33 measurement occasions of three days each. The advantage of this grouping is twofold. First, it circumvents the problem of large differences in the number of completed surveys per assessment. Second, given that the full surveys were sent out randomly twice a week, dividing the assessments into three-day intervals makes the time between two completed surveys more equidistant. In case participants completed multiple surveys within one measurement occasion, we averaged their responses. Following the recommendations for estimating a multilevel VAR model (Jordan et al., 2020), we selected participants who completed surveys for at least 20 measurement occasions, resulting in a final sample size of N=228 participants (16.8%). Thus, for each included participant, we have completed data for a minimum of 20 and a maximum of 33 measurement occasions that span 9-14 weeks.

To explore whether the evolution of depressive complaints (i.e., aggravation or alleviation of complaints) is accompanied by general trends in affect fluctuations over the same period we defined different groups of depressive evolution. First, we defined two groups of participants that experienced a clinically meaningful change in their depressive complaints, defined by a 5-point difference in their PHQ-9 total score (Lowe et al., 2004; Round et al., 2020), and indicative of an aggravation or alleviation of their complaints. Second, we defined groups based on the overall level of experienced complaints throughout the studied time period: participants who are consistently without depressive complaints, defined as a PHQ-9 score of four or lower on all assessments; participants who experience occasional depressive complaints, defined as a PHQ-9 score higher than four on at least one assessment (and lower than four on at least one other assessment); and participants who experience consistent depressive complaints, defined as a PHQ-9 score consistently higher than four on all assessments (Kroenke et al., 2001).

To visually explore whether general trends can be observed, we plotted the smoothed means of each affect over time using locally estimated scatterplot smoothing (i.e., a loess curve). We did this for all participants together as well as for each of the defined groups to explore whether trends in affect trajectories accompany differences in evolution of depressive complaints. We used loess regression to facilitate the exploration of general trends. Loess regression is a non-parametric method that fits least squares regressions in localized subsets of the data (Cleveland, 1979). The amount of smoothing that is applied depends on the number of data points that are used in each local regression (i.e., the neighborhood) and is controlled by setting the smoothing parameter α between 0 and 1. The larger the values for α, the more data points are being selected in the neighborhood (i.e., nα data points are selected, where n represents the total number of datapoints). More datapoints in the local regression results in smoother functions, that are more robust to fluctuations in the data. We set the smoothing parameter to α = 0.2 in order to aid the visualization of patterns in the data, without losing sensitivity of fluctuations in the data.

To investigate the interplay between depressive complaints and affect directly we estimated a multi-level network model including both affect and depressive complaints. To estimate the relations while taking the longitudinal structure of the data into account, we estimated a two-step multi-level GVAR model as implemented in the mlVAR package (Epskamp et al., 2017). A multi-level GVAR has two major benefits: (1) a single model can be estimated, leading to an adequately powered analysis, and (2) the within-person effects can be separated from between-person effects (Epskamp, van Borkulo, et al., 2018).

The edges in the multi-level GVAR network are computed from partial correlations, meaning they portray the unique association among two variables after controlling for all other variables in the network. Edges can be positive or negative, indicating the corresponding nature of the associations between the nodes (Epskamp, Borsboom, et al., 2018).

The mlVAR package estimates three types of network structures: (a) a temporal network (both a fixed-effect structure over all persons and a random-effect structure per person), (b) a contemporaneous network (both a fixed-effect structure over all persons and a random-effect structure per person), and (c) a between-persons network (Epskamp, van Borkulo, et al., 2018). In line with Epskamp, Waldorp, et al. (2018), for the remainder of this paper, the term within-person will refer to the fixed-effect within-person network (either temporal or contemporaneous) and the term person-specific will refer to random-effects within-person networks (either temporal or contemporaneous).

The temporal network indicates how well a variable predicts another variable at the next time point while controlling for all variables at the current time point. For example, a direct association from the depressive complaint ‘trouble concentrating’ to ‘feeling distressed’ indicates that having ‘trouble concentrating’ now predicts ‘feeling distressed’ at the next time point, taking into account all other current affects and depressive complaints. In the contemporaneous network we controlled for all these temporal effects, which shows the unique association among variables within the same time window. For example, a direct and positive edge between ‘feeling distressed’ and ‘feeling afraid’ indicates that these two negative affects are positively associated after removing the lagged effects. Finally, the between-persons network shows the relationships among the means of persons in the data. For example, a positive edge between ‘trouble concentrating’ and ‘feeling distressed’ would indicate that persons who have, on average, more trouble concentrating, also, on average, feel more distressed.

Based on the theorized role of affect dynamics in the course of depressive complaints, we expected the network structure to differ between people who experienced a meaningful change in their depressive complaints (either aggravation or alleviation) and people who did not experience such a meaningful change. Specifically, we expected that more densely connected networks result in more depressive complaints over time. To explore this hypothesis we computed, for each participant, the average absolute strength of their temporal associations (i.e., their density; Oreel et al., 2019). Subsequently, we correlated their density to the (absolute) maximum change in their PHQ-9 score over the study period: the difference between an individual’s maximum PHQ-9 score, and their minimum PHQ-9 score over the course of 4 months. To compute the maximum difference, we took the order in which the scores occur into account, to capture whether an aggravation or alleviation was experienced. For example, two individuals A (PHQ-9 scores: 3,7,6,8,6) and B (7,8,4,4,3) who both have a minimum PHQ-9 score of 3 and a maximum PHQ-9 score of 8 but differ in their PHQ-9 difference score reflecting their aggravation (A: maximum difference 3-8=-5) or alleviation (B: maximum difference 8-3=5). The PHQ-9 difference score indicates the maximum change in depressive complaints this individual reported over the study period. In addition, we inspected the relation between network density and change in PHQ-9 regardless of an aggravation or alleviation in depressive complaints by correlating person-specific network density to an individual’s absolute maximum change in PHQ-9.

To examine the extent to which the observed correlation between person-specific temporal network density and their change in PHQ-9 score could be dependent on methodological choices we conducted two sensitivity checks. First, we determined whether potential trends in the data could have affected our observed correlation. We checked for trends in the variables and detrended any variables with significant trends. Then we followed the same procedures and computed the correlation. Second, since the network includes the PHQ-9 items we examined if the correlation between network density and maximum change in PHQ-9 score might be driven by this overlap. Given this overlap, the temporal density is in part based on the same information (i.e., the PHQ-9 items) as the change score in PHQ-9 items. To that end, we re-estimated a mlVAR network including only the affect states as nodes and correlated the person-specific temporal network density of the affect states with participants’ change in PHQ-9 scores.

To examine the stability of the correlation between person-specific temporal network density and change in PHQ-9 score we followed a similar procedure as described in Jongeneel et al. (2020). We re-estimated 100 mlVAR networks that included a random selection of 80% of the original data (i.e., using the data of ~182 participants). For each of these re-estimated networks we followed the same procedure as described above, (i.e., we calculated person-specific network density and correlated the network density to change in PHQ-9 score).

All analyses were performed in R (version 4.0.5) using the packages ‘ggplot2’ (version 3.4.0), ‘mlVAR’ (version 0.5), ‘qgraph’ (version 1.9.3), ‘dplyr’ (version 1.0.10), ‘nnet’ (version 7.3-18), and ‘ggpubr’ (version 0.5.0). The derived data used for analyses and corresponding code can be found on the Open Science Framework (OSF): https://osf.io/2zh4f/.

We included 228 participants that completed assessments on at least 20 measurement occasions within March 20^th 2020 and June 26^th 2020. Over the course of four months, on average, participants completed 23±3 (mean±SD) assessments. On average, 9±3(mean±SD) assessments were missing, thus, on average, 28% waves of data are missing per individual. As responses for all variables of interest where forced, there is no within wave missingness. For missing data patterns per participant, see Figure A3 in Supplement A. The majority of the participants were female (n =186, 81.58%) and, on average, 45±19 (mean±SD) years old. For more details on sample characterization, see Table 1. Table 2 shows the means, standard deviations, and range for each variable used in this study. Our sample (n=228) differed from the original full sample (n=1355) in their mean PHQ-9 score at baseline; the original full sample had a higher PHQ-9 mean at baseline. Our sample did not differ in their PANAS scores at baseline from the original full sample (n=1355), see Supplement B, Table B1 for more details on sample characterization, means standard deviations, and range for the PHQ-9 and PANAS for the original full sample.

Figure 1(a) shows the affect trajectories across the 33 measurement occasions for all participants concertedly. Averaging over all participants shows that PA is generally rated somewhat higher than NA, which is consistent over the entire study period.

We inspected the affect trajectories for participants based on (i) their change in depressive complaints over time, and (ii) the severity of the experienced depressive complaints. Over the studied time period, 116 participants experienced a meaningful change in their depressive complaints: 56/116 participants (48.3%) experienced an aggravation in their depressive complaints (i.e., an increase in PHQ-9 score of at least 5), and for 60/116 participants (51.7%) their depressive complaints alleviated (i.e., a decrease in PHQ-9 score of at least 5). Splitting the affect trajectories for each of these individuals, shown in Figure 1(b), indicates that both for participants who experienced an aggravation and an alleviation of their depressive complaints over time, on average, their affects are more intertwined. Interestingly, no marked differences are seen in participants who experienced an aggravation of their complaints compared with participants who experienced an alleviation of their complaints.

Inspecting the consistency of participants depressive complaints, we found that 50/228 participants (21.9%) were consistently without depressive complaints, 137/228 participants (60.1%) experienced depressive complaints at least occasionally, and 41/228 participants (18.0%) experienced depressive complaints consistently. Figure 1(c) shows the affect trajectories for each of these participants. Here, we see clear differences in affect trajectories across participants: in people without depressive complaints there is, on average, a clear distinction between the PA scores, which are rated relatively high, and the NA scores, which are all rated consistently low. In the people with occasional depressive complaints, the PA scores are, on average, still rated higher than the NA scores, but the distinction is less clear. Finally, in people who consistently experience depressive complaints, the ratings of positive and NA have flipped, as NA is, on average, rated higher than PA.

We investigated the dynamical relations among affect and depressive complaints by estimating a multilevel network model. Figure 2 shows the three estimated network structures: (a) the average temporal associations for all participants; (b) the average contemporaneous associations for all participants; and (c) the between-persons network structure. For the interpretation of the relations within each network structure, it is important to note that questions regarding affect and depressive complaints were prompted differently. Affect was prompted: “Indicate which description best describes how you currently feel, right now in the moment” while questions regarding depressive complaints were prompted: “In the last several days, how often have you been bothered by any of the following problems”. This difference in phrasing, as well as the temporal scaling of our study (roughly 3 days between assessments) influences the interpretation of the estimated networks. The relations in the average contemporaneous network reflect how depressive complaints over the days prior to the assessment and affect at the moment of assessment co-occur (e.g., how depressive complaints over a period of 3 days, let’s say Monday through Wednesday, co-occurs with affect on Wednesday). Relations in the average temporal network reflect how depressive complaints over the days prior to the assessment predicts affect at the next assessment occasion (e.g., how depressive complaints over a period of 3 days, for example, Monday through Wednesday, predicts affect on Friday) and vice versa. The relations in the between-person network reflect how, on average, depressive complaints relate to affect.

The average temporal network shows strong autoregressive effects for depressive complaints, indicating depressive complaints have a (relatively) strong reinforcing tendency, see the left panel of Figure 2. We found positive relations within the three domains (e.g., the positive relations between the NA variables “afraid” and “nervous”, or the positive relations between the PA variables “determined” and “enthusiastic”), while negative and positive associations can be found between the three domains (e.g., a negative relation from the PA variable “inspired” to the depressive complaint “loss of interest”). Interestingly, the temporal network shows many relations between affect states and depressive complaints (e.g., the negative relation from “afraid” to “depressed mood” or the negative relation from “excited” to “depressed mood”), whereas only one relation between PA and NA can be found (i.e., from “determined” to “afraid”).

Compared to the temporal network, the average contemporaneous network shows clearer demarcations between PA, NA, and depressive complaints: edges are comparatively stronger within than between the three domains. This can be seen in the middle panel of Figure 2, and is confirmed by inspecting the edge weights of the contemporaneous network. For example, strong positive associations can be found between NA items such as “distressed” and “upset”, and “afraid” and “scared”, as well as between PA items such as “enthusiastic” and “excited”, and between depressive complaints such as “loss of interest” and “depressed mood”. Weak associations can be found between the three domains (e.g., the negative association between “alert” and “depressed mood”).

The between-persons network portrays stronger relations between the different domains compared to the contemporaneous network, indicating that the average affect people experience is related to their average depressive complaints. As expected, we found positive associations within PA, NA, and depressive complaints, see the right panel Figure 2. In addition, we found positive associations between the NA items and depressive complaints (e.g., the association between “afraid” and “feelings of worthlessness”), and negative associations between the PA items and depressive complaints (e.g., the association between “alert” and “fatigue”). However, contradictory to our expectations, we also found negative associations between the NA items and depressive complaints (e.g., the association between “distressed” and “loss of interest”) and positive associations between PA and depressive complaints (e.g., the association between “alert” and “fatigue”).

To further investigate the relations between affects and depressive complaints, we correlated person-specific network density of the temporal network to their (absolute) maximum change in PHQ-9 score. As shown from the correlation plot in Figure 3 (right panel), there is a strong correlation (R = 0.77) between person-specific network density and absolute change in PHQ-9 score. Interestingly, and contrary to our expectations, stronger network densities relate to both a more substantial aggravation and to a more substantial alleviation in depressive complaints, as can be seen in Figure 3 (left panel) illustrating the correlation between person-specific network density and maximum change in PHQ-9 score. Network density is thus related to clinically meaningful change in both directions.

Sensitivity checks showed the strong correlation between absolute change in PHQ-9 and person-specific temporal network density was still present when data was detrended (R = 0.79). When only affect states were included in the network structure, the strength of the observed correlation decreased (R = 0.4), however we still observed a positive association between person-specific network density and absolute maximum change in PHQ-9 score.

In addition, the relation between network density to both a more substantial aggravation and a more substantial alleviation in depressive complaints was still present when data was detrended, or when only affect states were included in the network structure. See Supplement C for more details on the performed sensitivity checks.

Stability checks indicated the originally found correlation between person-specific network density and maximum change in PHQ-9 score was stable. The strong correlation between absolute maximum change in PHQ-9 and persons-specific temporal network density was still present when recursively re-estimating networks using 80% of the data at a time. The median correlation between person-specific network density and their maximum change in PHQ-9 score using 80% of the data was R = 0.77 and ranged from 0.72-0.81. In addition, we found stable results for the relation between person-specific network density and the aggravation and alleviation of depressive complaints; when looking at the correlation between maximum change in PHQ-9 and person-specific network density, we found a median correlation of R = -.04, ranging from -0.15 to 0.07. (See Figure C3 in Supplement C for the correlation plots).

Depressive complaints have often found to be associated with the regulation of affect while facing stress. To better understand the mechanisms between affect and depression, in the current paper, we set-out to investigate the dynamic interplay between affect and depressive complaints during a prolonged and eventful time imposed by the start of the COVID-19 pandemic. First, we explored whether the evolution of depressive complaints is accompanied by general trends in affect fluctuations using a longitudinal dataset from the early phase of the COVID-19 pandemic. We divided the sample into groups of people who experienced a clinically meaningful change during the studied period (i.e., a substantive aggravation or alleviation of complaints), and groups whose scores can consistently be classified as no complaints, mild complaints, or severe complaints. We then visually inspected affect fluctuations of these groups. As expected, we found differences in the affect trajectories among people who experienced consistently no depressive complaints (showing higher PA than NA) compared with consistent depressive complaints (showing the reversed pattern). Crucially, these differences pertained to both PA and NA trajectories, showing that there is a clear link between consistent severity levels of depressive complaints and both positive and negative affect. Contrary to our expectations, affect trajectories were similar for people experiencing either an aggravation or alleviation of depressive complaints, meaning that we reject our hypothesis that people who experienced an increase in depressive complaints would show high levels of NA, whereas people who reported a decline in complaints would show high levels of PA.

Second, we investigated the link between affect and depressive complaints in more detail by estimating a multilevel longitudinal network model. The network showed many and strong relations between affect and depressive complaints, indicating that these elements are indeed directly linked, both at the average intraindividual level, as indicated by the contemporaneous and temporal networks (i.e., fixed effects), and on the interindividual level as indicated by the between-person network. At the interindividual level, we found mostly positive associations within the PA and NA elements, and within the depressive complaints, as expected. We hypothesized to find positive associations between the NA elements and depressive complaints and negative associations between the PA elements and depressive complaints in the between-subjects network. This was true for some (e.g., positive association between the NA ‘feeling distressed’ and the depressive complaint ‘depressed mood’) but not all associations in the network (e.g., positive association between the PA ‘feeling inspired’ and the depressive complaint ‘concentration problems’). The average temporal and contemporaneous networks did show the expected associations, with negative (positive) associations between PA (NA) and depressive complaints.

Furthermore, we investigated the relation between the strength of the associations in the temporal person-specific networks. As expected, we found a strong relation between the density of the person-specific networks and the individuals’ absolute change in PHQ-9 score. Interestingly, similar network densities are related to both a worsening and improvement in depressive complaints. This indicates that, contrary to our expectations, strong associations between affect and depressive complaints are related to change in any direction: both to an improvement of complaints as well as to a worsening of complaints. The second part of our hypothesis should, therefore, be rejected, as the expectation was that a higher density would specifically be related to an aggravation in PHQ-9.

The strong relation between network density and change in depressive complaints may have important implications for the clinical interpretation of networks, as network density has generally been related to more severe psychopathology (e.g., see Calugi et al., 2021; van Borkulo et al., 2015). However, our study shows an alternative situation in which a larger density of networks indicates more fluctuations and potential for flexibility (Hayes et al., 2015). One possible explanation for our finding is that phase transitions in a wide variety of systems (e.g., transitioning from mild depressive complaints to severe complaints) are often characterized by a period of instability, in which the behavior of system shows many fluctuations (Olthof et al., 2020; Wichers et al., 2016). This increase in fluctuations before a clinically meaningful alleviation or aggravation could be reflected in the increased network density. This explanation possibly corroborates findings in other studies that found larger densities in networks to be related with a decrease in psychopathology symptoms over time (McElroy et al., 2019).

While the relation between network density and change in depressive complaints may signal relevant clinical importance, it is important to note that this strong relation could also merely reflect a well-known property of test reliability, namely that the variance of a total score (in our case the change in depressive complaints) consists of the sum over the variance in all items (in our case the individual affects and individual PHQ-9 items) and the sum over their covariances (Cronbach, 1951). Since denser networks reflect stronger covariances, it is a statistical necessity that denser networks are accompanied by larger variations in the total score (i.e., the variation in PHQ-9 score). However, our finding does not merely reflect a methodological artefact, as our sensitivity analysis indicate that only including the affect states in the network structure (thus, removing the PHQ-9 items from the network) still rendered a positive association between network density and absolute change in PHQ-9 score (see Supplement C for more details on the performed sensitivity checks). Additionally, it is important to note that one should be cautious to compare network densities of different types of network models. Thus, our finding cannot simply be generalized to different types of network models (e.g., non-linear models such as the Ising network models which are estimated from cross-sectional data; Cramer et al., 2016).

Some limitations in the current study warrant attention. First, the questionnaires were sent out twice a week at random intervals, thereby violating the assumption of equidistant measures for longitudinal analyses. We partly addressed this problem by defining measurement occasions as three-day periods. Second, due to the high amount of missing data in the original full sample, we selected participants that answered enough measurement occasions. The selected sample of participants who completed at least 20 measurement occasions had a lower PHQ-9 mean at baseline than the original full sample. Therefore, it can be that we may have missed more severe depression cases. At the same time, our selected sample still included 129 (56.6%) participants who scored above the cut-off of severe depressive complaints at baseline and 41 participants (18.0%) who consistently scored above the cut-off for severe depressive complaints.

Third, affect and depressive complaints may operate on different timescales. To the best of our knowledge, there are no current techniques available that can account for these possible time-scale differences in processes in the field of network analysis (Bringmann et al., 2022). However, the time-scale differences in affect and depressive complaints were reflected in the phrasing of the questionnaire. Questions regarding affect were prompted: “Indicate which description best describes how you currently feel, right now in the moment” while questions regarding depressive complaints were prompted: “In the last several days, how often have you been bothered by any of the following problems”. This difference in phrasing captures the idea that affect operates on a faster time-scale (e.g., fluctuates within a day), while depression operates on a slower time-scale (e.g., fluctuates from day-to-day).

Fourth, it should be noted that while we are interested in mechanisms of change, the current available network estimation techniques assume that the mean and variance of the time series data remains the same (i.e., stationarity) (Jordan et al., 2020). Unfortunately, alternative time-varying network models require many more datapoints than present in the current dataset (e.g., see Haslbeck et al., 2020). Therefore, there is a mismatch between our data, our interest in change, and the available statistical models. In addition, although multilevel VAR estimation allows for the estimation of person-specific networks, these networks are not purely idiographic. With multilevel estimation, a shared underlying distribution for all model parameters is assumed. This means person-specific edge weights are “shrinked” to follow the same underlying group distribution. As a result, individual differences are smoothed out in the estimation process. Therefore, future research should indicate whether the result found here hold on a purely idiographic level.

To conclude, we found that affect fluctuations and the evolution of depressive complaints are strongly related during stressful periods. We showed this relation in three ways. First, we visually explored the trajectories of PA, NA, and depressive complaints at the beginning of the COVID-19 pandemic. Individuals who consistently reported no or mild depressive complaints showed higher levels of PA, while people with consistently severe depressive complaints showed higher levels of NA. Second, we zoomed in on these relations by estimating longitudinal networks, both at the individual and group level. These network models showed many direct relations between PA, NA, and depressive complaints. Third, we found that individuals with an alleviation or aggravation of depressive complaints have more densely connected networks. This means that the stronger affect and depressive complaints are connected over time, as indicated by the network model, the larger the change is in depressive complaints. Together, these findings shed light on the potential underlying mechanisms of change and the development of mental disorders.

All authors contributed equally to the manuscript:

• Contributed to conception and design: GL, TFB, RHAH

• Contributed to acquisition of data: GL, TFB, RHAH

• Contributed to analysis and interpretation of data: GL, TFB, RHAH

• Drafted and/or revised the article: GL, TFB, RHAH

• Approved the submitted version for publication: GL, TFB, RHAH

Nothing to declare.

We would like to thank Tony Cunningham and colleagues from Boston College for collecting and sharing this unprecedented dataset. Furthermore, we would like to thank Denny Borsboom, Jonas Haslbeck, and the editor and reviewers for their very helpful comments on earlier drafts of this manuscript.

GL is supported by the University of Amsterdam Data Science Centre.

RHA is supported by NWO Research Talent Grant no. 406-18-532 awarded to RHA Hoekstra.

TFB is supported by BIAL Foundation Grant no. 284/20 awarded to TF Blanken, and by the University of Amsterdam Data Science Centre.

The derived data used for analyses and corresponding R-code containing analysis scripts can be found on the Open Science Framework (OSF): https://osf.io/2zh4f/.