We used data from the Netherlands Study of Depression and Anxiety (NESDA, Details about NESDA were given extensively before34. In short, NESDA is an ongoing multisite naturalistic longitudinal cohort study among 2981 adults (18–65 years) at baseline, including individuals with depressive disorders and/or anxiety disorders, as well as healthy controls, which were recruited from community, primary care, and specialized mental health-care settings. Patients with a psychotic disorder, bipolar disorder, obsessive–compulsive disorder, or substance use disorder, and persons not fluent in Dutch were excluded from the baseline assessment. The NESDA study was approved by the VUmc ethical committee (reference number 2003/183) and all respondents gave informed consent.

In the current study, we used a subsample (n = 121) from wave 6 (9-year follow-up), which contained a 14-day period of ecological momentary assessments. Details of the actigraphy measurements are given elsewhere18. Of the 384 initially included participants, 14 had no available actigraphy data for several reasons, such as technical failure, 10 individuals had less than 1 weekday or weekend-day data available. For this study, we examined the effect of a depressive episode; therefore, participants with remitted depressive disorder (n = 152), with anxiety disorder only (n = 67) and the siblings of NESDA participants (n = 20) were excluded. As comorbidity of anxiety and depressive disorders is common35, this was not an exclusion criterion. Actigraphy data from individuals with an episode of depressive disorder (major depressive disorder and/or dysthymia) in the past 6 months (called “total depression group”, n = 58) and from individuals with no lifetime depressive and anxiety disorder (controls, n = 63) were used. Additionally, we created a subgroup of the total depression group consisting of people with a depressive episode in the past month (called “acute depression group”, n = 43).


Participants completed 14 consecutive days of all-day actigraphy activity monitoring. They wore a GENEActiv actigraphy device on a nondominant wrist. Participants were instructed to wear the watch day and night and only taking it off when going to the sauna or when playing a contact sport in which wearing a wristband is unsafe. The device is a triaxial wrist accelerometer recording data in SI units represented as acceleration in three axes (x, y, and z). GENEActiv validity studies have demonstrated strong correlations for criterion validity (Pearson r = 0.79–0.98)36 and a good ability to determine sedentary behavior in adults (18–55 years) (Pearson r = 0.81)37. To pre-process the raw data, an open-source R package, GGIR (version 1.5-12)38 was used. At the first stage of pre-processing, we verified sensor calibration error using local gravity as a reference, checked for abnormally high values, non-wear periods, and then extracted objective PA measures. During raw data cleaning, missing data (e.g. suspected of monitor non-wear) or invalid data were imputed by the averages at similar time points on different days of the week, as GGIR does by default. In this way, within-day variability of the data should not be affected, unlike the situation when the data are imputed with the average of close by time points of the same day. Only respondents with valid actigraphy data for at least 16 h per day, and at least 10 days were included in the analysis (n = 121). To estimate PA, we calculated a metric ENMO (Euclidean Norm Minus One) using the formula (sqrt {x^2 + y^2 + z^2} – 1g) with any negative values rounded up to zero and by reducing such measure over 1-min epochs39. We adapted and executed the R-script for the current study.

Predictor and outcome measures

The main predictor in the study was clinical group status (depression diagnosis versus healthy control status). Diagnoses were acquired with the Composite International Diagnostic Interview (CIDI, version 2.1) based on the DSM-IV40. The CIDI has high interrater reliability, high test–retest reliability, and high validity for depressive disorders40. Participants were divided into the following groups based on the results of a CIDI assessment: (1) a group with no lifetime history of psychiatric disorders (healthy controls) (n = 63), and (2) a group with a depressive disorder (either with or without anxiety comorbidity in the past 6 months, n = 58). In a next step, to assess whether there was an acute state effect of depression on PA, we selected a subgroup of the 6-month depressed group with a depressive disorder present in the past 1 month (termed “acute”, n = 43) in the analysis to compare to the control group.

The main outcome variables were three actigraphy-derived PA variables: mean levels of activity (MESOR, an acronym for Medline Estimated Statistic of Rhythm), height of peak with respect to the mean level (amplitude), and the timing of the activity peak (acrophase). MESOR, as a measure of the mean level of the curve between the highest and the lowest point, represents the average activity level of the day. Amplitude, as a measure of the difference between activity during the day and the night, indicates the robustness of the diurnal rhythm. Acrophase, as a measure of the time of the day when the activity peak occurred, indicates timing preferences of an individual during the day. MESOR and amplitude were measures for the level of PA, and acrophase—a measure for the timing of PA. To calculate these variables, circadian rhythms were estimated by fitting individual 1-min epoch ENMO data to a cosine curve of a 24-h activity rhythm, which was obtained by the cosinor method41,42 using the following equation: y(t) =M+Acos(2πt/τ+ϕ) +e(t), where y is the selected variable, t is the elapsed time, M is MESOR, a rhythm-adjusted mean, A is the amplitude of circadian rhythm, τ is the period of the circadian rhythm (the 24-h period), ϕ is the acrophase, and e(t) is the error term.

Cosinor analysis was performed in R statistical software version 1.1.38343. For this purpose, we created an R script, which obtains daily values (i.e. 14 values per person) of the cosinor parameters MESOR, amplitude, and acrophase, and automates this process across persons and days and exporting the values to a new data set. The script is given in Supplementary Materials 1.


Chronotype was assessed with the Munich Chronotype Questionnaire (MCTQ)44. The MCTQ is not a scale; therefore, its reliability cannot be assessed. The MCTQ highly correlates with the Morningness−Eveningness Questionnaire (MEQ) (r = –0.73)45, which in turn, consistently report high levels of reliability (>0.80)46. Chronotype was defined as the midpoint in time between falling asleep and waking up on free days corrected for “oversleep” due to the sleep debt that individuals accumulate over the workweek (MSFsc).

Potential confounding variables

Demographics (gender, age, marital status), socioeconomic status (education, employment)12, smoking status, BMI47, current medication use (e.g. benzodiazepine, antidepressants)13, and depression severity were included in the analysis as additional covariates. Depression severity was assessed with the Inventory of Depressive Symptomatology (IDS)48. Internal consistency of the IDS-SR is high, and the Cronbach α ranged from 0.67 to 0.9448,49. Current medication use was based on drug container inspection and medications were coded according to the World Health Organization Anatomical Therapeutic Chemical (ATC) classification and considered present if participants reported currently using psychopharmacological medication.

Statistical analysis

Baseline characteristics were compared pairwise between the control group and the total depression (diagnosis in the past 6 months) or acute depression (diagnosis in the past 1 month) group, respectively. The independent t test was used for continuous variables (age, BMI, education, IDS scores) and Chi-squared test was used for categorical variables (gender, marital status, employment, smoking status, IDS severity groups, and psychopharmaca use).

As the data are hierarchically structured, we used multilevel linear modeling to assess whether the MESOR, amplitude, and acrophase of the PA rhythm differed between individuals with and without the diagnosis of depression. A two-level data structure was used where days of actigraphy were defined as the first-level unit and individuals as the second-level unit. The first set of analyses was performed using the 6-month cut-off for an episode of depressive disorder (total depression group). Two models were run for each outcome variable. The first model assessed the association between group (non-depressed = 0, depressed = 1) and PA levels and timing (Model 1). Next, this association was tested while including all covariates (Model 2). Only for acrophase, a third model (Model 3) was tested as an additional check whether chronotype (also an indicator of the phase of the circadian rhythm) explains (part of) the association between acrophase and group status. A significant association would support the hypothesis that a shift in acrophase indeed reflects a shift in the circadian rhythm.

For the second set of analyses, all models for three outcome variables (MESOR, amplitude, and acrophase) were repeated in the subsample of acutely depressed (1-month cut-off, n = 43) versus controls.

Besides the fixed effects (group status and the covariates), a random intercept and time trend (day number) were included if they improved the model fit. Model fit was assessed using the Akaike Information Criterion (AIC) (lower is better). For all models, a random intercept, but not a slope for day number, improved the model fit. The best fitting covariance structure for the random effects was the variance components (VC) structure. In addition, to correct for potential autocorrelation of the residuals, which is often present in repeated assessments data, we examined several possible covariance structures for the residuals. The best structure was an autoregressive heterogeneous (AR(1)H). Multilevel analyses were conducted using SPSS Statistics version 25. Significance levels were set at p < 0.05.

The residuals of the final models were not normally distributed. To overcome this issue, we repeated all analyses with a bootstrap approach, as bootstrapping is robust against violations of normality50. Due to model complexity, we needed to slightly adapt the model; instead of specifying an AR(1)H covariance structure for the residuals, we included a lag(1) outcome variable as a fixed and random predictor. This way, potential autocorrelation was still addressed. Even though the bootstrap analysis gave similar results, and in three cases borderline significant findings even became significant, we decided to keep a more conservative approach to prevent overstating the results. Therefore, we reported the results of the original models only. However, the syntax and the results of the bootstrap analyses are given in the Supplementary Materials 24 (Tables S1 and S2).

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *