both reminders, the questionnaire would expire and not be taken into account. Participants would also maintain a sleep diary each morning, where they were asked to fill in a short sleep diary following their night’s sleep.
Measures
VariablesGender, age, education level, working hours, actual working hours and work sector were measured using one question, respectively. Work engagement does not seem to be significantly influenced by these variables, so therefore they were not included in the subsequent analyses (Becker, 2005).
Momentary job demands. Momentary job demands were measured using the Job Content
Questionnaire (JCQ; Karasek et al., 1998). There was one item where job demands were assessed with the question: “Right now, my activity requires that I work hard”. The response format ranges from 1 (I disagree) to 7 (I agree).
Momentary job autonomy. Momentary job autonomy was measured using the Questionnaire on the Experience and Evaluation of Work (QEEW; Van Veldhoven & Broersen, 2003). One item from the QEEW was used to assess momentary levels of job autonomy with the question ‘Right now, I can decide what I do’ and scored on a seven-point Likert scale (1 = I disagree, 7 = I agree).
Daily sleep quality. Sleep quality is a conglomerate variable which is researched in sub-variables.
Self-perceived sleep quality. One item was used to measure the self-perceived sleep quality of the participant with the question: “How would you rate the quality of your sleep?”. The response format ranges from 1 (very bad) to 5 (very good).
Self-perceived recovery. One item measured the self-perceived recovery of the participant. This was assessed with the item: “How rested do you feel?”. Here the response format ranged from 1 (not rested at all) to 5 (completely rested).
Observed sleep latency, observed sleep efficiency and observed total sleep. Data on sleep latency, sleep efficiency and total sleep were collected using the Actigraphy (activity watch), that measured physical activity at the wrist and allows for quantification of sleep-wake patterns. These measurements were therefore collected objectively. In the tested hypotheses where sleep quality was included (hypothesis 2 and 5), all the sub-variables were added separately in the model to include the full spectrum of sleep quality.
Momentary work engagement. It is assumed that the dimensions of burnout and work engagement are bipolar dimensions (Demerouti, Bakker, Vardakou, & Kantas, 2003; Demerouti, Mostert, & Bakker, 2010;
Demerouti et al., 2001; González-Romá et al., 2006). This assumption is also reflected in the OLBI where both positively worded items as well as negatively worded items are included. In order to be assess work engagement using the OLBI, the negatively frames items need to be recoded (Demerouti & Bakker, 2008;
Demerouti et al., 2003). Therefore, momentary levels of work engagement were assessed using the Oldenburg Burnout Inventory (OLBI; Demerouti, Bakker, Vardakou, & Kantas, 2003). The Utrecht Work Engagement Scale (UWES; Schaufeli et al., 2002) is a commonly used instrument for measuring work engagement. However, for this research the OLBI was used for assessing work engagement, because the OLBI essentially captures the same constructs as the UWES (Demerouti et al., 2010). Therefore, the OLBI is a reasonable alternative for measuring work engagement.
The two dimensions of the OLBI (dedication and vigor) were measured in two items each, of which one was asked in a reversed manner. This means that the dimensions of dedication and vigor had one positive and one negative question. In accordance with previous research, this thesis followed the same strategy of measuring work engagement. The two items used to measure the items of dedication were as
follows: (1) ‘At this moment, I am experiencing my activity as a real challenge’) where one was asked in a reversed manner (2) ‘I am sickened by my current activity’), and both were scored on a seven-point Likert scale (1 = I disagree, 7 = I agree). The two items used to assess vigor were as follows: (1) “I have enough energy for my current activity”, where one was asked reversed (2) “Right now, I feel emotionally drained”. The response format for these two questions were also scored on a seven-point Likert scale (1
= I disagree, 7 = I agree). Since research has shown that burnout and work engagement can be seen as each other’s opposites, the measured items were inverted so that they would represent the dimensions of dedication (opposite of disengagement) and vigor (opposite of emotional exhaustion) (González-Romá et al., 2006).
Questionnaires and measurement types
This research employed three different measurement types. The first type was the intake questionnaire, where the participant was asked on his demographics (age, gender, education level, weekly working hours on contract, actual weekly working hours, work sector and working days per week). Furthermore, the intake questionnaire measured the general work engagement levels using the Oldenburg Burnout Inventory (OLBI; Demerouti et al., 2003), the general job resources and demands using the Job Content Questionnaire (JCQ; Karasek et al., 1998), and the general sleep quality was assessed using the Pittsburgh Sleep Quality Index (PSQI; Buysse et al., 1989).
The second measurement type used was the wearable sensors. Participants wore two devices during the data collection period. The first device is a light logger that measures the amount of light a subject gets exposed to. The participant was therefore asked to wear the light logger as close to the eyes as possible.
The data collected by the light logger was not used in this research and will therefore not be mentioned further. The activity watch was the second device worn by the participants. This watch measured the physical activity at the wrist and allows for quantification of sleep wake patterns. Because it measured sleep wake patterns, it was necessary for the subject to wear the activity watch for the entirety of the experiment.
The third and final measurement type was the ESM items. Four ESM items measured momentary work engagement adapted from the OLBI (Demerouti et al., 2003) and two items adapted from the
Questionnaire on the Experience and Evaluation of Work (QEEW) were used to assess the current level of autonomy (Van Veldhoven & Broersen, 2003). In addition to ESM items, participants were asked to complete the Consensus sleep diary by Carney et al. (2012) every morning pertaining to the previous night’s sleep. An item on type of day (workday or work-free day for previous day) and use of alarm clock in the morning are added to the diary. The level of sleep quality was measured in the sleep diary with one item (‘How would you rate the quality of your sleep?’) and was scored on a five-point Likert scale (1
= very bad, 5 = very good).
Statistical Analyses
The collected data, obtained from the questionnaires and diaries, was first mapped in a dataset using Microsoft Excel software. Thereafter, the ESM variables used to test the hypotheses were divided into datasets and structured using SPSS software. SPSS was further used to do find the descriptive statistics of the variables and to transform the data, such that it could be used for the analyses. This transformation was needed so that the analyses software could read the data. In order to analyze the data, multilevel analyses were performed using MLwiN software (Rasbash, Browne, Healy, Cameron, & Charlton, 2000).
Moreover, for interpretation of the interaction effects for the moderator analyses, simple slopes analyses was conducted using Microsoft Excel software (Dawson, 2014; Dawson & Richter, 2006;
Toothaker, Aiken, & West, 1994). The collected data has a hierarchical structure with three levels where ESM variables measured at a moment were nested within days and days nested within persons.
Multilevel analysis was found to be the best way to analyze such interdependent observations using the
hierarchical linear modelling approach (Raudenbush & Bryk, 2002). A null model was developed to randomize intercepts at all levels. Independent variables were person mean-centered in order to avoid multicollinearity, which is especially valuable for measuring interactions between variables. (Bakker &
Bal, 2010). In order to test the hypotheses a model was created for each hypothesis separately. Before testing the models and analyzing the data, the data was checked for missing values and structured so that it would be appropriate to use. When creating the models in order to test the hypotheses, momentary work engagement was the construct variable and momentary job demands and job
autonomy were the explanatory variables. Sleep quality with underlying variables acted as a moderator in the models. Job autonomy also acted as a moderator in one of the models.
For hypothesis 1, the tested model (model 1) included the dependent variable work engagement, and the independent variable job demands both on a momentary level. For hypothesis 2, a model (model 2) including dependent variable work engagement and independent variable job demands both on
momentary level, but now also sleep quality was included on a day level. For the third hypothesis, job autonomy was the independent variable and direct effect on work engagement was assessed (model 3).
In hypothesis 4, the effect of the independent variable job demands on the dependent variable work engagement was measured with job autonomy as a moderator (model 4). Lastly, hypothesis 5 was tested by calculating the effect of the independent variable job autonomy on the dependent variable work engagement, with sleep quality as a moderator (model 5).
Data
Construct variable
Because the construct variable momentary work engagement is built up out of four items, it needed to be determined whether they could be taken together so that they would form one single construct.
Because the data is multilevel it was not possible to determine correlations with the raw data alone.
Moreover, since the data was multilevel and there were only four items, it was also not possible to perform a factor analysis or find reliable Cronbach Alpha’s. Therefore, in order to determine the
construct variable(s), the data needed to be altered in order to analyze the correlations of the variables, while respecting the multilevel nature of the data. The data was altered by first calculating the mean scores for each person (person mean) for DED1, DED2, VIG1 and VIG2. In order to calculate the within-person mean-centered scores, the within-person means were subtracted from all the raw scores resulting in N=1298 scores. Next, in order to calculate the person-centered scores, the person mean was taken once for every person (N=51). After these scores were calculated, the construct item correlations were calculated for the within-person (above the diagonal) and person centered (below the diagonal) scores.
The results are described in Table 1. These correlations show that there exists a very weak correlation between the items of dedication DED1 and DED2. Moreover, it shows a similar weak correlation between the items of vigor, VIG1 and VIG2. Other correlations were also too weak, which is why it was not
possible to combine items. Therefore, the four items together were considered to represent the construct work engagement.
Table 1: Means, standard deviations and correlations of construct items
** Correlation is significant at the 0.01 level (2-tailed). * Correlation is significant at the 0.05 level (2-tailed). Correlations below the diagonal are person-centered correlations (N= 51). Correlations above the diagonal are within-person mean-centered correlations (N=1298).
Null model
The null model is meant to provide a base model on which can be built further in the analyses, which includes momentary, day and person level random effects nested in a constant variable.
This means that only the dependent variables DED1, DED2, VIG1 and VIG2 were measured over the three levels. Mathematically, this model is described as follows:
𝐷𝐸𝐷1!"# = 𝛽$+ 𝑣$#+ 𝑢$"#+ 𝑒$!"#
𝐷𝐸𝐷2!"# = 𝛽$+ 𝑣$#+ 𝑢$"#+ 𝑒$!"#
𝑉𝐼𝐺1!"# = 𝛽$+ 𝑣$#+ 𝑢$"#+ 𝑒$!"#
𝑉𝐼𝐺2!"# = 𝛽$+ 𝑣$#+ 𝑢$"#+ 𝑒$!"#
Where:
In the model, 𝛽! represents the intercept for the constant, 𝑣!" represents person level variance, 𝑢!#"
represents the day level variance and 𝑒!$#" represents the momentary level variance.
To determine whether multilevel analysis is appropriate for this data, two tests were performed. First, a test to see whether unexplained variance was present at all levels. For DED1 the null model showed 14.09% unexplained variance on the person level, 4.87% unexplained variance the day level and 81.04%
unexplained on the momentary level. For DED2 the null model showed 17.20% unexplained variance on the person level, 2.28% unexplained variance the day level and 80.52% unexplained on the momentary level. The null model for VIG1 showed unexplained variance of 28.69%, 4.14% and 67.18% on the person, day and momentary level, respectively. Finally, for VIG2 the null model showed 37.31% unexplained variance on the person level, 5.76% unexplained variance the day level and 59.93% unexplained on the momentary level. This means that the first test shows evidence for variances on multiple levels, which would suggest multilevel analysis would be appropriate.
The second test performed to determine multilevel analysis was, by determining if the three levels are also necessary in order to perform the analysis. By calculating the deviance scores (-2*log(lh)) on all three levels the chi-square difference (D-2*log(lh)) could be found. This D-2*log(lh) could then determine if there was a significant difference between the models of the different levels. The -2*log(lh) of the
𝑣$#~𝑁(0, 𝜎%$& ) 𝑢$"#~𝑁(0, 𝜎'$& ) 𝑒$!"#~𝑁(0, 𝜎($& )
multiple levels were compared and subtracted in order to determine if multilevel analysis would be appropriate for this data. As can be seen in
Table 2, all the null models have significantly better models when fitted on three levels instead of one or two levels. This means that the second test also suggests multilevel modeling. Therefore, the analysis could be continued in a multilevel fashion.
Table 2: Deviance scores and Chi-square tests of null models
DED1 DED2 VIG1 VIG2
-2*log(lh) three levels 5310.38 4459.29 3760.98 3699.30 -2*log(lh) two levels 5363.64 4538.87 3894.30 3852.37 -2*log(lh) one level 5453.91 4627.47 4109.43 4175.65
D-2*log(lh) level two and three 53.26 79.57 133.32 153.08
D-2*log(lh) level one and two 90.27 88.60 215.13 323.28
P-value for all D-2*log(lh) <.001 <.001 <.001 <.001
Descriptive Statistics
The means, standard deviations and correlations of the observed variables are reported in
Table 3. In order to determine how the momentary variables correlate with each other, the within-person mean-centered correlations for all moments (N=1298) were assessed, which can be seen in Table 3 above the diagonal. Furthermore, the correlation between the four items of work engagement were insufficient. Below the diagonal, the variables were centered on the highest level (person level; N=51) where the momentary variables were person-mean scores, and the day level variables were raw scores.
However, a strong correlation exists between DED1 and job demands. There also seems to be a strong correlation between total sleep time and sleep efficiency.
Table 3: Means, standard deviations and correlations of the independent variables.
Mean Standard deviation
1 2 3 4 5 6 7 8 9 10 11
1. DED1 3.23 .86 - .18** -.07** .05 .73** -.39**
2. DED2 5.71 .66 -.07 - .21** .12** .14** .13**
3. VIG1 6.07 .67 -.32* .50** - .34** -.03 .11**
4. VIG2 6.18 .76 -.32* .49** .64** - .02 .05
5. Job
Demand 2.99 .91 .84** -.11 -.34* -.29* - -.39**
6. Job
Autonomy 5.15 .97 -.18 .04 .26 0.11 -.24 -
7. Sleep
quality 3.67 .78 .02 .23 .19 .42** .15 .08 -
8.
Recovery 3.47 .74 -.06 .04 .39** .28** -.05 .16 .51** -
9. Total
sleep time 368.95 111.65 -.18 .02 .12 -.01 -.17 .15 -.21 -.13 - 10. Sleep
latency 15.93 34.38 .03 -.16 -.13 -.13 .22 -.14 -.27 -.06 .20 -
11. Sleep efficiency
.78 .22 -.13 .01 .15 -.04 -.17 .04 -.14 .04 .83** .11 -
** Correlation is significant at the 0.01 level (2-tailed). * Correlation is significant at the 0.05 level (2-tailed). Correlations below the diagonal are person-mean centered for variables 1-6 and person centered for variables 7-11 (N= 51). Correlations above the diagonal are within-person mean-centered correlations (N=1298). Means and standard deviations for 9 and 10 are reported in minutes.