Intraindividual Variability and Temporal Stability of Mid-Sleep on Free and Workdays

People differ in their sleep timings that are often referred to as a chronotype and can be operationalized as mid-sleep (midpoint between sleep onset and wake-up). The aims of the present studies were to examine intraindividual variability and longer-term temporal stability of mid-sleep on free and workdays, while also considering the effect of age. We used data from a 2-week experience sampling study of British university students (Study 1) and from a panel study of Estonian adults who filled in the Munich Chronotype Questionnaire twice up to 5 years apart (Study 2). Results of Study 1 showed that roughly 50% of the variance in daily mid-sleep scores across the 14-day period was attributed to intraindividual variability as indicated by the intraclass correlation coefficient. However, when the effect of free versus workdays was considered, the intraindividual variability in daily mid-sleep across 2 weeks was 0.71 the size of the interindividual variability. In Study 2, mid-sleep on free and workdays showed good levels of temporal stability—the retest correlations of mid-sleep on free and workdays were 0.66 and 0.58 when measured twice over a period of 0-1 to 5 years. The retest stability of mid-sleep scores on both free and workdays sharply increased from young adulthood and reached their peak when participants were in late 40 to early 50 years of age, indicating that age influences the stability of mid-sleep. Future long-term longitudinal studies are necessary to explore how age-related life circumstances and other possible factors may influence the intraindividual variability and temporal stability of mid-sleep.

We excluded 21 instances due to several reasons in the following order: Six instances because participants had indicated the same wake up and going to bed times, one because they went to bed before trying to fall asleep, one because they needed more than 5 h to fall asleep, six because their sleep duration was less than or equal to 1 h, one because their sleep duration was more than 15 h, three because their mid-sleep score was more than 15, and finally three because there was no information available on whether it was a work or free day.

Supplemental Material S3
If we also consider the strong correlations between the mid-sleep scores on free and workdays (both at the level of retrospective [i.e., MCTQ] and daily average measurements, rs = 0.84 and 0.87, respectively), our findings indicate that people seem to have a general disposition which makes them go to bed either earlier or later, regardless of whether it is a work-or free day. Our final model also showed that participants had a later mid-sleep score when they woke up on a free day and went to bed on a free day. Seizing the opportunity to sleep in on a free day might lead to going to bed later which in turn might also influence one's wake-up time as suggested by the models examining sleep onset and wake-up time.
Therefore, when asking about one's MSF, a free day should be defined as a day when one can go to bed and get up on a free day, which would be a Sunday in a typical European workweek. In the MCTQ (Roenneberg et al., 2003), for example, participants are asked to differentiate between free and workdays when talking about their sleeping patterns but it is not properly defined what a free day actually means.

Figure S4.1. Flowchart of the Study 2 sample selection process (Estonian Biobank).
Participants who had at least two measurements of chronotype assessed with the MCTQ (N = 1111) Participants with valid data both at T1 and T2 (N = 681), including 564 participants who did not use an alarm clock on weekends; • 118 participants who used an alarm clock at least once either at T1 or T2.
296 participants were excluded at T1 because of … • taking medications that influence sleep (n = 174) • doing shift work (n = 104) • both doing shift work and taking medications (n = 12) • irregularities in the data (n = 6) • having a sleep duration of less than four hours (n = 2) 134 participants were excluded at T2 because of … • taking medications that influence sleep (n = 69) • doing shift work (n = 48) • the second measurement took place more than 5 years later, i.e., between 6 to 9 years (n = 10) • doing both shift work and taking medications (n = 3) • irregularities in the data (n = 2) • having a sleep duration of less than four hours (n = 2) Participants with valid data at T1 (N = 815)

Days Corrected for Sleep Debt (MSF sc ) and Mid-Sleep on Free Days (Study 2)
As chronotype is often operationalized as mid-sleep on free days corrected for sleep debt (MSFsc; Roenneberg et al., 2003;Roenneberg, 2015), we wanted to ensure that our analyses were not influenced by the decision to only investigate mid-sleep on free days (MSF). MSFsc can only be calculated if participants do not use an alarm clock on weekends.
If the sleep duration on free days is smaller than or equal the sleep duration on workdays, MSF does not need to be corrected. When the sleep duration on free days is greater than the sleep duration on workdays, it is calculated as such: where SDfree is equal sleep duration on free days and SDweek is equal sleep duration on workdays.
We had to exclude 118 participants because they had used an alarm clock on free days at either the first or second time of assessment. Therefore, our final sample consisted of 563 participants. As the exclusion of participants might have impacted our results, we re-ran the analyses for both MSF and MSFsc. The mean age of the participants at T1 was 49.57 (SD = 15.51). Around half of the participants identified as female (272; 48.31%). At T1, 58 (10.30%) participants had basic education, 309 (54.88%) had completed secondary education/secondary vocational education, and 196 (34.81%) had higher education.

Intervals
The test-retest correlations for this sample were r = .67 for MSF and r = .59 for MSFsc. All correlations significant at p < .001. The test-retest correlations for groups with different test intervals ranging from 0-1 to 5 years of MSF and MSFsc are depicted in Figure   S5.1. In general terms, the stability of the two variables remained quite high over the course of the years. The overall highest test-retest correlations were found for MSF which ranged from r = .64 (2 years) to r = .74 (0-1 year) which were slightly higher than the test-retest correlations of MSFsc with rs ranging from .57 (2 years) to r = .64 (0-1 and 3 years).
However, the test-retest correlations for MSF and MSFsc according to the retest interval did not differ from each other significantly at p < .05.

Individual and Group-Level Stability of Mid-Sleep Across the Life Span
Next, we examined how the stability coefficients of MSF and MSFsc depend on age. Adding the quadratic terms to both models explained about three percent more of the variance.
To further elaborate on how the rank-order stability of MSF and MSFsc is influenced by age, we divided participants into six age categories at T1: 18-25 (n = 47), 26-35 (n = 76), 36-45 (n = 110), 46-55 (n = 108), 56 to 65 (n = 112), and 66-87 (n = 110). We then calculated test-retest correlations for MSF and MSFsc for each group. Figure S5.4 illustrates these testretest correlations by age group. The rank-order stability for all three variables seems to reach its peak when participants are around 46-55-years old (rs ranging from .71 to .74, ps < .001) and then slightly decreases and reaches a plateau until older age. The test-retest correlations of MSF and MSFsc of each age group did not differ from each other significantly.
As in the main study, we again ran a series of hierarchical regression analyses where we predicted individual stability coefficients (MSF and MSFsc in separate models) from participant's age and the square of age at T1 (in order to account for both linear and nonlinear relationships) when also controlling for retest interval. The results of the hierarchical multiple regression analysis for the individual stability of MSF and MSFsc are shown in Table   S5.1 and Table S5.2. In the models explaining the stability of MSF and MSFsc, age and age square had a significant effect on the stability of mid-sleep and the addition of the variables resulted in significant improvements of the models. However, the time difference in the retest interval was not a significant predictor of stability.

Conclusion
We repeated all the analyses using both MSF and MSFsc in order to ensure that our analyses were not impacted by the decision to solely look into MSF and not MSFsc. As MSFsc can only be calculated when participants are not using an alarm clock on weekends, we had to exclude 118 participants who either had used an alarm clock at T1 or T2.
The test-retest correlations of MSF and MSFsc when looking into the effect of age and time interval of filling in the questionnaire show very similar patterns. However, the coefficients seem to be slightly, but not significantly, higher in MSF than MSFsc, indicating that participants' chronotypes change more when sleep debt is accounted for. This might be because even if sleep times remain the same, work hours or social demands might have changed over time.  Asendorpf's (1992)  Note. CI = A 95% confidence interval. Note. CI = A 95% confidence interval.  Asendorpf's (1992)  Note. CI = A 95% confidence interval