We used a large mobile EEG-based sample to identify biological, demographic, behavioral and psychological factors underlying sleep variability. We systematically explored the effect of controlling age and sleep metric means on the associations, as well as the use of different variability metrics.
Uncorrected correlations (Supplementary figure S1-S4) replicated patterns familiar from the previous literature. Younger participants and students had less regular sleep, reflected in more variable timing, duration, and composition of sleep and reduced Sleep Regularity Index. In line with previous findings [34, 35], we found no associations between sleep variability and cognition (school grades or intelligence) or BMI. We found no associations between affective symptoms and sleep variability, although less regular sleep was associated with dissociative symptoms. Sleep was more variable in those with a later chronotype [7] and those who napped frequently, but co-sleeping with people or animals was associated with more regular sleep patterns [36, 37]. To the best of our knowledge, ours is the largest study exploring the association between personality and sleep regularity (see [22, 23] for previous work). Findings were sporadic, but impulsivity and sociability had some associations with less regular sleep. Although the choice of the specific variability metric is very important in theory [27], in practice we found that the individual standard deviation and two alternative variability metrics were very highly correlated in the current sample and the choice of variability metric has little influence on the results.
There are two factors, however, which may confound these results. First, age is associated with less variable sleep (Fig. 1, Supplementary figure S1, see also [7, 19]), suggesting that it may confound the association of other age-variant human characteristics with sleep variability. A statistical control for age eliminated the apparent higher sleep variability in those who are students, have higher Sociability (BFI) and Impulsive Sensation Seeking (ZKPQ) personality scores, more dissociative symptoms, a later chronotype, and those who sleep alone more frequently, indicating that the associations were primarily driven by age and not by the psychosocial characteristics per se.
Second, the standard deviation of a person’s sleep metrics across days was strongly correlated with the means (Supplementary table S1, see also [7, 38] for theoretical treatises). As expected, given that sleep variability metrics are often bounded by an ideal value such as zero sleep latency or perfect sleep efficiency, bounded variables had the strongest mean-SD correlations and the sign of the correlation corresponded to the direction of skew in these variables. This suggests that variability is strongly driven by outlying values, and these outlying values also affect the mean. The correlation between means and standard deviations suggests that correlations with variability in sleep characteristics may simply reflect a correlation with mean values. For example, while participants scoring higher on the Impulsive Sensation Seeking ZKPQ personality scale have more variable sleep onset times (r = 0.25, p < 0.001), they also have later sleep onset times (r = 0.21, p < 0.001), mean sleep onset times and their variability are correlated (r = 0.46, p < 0.001), and controlling for means reduces the correlation between the personality trait and sleep onset variability to insignificance (r = 0.1, p = 0.11).
After further controls for the means of sleep metrics, the pattern of correlations with sleep variability changed substantially. Older age was still associated with less variable sleep after controlling for means. Of personality traits, only Openness was associated with more variable sleep (WASO, N2 sleep percentage and awakenings). Isolated correlations were seen between more variable sleep efficiency and the Agreeableness personality trait and N1 sleep percentage and dissociative symptoms. A later chronotype was only associated with increased variability in sleep timing, mirroring its well-known association with social jetlag [39]. More frequent napping was associated with more variable sleep composition, likely reflecting variable sleep pressure during nightly sleep [40]. Of these nominally significant (p < 0.01) correlations, only age effects survived a stringent control for multiple testing, a finding replicated using the two alternative sleep variability metrics as well.
In sum, we found age to be the only factor unambiguously associated with sleep regularity: older participants had more regular sleep schedules. This finding survived corrections for sleep metric means and stringent controls for multiple testing. Some correlations with personality and mental health were significant using the p < 0.01 threshold but did not survive a more stringent control for age and multiple testing and are possibly spurious. Sleep habits – especially napping and a late chronotype – also emerged as correlates of higher sleep variability, but their formal significance after multiple comparisons corrections depended on whether the means of sleep metrics were controlled.
We consider controlling for age to be unambiguous, as age is a clear confounder of the relationship between sleep and other characteristics. It is our view, however, that despite recommendations for it [7], controlling for means may be unnecessary and may in fact mask a genuine association between sleep variability and other traits. This is especially true for sleep metrics bounded at ideal value, such as sleep onset latency or sleep efficiency, which have mean values close to ideal (e.g. in the current sample MSOL=14.99, SD = 12.61; MSE=90.92, SD = 5.51). For these variables, the room for variation is asymmetric and the variables are skewed, approximately following a gamma distribution. In such a distribution, means and standard deviations are mathematically related and there is a plausible causal route from both to the other. While a mean value more distant from the ideal increases the room for variation [38], increased variability necessarily results in more bound-distant values which ultimately affect the mean. (See the Supplementary note for a simulation and Supplementary table S1 which demonstrates that bounded and skewed sleep metrics exhibit the higher correlations between means and standard deviations). Thus, we suggest that mean differences in certain sleep metrics may be the mathematical consequence, rather than the cause, of sleep variation, and statistically controlling it introduces overadjustment bias [41].
In the current study, the question of controlling for means was critical in the case of napping effects on the variability of sleep onset latency, WASO and sleep efficiency. Without controlling for means, the increased variability of these variables (in addition to sleep stage percentages), were associated with longer naps (Fig. 4). These effects survived corrections for multiple comparisons. However, with controls for means only the correlation with stage percentages persisted and did not survive corrections for multiple comparisons (Supplementary figure S8). We believe that this illustrates our point about the problematic nature of controlling for means. Naps during the day reduce sleep pressure, affecting the characteristics of nightly sleep [42, 43]. More frequent napping changes the means of sleep characteristics (because some, if not all, nightly sleeps are affected by the preceding nap), but also introduce variability as a nap occurs on some days but not others. Changing means in this case are not an artifact to be controlled but the consequence of the very same phenomenon that increases variability. Controlling for means effectively controls for a large part of the mechanism that caused variability.
Our study has a number of limitations. First, for the purposes of the current study BSETS is a cross-sectional study and we cannot establish the route of causation between sleep variability and other factors. Longitudinal studies are needed to clarify these. Second, the use of self-reported variables may introduce bias, typically in the downward direction as error is introduced due to the inaccuracy of self-reports. Third, although the findings cohere well with the literature, the generalizability of the findings may be affected by the demographics of the BSETS sample which predominantly consists of young, well-educated individuals.
In conclusion, we have shown age and sleep habits to be the most important drivers of sleep variability, with a comparatively smaller role of normal and pathological psychological characteristics. Controlling for age as an important driver of reduced sleep variability, many zero-order correlations (such as those with student status or sleeping alone) are reduced to non-significance. Conversely, we claim that statistically adjusting associations for variable means is not obviously the right choice and may introduce overcontrol bias. Our findings highlight napping habits as the most modifiable factor affecting sleep regularity.