The initial sequence of seven steps was estimated under the simplest assumptions about the error structure over time (i.e. homoscedasticity and independence of errors). As Table 3 shows for positive affect, the fit (−2LL) of this model was 7800.90. We re-estimated this model under alternative error structure assumptions, including diagonal (heteroscedasticity and independence of observations; −2LL = 7778.17), compound symmetry (homoscedasticity and fixed correlation between all occasions; model would not converge), and autoregressive error (AR(1); homoscedasticity and interoccasion correlations that are reduced, exponentially, by the temporal lag between them; −2LL = 7761.30). Thus, homoscedasticity and independence of errors are not tenable assumptions in these data, and AR(1) provided the best specification of the repeated measures error of the three approaches tested. It is important to note, however, that all fixed and random parameter estimates remained essentially unchanged under different error structure assumptions, and the pattern of significance was not altered.
As a concluding step, for negative affect, we again examined the tenability of the homoscedasticity and errors independence assumptions for the repeated measures, for which the −2LL (Table 3) was 6660.12. As with Positive Affect, alternative error structures provided better fits to the data: diagonal,−2LL = 6598.45; compound symmetry, model again would not converge; and autoregressive error,−2LL = 6628.23. Thus, for Negative Affect, of the three specifications of repeated measures error tested, a diagonal structure provided the best fit to the data. Again, as with Positive Affect, different error structure specifications had little effect on fixed and random parameter estimates or their pattern of significance.
Christina S. McCrae, PhD, Assistant Professor, Department of Clinical & Health Psychology, University of Florida, PO Box 100165, HSC, Gainesville, FL 32610-0165, USA. Tel.: +1 352 273 6053; fax: +1 352 273 6156; e-mail: email@example.com
The main objective of the present study was to examine daily associations (intraindividual variability or IIV) between sleep and affect in older adults. Greater understanding of these associations is important, because both sleep and affect represent modifiable behaviors that can have a major influence on older adults’ health and well-being. We collected sleep diaries, actigraphy, and affect data concurrently for 14 days in 103 community-dwelling older adults. Multilevel modeling was used to assess the sleep–affect relationship at both the group (between-persons) and individual (within-person or IIV) levels. We hypothesized that nights characterized by better sleep would be associated with days characterized by higher positive affect and lower negative affect, and that the inverse would be true for poor sleep. Daily associations were found between affect and subjective sleep, only and were in the hypothesized direction. Specifically, nights with greater reported awake time or lower sleep quality ratings were associated with days characterized by less positive affect and more negative affect. Gender was not a significant main effect in the present study, despite previous research suggesting gender differences in the sleep–affect relationship. The fact that self-ratings of sleep emerged as the best predictors of affect may suggest that perceived sleep is a particularly important predictor. Finally, our results suggest exploration of affect as a potential intervention target in late-life insomnia is warranted.
Multilevel modeling (MLM) was used to examine daily associations between sleep and affect in older adults. Although sleep and affect are frequently measured on a daily basis, traditional analytic approaches aggregate these data to reduce error. However, studies now show that transient fluctuations in an individual’s responses over time [often referred to as within-person variation or intraindividual variability (IIV)] provide unique, meaningful information that can be reliably distinguished from error (Hultsch et al., 2000; Li et al., 2001). Using MLM, sleep researchers can examine the impact of daily fluctuations (IIV) within a data set and can gain a more precise understanding of the dynamic relationships among variables, particularly when compared with traditional mean level analyses.
An additional complication to understanding the sleep–affect relationship involves the method used to measure sleep. Sleep can be assessed subjectively (e.g. self-report sleep diaries) and objectively (e.g. actigraphy, polysomnography). Unfortunately, quite often these different modes of measurement are poorly correlated within individuals (Espie et al., 1989; McCrae et al., 2005; Means et al., 2003). Consequently, the literature on the importance of sleep to various constructs may vary depending on how sleep was measured. Further, there may be unique temporal associations between (subjective and/or objective) measures of sleep and other constructs (e.g. affect) that are individually specific. These may be obscured by excluding one type of sleep measurement or through a comparison of mean level performance between groups.
According to State-Trait theory (Cattell, 1963; Mischel and Shoda, 1995; Nesselroade, 1988), affect (frequently referred to as mood) is the temporal (state) portion of human emotional life, which fluctuates with present experiences (Kraemer et al., 1994). Historically, the literature conceptualized affect as a unitary, bipolar construct, bracketed by high levels of either positive or negative scores (Meddis, 1972; Russell, 1979). By definition, a shift toward, or increase in one dimension (e.g. negative) necessitates a decrease and shift away from the other polar dimension (e.g. positive) (Russell, 1979). However, studies have shown that positive and negative affect scores are not mutually exclusive, and thus are relatively distinct constructs (Watson et al., 1984, 1988). Moreover, studies have argued that positive and negative affective systems operate independently (Cacioppo and Bernston, 1994; Cacioppo et al., 1999). Recent neuroimaging studies have effectively identified and manipulated disparate positive and negative affective neural networks (Canli et al., 1998, 2001; Davidson et al., 2000), lending credence to this view. Given the conceptual and biological support for distinct positive and negative affect constructs, the present study examined the relationship between sleep and positive–negative affect, independently.
Berry and Webb (1983) examined correlations between sleep and affect using two nights of sleep recordings (18 variables) and one night of affect data [five mood scales from the Lorr Mood test (Lorr et al., 1967) collected on the second night] from middle aged adults (aged 50–60). They analyzed their data separately for men and women and found a greater number of significant correlations for women compared with men (24 versus 3). They did not interpret the correlations for males (only three), attributing them to chance alone. For women, they concluded that positive affective states were related to increasing sleep efficiency and total sleep time, while negative affective states were positively correlated with waking after sleep onset.
In a follow-up study, Berry et al. (1985) examined four nights of laboratory recorded sleep and three nights of affect information (collected on the second, third, and fourth nights) from 25 women aged 56–66. Results from this study found that sleep efficiency and the latency to the first rapid eye movement (REM) period were significantly related to affect. Combined, these two studies suggest that the relationship between sleep and affect might be stronger for women than men, and this relationship may be mediated by objective measures of sleep (i.e. sleep efficiency and latency to first REM period).
Using subjective measures of sleep, Totterdell et al. (1994) examined 14 concurrent days of sleep diaries and well-being measures (affect, physical and cognitive symptoms, and social interaction) from 30 healthy, employed adults (aged 20–59). Their results showed small, but significant correlations among most of the subjective sleep and well-being variables (including affect).
Unfortunately, the interpretation and contextualization of these findings is complicated by methodological differences between these studies (e.g. objective versus subjective sleep measures). Moreover, Totterdell and colleagues’ sample was younger (mean age 31.6 versus 59.5 years) and of mixed gender (versus women only). These studies also used different tools to measure affect [items from the University of Wales Institute of Science and Technology Mood Adjective List (Matthews et al., 1990) versus the Loor Mood Test (Lorr et al., 1967)] and had different lengths of data collection (14 versus 4 days). Consequently, ambiguity remains in understanding the sleep–affect relationship.
However, these three studies suggest that sleep and affect may fluctuate on a daily basis, making them ideal constructs for MLM. This type of analysis offers an important next step in understanding the dynamic relationship between sleep and affect for at least two reasons. First, because sleep and affective disorders are amenable to behavioral modification, a more detailed understanding of the relationship between the two would lead to better intervention outcomes. Second, enhanced intervention efficacy would improve the quality of life in older adults as both sleep and affect are known to have a major influence on their health and well-being.
As discussed above, few studies have focused on the relationship between sleep and affect in older adults. Those that have are not conclusive, in part, because of methodological and/or analytic disparities. Increased clarity regarding the sleep–affect relationship has broad implications for our knowledge of sleep in later life and specific implications for the treatment of late-life insomnia.
This study addressed the aforementioned limitations by using a multifaceted approach to assess sleep and then analyzing its relationship with affect over time using MLM. We hypothesized that nights characterized by better sleep would be associated with days characterized by higher positive affect and lower negative affect, and that the inverse would be true for nights characterized by poor sleep. Additionally, we hypothesized that the nature of this relationship would be consistent for both objective and subjective sleep variables.
Subjective sleep measures
Participants completed sleep diaries (Lichstein et al., 1999) each morning for 14 days, providing subjective estimates of the following sleep–wake parameters: (i) total wake time (TWTs)-total unwanted awake time in bed; and (ii) sleep quality rating (SQR), scaled from 1 (very poor) to 5 (excellent). Because TWT was also collected objectively using actigraphy (see below), an ‘s’ subscript indicates the variable was measured subjectively, and an ‘o’ subscript indicates it was measured objectively. Compliance with diary completion was exceptionally high, and only minimal data were lost. Out of 5768 possible data cells (four sleep diary variables × 14 days × 103 participants), only 76 were missing (1.32%).
A mean was computed for each of these variables for the 2-week recording period. Centered variables, which give the daily deviation from the participant’s mean, were also computed for each sleep-wake parameter (Kreft et al., 1995; Singer, 1998). These centered variables were used to examine daily fluctuations or IIV (see Statistical analysis).
Objective sleep measures
Participants wore an actigraph, the Actiwatch-L, which has an integral ambient light sensor (Mini Mitter Co., 2001), on their non-dominant wrist for 14 consecutive days, concurrent with the sleep diary period. The Actiwatch-L monitors ambient light exposure and gross motor activity and contains an omni-directional, piezoelectric accelerometer with a sensitivity of ≥0.01 g-force and a light sensor with a recording range of 0.1–150 000 Lux.
The sensors of the Actiwatch-L are sampled 32 times/s and record the peak value for each second. These peak values are then summed into 30-s ‘activity’ counts. These activity counts are then downloaded to a PC and analyzed using Actiware-Sleep v. 3.3 (Mini Mitter Co., 2001), which uses a validated algorithm to identify each epoch as either sleep or wake (Oakley, 1997). The software provides three default sensitivity settings (high, medium, low). This study utilized medium sensitivity. On medium sensitivity, the threshold is set at 40 activity counts. If the total activity for an epoch was ≥40, it was scored as wake. If the total activity was ≤40, the final activity count for the epoch was based on the level of activity in the surrounding 2 min (Eq. 1).
where A = activity count for the epoch being scored and EA±1–4 = activity count in adjacent epochs. If Epoch A total activity (i.e. weighted sum of activity counts) exceeded the threshold of 40, then Epoch A was scored as wake; otherwise, it was scored as sleep.
Bedtime and time out of bed in the morning were based on sleep diary entries as recommended in the software manual (Mini Mitter Co., 2001). Actiware-Sleep determined sleep start automatically by searching for the first 10 min during which no more than one epoch was scored as wake. Likewise, sleep end was the last 10 min during which no more than one epoch was scored as wake. As previously mentioned, Actiware-Sleep provides objective estimates for TWT, a variable also provided by the sleep diaries. When measured objectively by actigraphy TWTo represents the sum of all wake epochs within the sleep period. Similar to the sleep diaries, TWTo was averaged over the 14 days. Centered variables, which give the daily deviation from the participant’s mean, were also computed for this objective sleep–wake parameter (Kreft et al., 1995; Singer, 1998). This centered variable was used to examine daily fluctuations or IIV (see Statistical analysis).
Data loss was minimal. There were no equipment failures. Three participants reported taking their watch off during the day for several hours (<3 h). However, in each case, the participant reported putting the watch back on several hours before bedtime. Another three participants reported leaving their watches off for an entire day (24 h). To make up for this lost day, these participants wore their watches and completed their sleep diaries for an additional day immediately following the study period. One participant forgot to wear the watch during week 2 and was excluded for the present study. Thus, concurrent sleep diary and actigraphy data are available for 14 days for 103 participants.
Participants completed the Positive and Negative Affect Schedule (PANAS; Watson et al., 1988) following completion of the sleep diary each morning for 14 days. The PANAS consists of 20 self-report items; 10 items measure positive affect (PA), and 10 items measure negative affect (NA). Participants rated the degree to which each mood was experienced at that time using a 1 (very slightly or not at all) to 5 (extremely) Likert scale. Daily PA and NA scores were calculated by summing the scores of the 10 PA and 10 NA items, respectively.
Participants and procedure
A convenience sample of 116 adults aged 60 years and over (M =72.81 years; SD = 7.12) who resided in North Central Florida were recruited through media advertisements, community groups, and flyers to participate in a study of sleep patterns in the elderly. Interested individuals were screened in two phases: a brief telephone interview (15–20 min) that was followed by a 1–1½ hour in-person interview conducted either in the individual’s home (76%) or at a local continuing care retirement center (24%). Exclusionary criteria included: (i) age-younger than 60 years; (ii) presence of sleep disorders other than insomnia (e.g. sleep apnea, narcolepsy); (iii) severe psychiatric disorders (e.g. thought disorders, depression); (iv) cognitive impairment [scoring in the impaired range on three or more subtests of the Cognistat (Mueller et al., 2001)]; and (v) psychotropic or other medications (e.g. beta-blockers) known to alter sleep. Twelve individuals were ineligible following the initial telephone screening for reasons including age, cognitive impairment, and sleep apnea diagnosis; no individuals were ineligible following the in-person interview; one individual was excluded for failure to wear the actigraph during week 2. Thus, the final sample consisted of 103 participants. All were living in their own homes during the study. The majority of participants were Caucasian (96%) and had some college coursework or a college degree (75%; M =16.34 years, SD = 2.92).
During the 1–1½ interview session, participants read and signed an informed consent form approved by the University of Florida’s Institutional Review Board, and a member of the research team administered the Cognistat (Mueller et al., 2001), conducted a sleep history interview, and explained how to complete the sleep diaries and PANAS. The research team member also gave the participant an actigraph and explained how it works and how it should be worn. Participants were asked to complete the sleep diaries and PANAS every morning and to wear the actigraph continuously for 2 weeks. After the first week, a research team member visited the participants again to determine how the study was going, answer any questions, and download the first week’s data from the actigraph. At the end of the 2 weeks, a research team member visited the participants a third time to collect the actigraph and daily measures (sleep diaries and PANAS). All participants received $30 compensation.
The current study used daily data from the objective and subjective sleep measures (TWT) along with the subjective sleep quality rating (SQR) to predict both positive and negative affect applying a MLM approach. MLM, also referred to as hierarchical linear modeling (HLM, Bryk and Raudenbush, 1992), is an extension of the general linear model, and does not require observations to be independent. Thus, MLM is very flexible and especially suited for daily data because of its autoregressive nature and hierarchical structure with daily observations nested within each participant (Singer et al., 1998a,b; Willett et al., 1998; Zautra et al., 2005).
Fixed and random effects can be estimated with MLM. Fixed effects refer to ‘average effects’, or effects that hold over all persons. In our models, fixed effects are examined at two levels. Level 2 fixed effects assess the between-person association, across all participants, of individual differences in predictor and criterion variables. Level 1 fixed effects assess the daily association, across all participants and occasions, between day-level variation in a predictor (i.e. SQR) and day-level variation in an outcome (i.e. PA). Random effects test whether there are significant individual differences in obtained fixed effects. For example, if there is a generally small positive within-person relationship between SQR and PA, a significant random variance term would indicate that the magnitude of that within-person relationship may differ substantially across individuals. Fig. 1 illustrates these three levels/types of effects schematically.
Because of the hierarchical nature of our data (14 consecutive days nested within 103 participants) and in order to increase the precision of predicting fluctuations in PA and NA with changes in sleep patterns, we modeled the data with an MLM approach. This provided the opportunity to examine how well sleep predicted affect both within- (level 1) and between- (level 2) persons. Level 1 analyses addressed questions such as: ‘On days in which a person reports above-average sleep quality, does s/he also experience higher levels of positive affect?’. Level 2 analyses examined questions like: ‘Do people who are generally better sleepers report higher levels of positive affect?’.
Analytical framework-MLM analysis
Both objective and subjective sleep measures were used to predict affect (positive and negative, tested separately) using a multistep MLM approach, using the MIXED procedure in SPSS 14.0. All models were estimated using the Maximum Likelihood (ML) method. The ability of a model to predict affect better than a baseline (null) model was used as an index of Goodness of Fit. Improvements in predictability were determined by the proportional reduction of within- and between-person residual variances compared with this baseline model (Bryk and Raudenbush, 1992). Decreases in residual variances represent a proportional reduction of the prediction error, which is analogous to R2, and used as an estimate of effect size. One issue that has received substantial attention in work on longitudinal MLM is the issue of optimal modeling of repeated measures error structures (e.g. Singer and Willett, 2003); issues such as homoscedasticity and autocorrelation of errors over time warrant consideration. While these issues go beyond the scope of the current paper, we examined the effect of different error structure specifications on model fit. The different error structures tested had little effect on the fixed and random parameter estimates or their pattern of significance (Singer and Willett, 2003).
Several features of the modeling are briefly described here. First, level 2 effects were estimated using predictors, which varied between persons, and for which there was only one value per person. Thus, for the variables of TWTo, TWTs, and SQR, the level 2 predictor was each person’s mean. Second, level 1 effects were estimated using predictors which varied within persons, and for which there was a new value at each occasion of measurement. For the variables of TWTo, TWTs, and SQR, the level 1 predictor was each person’s centered daily score (i.e. daily score–person’s mean). Thus, the level 1 predictors represented each person’s daily deviation from his or her average value on the predictor.
We built our models in a series of seven steps. Details of these steps, including specific parameters estimated, and the corresponding MLM equations, can be found in the statistical Appendix. Briefly, the seven steps estimated were as follows: step 1: null model (no predictors), which defined the variance to be explained, and which served as the baseline against which subsequent models would be used to calculate (i) variance explained, and (ii) improvement in fit; step 2: fixed (overall effect, averaged across all subjects) and random (estimation of the size of individual differences in the strength of the effect) effects of Day (Occasion, coded 0–13) were estimated. Introducing Day as a fixed effect into the models anchored all subsequently added fixed level 2 effects at the first occasion of measurement (i.e. day 1, coded as 0). This step controlled for any overall temporal trends in the data which might spuriously create time-varying relationships between predictors and the affect outcome; step 3: fixed effect of gender estimated; steps 4–6: effects of three sleep variables (objective, accelerometer-measured total wake time; subjective, sleep-diary measured total wake time; subjective sleep quality rating) were estimated, respectively. For each of these sleep variables, we estimated a fixed level 2 effect (i.e. the association between a participant’s mean level of sleep and their level of affect at the beginning of the assessment period), a fixed level 1 effect (i.e. the association between participants’ day-to-day deviations from their mean level of sleep and their day-to-day variation in affect), and a random level 1 effect (i.e. whether there were significant between-person differences in the strength of the day-to-day sleep and affect relationship); step 7: a reduced form equation including only Occasion, TWTs and SQR was conducted (because only these predictors were significant at all levels in both equations). This test was conducted to verify that the significant predictors in the preceding model were not conditional on the inclusion of the other, non-significant predictors.
Table 1 shows the means and standard deviations for the objective and subjective sleep variables. Nineteen participants reported insomnia consistent with the American Academy of Sleep Medicine Work Group’s published Research Diagnositic Criteria for insomnia (Edinger et al., 2004). Average duration of insomnia was 8.41 years (SD = 11.33; range .5–50).
Table 1. Means and standard deviations of objective and subjective sleep variables
TWT, total wake time; SQR, sleep quality rating. A subscript ‘s’ indicates the variable was measured subjectively (Sleep Diary); a subscript ‘o’ indicates it was measured objectively (Actigraphy).
Multicollinearity is a situation in which the data from two or more variables are perfectly or near perfectly correlated (Pedhazur, 1997). In the presence of multicollinearity, slight fluctuations in the data may lead to substantial fluctuations in the sizes of such estimates and may even change their signs. Consequently, ‘multicollinearity may have devastating effects on regression statistics to the extent of rendering them useless or highly misleading (Pedhazur, 1997).’
Because this study assessed sleep repeatedly, and used two different methods (sleep diaries and actigraphy), issues of multicollinearity were a concern. Table 2 shows the bivariate correlations of the predictor variables at both the between- and within-person levels. These values were obtained by estimating a multivariate mixed-effects null model in SPSS 14.0. This procedure produced both ‘G’ (between-persons) and ‘R’ (within-persons) covariance matrices, which were subsequently rescaled into correlations.
Table 2. Correlations among predictor variables and positive and negative affect between-persons (below diagonal) and within-persons (above diagonal)
TWT, total wake time; SQR, sleep quality rating. A subscript ‘s’ indicates the variable was measured subjectively (Sleep Diary);a subscript ‘o’ indicates it was measured objectively (Actigraphy).
*Correlation is significant at the 0.05 level (2-tailed).
**Correlation is significant at the 0.001 level (2-tailed).
4. Positive Affect
5. Negative Affect
In terms of absolute values, significant between-person bivariate correlations among predictor variables ranged from 0.20 to 0.50. TWTo was significantly, negatively correlated with SQR (r = −0.25, P < 0.05) only. In contrast, TWTs was significantly, negatively correlated to SQR and PA and positively correlated with NA (see Table 2). Both objective and subjective TWT were significantly, positively correlated.
Again, in terms of absolute values, significant within-person bivariate correlations among predictor variables ranged from 0.07 to 0.55. Consistent with the between-person analyses, there was evidence for convergent validity (though somewhat weaker) for subjective and objective TWT (see Table 2, upper off-diagonal). Likewise, there was a significant negative correlation between TWT (both objective and subjective) and SQR. TWTs was significantly, positively correlated with positive affect and negatively with negative affect, while TWTo was significantly, positively correlated with negative affect only. Collectively, these results suggest that some of the predictor variables shared up to 30% of their variance. Consequently, we orthogonalized SQR in the later models based on a regression model, effectively isolating the unique component of SQR that is independent of TWTs. The orthogonalized SQR represents the day-to-day variations in sleep quality that are not explained by day-to-day variations in total wake time. As with all regressions, regression parameter effects represent partial effects, and should be interpreted as the unique effects of the predictors, controlling for all other predictors.
MLM analysis and results
An initial step in MLM is determining how much variance can be accounted for by improving upon the null (baseline) model of predicting affect with sleep variables; which includes estimates of level 1 and level 2 variability. Intraclass correlation coefficients (ICC) serve as an index of these variability estimates (Bryk and Raudenbush, 1992). For positive affect, 27% of the overall variability was within-person and 73% was between-person. For negative affect, 48% was within-person and 52% was between-person. Therefore, our initial analyses revealed that there was a significant amount of variability in both level 1 and level 2 estimates which could be explained by our models.
Sleep and positive affect
Table 3 shows the model fit results for each added predictor block. The left side of the figure shows results for positive affect. Each modeling step resulted in a reduction in −2LL for the model, indicating better fit. However, in step 2 (fixed level 2 effect of gender), step 3 (fixed level 2 effect of mean TWTo; fixed and random level 1 effect of centered daily TWTo), and step 7 (reduced-form equation), changes in −2LL were negligible (when a nested χ2-test was used), and did not represent a significant change in fit.
Table 3. Model construction procedure
*Reduced model retained Time, TWTs, and SQR only.
−2LL, −2 log likelihood; Δ−2LL, change in −2LL relative to preceding model; , unexplained intercept-related (between subjects) variance; , unexplained residual-related (within subjects) variance; , between-subjects pseudo R-squared, an estimate of the amount of between subjects variance (estimated from null model) explained by fixed and random predictors; , within-subjects pseudo R-squared, an estimate of the amount of within subjects variance (estimated from null model) explained by fixed and random predictors.
(2) Time added
(3) Gender added
(4) TWTo added
(5) TWTS added
(6) SQR added
(7) Reduced form*
The final model, the reduced-form step 7, which retained only the variables which were significant (generally at all levels) in the preceding step 6, is summarized in Table 4 (left panel). The level 1 (within person) results suggested that only TWTs and SQR had significant within-person associations with positive affect. On those days where perceived TWT was above-average, or the residualized SQR was below-average, positive affect was also lower. The level 1 residualized SQR effect was qualified by a random effect, suggesting that there were significant individual differences in the strength of the residualized SQR-positive affect relationship. In addition, while there was no overall linear temporal trend in positive affect, a significant random effect for occasion (day) suggested that there were individual differences in the 14-day affect trend.
Table 4. Sleep variables predicting positive and negative affect
Fixed effects – predictor variable
Random effects – Covariance parameter estimate
Within pseudo R2
Within pseudo R2
Between pseudo R2
Between pseudo R2
With regard to the level 2 effects, there was a single significant effect of TWTs, suggesting that persons with higher average self-reported wake times experienced lower positive affect. The level 2 model explained about 28% of the between person variation in positive affect, and the level 1 model explained about 18% of the within person (daily) variation in positive affect. Finally, different repeated measures error structures were tested, but had little effect on fixed and random parameter estimates or their pattern of significance (Singer and Willett, 2003).1
Sleep and negative affect
Model fit results for negative affect are shown on the right side of Table 3. As for positive affect, each modeling step resulted in a reduction in −2LL for the model, indicating better fit. However, in step 2 (fixed level 2 effect of gender), step 3 (fixed level 2 effect of mean TWTo; fixed and random level 1 effect of centered daily TWTo), and step 7 (reduced-form equation), changes in −2LL were negligible (when a nested χ2 test was used), and did not represent a significant change in fit.
The right side of Table 4 shows the final model, step 7. The level 1 (within person) results suggested that there was a significant negative linear temporal trend in negative affect (such that persons became more negative over time). This time effect was qualified by a significant random effect, suggesting that there were individual differences in negative affect change over time. As with positive affect, both the level 1 effects of TWTs and residualized SQR were significant; on days where persons had above-average TWTs or below-average residualized SQR, negative affect was higher. There were significant individual differences in this TWTs-negative affect relationship, as qualified by a significant random effect. The level 2 model suggested that persons with higher TWTs, and lower residualized SQR reported higher negative affect at the beginning of the study. It should be noted that in step 6 there was also a significant negative effect of TWTo, such that persons with lower mean TWTo had higher negative affect. The step 7 model shows, however, that this term could be removed without a significant loss of model fit; it probably represented a statistical artifact (i.e. suppressor effect). In this final reduced-form model, the level 2 model explained about 26% of the between person variation in negative affect, and the level 1 model explained about 3% of the within person (daily) variation in negative affect. As with positive affect, different error structures were tested, but had little effect on fixed and random parameter estimates or their pattern of significance (Singer and Willett, 2003).2
Only the subjective measures of sleep exhibited daily associations with positive affect. Specifically, on days on which individuals reported above average sleep quality (shared variance with total wake time controlled) or below average wake time, they also reported experiencing higher levels of positive affect. For sleep quality, this was qualified by a significant random effect, suggesting that the average daily association between positive affect and residualized sleep quality rating varied between individuals. At the between persons level, individuals who reported greater wake time during the night also experienced lower positive affect.
Similar to positive affect, only the subjective measures of sleep exhibited daily associations with negative affect. On days on which individuals reported below average sleep quality (shared variance with total wake time controlled) or above average wake time, they also reported experiencing higher levels of negative affect. For total wake time, this relationship was qualified by a significant random effect, suggesting the average daily association between negative affect and reported wake time varied between individuals. At the between persons level, both the objective and subjective measures of sleep were associated with negative affect. Thus, on average, persons with greater subjective wake time during the night and lower sleep quality experienced greater negative affect already at the beginning of the study.
Overall, our results support daily associations between affect and subjective sleep only. These associations were in the hypothesized direction, indicating that nights with lower residualized quality ratings or greater reported wake time were associated with days characterized by less positive affect and greater negative affect. Interestingly, despite previous research suggesting gender differences in the sleep–affect relationship, gender was not associated with either positive or negative affect. Additional analyses (not shown here) further confirmed that gender did not moderate any of the effects of sleep on affect in the present study.
Our results extend the existing literature in this area in several important ways. First, we examined both subjective and objective sleep over a relatively long time period; whereas previous researchers examined affect in relation to either subjective (Totterdell et al., 1994) or objective (Berry and Webb, 1983, 1985) sleep alone. Because subjective and objective sleep measures are poorly correlated for some individuals (Espie et al., 1989; McCrae et al., 2005; Means et al., 2003), daily associations between subjective sleep and affect apparently differ significantly from those between objective sleep and affect. Second, we included a broader range of older adults (60–89 years) than did Berry and Webb, increasing the generalizability of our results. Third, the inclusion of both men and women allowed for further exploration of the interesting gender differences found in Berry and Webb’s (1983) initial study. This study failed to support the importance of gender in predicting affect or moderating sleep’s effects on affect.
The subjective sleep–affect relationships identified in this study are generally consistent with those identified by Totterdell et al. (1994). Our failure to find daily associations between objective sleep and affect is inconsistent with the earlier study of Berry and Webb (1983). However, they are generally consistent with their follow up work (Berry and Webb, 1985) which found associations between affect and only two out of 18 polysomnographic sleep variables studied (sleep efficiency and latency to first REM period). They reported total sleep time (but not total wake time), and it was not significantly associated with their measure of affect. It is important to note that direct comparison of the Berry and Webb’s studies with the present study is complicated by several important methodological differences. Specifically, sleep and affect were measured concurrently using 14 days of actigraphy and affect data in the present study compared with only 1 and 3 days of polysomnography and affect data, respectively (Berry and Webb, 1983, 1985).
The observed differences between affect and actigraphic versus self-reported sleep data in the present study are not surprising. The importance of utilizing different types of measurement to fully understand and explain sleep has been well-documented, and discrepancies between self-report and polysomnographic measures of sleep have been identified. Such discrepancies are typically greater for individuals with insomnia than they are for normal sleepers (Espie et al., 1989; Means et al., 2003). For many individuals, sleep perceptions are more important in terms of sleep satisfaction than are objective sleep parameters. For individuals with insomnia, cognitive-behavioral therapy for insomnia (CBTi) often improves their perceptions of specific sleep–wake parameters, including total wake time, as well as their overall satisfaction with sleep despite the fact that improvements in objective sleep parameters are generally much smaller or in some cases, not significant (Morin et al., 1999; Murtagh and Greenwood, 1995). Harvey’s cognitive model (Harvey, 2002) of insomnia captures the importance of perceptions for sleep. According to this model, insomnia is maintained by excessively negatively toned cognitive processes (attention and perception). One implication of the daily subjective sleep–affect associations identified in this study is that treatment of insomnia should not focus solely on improving sleep perceptions per se, but should also focus on reducing negative affect and increasing positive affect. Because the present study utilized a community-based sample, future research examining the sleep–affect link in the context of insomnia is warranted. In particular, whether insomnia explains the residual variance between people in the relationship between self-reported sleep and affective outcomes would be an important hypothesis to test. Additionally, exploration of a potential causal connection between sleep and affect would be of particular interest.
Sleep and emotional inhibition
Mood regulation refers to a basic drive to feel more pleasant than unpleasant affect (Erber and Erber, 2001; Larsen, 2000). Our finding of an association between poorer sleep and greater emotional disregulation (e.g. less positive and more negative affect) is consistent with Totterdell et al. (1994) and Berry and Webb (1983, 1985). Interestingly, our results contradict evidence suggesting that poorer sleep (e.g. decreased stages 3 and 4 and increased stage 1) is associated with alexithymia (a condition for which a key symptom is the inability to identify emotions; Bazydlo et al., 2001). However, it is important to note that these contradictory findings have not always been replicated (De Gennaro et al., 2002, 2004). Thus, questions about the relationships between quality of sleep, affective regulation, and emotional inhibition remain.
Potential limitations of the current study deserve mention. For example, participants were generally healthy, community-dwelling older adults who volunteered to participate. Therefore, the generalizability of these results to a clinical population of older adults may not be straightforward. Affect was measured only once per day (in the morning). Thus, it is possible that the relationship between sleep and affect may differ with multiple measurements and/or measurement later in the diurnal cycle. Because affect and subjective sleep were assessed using ‘paper and pencil’, it is possible that shared method variance contributed to their association. Additionally, only one measure of affect was used in the present study. While the PANAS is both highly reliable and valid (Watson et al., 1988), the inclusion of multiple measures of affect may have provided additional information. Furthermore, because mood is highly correlated with other subjective measures, such as alertness, sleepiness, and fatigue, it is unclear whether the results of the present study could be interpreted solely in the terms of general subjective well-being rather than specifically in terms of affect. Thus, future research employing a broader range of subjective daytime functioning measures is needed. Also, the current study lacks in the measurement of possible covariates of the sleep–affect relationship. Variables such as physical activity (Montgomery and Dennis, 2004; Pretty et al., 2005) and daily stress (Morin et al., 2003) that are known to be associated with both sleep and affect could prove useful in interpretation of the results. Consequently, future studies might assess a broader range of daytime functioning such as physical activity (Montgomery and Dennis, 2004; Pretty et al., 2005) and daily stress (Morin et al., 2003). Finally, the design of the study prohibits any causal conclusions.
Our results highlight the need for greater focus on intraindividual variability in sleep research and reinforce the importance of perceptions–not only in terms of sleep–wake parameters, but also in terms of affect and other daytime functioning factors. Finally, our results suggest exploration of affect as a potential intervention target in late-life insomnia is warranted.
Preliminary results from this study were presented in December 2006 at the annual meeting of the Gerontological Society of America, Dallas, TX.
Level 1 and level 2 equations for each model.
Level 1 equation
Level 2 equation
For each model step, following the conventions of hierarchical linear modeling (Bryk and Raudenbush, 1992) a level 1 (prediction of within-person processes) and a level 2 (prediction of between-person differences) equation were estimated. In SPSS MIXED, level 1 and level 2 equations not separately specified, but we have retained this convention in presenting the models to allow comparability with other MLM research. In step 1, we specified an initial level 1 null model, specifying that affect for person j on day I is a function of his or her mean level of affect (β0j) and a random residual component (eij). The level 2 null model specifies that individuals’ level 1 coefficients (β0j) or average affect levels) reflect an overall grand mean (γ00) and a between-persons error term (u0j). In step 2, we estimated the unconditional growth model (to control for any general linear temporal trends in affect), allowing both fixed and random effects of time. In the level 1 model, the estimation of each person’s daily affect now included the effect of the day of measurement (β1j), and the level 2 model specified that each person’s day effect reflected an overall relationship (γ10) and between-person variation in the effect of day (u1j). In step 3, a fixed level 2 effect of Gender was added, such that each person’s mean affect (β0j) was estimated to be a partial function (γ01) of that person’s gender. In steps 4, 5 and 6, the fixed level 2 effect, and fixed and random level 1 effects of three sleep related variables (TWTo, TWTs, SQR, respectively) were added. In the level 1 model, each person’s daily variation in affect was now also modeled as a partial function of that day’s person-centered effect of TWTo (β2j), TWTs (β3j), and SQR (β4j). In the corresponding level 2 models, we specified that individuals’ level 1 coefficients (β2j, β3j, β4j, representing the person-level relationship between daily variation in affect and sleep) reflected an overall slope (γ20, γ30, γ40) and random between-person variation in that slope (u2j, u3j, u4j). Prior to running these models, to reduce multicollinearity between TWTs and SQR, a residualized SQR term was produced (with the corresponding day’s TWTs regressed out); this residualized SQR term is used in all models that follow. In the final step 7, we investigated reduced-form equations based on the step 6 model; only Occasion, TWTs and SQR were retained, because they were significant (generally at all levels) in the preceding models.
Reduced form equation: retain only significant terms from the model above:
Reduced form equations: retain only significant terms from the model above: