- Email: [email protected]
Psychometric evaluation of daytime sleepiness and nocturnal sleep onset scales in a representative community sample
Psychometric Evaluation of Daytime Sleepiness and Nocturnal Sleep Onset Scales in a Representative Community Sample Eric O. Johnson, Naomi Breslau, Thomas Roth, Timothy Roehrs, and Leon Rosenthal Background: The public health importance of daytime sleepiness as a risk factor for accidents, interpersonal problems, and decreased productivity has been recognized. However, epidemiologic research on this topic has been limited by the reliance on laboratory measures (i.e., the Multiple Sleep Latency Test—MSLT). Two scales, daytime sleepiness and nocturnal sleep onset, have been identified from the self-report Sleep–Wake Activity Inventory (SWAI) in a clinic sample and validated against the MSLT. This study evaluates the replicability of the two scales in a population sample and assesses potential thresholds in scale scores that distinguish normal from pathologic levels of daytime sleepiness and difficulty falling asleep. Methods: The sample consisted of 2181 subjects 18 – 45 years old in the Detroit metropolitan area. All sleep characteristic information covered the 2 weeks prior to interview. Split-half sample factor analyses were conducted to assess replicability of the results. Distribution of scale scores and their relation to construct validity variables were used to evaluate possible thresholds. Results: A two-factor model appeared to best account for the variation among the 12 items from the SWAI. The two factors accounted for 50% of the variance in both splithalf sample analyses. The revised eight-item daytime sleepiness and two-item nocturnal sleep onset scales showed good and fair internal consistency respectively across both split-half samples. There appeared to be a “natural break” in daytime sleepiness scale scores that was associated with a substantial and consistent change in number of hours slept. No breaks appeared in nocturnal sleep onset scores. Conclusions: This study replicated the results of the clinic-based study and suggested a potentially useful
From the Department of Psychiatry and the Sleep Disorder Center, Henry Ford Health Sciences Center, Detroit, Michigan (EOJ, NB, TRoth, TRoehrs, LR); Department of Psychiatry, Case Western Reserve University School of Medicine, Cleveland, Ohio (NB); and Department of Psychiatry, University of Michigan School of Medicine, Ann Arbor, Michigan (NB, TRoth). Address reprint requests to Eric O. Johnson, PhD, Henry Ford Health Sciences Center, Department of Psychiatry, 1 Ford Place, 3a, Detroit, MI 48202-3450. Received July 22, 1997; revised January 12, 1998; revised February 19, 1998; accepted February 23, 1998.
© 1999 Society of Biological Psychiatry
diagnostic threshold for self-report excessive daytime sleepiness. Epidemiology of sleep depends on the ability to move from the laboratory to population surveys in reliable and valid ways. Development of self-report is a step in that direction. Biol Psychiatry 1999;45:764 –770 © 1999 Society of Biological Psychiatry Key Words: Daytime sleepiness, sleep problems, Sleep– Wake Activity Inventory
Introduction
A
lthough everyone experiences occasional daytime sleepiness resulting from insufficient sleep, the clinical and public health importance of chronic excessive daytime sleepiness has been increasingly recognized. Excessive daytime sleepiness has been associated with increased risk of motor vehicle and industrial accidents, decreased productivity, and interpersonal problems (Roth et al 1989). In spite of growing public health concerns about the effects of excessive daytime sleepiness, there has been little epidemiologic research on its prevalence and etiology in the U.S. population. Several community-based studies of Scandinavian populations have been done. In a study of the Finnish population, 11% of women and 7% of men reported daytime sleepiness almost every day (Hublin et al 1996). In another survey, of a large area of Sweden, insufficient sleep was the focus of the questions, and 12% of the respondents felt their sleep was insufficient— leading to daytime sleepiness (Broman et al 1996). However, none of the community-based studies have used validated self-report measures of excessive daytime sleepiness. Indeed, one reason for the scarcity of epidemiologic studies has been the need to rely on laboratory measures for valid estimates of excessive daytime sleepiness. Specifically, the Multiple Sleep Latency Test (MSLT), which measures time to sleep by a polysomnogram at 2-hour intervals throughout the day following a forced 8 hours in 0006-3223/99/$19.00 PII S0006-3223(98)00111-5
Daytime Sleepiness and Nocturnal Sleep Onset
bed the prior night, has become the gold standard of daytime sleepiness (Carskadon et al 1986). The reliability and validity of the MSLT have been well documented (Carskadon and Dement 1982). However, the burden of a minimum 24-hour stay in a sleep laboratory, the accompanying self-selection of research subjects, and the high research costs have limited the utility of this method in large epidemiologic studies. Two self-report measures of daytime sleepiness have been developed and validated relative to the MSLT, the Epworth Sleepiness Scale and the Sleep–Wake Activity Inventory (SWAI) (Johns 1991; Rosenthal et al 1993). The Epworth and SWAI ask subjects to rate their likelihood of falling asleep during the day in different situations. The SWAI also assesses an important correlate of daytime sleepiness, difficulty of falling asleep at night or nocturnal sleep onset. Assessing the SWAI in a clinic sample (76% of whom reported sleep problems), Rosenthal et al (1993) identified six scales using a principal components analysis. Two scales, a daytime sleepiness scale (DSS) and a nocturnal sleep onset scale (NSOS) significantly predicted MSLT scores (p , .05). For the DSS significant mean differences were found between those with mean MSLT scores of 5 min or less, 6 –10 min, and greater than 10 min, both for those with and those without sleep complaints. The purpose of the present study was to assess the daytime sleepiness and nocturnal sleep onset scales in a representative sample of the population. Two primary questions were addressed: 1) do the daytime sleepiness and nocturnal sleep onset scales identified from data on a clinic sample reliably replicate in a population representative sample; and 2) are there “naturally occurring” thresholds in these scales that may help identify those in the general population with pathological levels of daytime sleepiness or difficulty with nocturnal sleep onset?
Methods and Materials
765
BIOL PSYCHIATRY 1999;45:764 –770
Table 1. Sociodemographic Characteristics (N 5 2181)
Sex Male Female Race White Nonwhite Age (years) 18 –25 26 –35 36 – 45 Marital status Married Separated/divorced/widowed Never married Education ,High school High school graduate Some college 4-year college graduate1 Employment Employed Not employed Income ,10,000 10,000 –14,999 15,000 –19,999 20,000 –24,999 25,000 –34,999 35,000 – 49,999 50,000 –74,999 75,0001 County Lapeer Macomb Monroe Oakland St. Clair Wayne City of Detroit Other
n
Unweighted %
U.S. Census %a
1107 1074
50.8 49.2
48.6 51.4
1549 608
71.8 28.2
76.3 23.7
488 838 843
22.5 38.6 38.9
26.2 39.7 34.1
972 335 868
44.7 15.3 39.9
49.8 12.9 37.3
166 744 707 563
7.6 34.1 32.4 25.8
16.2 29.6 36.3 17.9
1726 445
79.5 20.5
72.3 27.6
178 155 166 170 290 385 396 322
8.6 7.5 8.1 8.2 14.1 18.7 19.2 15.6
9.3 4.0 5.5 5.3 13.6 21.8 24.9 15.7
58 416 65 547 101 994 464 530
2.7 19.1 3.0 25.1 4.6 45.6 21.3 24.3
1.7 16.9 3.1 26.4 3.3 48.7 23.5 25.2
a U.S. Bureau of the Census: Public Use Microdata Sample—1% sample (1990).
Sample The 1996 Detroit Area Survey collected information from a representative sample of 2181 people 18 – 45 years of age in the Detroit primary metropolitan statistical area. The survey used random digit dialing and computer-assisted telephone interviewing techniques. Of the 6110 households identified, 76.2% completed screening, and 64.1% of screened households contained an age-eligible respondent. An intensive recruitment effort with financial incentives was made in a randomly selected subsample of initial nonrespondents. Thirty percent of this subsample of initial nonrespondents (n 5 295) were interviewed. These respondents were added to the final sample. The overall response rate among eligible households of was 86.8%. This response rate was weighted to adjust for the underrepresentation of initial nonresponders.
Sample Representativeness Table 1 presents the sociodemographic characteristics of the sample and estimates of those characteristics for the primary metropolitan statistical area of Detroit from the 1990 U.S. Census. Since sampling and poststratification weights could not be used in the evaluation of items in the factor analyses, sample percentages are shown unweighted. Unweighted, the sample remained highly representative of the population from which it was drawn. Most sociodemographic characteristics of the sample and the Census estimates were within a few percent. However, relative to the population, the sample had a lower percent not completing high school, a higher percent completing college, and a greater percent employed.
766
E.O. Johnson et al
BIOL PSYCHIATRY 1999;45:764 –770
Table 2. Distributions of Daytime Sleepiness and Nocturnal Sleep Onset Self-Ratings during the Past 2 Weeks (N 5 2181)
Daytime sleepiness items I fell asleep when riding as a passenger I dozed off while watching TV I got drowsy within 10 min when I sat still I fell asleep when visiting with friends I could nap anywhere I fell asleep during conversations I got drowsy after driving a few minutes I got sleepy after reading for 15 min I dozed off when I relaxed Nocturnal sleep onset items Even if I took a nap, I slept well at night I had difficulty falling asleep It took me less than 5 min to fall asleep
Never
Rarely
Sometimes
Often
Missing
64.6 32.2 44.5 89.5 56.2 91.3 65.9 38.5 29.0
14.5 17.9 24.1 6.6 18.1 5.5 14.6 14.3 20.4
14.4 28.5 21.8 3.1 13.5 2.5 14.7 29.4 35.0
6.3 21.0 9.5 0.8 12.0 0.6 4.7 17.6 15.3
0.2 0.4 0.1 0.0 0.2 0.0 0.1 0.2 0.3
34.3 42.0 28.7
9.2 20.0 17.0
16.4 23.4 25.1
39.2 14.4 28.7
0.9 0.2 0.6
Measures The self-report measures of daytime sleepiness and nocturnal sleep onset examined in this study were originally developed as part of the SWAI by Rosenthal et al (1993). The 59-item SWAI was designed to measure several other dimensions as well, including psychic distress, social desirability, and behavioral activation. Only the nine items identified as measuring daytime sleepiness and the three measuring nocturnal sleep onset (i.e., difficulty falling asleep at night) were included in these analyses (see Table 2). Respondents rated items for the 2-week period preceding the interview on a four-point ordinal scale: never, rarely, sometimes, or often. The average amount of sleep gotten, the typical time gone to bed and gotten up from bed, and whether the respondent took naps were separately assessed in the study interview. Each question concerned weekdays for the 2 weeks prior to interview. Sleep efficiency was calculated by dividing the average amount of sleep by the time spent in bed (derived from average time gone to bed and gotten up from bed reports).
Factor Analyses To assess whether the proposed self-report items reliably measured dimensions of daytime sleepiness and nocturnal sleep onset, factor analyses were conducted. The sample was randomly split into two equal-sized subsamples. Repeating the factor analysis in each subsample provided a measure of the stability of the solutions, in terms of number of factors and the relationship of specific items with the factors. This strategy reduced the risk that final results were effected by chance features of a particular sample. Since the items were measured on a four-point ordinal scale, standard principle components and factor analytic techniques based on Pearson product–moment correlation coefficients were inappropriate. The LISCOMP computer program (Muthen 1988), which provides factor analysis based on the appropriate polychoric correlations, was used in these analyses. Factor models were estimated using unweighted least squares (ULS). When a large proportion of items are highly skewed, ULS is a more stable estimator than a generalized least squares, although it does
not generate a chi-square test of model fit (Muthen 1989). The number of factors was determined by examining the scree plot of eigenvalues, the relative change in the root mean square residual, and the stability of results across the two subsamples. Promax rotation, which allows for correlated factors, was used in these analyses due to the expected inverse association between daytime sleepiness and nocturnal sleep onset. Factor analysis requires complete data on the evaluated items for each subject. Fifty-three cases (2.4% of the total sample) were deleted because of missing data on one or more of the 12 items. Incomplete cases did not differ from complete cases in sex, race, marital status, level of education, or income (p values from .23 to .96). Incomplete cases did have a slightly higher mean age (33.8 years, SD 5 7.0) than complete cases (31.6 years, SD 5 7.7, p 5 .04). The means for items among subjects with incomplete data who did respond to a particular item did not differ significantly from the means for those subjects with complete data (p . .05). Thus the missing data appeared to be largely at random, and elimination of cases without complete data were unlikely to bias results (Little and Schenker 1995).
Reliability Analysis Analyses of the internal consistency of the proposed scales derived from the factor analyses were conducted using Cronbach’s alpha. Because sampling variability and measurement error often cause specific item lambdas (factor loadings) to vary from sample to sample, we used the conservative scale scoring strategy of selecting those items that loaded strongly and consistently on a factor and adding the item score (0 –3) to a subject’s scale score for each item. Reliability analyses were conducted as a means to assess the appropriateness of using factors as simple sum scales, that is, without weighting for factor loadings. These analyses were conducted on the same split-half samples as the factor analyses.
Threshold Analysis The Multiple Sleep Latency Test, the gold standard of daytime sleepiness assessment, is a continuous measure of the average
Daytime Sleepiness and Nocturnal Sleep Onset
amount of time it takes people to fall asleep over 5 trials within a day. A threshold of 5 min or less has been identified as an indicator of a clinically important level of daytime sleepiness. The distribution of scale scores were examined for “natural breaks,” which may indicate thresholds in the self-report measures analogous to the 5-min threshold for the MSLT. Potential threshold values for the DSS were examined against the reported average number of hours slept on weekdays during the past 2 weeks. Daytime sleepiness is in large part a function of the amount of sleep one has had recently (Roth et al 1989). Thus, a measure such as the DSS should be related to average sleep time, and nonlinear changes in this relationship may reveal clinically significant natural breaks that separate normal from pathologic levels of daytime sleepiness. A similar approach was taken to assess whether there was a threshold in the distribution of NSOS scores that would indicate abnormal levels of difficulty in falling asleep. For this analysis potential threshold values of the NSOS where evaluated against sleep efficiency (sleep time/total time in bed). Difficulty falling asleep should be inversely related to the ratio of sleep time to total time in bed: the more difficulty falling asleep the lower the sleep efficiency. Again, nonlinear changes in this relationship may indicate a diagnostic threshold.
Results Item Response Distributions of responses to the nine daytime sleepiness and three nocturnal sleep onset items are shown in Table 2. For the nine daytime sleepiness items, the more frequent the occurrence the higher the level of daytime sleepiness. Greater difficulty falling asleep at night (nocturnal sleep onset) was indicated by less frequent occurrence of the items “Even if I took a nap, I slept well at night” and “It took me less than 5 minutes to fall asleep,” but more frequent occurrence of the item “I had difficulty falling asleep.” Response distributions for nine of the items were significantly skewed (p # .05). The three items not skewed were: 1) “I dozed off while watching TV,” 2) “I dozed off when I relaxed,” and 3) “It took me less than 5 minutes to fall asleep.” Almost all of the skewed items showed a skew to the right, indicating relative infrequent occurrence of the daytime sleepiness or sleep difficulty manifestations during the prior 2-week period. Occurrence of two indicators of daytime sleepiness were particularly infrequent: falling asleep when visiting friends or during conversations.
Factor Analyses The scree plot of eigenvalues for each of the split-half samples indicated two but not more than three factors could be extracted from these items (see Figure 1). The scree in each analysis began with factor three, and the
BIOL PSYCHIATRY 1999;45:764 –770
767
Figure 1. Scree plot of latent roots from split-half factor analysis (n 5 1062; n 5 1066). The number of subjects in split-half subsamples do not sum to the total subject number (2181) due to missing data; see Factor Analyses portion of Methods section.
eigenvalues for the third factor were approximately one. The two-factor solution was chosen over the three-factor solution as the most reliable and parsimonious model. The three-factor solution was less stable across the two splithalf subsamples and only marginally improved fit [root mean square residual (rmsr) reduced by less than .020]. Additionally, the third factor was substantially correlated with the first factor in each subsample (r 5 .62 and .69). The two-factor solution fit the data reasonably well across the two split-half samples (see Table 3). The root mean square residuals were consistent and low, .058 and .056, respectively. Nearly all of the items loaded strongly on one factor and weakly on the other. In both subsamples these two factors were modestly correlated and accounted for nearly 50% of the variance of the items. The first factor was well measured by eight of the items of the daytime sleepiness scale. The item “I could nap anywhere” did not appear to be a reliable indicator of daytime sleepiness, as its highest loading switched factors across the two subsamples. The second factor was well measured by two of the three nocturnal sleep onset items. The item “Even if I took a nap, I slept well at night” loaded on both factors to a moderate degree and thus was an unreliable indicator.
Reliability The eight items that consistently measured the daytime sleepiness factor showed good reliability as measured by Cronbach’s alpha in each split-half sample (a 5 .71 and .69). The addition of the item “I could nap anywhere,” which was deleted based on the factor analyses, did not improve reliability (a 5 .73 and .69). Reliability of the nocturnal sleep onset scale was improved in each split-half
768
E.O. Johnson et al
BIOL PSYCHIATRY 1999;45:764 –770
Table 3. Factor Analysis Results: Two-Factor Solutions for Split-Halves (n 5 1062, n 5 1066)a First split-half factor loadings
Second split-half factor loadings
In the past 2 weeks . . .
Factor 1
Factor 2
Factor 1
Factor 2
I fell asleep when riding as a passenger I dozed off while watching TV I got drowsy within 10 min when I sat still I fell asleep when visiting with friends I could nap anywhere I fell asleep during conversations I got drowsy after driving a few minutes I got sleepy after reading for 15 min I dozed off when I relaxed Even if I took a nap, I slept well at night I had difficulty falling asleep It took me less than 5 min to fall asleep
.595 .535 .585 .687 .304 .719 .517 .432 .659 2.208 .264 .006
2.059 2.049 2.012 2.027 .498 2.002 .005 .031 .008 2.418 2.691 2.753
.607 .581 .646 .748 .514 .705 .533 .495 .673 2.315 .220 2.076
2.023 2.015 2.010 2.030 .276 2.087 2.142 2.011 2.013 2.319 2.786 2.656
First-half two-factor solution: root mean square residual 5 .058; factor correlation 5 .22; R2 5 .47. Second-half two-factor solution: root mean square residual 5 .056; factor correlation 5 .19; R2 5 .49. a Number of subjects in split-half subsamples do not sum to total subject number (2181) due to missing data; see Factor Analyses portion of Methods section.
sample (a 5 .58 and .60) by dropping the item “Even if I took a nap, I slept well at night,” leaving only two items.
Thresholds Figure 2 presents distributions of the eight-item DSS and two-item NSOS scores; each item was scored 0 –3 from “never” to “always” and then summed. For the DSS a
“natural break” in the distribution of scores appeared to occur at a score of 11. The full distribution of DSS scores (0 –24) showed a significant positive skew (p , .05). However, DSS scores 0 –10 were normally distributed (no significant skew or kurtosis p . .05). Thus, it is the tail of the distribution, scores of 11 and higher, that produced the significant skew of the total scale distribution (p , .05). Approximately 17% of the sample had DSS scores of 11 or higher. No obvious break occurred in the distribution of NSOS scores, although a score of 4 or higher may constitute such a threshold. The top panel of Figure 3 shows box plots of average number of hours slept on weekdays in the past 2 weeks by DSS score. For DSS scores 0 –10, the median hours slept, indicated by the line through each box, was, with one exception, uniform at 7 hours. DSS scores of 11 or higher were all below 7 hours, generally at 6 hours of sleep. The lower panel of Figure 3 shows boxplots of sleep efficiency (hours slept/total time in bed) by NSOS scores. No thresholdlike change occurred; the relationship appeared largely linear. The correlation (r) between NSOS and sleep efficiency was 2.27 (p , .001): the greater one’s difficulty falling asleep the lower one’s sleep efficiency.
Discussion
Figure 2. Distributions of revised daytime sleepiness and nocturnal sleep onset scales.
A two-factor model appeared to best account for the variation in the 12 items from the SWAI across the split-half samples. The first factor was well measured by eight of the nine items proposed to measure daytime sleepiness. The second factor was well measured by two of the three items that constitute the nocturnal sleep onset scale. These two factors accounted for approximately 50% of the variance in both split-half sample analyses. The
Daytime Sleepiness and Nocturnal Sleep Onset
BIOL PSYCHIATRY 1999;45:764 –770
769
topic of napping. Responses to the questions “I could nap anywhere” and “Even if I took a nap, I slept well at night” were found to differ significantly between those who did and did not take naps during the 2-week period covered by the questions (x2 5 89.7, df 5 3, p , .001; x2 5 440.5, df 5 3, p , .001, respectively). Those who did not nap tended to report that they never or rarely could “nap anywhere,” while those who did nap tended to report that they sometimes or often could “nap anywhere.” There was a similar but reversed discrepancy between those who did and did not take naps in response to “Even if I took a nap, I slept well at night”; nonnappers indicated sometimes or often, while nappers indicated never or rarely did they sleep well despite napping. Having subjects respond to these questions even if they did not take a nap appears to be the reason for the unreliability of the measures. It was statistically and rationally appropriate to drop these items from future use of the DSS.
Threshold of Daytime Sleepiness
Figure 3. Boxplots of construct validity variables across daytime sleepiness and nocturnal sleep onset scale scores.
eight-item daytime sleepiness scale and the two-item nocturnal sleep onset scale showed good internal consistency. The items that showed inconsistent factor analysis results (“I could nap anywhere” and “Even if I took a nap, I slept well at night”) weakened or did not improve the internal consistency of the proposed scales. There appeared to be a “natural break” or threshold in daytime sleepiness scale scores that was associated with a substantial and consistent change in number of hours slept, indicating a possible point of distinction between normal and abnormal levels of daytime sleepiness. No breaks or thresholds appeared in nocturnal sleep onset scale scores.
The revised DSS had range of 0 –24. The distribution of scores within that range indicated an empirical “natural break” between scores from 0 –10 and 11 or greater. The distribution of scores 0 –10 appears to be normal (no significant skew or kurtosis), while that of scores of 11 or higher produced the significant skewedness for the whole distribution. This proposed threshold showed construct validity. Because nocturnal sleep is related to daytime sleepiness (Roth et al 1989), a threshold for some measure of daytime sleepiness should show differences in average amount of sleep. As demonstrated by the boxplots in Figure 3, a substantial and consistent shift in median number of hours slept takes place between scores of 10 and 11. In this study we could not compare the DSS, nor the proposed threshold, to MSLT values. Although a one-toone correspondence between the DSS and the MSLT is not expected, it is of interest that the threshold of pathological levels of daytime sleepiness generally accepted for the MSLT (#5-min sleep latency) describes 10 –15% of normal healthy clinic-ascertained volunteers (Levine et al 1988), and that the proposed threshold in the DSS identifies approximately 17% of the population as potentially having abnormal levels of daytime sleepiness. Future epidemiologic studies that also collect laboratory data are necessary to test the validity of this threshold.
No Threshold for Nocturnal Sleep Onset SWAI Items Dropped A characteristic shared by the two items from the DSS that were dropped from the scale due to unreliability was the
The revised NSOS had a range of 0 – 6. The distribution of scores within that range did not show evidence of a “natural break” indicative of a threshold. Consistent with
770
E.O. Johnson et al
BIOL PSYCHIATRY 1999;45:764 –770
the lack of a clear threshold in the distribution of NSOS scores, the relationship between sleep efficiency and NSOS appeared to be linear in the boxplot analysis. The correlation (r) between NSOS and sleep efficiency was 2.27, p , .001.
Limitations The most significant limitation of this study was that MSLT data could not be used to assess directly the validity of the DSS. However, the study that originally derived the DSS did indicate that the MSLT “gold standard” was related in expected ways to this self-report measure, albeit in a clinic-ascertained sample of which 76% reported sleep problems.
Conclusion This study largely replicated the results of the clinic-based study that identified self-report scales from the SWAI that appeared to measure daytime sleepiness and nocturnal sleep onset. The further development of an epidemiology of sleep, including daytime sleepiness and sleep disturbances, depends on the ability to move from the laboratory results to population surveys in reliable and valid ways. Development of self-report measures such as the daytime sleepiness scale and nocturnal sleep onset scale is a step in that direction. Supported in part by grant MH8802 from the National Institute of Mental Health and HL42215 from the National Institutes of Health.
References Broman JE, Lundh LG, Hetta J (1996): Insufficient sleep in the general population. Neurophysiol Clin 26:30 –39. Carskadon M, Dement W (1982): Nocturnal determinants of daytime sleepiness. Sleep 5:S73–S81. Carskadon M, Dement W, Mitler M, et al (1986): Guidelines for the multiple sleep latency test (MSLT): A standard measure of sleepiness. Sleep 9:519 –524. Hublin C, Kaprio J, Partinen M, Heikkila K, Koskenvuo M (1996): Daytime sleepiness in an adult Finnish population. J Intern Med 239:417– 423. Johns M (1991): A new method for measuring daytime sleepiness: The Epworth Sleepiness Scale. Sleep 14:540 –545. Levine B, Roehrs T, Zorick F, Roth T (1988): Daytime sleepiness in young adults. Sleep 11:39 – 46. Little R, Schenker N (1995): Missing data. In: Arminger G, Clogg C, Sobel M, editors. Handbook of Statistical Modeling for the Social and Behavioral Sciences. New York: Plenum Press, pp 39 –76. Muthen B (1988): LISCOMP. Analysis of linear structural equations with a comprehensive measurement model: A program for advanced research, 2nd ed. Mooresville, IN: Scientific Software. Muthen B (1989): Dichotomous factor analysis of symptom data. Sociol Methods Res 18:19 – 65. Rosenthal L, Roehrs T, Roth T (1993): The sleep-wake activity inventory: A self-report measure of daytime sleepiness. Biol Psychiatry 34:810 – 820. Roth T, Roehrs T, Carskadon M, Dement W (1989): Daytime sleepiness and alertness. In: Kryger M, Roth T, Dement W, editors. Principles and Practice of Sleep Medicine. Philadelphia: Saunders, pp 14 –23.
From the Department of Psychiatry and the Sleep Disorder Center, Henry Ford Health Sciences Center, Detroit, Michigan (EOJ, NB, TRoth, TRoehrs, LR); Department of Psychiatry, Case Western Reserve University School of Medicine, Cleveland, Ohio (NB); and Department of Psychiatry, University of Michigan School of Medicine, Ann Arbor, Michigan (NB, TRoth). Address reprint requests to Eric O. Johnson, PhD, Henry Ford Health Sciences Center, Department of Psychiatry, 1 Ford Place, 3a, Detroit, MI 48202-3450. Received July 22, 1997; revised January 12, 1998; revised February 19, 1998; accepted February 23, 1998.
© 1999 Society of Biological Psychiatry
diagnostic threshold for self-report excessive daytime sleepiness. Epidemiology of sleep depends on the ability to move from the laboratory to population surveys in reliable and valid ways. Development of self-report is a step in that direction. Biol Psychiatry 1999;45:764 –770 © 1999 Society of Biological Psychiatry Key Words: Daytime sleepiness, sleep problems, Sleep– Wake Activity Inventory
Introduction
A
lthough everyone experiences occasional daytime sleepiness resulting from insufficient sleep, the clinical and public health importance of chronic excessive daytime sleepiness has been increasingly recognized. Excessive daytime sleepiness has been associated with increased risk of motor vehicle and industrial accidents, decreased productivity, and interpersonal problems (Roth et al 1989). In spite of growing public health concerns about the effects of excessive daytime sleepiness, there has been little epidemiologic research on its prevalence and etiology in the U.S. population. Several community-based studies of Scandinavian populations have been done. In a study of the Finnish population, 11% of women and 7% of men reported daytime sleepiness almost every day (Hublin et al 1996). In another survey, of a large area of Sweden, insufficient sleep was the focus of the questions, and 12% of the respondents felt their sleep was insufficient— leading to daytime sleepiness (Broman et al 1996). However, none of the community-based studies have used validated self-report measures of excessive daytime sleepiness. Indeed, one reason for the scarcity of epidemiologic studies has been the need to rely on laboratory measures for valid estimates of excessive daytime sleepiness. Specifically, the Multiple Sleep Latency Test (MSLT), which measures time to sleep by a polysomnogram at 2-hour intervals throughout the day following a forced 8 hours in 0006-3223/99/$19.00 PII S0006-3223(98)00111-5
Daytime Sleepiness and Nocturnal Sleep Onset
bed the prior night, has become the gold standard of daytime sleepiness (Carskadon et al 1986). The reliability and validity of the MSLT have been well documented (Carskadon and Dement 1982). However, the burden of a minimum 24-hour stay in a sleep laboratory, the accompanying self-selection of research subjects, and the high research costs have limited the utility of this method in large epidemiologic studies. Two self-report measures of daytime sleepiness have been developed and validated relative to the MSLT, the Epworth Sleepiness Scale and the Sleep–Wake Activity Inventory (SWAI) (Johns 1991; Rosenthal et al 1993). The Epworth and SWAI ask subjects to rate their likelihood of falling asleep during the day in different situations. The SWAI also assesses an important correlate of daytime sleepiness, difficulty of falling asleep at night or nocturnal sleep onset. Assessing the SWAI in a clinic sample (76% of whom reported sleep problems), Rosenthal et al (1993) identified six scales using a principal components analysis. Two scales, a daytime sleepiness scale (DSS) and a nocturnal sleep onset scale (NSOS) significantly predicted MSLT scores (p , .05). For the DSS significant mean differences were found between those with mean MSLT scores of 5 min or less, 6 –10 min, and greater than 10 min, both for those with and those without sleep complaints. The purpose of the present study was to assess the daytime sleepiness and nocturnal sleep onset scales in a representative sample of the population. Two primary questions were addressed: 1) do the daytime sleepiness and nocturnal sleep onset scales identified from data on a clinic sample reliably replicate in a population representative sample; and 2) are there “naturally occurring” thresholds in these scales that may help identify those in the general population with pathological levels of daytime sleepiness or difficulty with nocturnal sleep onset?
Methods and Materials
765
BIOL PSYCHIATRY 1999;45:764 –770
Table 1. Sociodemographic Characteristics (N 5 2181)
Sex Male Female Race White Nonwhite Age (years) 18 –25 26 –35 36 – 45 Marital status Married Separated/divorced/widowed Never married Education ,High school High school graduate Some college 4-year college graduate1 Employment Employed Not employed Income ,10,000 10,000 –14,999 15,000 –19,999 20,000 –24,999 25,000 –34,999 35,000 – 49,999 50,000 –74,999 75,0001 County Lapeer Macomb Monroe Oakland St. Clair Wayne City of Detroit Other
n
Unweighted %
U.S. Census %a
1107 1074
50.8 49.2
48.6 51.4
1549 608
71.8 28.2
76.3 23.7
488 838 843
22.5 38.6 38.9
26.2 39.7 34.1
972 335 868
44.7 15.3 39.9
49.8 12.9 37.3
166 744 707 563
7.6 34.1 32.4 25.8
16.2 29.6 36.3 17.9
1726 445
79.5 20.5
72.3 27.6
178 155 166 170 290 385 396 322
8.6 7.5 8.1 8.2 14.1 18.7 19.2 15.6
9.3 4.0 5.5 5.3 13.6 21.8 24.9 15.7
58 416 65 547 101 994 464 530
2.7 19.1 3.0 25.1 4.6 45.6 21.3 24.3
1.7 16.9 3.1 26.4 3.3 48.7 23.5 25.2
a U.S. Bureau of the Census: Public Use Microdata Sample—1% sample (1990).
Sample The 1996 Detroit Area Survey collected information from a representative sample of 2181 people 18 – 45 years of age in the Detroit primary metropolitan statistical area. The survey used random digit dialing and computer-assisted telephone interviewing techniques. Of the 6110 households identified, 76.2% completed screening, and 64.1% of screened households contained an age-eligible respondent. An intensive recruitment effort with financial incentives was made in a randomly selected subsample of initial nonrespondents. Thirty percent of this subsample of initial nonrespondents (n 5 295) were interviewed. These respondents were added to the final sample. The overall response rate among eligible households of was 86.8%. This response rate was weighted to adjust for the underrepresentation of initial nonresponders.
Sample Representativeness Table 1 presents the sociodemographic characteristics of the sample and estimates of those characteristics for the primary metropolitan statistical area of Detroit from the 1990 U.S. Census. Since sampling and poststratification weights could not be used in the evaluation of items in the factor analyses, sample percentages are shown unweighted. Unweighted, the sample remained highly representative of the population from which it was drawn. Most sociodemographic characteristics of the sample and the Census estimates were within a few percent. However, relative to the population, the sample had a lower percent not completing high school, a higher percent completing college, and a greater percent employed.
766
E.O. Johnson et al
BIOL PSYCHIATRY 1999;45:764 –770
Table 2. Distributions of Daytime Sleepiness and Nocturnal Sleep Onset Self-Ratings during the Past 2 Weeks (N 5 2181)
Daytime sleepiness items I fell asleep when riding as a passenger I dozed off while watching TV I got drowsy within 10 min when I sat still I fell asleep when visiting with friends I could nap anywhere I fell asleep during conversations I got drowsy after driving a few minutes I got sleepy after reading for 15 min I dozed off when I relaxed Nocturnal sleep onset items Even if I took a nap, I slept well at night I had difficulty falling asleep It took me less than 5 min to fall asleep
Never
Rarely
Sometimes
Often
Missing
64.6 32.2 44.5 89.5 56.2 91.3 65.9 38.5 29.0
14.5 17.9 24.1 6.6 18.1 5.5 14.6 14.3 20.4
14.4 28.5 21.8 3.1 13.5 2.5 14.7 29.4 35.0
6.3 21.0 9.5 0.8 12.0 0.6 4.7 17.6 15.3
0.2 0.4 0.1 0.0 0.2 0.0 0.1 0.2 0.3
34.3 42.0 28.7
9.2 20.0 17.0
16.4 23.4 25.1
39.2 14.4 28.7
0.9 0.2 0.6
Measures The self-report measures of daytime sleepiness and nocturnal sleep onset examined in this study were originally developed as part of the SWAI by Rosenthal et al (1993). The 59-item SWAI was designed to measure several other dimensions as well, including psychic distress, social desirability, and behavioral activation. Only the nine items identified as measuring daytime sleepiness and the three measuring nocturnal sleep onset (i.e., difficulty falling asleep at night) were included in these analyses (see Table 2). Respondents rated items for the 2-week period preceding the interview on a four-point ordinal scale: never, rarely, sometimes, or often. The average amount of sleep gotten, the typical time gone to bed and gotten up from bed, and whether the respondent took naps were separately assessed in the study interview. Each question concerned weekdays for the 2 weeks prior to interview. Sleep efficiency was calculated by dividing the average amount of sleep by the time spent in bed (derived from average time gone to bed and gotten up from bed reports).
Factor Analyses To assess whether the proposed self-report items reliably measured dimensions of daytime sleepiness and nocturnal sleep onset, factor analyses were conducted. The sample was randomly split into two equal-sized subsamples. Repeating the factor analysis in each subsample provided a measure of the stability of the solutions, in terms of number of factors and the relationship of specific items with the factors. This strategy reduced the risk that final results were effected by chance features of a particular sample. Since the items were measured on a four-point ordinal scale, standard principle components and factor analytic techniques based on Pearson product–moment correlation coefficients were inappropriate. The LISCOMP computer program (Muthen 1988), which provides factor analysis based on the appropriate polychoric correlations, was used in these analyses. Factor models were estimated using unweighted least squares (ULS). When a large proportion of items are highly skewed, ULS is a more stable estimator than a generalized least squares, although it does
not generate a chi-square test of model fit (Muthen 1989). The number of factors was determined by examining the scree plot of eigenvalues, the relative change in the root mean square residual, and the stability of results across the two subsamples. Promax rotation, which allows for correlated factors, was used in these analyses due to the expected inverse association between daytime sleepiness and nocturnal sleep onset. Factor analysis requires complete data on the evaluated items for each subject. Fifty-three cases (2.4% of the total sample) were deleted because of missing data on one or more of the 12 items. Incomplete cases did not differ from complete cases in sex, race, marital status, level of education, or income (p values from .23 to .96). Incomplete cases did have a slightly higher mean age (33.8 years, SD 5 7.0) than complete cases (31.6 years, SD 5 7.7, p 5 .04). The means for items among subjects with incomplete data who did respond to a particular item did not differ significantly from the means for those subjects with complete data (p . .05). Thus the missing data appeared to be largely at random, and elimination of cases without complete data were unlikely to bias results (Little and Schenker 1995).
Reliability Analysis Analyses of the internal consistency of the proposed scales derived from the factor analyses were conducted using Cronbach’s alpha. Because sampling variability and measurement error often cause specific item lambdas (factor loadings) to vary from sample to sample, we used the conservative scale scoring strategy of selecting those items that loaded strongly and consistently on a factor and adding the item score (0 –3) to a subject’s scale score for each item. Reliability analyses were conducted as a means to assess the appropriateness of using factors as simple sum scales, that is, without weighting for factor loadings. These analyses were conducted on the same split-half samples as the factor analyses.
Threshold Analysis The Multiple Sleep Latency Test, the gold standard of daytime sleepiness assessment, is a continuous measure of the average
Daytime Sleepiness and Nocturnal Sleep Onset
amount of time it takes people to fall asleep over 5 trials within a day. A threshold of 5 min or less has been identified as an indicator of a clinically important level of daytime sleepiness. The distribution of scale scores were examined for “natural breaks,” which may indicate thresholds in the self-report measures analogous to the 5-min threshold for the MSLT. Potential threshold values for the DSS were examined against the reported average number of hours slept on weekdays during the past 2 weeks. Daytime sleepiness is in large part a function of the amount of sleep one has had recently (Roth et al 1989). Thus, a measure such as the DSS should be related to average sleep time, and nonlinear changes in this relationship may reveal clinically significant natural breaks that separate normal from pathologic levels of daytime sleepiness. A similar approach was taken to assess whether there was a threshold in the distribution of NSOS scores that would indicate abnormal levels of difficulty in falling asleep. For this analysis potential threshold values of the NSOS where evaluated against sleep efficiency (sleep time/total time in bed). Difficulty falling asleep should be inversely related to the ratio of sleep time to total time in bed: the more difficulty falling asleep the lower the sleep efficiency. Again, nonlinear changes in this relationship may indicate a diagnostic threshold.
Results Item Response Distributions of responses to the nine daytime sleepiness and three nocturnal sleep onset items are shown in Table 2. For the nine daytime sleepiness items, the more frequent the occurrence the higher the level of daytime sleepiness. Greater difficulty falling asleep at night (nocturnal sleep onset) was indicated by less frequent occurrence of the items “Even if I took a nap, I slept well at night” and “It took me less than 5 minutes to fall asleep,” but more frequent occurrence of the item “I had difficulty falling asleep.” Response distributions for nine of the items were significantly skewed (p # .05). The three items not skewed were: 1) “I dozed off while watching TV,” 2) “I dozed off when I relaxed,” and 3) “It took me less than 5 minutes to fall asleep.” Almost all of the skewed items showed a skew to the right, indicating relative infrequent occurrence of the daytime sleepiness or sleep difficulty manifestations during the prior 2-week period. Occurrence of two indicators of daytime sleepiness were particularly infrequent: falling asleep when visiting friends or during conversations.
Factor Analyses The scree plot of eigenvalues for each of the split-half samples indicated two but not more than three factors could be extracted from these items (see Figure 1). The scree in each analysis began with factor three, and the
BIOL PSYCHIATRY 1999;45:764 –770
767
Figure 1. Scree plot of latent roots from split-half factor analysis (n 5 1062; n 5 1066). The number of subjects in split-half subsamples do not sum to the total subject number (2181) due to missing data; see Factor Analyses portion of Methods section.
eigenvalues for the third factor were approximately one. The two-factor solution was chosen over the three-factor solution as the most reliable and parsimonious model. The three-factor solution was less stable across the two splithalf subsamples and only marginally improved fit [root mean square residual (rmsr) reduced by less than .020]. Additionally, the third factor was substantially correlated with the first factor in each subsample (r 5 .62 and .69). The two-factor solution fit the data reasonably well across the two split-half samples (see Table 3). The root mean square residuals were consistent and low, .058 and .056, respectively. Nearly all of the items loaded strongly on one factor and weakly on the other. In both subsamples these two factors were modestly correlated and accounted for nearly 50% of the variance of the items. The first factor was well measured by eight of the items of the daytime sleepiness scale. The item “I could nap anywhere” did not appear to be a reliable indicator of daytime sleepiness, as its highest loading switched factors across the two subsamples. The second factor was well measured by two of the three nocturnal sleep onset items. The item “Even if I took a nap, I slept well at night” loaded on both factors to a moderate degree and thus was an unreliable indicator.
Reliability The eight items that consistently measured the daytime sleepiness factor showed good reliability as measured by Cronbach’s alpha in each split-half sample (a 5 .71 and .69). The addition of the item “I could nap anywhere,” which was deleted based on the factor analyses, did not improve reliability (a 5 .73 and .69). Reliability of the nocturnal sleep onset scale was improved in each split-half
768
E.O. Johnson et al
BIOL PSYCHIATRY 1999;45:764 –770
Table 3. Factor Analysis Results: Two-Factor Solutions for Split-Halves (n 5 1062, n 5 1066)a First split-half factor loadings
Second split-half factor loadings
In the past 2 weeks . . .
Factor 1
Factor 2
Factor 1
Factor 2
I fell asleep when riding as a passenger I dozed off while watching TV I got drowsy within 10 min when I sat still I fell asleep when visiting with friends I could nap anywhere I fell asleep during conversations I got drowsy after driving a few minutes I got sleepy after reading for 15 min I dozed off when I relaxed Even if I took a nap, I slept well at night I had difficulty falling asleep It took me less than 5 min to fall asleep
.595 .535 .585 .687 .304 .719 .517 .432 .659 2.208 .264 .006
2.059 2.049 2.012 2.027 .498 2.002 .005 .031 .008 2.418 2.691 2.753
.607 .581 .646 .748 .514 .705 .533 .495 .673 2.315 .220 2.076
2.023 2.015 2.010 2.030 .276 2.087 2.142 2.011 2.013 2.319 2.786 2.656
First-half two-factor solution: root mean square residual 5 .058; factor correlation 5 .22; R2 5 .47. Second-half two-factor solution: root mean square residual 5 .056; factor correlation 5 .19; R2 5 .49. a Number of subjects in split-half subsamples do not sum to total subject number (2181) due to missing data; see Factor Analyses portion of Methods section.
sample (a 5 .58 and .60) by dropping the item “Even if I took a nap, I slept well at night,” leaving only two items.
Thresholds Figure 2 presents distributions of the eight-item DSS and two-item NSOS scores; each item was scored 0 –3 from “never” to “always” and then summed. For the DSS a
“natural break” in the distribution of scores appeared to occur at a score of 11. The full distribution of DSS scores (0 –24) showed a significant positive skew (p , .05). However, DSS scores 0 –10 were normally distributed (no significant skew or kurtosis p . .05). Thus, it is the tail of the distribution, scores of 11 and higher, that produced the significant skew of the total scale distribution (p , .05). Approximately 17% of the sample had DSS scores of 11 or higher. No obvious break occurred in the distribution of NSOS scores, although a score of 4 or higher may constitute such a threshold. The top panel of Figure 3 shows box plots of average number of hours slept on weekdays in the past 2 weeks by DSS score. For DSS scores 0 –10, the median hours slept, indicated by the line through each box, was, with one exception, uniform at 7 hours. DSS scores of 11 or higher were all below 7 hours, generally at 6 hours of sleep. The lower panel of Figure 3 shows boxplots of sleep efficiency (hours slept/total time in bed) by NSOS scores. No thresholdlike change occurred; the relationship appeared largely linear. The correlation (r) between NSOS and sleep efficiency was 2.27 (p , .001): the greater one’s difficulty falling asleep the lower one’s sleep efficiency.
Discussion
Figure 2. Distributions of revised daytime sleepiness and nocturnal sleep onset scales.
A two-factor model appeared to best account for the variation in the 12 items from the SWAI across the split-half samples. The first factor was well measured by eight of the nine items proposed to measure daytime sleepiness. The second factor was well measured by two of the three items that constitute the nocturnal sleep onset scale. These two factors accounted for approximately 50% of the variance in both split-half sample analyses. The
Daytime Sleepiness and Nocturnal Sleep Onset
BIOL PSYCHIATRY 1999;45:764 –770
769
topic of napping. Responses to the questions “I could nap anywhere” and “Even if I took a nap, I slept well at night” were found to differ significantly between those who did and did not take naps during the 2-week period covered by the questions (x2 5 89.7, df 5 3, p , .001; x2 5 440.5, df 5 3, p , .001, respectively). Those who did not nap tended to report that they never or rarely could “nap anywhere,” while those who did nap tended to report that they sometimes or often could “nap anywhere.” There was a similar but reversed discrepancy between those who did and did not take naps in response to “Even if I took a nap, I slept well at night”; nonnappers indicated sometimes or often, while nappers indicated never or rarely did they sleep well despite napping. Having subjects respond to these questions even if they did not take a nap appears to be the reason for the unreliability of the measures. It was statistically and rationally appropriate to drop these items from future use of the DSS.
Threshold of Daytime Sleepiness
Figure 3. Boxplots of construct validity variables across daytime sleepiness and nocturnal sleep onset scale scores.
eight-item daytime sleepiness scale and the two-item nocturnal sleep onset scale showed good internal consistency. The items that showed inconsistent factor analysis results (“I could nap anywhere” and “Even if I took a nap, I slept well at night”) weakened or did not improve the internal consistency of the proposed scales. There appeared to be a “natural break” or threshold in daytime sleepiness scale scores that was associated with a substantial and consistent change in number of hours slept, indicating a possible point of distinction between normal and abnormal levels of daytime sleepiness. No breaks or thresholds appeared in nocturnal sleep onset scale scores.
The revised DSS had range of 0 –24. The distribution of scores within that range indicated an empirical “natural break” between scores from 0 –10 and 11 or greater. The distribution of scores 0 –10 appears to be normal (no significant skew or kurtosis), while that of scores of 11 or higher produced the significant skewedness for the whole distribution. This proposed threshold showed construct validity. Because nocturnal sleep is related to daytime sleepiness (Roth et al 1989), a threshold for some measure of daytime sleepiness should show differences in average amount of sleep. As demonstrated by the boxplots in Figure 3, a substantial and consistent shift in median number of hours slept takes place between scores of 10 and 11. In this study we could not compare the DSS, nor the proposed threshold, to MSLT values. Although a one-toone correspondence between the DSS and the MSLT is not expected, it is of interest that the threshold of pathological levels of daytime sleepiness generally accepted for the MSLT (#5-min sleep latency) describes 10 –15% of normal healthy clinic-ascertained volunteers (Levine et al 1988), and that the proposed threshold in the DSS identifies approximately 17% of the population as potentially having abnormal levels of daytime sleepiness. Future epidemiologic studies that also collect laboratory data are necessary to test the validity of this threshold.
No Threshold for Nocturnal Sleep Onset SWAI Items Dropped A characteristic shared by the two items from the DSS that were dropped from the scale due to unreliability was the
The revised NSOS had a range of 0 – 6. The distribution of scores within that range did not show evidence of a “natural break” indicative of a threshold. Consistent with
770
E.O. Johnson et al
BIOL PSYCHIATRY 1999;45:764 –770
the lack of a clear threshold in the distribution of NSOS scores, the relationship between sleep efficiency and NSOS appeared to be linear in the boxplot analysis. The correlation (r) between NSOS and sleep efficiency was 2.27, p , .001.
Limitations The most significant limitation of this study was that MSLT data could not be used to assess directly the validity of the DSS. However, the study that originally derived the DSS did indicate that the MSLT “gold standard” was related in expected ways to this self-report measure, albeit in a clinic-ascertained sample of which 76% reported sleep problems.
Conclusion This study largely replicated the results of the clinic-based study that identified self-report scales from the SWAI that appeared to measure daytime sleepiness and nocturnal sleep onset. The further development of an epidemiology of sleep, including daytime sleepiness and sleep disturbances, depends on the ability to move from the laboratory results to population surveys in reliable and valid ways. Development of self-report measures such as the daytime sleepiness scale and nocturnal sleep onset scale is a step in that direction. Supported in part by grant MH8802 from the National Institute of Mental Health and HL42215 from the National Institutes of Health.
References Broman JE, Lundh LG, Hetta J (1996): Insufficient sleep in the general population. Neurophysiol Clin 26:30 –39. Carskadon M, Dement W (1982): Nocturnal determinants of daytime sleepiness. Sleep 5:S73–S81. Carskadon M, Dement W, Mitler M, et al (1986): Guidelines for the multiple sleep latency test (MSLT): A standard measure of sleepiness. Sleep 9:519 –524. Hublin C, Kaprio J, Partinen M, Heikkila K, Koskenvuo M (1996): Daytime sleepiness in an adult Finnish population. J Intern Med 239:417– 423. Johns M (1991): A new method for measuring daytime sleepiness: The Epworth Sleepiness Scale. Sleep 14:540 –545. Levine B, Roehrs T, Zorick F, Roth T (1988): Daytime sleepiness in young adults. Sleep 11:39 – 46. Little R, Schenker N (1995): Missing data. In: Arminger G, Clogg C, Sobel M, editors. Handbook of Statistical Modeling for the Social and Behavioral Sciences. New York: Plenum Press, pp 39 –76. Muthen B (1988): LISCOMP. Analysis of linear structural equations with a comprehensive measurement model: A program for advanced research, 2nd ed. Mooresville, IN: Scientific Software. Muthen B (1989): Dichotomous factor analysis of symptom data. Sociol Methods Res 18:19 – 65. Rosenthal L, Roehrs T, Roth T (1993): The sleep-wake activity inventory: A self-report measure of daytime sleepiness. Biol Psychiatry 34:810 – 820. Roth T, Roehrs T, Carskadon M, Dement W (1989): Daytime sleepiness and alertness. In: Kryger M, Roth T, Dement W, editors. Principles and Practice of Sleep Medicine. Philadelphia: Saunders, pp 14 –23.