Abstract
Introduction
Lack of sleep, or poor sleep quality, has societal impacts for older adults. Poor sleep can have detrimental consequences on many aspects of vitality and resilience required for successful aging (McCall, 2004; Neikrug & Ancoli-Israel, 2009). Sleep disturbances have been shown to be associated with greater frailty in the elderly (Ensrud et al., 2009). Inability to sleep can also lead to delayed response time, which may affect driving ability as well as increase the risk of falls (Ancoli-Israel & Cooke, 2005). Insomnia has also been linked to increased consumption of health care resources (Ancoli-Israel & Cooke, 2005). In America, hundreds of billions of dollars are spent each year on doctor visits, hospital services, prescriptions, and over-the-counter-medications, associated with chronic disorders of sleep and wakefulness (Institute of Medicine [US] Committee on Sleep Medicine and Research, 2006). A recent report found that the economic impact of sleep disorders costs Australia $Aus5.1 billion per year, spread over health care costs, costs associated with medical conditions attributable to sleep disorders, and non-medical costs resulting from sleep loss-related accidents (Hillman & Lack, 2013). Early detection and intervention for sleep disturbance in older adults can help reduce the financial burden associated with sleep-related accidents, depression, and illness, and promote better quality lives (Smyth, 2008).
Insufficient sleep and sleep disorders have been associated with a number of chronic diseases (Ancoli-Israel, 2009; Magee, Kritharides, Attia, McElduff, & Banks, 2011), such as arthritis (Anderson et al., 2014), cardiovascular disease (Suzuki et al., 2009), diabetes (Gangwisch et al., 2007), hypertension (Gottleib et al., 2006), and stroke (Palomäki, Partinen, Erkinjuntti, & Kaste, 1992). The pain associated with many of these diseases has also been linked to sleep disturbance (Zarit, Griffiths, & Berg, 2004). There is a strong bidirectional relationship between sleep disorders and serious medical problems in adults, many diseases being more likely to develop in individuals with sleep disorders and individuals with sleep disorders at greater risk of developing these diseases (Bloom et al., 2009).
This study, therefore, aims to investigate the relationship between sleep difficulty and a number of chronic diseases, exclusively in a cohort of very old women who have participated in the Australian Longitudinal Study on Women’s Health (ALSWH).
Previous analysis of this cohort found that, unadjusted for other covariates, older women with low levels of sleep difficulty had reduced hazard of death compared with women reporting no sleep difficulty, and older women with the most sleep difficulty had greater hazard of death (Leigh, Hudson, & Byles, 2015). An interaction between disease count (of baseline diabetes, arthritis, heart disease, hypertension, asthma, bronchitis/emphysema, stroke, osteoporosis, and cancer) and sleep difficulty on mortality rate was also uncovered. The association between mortality rate and having more diseases was enhanced for older women with low levels of sleep difficulty and reduced for older women with the greatest sleep difficulty. Furthermore, after adjusting for disease and other covariates, the increased hazard of death associated with greater sleep difficulty was no longer significant. Part of the reason for this change after covariate adjustment may be due to correlation between sleep difficulty and disease. Unclear from the previous analysis was whether the relationship between sleep difficulty and mortality could be explained by a correlation between sleep difficulty and the number of chronic diseases. Therefore, to further understand the relationship between sleep difficulty and disease, in this subsequent analysis, we investigate which of the previously (if any) included diseases predict sleep difficulty in older women and the strength of this relationship. Identifying which chronic diseases are more likely to be associated with sleep difficulty may be useful to health practitioners and carers in identifying older women who are at greater risk of experiencing sleep difficulty.
Method
Data
Data analyzed were part of the ALSWH, a nationally representative, prospective study of more than 40,000 participants, which commenced in 1996. Cohorts of women, born in 1973-1978, 1946-1951, and 1921-1926, were sampled from the Medicare Australia database (Medicare Australia, n.d.) and invited to complete the baseline postal survey. Since then, women have been re-surveyed on a 3-yearly basis. Further details on the establishment of the cohorts, and follow-up, have been published elsewhere (Brown et al., 1996; Lee et al., 2005). This study presents data collected on the 1921-1926 cohort who completed the baseline survey in 1996 (Survey 1) when aged 70 to 75 years, and who have since completed sleep questionnaires for at least one of the following five surveys: 1999 (Survey 2, 73-78 years), 2002 (Survey 3, 76-81 years), 2005 (Survey 4, 79-84 years), 2008 (Survey 5, 82-87 years), and 2011 (Survey 6, 85-90 years) (n = 10,721). Ethical approval was obtained from the Universities of Newcastle and Queensland (Ethics approvals H0760795 and 2004000224).
Previous analyses of data from this cohort (Leigh et al., 2015) uncovered four latent classes of sleep difficulty based on repeated measures of four items from the Nottingham Health Profile (NHP) Sleep Subscale (Hunt, McKenna, McEwen, Williams, & Papp, 1981). The four items were as follows: Do you wake in the early hours of the morning? Do you sleep badly at night? Do you take a long time to get to sleep? and Do you lie awake most of the night? The latent classes were identified using repeated measures latent class analysis (RMLCA), which is a longitudinal application of latent class analysis (LCA), and is appropriate for investigating latent longitudinal patterns (Goodman, 1974; Lanza & Collins, 2006; Silverwood, Nitsch, Pierce, Kuh, & Mishra, 2011). RMLCA is therefore appropriate for identifying potential subgroups exhibiting sleep difficulty. Repeated measurement of four items could produce a vast number of unique response vectors, but with some occurring more frequently than others. RMLCA allows this complex array of responses to be represented in a parsimonious and easily interpretable fashion (Collins & Lanza, 2010), representing both the commonly and uncommonly observed response patterns. In our previous analysis, between one and eight latent classes were trialed, and the optimal class structure was chosen using the Bayesian information criterion (BIC; Nylund, Asparouhov, & Muthén, 2007), entropy (Asparouhov & Muthén, 2013), and the adjusted likelihood ratio test (LRT; Nylund et al., 2007), as well as interpretability of the classes. These fit statistics are included in Table 1 of the online appendix and indicate that the four-class model adequately represents the patterns of sleeping difficulty.
The four classes represent “troubled sleepers” (T, 22.7%), “untroubled sleepers” (UT, 32.1%), “early wakers” (EW, 28.8%), and “trouble falling asleep” (TF, 16.5%) (Leigh et al., 2015). Figures 1 to 4, respectively, display the observed pattern of responses to the aforementioned sleep items, over Surveys 2 through 6. Class T was characterized by a high probability of sleep difficulty at each survey, as demonstrated by the top left graph in each of Figures 1 through 4. Class UT was characterized by a low probability of sleep difficulty over time, as shown in the bottom right graph of Figures 1 through 4. The top right graph in each of Figures 1 through 4 displays the pattern of responses over time for Class EW, which was characterized by a high probability of waking in the early hours of the morning, and low probability of the remaining three items. Finally, Class TF was characterized by a high probability of taking a long time to fall asleep and low probability of the remaining three items (bottom left graph in Figures 1 through 4).

Do you wake in the early hours of the morning? (% and incidence yes/no/deceased at each survey (2-6) for each latent sleep class).

Do you sleep badly at night? (% and incidence yes/no/deceased at each survey (2-6) for each latent sleep class).

Do you take a long time to get to sleep? (% and incidence yes/no/deceased at each survey (2-6) for each latent sleep class).

Do you lie awake most of the night? (% and incidence yes/no/deceased at each survey (2-6) for each latent sleep class).
Importantly, RMLCA allows classification of women according to the pattern of responses to all four sleep items simultaneously. For example, a woman reporting the sleep symptom “Do you wake in the early hours of the morning?” across all surveys is highly likely to belong to either the EW or the T class (Leigh et al., 2015). Responses to the remaining sleep items determine which of these two classes this particular woman most likely belongs to. For example, if she reports no further sleep symptoms, she has a high probability of belonging to Class EW; however, if other sleep symptoms are reported, she would be more likely to belong to Class T (Leigh et al., 2015). Further details of our RMLCA analysis and validation of the NHP Sleep Subscale have been published previously (Leigh et al., 2015).
Women were questioned at baseline about the following nine diagnosed medical conditions: diabetes, arthritis, heart disease, hypertension, asthma, bronchitis/emphysema, stroke, osteoporosis, and cancer. For each disease, a comorbidity count was created, calculated as the sum of the remaining diseases.
Statistical Analysis
Under RMLCA, a probability estimate is obtained for each subject for belonging to each of the latent classes. As such, there is a degree of uncertainty inherent in the model as to which latent class each subject belongs to. This uncertainty must be accounted for in subsequent analyses utilizing these latent classes, to avoid bias in the effect estimates (Asparouhov & Muthén, 2013; Vermunt, 2010). One-step regression methods incorporate this uncertainty directly by estimating the latent classes and the regression model simultaneously, but are cumbersome and impractical, and additionally the inclusion of covariates in the regression model can alter the meaning of the latent classes (Asparouhov & Muthén, 2013; Bolck, Croon, & Hagenaars, 2004; Vermunt, 2010). Two-step estimation, in which subjects are assigned to their most likely latent class, and the class variable is subsequently treated as an observed, non-random variable in subsequent analyses, introduces bias to the final results. The three-step regression method (Lanza, Tan, & Bray, 2013) is based around a correcting third step, in which the uncertainty of latent class membership (ignored in the two-step method) is accounted for in the regression model. This approach has been shown to have the same efficiency as the one-step approach when model entropy is 0.6 or higher (Asparouhov & Muthén, 2013), which was satisfied for the chosen four-class model (entropy = 0.665). Three-step latent class regression (Asparouhov & Muthén, 2013; Lanza et al., 2013) was therefore utilized in the current analysis to investigate the relationship between the latent sleep classes and disease, using multinomial logistic regression in the Mplus (Muthén & Muthén, 1998-2012) software (Version 7.3). Separate multinomial regressions were run for each chronic disease, adjusted for comorbidity count.
Results
Table 1 displays the comorbidity adjusted multinomial logistic regressions of the latent sleep difficulty classes on each of the nine diseases at baseline (and comorbidity count). Arthritis was associated with the greatest odds of belonging to Class T (odds ratio [OR] = 2.27, 95% confidence interval [CI] = [1.98, 2.61]), Class EW (OR = 1.64, 95% CI = [1.42, 1.9]), and Class TF (OR = 1.74, 95% CI = [1.46, 2.07]), compared with Class UT. The next strongest association was for heart disease, with ORs of 1.8 (95% CI = [1.5, 2.16]), 1.32 (95% CI = [1.08, 1.62]), and 1.5 (95% CI = [1.19, 1.89]), for Classes T, EW, and TF, respectively (compared with Class UT). The odds of belonging to a class characterized by sleep difficulty (Class T, EW, or TF), compared with the Class UT were also higher, to a lesser extent for asthma, bronchitis/emphysema, diabetes, hypertension, and osteoporosis. Both cancer and stroke were found not to be significantly associated with increased odds of membership to a sleep class characterized by greater sleep difficulty (compared with UT). Unsurprisingly, in each model increased comorbidity count was also associated with increased risk of belong to Class T, EW, or TF compared with Class UT.
ORs (and 95% CI) for Separate Multinomial Logistic Regressions of Latent Sleep Class on Each Baseline Disease (Comorbidity Adjusted).
Note. Bold values are significant at .05 level of significance. CI = confidence interval; OR = odds ratio; T = troubled sleepers class; EW = early wakers class; TF = trouble falling asleep class; UT = untroubled sleepers class; CC = comorbidity count (calculated as the sum of all other diseases).
To investigate post hoc the association between further incident disease cases (after baseline) and latent class membership, an “ever” disease variable was created for each chronic disease. Participants were assigned a “1” if they reported having the disease at any survey, and a “0” otherwise, indicating whether each given disease occurs at any time over the survey period. The comorbidity adjusted multinomial logistic regressions were then repeated using the new “ever” disease variables as a sensitivity analysis. The results of the sensitivity analysis are included in the Table 2 of the online appendix. There were few substantive differences between the main analysis and the sensitivity analysis. The association between asthma and Class EW (OR = 1.18, 95% CI = [0.98, 1.43]) and TF (OR = 1.2, 95% CI = [0.95, 1.51]) became non-significant, and the association between hypertension and Class TF (OR = 1.34, 95% CI = [1.12, 1.60]), osteoporosis and Class TF (OR = 1.23, 95% CI = [1.03, 1.48]), stroke and Class T (OR = 1.26, 95% CI = [1.01, 1.57]), and cancer and Class EW (OR = 1.26, 95% CI = [1.07, 1.48]) all became significant.
Discussion
Although sleep behavior changes with age (Roepke & Ancoli-Israel, 2010; Vaz Fragoso & Gill, 2007), few studies have examined sleep in very old populations (Hägg, Houston, Elmståhl, Ekström, & Wann-Hanson, 2014; Jacobs, Cohen, Ein-Mor, & Stressman, 2014; Sterniczuk, Rusak, & Rockwood, 2014). In our longitudinal study of women aged 70 to 75 years and followed until age 85 to 90 years, a majority of women had trouble sleeping, with only 30% classified in the untroubled class. The diseases that most strongly predicted sleep difficulty were arthritis and heart disease, followed by bronchitis/emphysema, osteoporosis, asthma, diabetes, and hypertension. Cancer and stroke were not predictors of latent sleep difficulty class. Therefore, health practitioners should be aware that in women aged 70 years and above, those who are diagnosed with these significant diseases may also be at greater risk of sleep difficulty. These women may need further care or counseling as to whether they require treatment for their sleep difficulty. Such care would be important to prevent depression (Fredman, Gordon, Heeren, & Stuver, 2014; Maglione et al., 2014), cognitive function decline (Blackwell et al., 2014), falls (Stone et al., 2008), frailty (Ensrud et al., 2012), and possibly increased mortality (Cappuccio, D’Elia, Strazzullo, & Miller, 2010; Gallicchio & Kalesan, 2009; Leigh et al., 2015), as well as greater risk of nursing home placement (Spira et al., 2012), and the associated health care costs and societal impacts poor sleep quality has for older adults. These conditions are well known to be exacerbated by sleep trouble.
Our findings add to the literature on sleep, as most research on sleep utilizes measures of sleep duration rather than sleep difficulty. Furthermore, our results using sleep difficulty may be more readily applicable than sleep duration in clinical or nursing home settings, as our subjective measure of sleep difficulty can be easily measured using four questions from the NHP Sleep Subscale. Conversely, sleep duration is a more objective measure, which may be estimated incorrectly by patient self-report, or else requires more accurate measurement approaches, such as polysomnography (Unruh et al., 2008) or actigraphy (Spira et al., 2012).
In alternative analyses that used disease prevalence over the entire study period to predict latent class membership, our results altered slightly. The question of whether to use baseline disease, or total disease prevalence over the study period, remains unclear. Using baseline measures only ignores the influence that incident disease cases may have on latent sleep class membership. However, by using disease prevalence over the entire study period, women who survive longer also have greater opportunity to report each disease, as well as greater opportunity to report sleep difficulty. This may contribute to some of the relationship between reporting each disease and the increased odds of belonging to the sleep classes characterized by some sleep difficulty (T, EW, and TF) in these alternative analyses. An underlying model assumption is that missing data on the predictor are Missing at Random (MAR). This assumption is unlikely to hold in the case of longitudinal disease data, where missingness due to death or dropout is likely to be informative. The baseline disease model was therefore chosen as the preferred model, as it makes fewer assumptions with regard to missingness.
Conclusion
Many older women have persistent difficulty sleeping, with three main patterns of sleeping problems identified, and one pattern showing few sleep problems. Patterns of sleeping difficulty were associated with arthritis, heart disease, respiratory conditions, diabetes, osteoporosis, and hypertension. The risk of sleeping problems, and advice on sleep hygiene, is therefore an important part of the care of older people with these chronic conditions.
Footnotes
Acknowledgements
The authors acknowledge the assistance of the Data Linkage Unit at the Australian Institute of Health and Welfare (AIHW) for undertaking the data linkage to the National Death Index (NDI).
Authors’ Note
The research on which this article is based was conducted as part of the Australian Longitudinal Study on Women’s Health, the University of Newcastle and the University of Queensland.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: We are grateful to the Australian Government Department of Health for funding and to the women who provided the survey data. This research was supported by infrastructure and staff of the Research Centre for Generational Health and Ageing, who are members of the Hunter Medical Research Institute.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
