Progress 08/15/12 to 08/14/15

Outputs
Target Audience: Nothing Reported Changes/Problems:Preparing this data set for analysisrequired more extensive data cleaning efforts than orginally anticipated. Also,the statistical analysis proved more complex to carry out (requiring significant new software development),while yeilding richer and more detailed results, than expected. As a result, the specfic objectives for this project were narrowed to include just the first two objectives and just the highest priority outcome variable, sleep efficiency. The project was also extended for an additional yearat no additional cost to NIFA. With these changes the project produced high rewardresults by generatingnovel findings on a topic of practical importance to much of the public and pioneering the use of powerful newdata science research methods. What opportunities for training and professional development has the project provided?This project provided advanced doctoral training for the project director in nutritional neuroscience and computational data science leading to the completion of her PhD in Nutritonal Biology. She plans to use the expertise gained from this training periodto develop a career researching topics of nutritional neuroscience that are of practical importance to the public using new health information technology tools. How have the results been disseminated to communities of interest?The results of this study have been disseminated to the research community through two conference presentations and the publication of the project director's dissertation. What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? This was an observational study of the relationship between bedtime hunger and sleep in users of a personal sleep measurement device. Theheadband was sold to the public during 2009 - 2013 to beworn while in bed and detectedelectrical potentials from the forehead. A computer application calculated, and recorded, the user's most probable stage of sleep or wake every 30 seconds. Users could upload these sleep records to their online accounts, where they couldalso track related nightly conditions, including hunger, by self-report. Summary data on each night of sleep was de-identified and aggregated into an IRB-approved research registry by the device company. We accessed this data set and extracted the sleep records with a bedtime hunger rating on a standard scale. The data set required more extensive data cleaning than expected to prepare it for analysis, and, as a result, we had to narrow the scope of the analysisto justthe first two specific objectives and justthe outcome variable of highest priority, sleep efficiency (the percent of time in bed spent asleep -- a commonly used indicator ofsleep quality). Weanalyzed the datawith logistic multilevel modeling to determine the relationship between hunger and sleep efficiency. The model was adjusted for age, gender, amount of sleep during the preceding three nights, and time of going to bed. Bedtime hunger ratings were decomposed into two variables, person-mean hunger and nightly hunger (the difference between each night's hunger rating and the person's mean), because thisyeilded amodel with greaterpredictive accuracy than using the raw bedtime hunger ratings. Results: 183 people (age 19-77, 68% male) provided 4,284 sleep records (mean: 23 records per person) with a bedtime hunger rating. Our model explained 60% of the variability in sleep efficiency. The model-derived expected sleep efficiency for an average person on an average night was 95%. Sleep efficiency was inversely related to nightly hunger (p = 0.01) but not related to person-mean hunger (p = 0.26). This means thatnights with high nightly hunger tended to have lower sleep efficiency than usual for that person. The magnitude of this effect varied accross the population. We represented the nightly hunger effect size for each person with the difference between their predicted sleep efficiency at a very low (1st percentile), versus very high (99th percentile), nightly hunger level. The effect size was -2.4 percentage points for the average person in the sample (range: -12-1.6). There were medium to strong correlations between each person's nightly hunger effect size and the person'stypical sleep efficiency, unexpalined variability in sleep efficiency, person-mean hunger, and age. The people whose sleep efficiency declined most on hungry nights were generally those with low typical sleep efficiency, high unexplained variability in sleep efficiency, low person-mean hunger, and older age. These findings indicate thatfor the people in this data set, on average, going to bed more hungrythan usual predicted reduced sleep efficiency that night. This effect was most marked in those who typically have low and variable sleep efficiency, those who are older, and those who rarely go to bed hungry. The resultsprovideevidence consistent with the ideathat hunger can disturbsleep, especially in poor sleepers andconfirm ourhypotheses 1 and 2 for the sleep efficiency outcome variable. Also, the finding that those most affected rarely go to bed hungry is suggestive that susceptible individuals may be avoiding bedtime hunger, such as by eating a bedtime snack. This underscores the potential conflict bedtween successful weight management and healthy sleep. This study establishesthe need for further research to elucidate the influencesof diet and appetite regulationon sleep, including studies to establish causality and mechanisms, and development of evidence-based guidance for dietary practices that simultaneously promote healthy sleep and metabolic health. The current findings also point to the need to develop effective weight management strategies that minimize hunger, especially at bedtime. The main accomplishments of this project were: 1) Produced the first sceintific evidence that bedtime hunger predicts poor sleep quality and discovered somecharacteristics that identify the popualtions most affected. 2) Demonstrated the potential of data from personal health tracking devicesto accelerate the discovery of patterns and relationships in the health ofthe popualtion. 3) Generated statistical code for cleaning and preparing this data set for analysis and for longitudinal modeling of the data. This code has potential to be re-used in future projectsor to serve as a prototype or example for future development of longitudinal data analysis tools. After further development, we hope to publish afuture version of it. 4) The project director completed her doctoral training, submitted her dissertation, and was awarded her PhD.

Publications

• Type: Theses/Dissertations Status: Published Year Published: 2015 Citation: Dixon, BM. Relationships Between Bedtime Hunger and Subsequent Sleep Efficiency in Users of a Personal Sleep Measurement Device: A Multilevel Modeling Analysis. University of California, Davis. ProQuest Dissertations & Theses. 2015.
• Type: Conference Papers and Presentations Status: Published Year Published: 2015 Citation: Dixon BM, Liu S, Nordahl TE, and German JB. Bedtime hunger predicts sleep quality in users of a personal sleep measurement device. 2015 UC Davis Interdisciplinary Graduate & Professional Student Symposium, Abstracts, 27. 2015.
• Type: Conference Papers and Presentations Status: Published Year Published: 2013 Citation: Dixon BM, Nordahl TE, Gold EB, Winkelman J, Fabregas SE, Hardin K, German JB. Bedtime hunger is inversely related to subsequent sleep duration in the DOZER registry. SLEEP, Vol. 36, Abstract Supplement, A67. 2013.

Progress 08/15/13 to 08/14/14

Outputs
Target Audience: Nothing Reported Changes/Problems: The data analysis for this project is yeilding more extensive, detailed, and sophisticated results than originally anticipated, but is also proving somewhat more complex to carry-out than anticipated. In order to fully realize the benefits of the high-reward results we are obtaining, the time frame for the project has been extended for another year at no additional cost to NIFA. The new end date of the project will be 08/14/2015. We expect the findings of this research to make an important contribution toward the effort to end the obesity epidemic. What opportunities for training and professional development has the project provided? During this reporting period, the PD continued to advance her skills in computational data science. To do this she audited courses in Regression, Longitudinal Analysis, and Data Science, and she obtained ongoing training and support from the Davis R Users' Group. Through indepedent study and mentorship, she advanced her statistics knowledge, especially on multimodel inference, and logistic multilevel modeling. She also learned additional R programming skills, especially for more efficient and professional data management and manipulation, generating polished output for communicating with collaborating scientists, and additional R tools for multilevel modeling. The PD also established and strengthened her scientific connections and collaborations during this period, including with the Informatics Research unit of the UC Davis Health System and the Initiative for Wireless Health and Wellness at UC Davis. Additionally, she served as an important informant to the Robert Wood Johnson Foundation funded, Health Data Exploration Project. Lessons learned from conducting this analysis were shared with the wider health research community in their report, “Personal Data for the Public Good”, which is available at <http://hdexplore.calit2.net/wp/project/personal-data-for-the-public-good-report/>. How have the results been disseminated to communities of interest? Nothing Reported What do you plan to do during the next reporting period to accomplish the goals? During the next reporting period, the PD will complete the final stages of this analysis of the sleep efficiency outcome and write her dissertation on this topic to complete her PhD degree. The code used in this analysis will be modified to run similar analyses on each of the other sleep outcome variables, and all results will be published in the scientific literature.

Impacts
What was accomplished under these goals? Impact: This project is part of the effort to stop the epidemic of obesity and related diseases. One barrier many people encounter when dieting, is the commonly reported effect of food restriction and hunger to disturb sleep and cause daytime dysfunction due to sleep deprivation. But there is no scientific documentation of this effect. We are addressing this gap in current knowledge by determining the relationship between bedtime hunger and sleep in people who use a sleep measurement device. Clarifying this relationship will better inform efforts to get a good night’s sleep while dieting, empowering more people to successfully achieve a healthy weight. Accomplishments: During this reporting period we analyzed the relationship between bedtime hunger and the most important outcome variable, sleep efficiency, the percent of time in bed spent asleep. This analysis is almost complete and will be the topic of the PD's dissertation. In conducting the analysis, we also developed and tested the protocol we will use to analyze all the outcome variables, and we created several complex new R functions to carryout specific steps. When this analysis is finished, we expect to be able to use the statistical programming code produced, with minor modifications, to conduct similar analyses very efficiently on each of the other outcome variables. In preparation for the analysis, we generated descriptive statistics on the whole data set, and on the subset of the data that includes bedtime hunger information, which was used in this analysis. After exclusions, the data included 7,967 observations (i.e. nights of recorded sleep) from 348 people. The population is 64% male, with mean age 37, and provided an average of 23 observations per person. The average bedtime hunger level (on a scale of 0 = “no hunger” to 3 = “greatest hunger”) was 0.4, consistent with anecdotal reports that most people find going to bed hungry unpleasant. Average sleep efficiency was 89%. We created plots and graphics to explore the relationships of sleep efficiency to hunger and potential covariates. The graphics suggested that sleep efficiency is lower at high hunger levels, and showed evidence of expected relationships between sleep efficiency and age, gender and reported satisfaction with sleep. The first step of the analysis was to create a "null model" of sleep efficiency, without predictors, to understand the variability in this outcome. The model showed that variability in sleep efficiency is 57% between-person and 43% within-person (i.e. night-to-night variability). The presence of substantial variability at both levels confirms the need for multilevel modeling, so we proceeded with creating predictive models of sleep efficiency. For each model, we first created several candidate models, with different combinations of predictors and their interactions, and then used the Akaike Information Criterion (AIC) to select the best one. Sleep efficiency was logit-transformed and model diagnostics were inspected to confirm that modeling assumptions were met. Initially, we created a model of sleep efficiency using the variables expected, based on existing knowledge, to be the best predictors, but not including bedtime hunger. This model included demographics (age and gender) and indicators of the factors theorized in the literature to be the main regulators of sleep (homeostatic sleep drive and circadian alerting force). The amount of sleep obtained during the past three days was used as an indicator of sleep homeostasis, and the time of going to bed (relative to one's usual bedtime) was used to indicate the circadian force. This model explained 61% of the variability in sleep efficiency, and the most important predictors were age (older people have lower sleep efficiency) and amount of sleep obtained in the past three days (little sleep obtained predicts higher sleep efficiency). The strength of each person’s relationship between sleep obtained in the past three days and sleep efficiency interacted with gender (the relationship is stronger in males than females) and was also positively related to person-mean sleep efficiency, suggesting that strong homeostatic regulation of sleep may be a key feature of those who sleep well. We made graphics illustrating the relationship of each predictor to sleep efficiency and used the model to generate predictions of sleep efficiency in different scenarios, showing great variability between expected sleep efficiencies for different age/gender people under different conditions. Specific objective 1) Determine the relationships between hunger at bedtime and subsequent sleep duration, sleep efficiency, and time spent in each stage of sleep. We added subject-reported bedtime hunger to the model as an additional predictor, revealing an inverse relationship between hunger and sleep efficiency (p = 0.01) in which, for the average person, each additional level of hunger predicts a 0.55 percentage point drop in sleep efficiency, for a total drop of 1.65 percentage points between the low, and high, ends of the hunger scale. This finding confirms Hypothesis 1 for this outcome. Because the within-person, and between-person, variability in a predictor can have different relationships to the outcome, we created another model in which we decomposed the variability in hunger, and included both types of variability as separate predictors. Between-person variability in hunger was represented by person-mean bedtime hunger, and within-person variability in hunger was represented by each night's difference in bedtime hunger from the person's mean. This model revealed that person-mean bedtime hunger is not a significant predictor of sleep efficiency, only within-person (i.e. night-to-night) variability in hunger is. On average, each additional level of within-person hunger predicted a 1.2 percentage point drop in sleep efficiency (p = 0.00) for a total drop of 3.6 percentage points between the low, and high, ends of the scale. Graphics were created to illustrate the relationships to sleep efficiency for bedtime hunger and within-person bedtime hunger. Specific objective 2) Identify any differences between sub-groups of the population in the relationships between hunger at bedtime and subsequent sleep measures. Examination and interpretation of the models to understand the variability between people in the relationship between hunger and sleep efficiency is almost complete. Variability is substantial, with some people dropping as much as 4 percentage points of sleep efficiency with each additional hunger level for a total drop of nearly 12 percentage points. A significant interaction between person-mean bedtime hunger and within-person bedtime hunger indicates that the relationship between within-person bedtime hunger and sleep efficiency is strongest in those who usually go to bed with little hunger. This may mean that most people know how much hunger disturbs their sleep, and behave accordingly, avoiding going to bed hungry if they are strongly negatively affected by it. Preliminarily, those with the strongest negative relationship between hunger and sleep efficiency also have below average sleep efficiency when not hungry. So, scenario predictions indicate that the demographic groups expected to have lowest sleep efficiency on average (i.e. older women) also have the strongest predicted relationship of hunger to their sleep efficiency, supporting our hypothesis for this outcome variable. Specific objective 3) Determine whether hunger-related differences in sleep measures are related to how subjects feel the next day. To be completed in the next reporting period. The change in knowledge generated by this analysis will be the topic of our upcoming paper and will inform the public’s efforts to simultaneously achieve healthy sleep and a healthy weight.

Publications

Progress 08/15/12 to 08/14/13

Outputs
Target Audience: Nothing Reported Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided? This phase of the project developed the project director’s skills in statistics, data manipulation, data analysis, and statistical programming. In particular, she learned how to conduct multilevel modelling using repeated measures regression, a powerful statistical technique for investigating within-individual variability in longitudinal data. She also developed statistical programming skills in the R programming language, a leading software tool for advanced data analysis. How have the results been disseminated to communities of interest? Nothing Reported What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? We began this research project by creating a data transfer agreement with Zeo, Inc., and transferring their aggregate user data to UC Davis. Initial exploration of the dataset revealed over 150 variables of summary data on approximately 1.5 million sleep recordings contributed by nearly 40,000 people in the US and Canada. We cleaned and curated this dataset to get it into a form that is amenable to scientific analysis. Then we extracted the sleep records that include a response to the journal question on bedtime hunger, resulting in a dataset of 8,717 sleep records from 361 people. We then created a detailed data analysis plan to guide the analysis. This involved determining: 1) the variables to be included, the necessary exclusions, and the adjustments to be made for potential confounding variables; 2) the sequence of statistical tests to be applied (repeated measures regression will be the main analysis technique); and 3) the workflow and file management system to be used for this large and complex analysis. Then we began the process of developing and testing the statistical programming code to be used for the analysis, and, as part of this process, we generated initial (unadjusted) findings on the first outcome variable, sleep duration. We found a highly significant inverse linear relationship between bedtime hunger and sleep duration within subjects, which provides initial evidence supporting the hypothesis that bedtime hunger may disturb sleep. We presented this finding in a poster presentation at the SLEEP 2013 meeting in Baltimore in June. During this first year of the project, we also made new contacts with several UC Davis faculty whose advising will be valuable for completing this project, and established new collaborations with multiple organizations, within and outside UC Davis, that are potential long-term partners in this line of research.

Publications

• Type: Conference Papers and Presentations Status: Published Year Published: 2013 Citation: Dixon BM, Nordahl TE, Gold EB, Winkelman J, Fabregas SE, Hardin K, German JB. Bedtime hunger is inversely related to subsequent sleep duration in the DOZER registry. SLEEP, Vol. 36, Abstract Supplement, A67. 2013.