Dikgale Health and Demographic Surveillance System site (DHDSS) sample[15]
A convenience sample of 167 adult participants was recruited from the AWI-Gen Phase 1 Study cohort (≥40 years), the methodological details of which are reported in detail elsewhere [14, 16]. Trained field workers collected self-reported, measured and biological data from participants by means of questionnaires, anthropometry, ultrasound scans, and venipuncture [14, 16]. We calculated body mass index (BMI, kg/m2) and conicity index (CI) [17]. The latter measure allows a single-measure of body shape to be used in multivariate analyses. Questionnaire data included behavioural, health and socio-demographic variables and is reported in detail elsewhere [14, 16]. The ultrasound scans and blood-derived variables were not considered for this analysis.
Accelerometer data collection and data reduction
Participants wore a small, light-weight, wrist-worn triaxial accelerometer for 9 days (ActiGraph wGT3X-BT, Actigraph, LLC, Pensacola, FL, 2013) [18-21]. Prior to use, the monitors were connected to an IBM-compatible computer via USB interface and initialized to sample at 30 Hz, using proprietary software (Actilife 6.13.4, Actigraph, LLC, Pensacola, FL, 2009-2015). The monitors were affixed with a proprietary woven nylon wristband on the non-dominant wrist, and unless there was to be a sustained period of water immersion, participants were requested not to remove the monitors. On the ninth day, the monitors were collected from the participants. Thereafter, the raw recorded data was downloaded from the monitors onto an IBM-compatible computer and stored for later analysis. Prior to re-use of the monitors, the batteries were fully recharged and the memory cleared of previous data. The wrist straps were washed using a disinfectant solution, rinsed in water and air-dried.
Using proprietary software (Actilife 6.13.4, Actigraph, LLC, Pensacola, FL, 2009-2015), valid data (at least one weekday and one weekend day) was obtained by first converting downloaded, raw data files to 60 second epochs. Thereafter, the Cole-Kripke sleep scoring algorithm was used to determine minute-by-minute asleep/awake status [22], and the Actilife-modified Tudor-Locke algorithm to identify sleep periods [23, 24]. Valid wear-time was evaluated using the Choi algorithm [25, 26], with sleep time marked as wear time, and a valid day requiring ≥10 hours of wear time. Vector Magnitude (VM)- and Ambulation-defined physical activity variables were defined as counts/day and counts/minute, and steps/day and steps/minute, respectively [27]. Sleep indices included Total Sleep Time (TST), Sleep Efficiency (SE), Wake After Sleep Onset (WASO) and Sleep Fragmentation Index (SFI) [28-30].
Valid physical activity and sleep data was downloaded in a summarised and detailed format in Microsoft Excel™ files and additional variables were extracted: diurnal and nocturnal sleep time, sleep periods during defined hours, and sleep variation across days (within-person total sleep time SD). Diurnal and nocturnal periods were defined as 06h00-18h00 and 18h01-05h59, respectively. The number of sleep periods initiated between 00h01-05h59 defined an additional sleep quality indicator. A sleep period falling completely within the period 06h00-18h00 was defined as potentially a “daytime napping” period. Sufficient sleep quantity and quality were defined as 7-9 hours [31] and SE≥ 85%, respectively [32]. Data was then imported into statistical software for further analyses.
Statistical analyses
Descriptive statistics comprised means (one standard deviation), medians (interquartile range) and frequencies. Relationships between categorical variables were examined through the Chi-square Test. For continuous data, independent and dependent t- tests examined differences between the sexes and weekday/weekend days and where required, the appropriate non-parametric test was employed. Due to non-normality, continuous variables were transformed to quantiles as required. Bivariate relationships were examined using correlation coefficients.
Multiple linear regression models were examined for predictors of sleep indices (TST, SE, WASO, SFI) and body composition measures (waist circumference and BMI) using selected socio-demographic, behavioural and biological variables. Forced and Backward selection (p in = 0.05, p out = 0.10) models were employed. Separate sex-specific analyses were run, specifically to include parity in the female analysis. For multivariate adiposity analyses TST and SE were used in forced models, while TST, SE, WASO, SFI were entered in selection models.
To analyze the trend across days for objectively-measured sleep indices, the day-by-day data was analysed by fitting a mixed-effects model (Fixed effect, Type III, Restricted Maximum Likelihood), using a compound symmetry covariance matrix. Missing values were considered missing completely at random. The Geisser-Greenhouse correction was adopted throughout. Multiple comparisons tests (Tukey) compared sleep indices across each day.
Data were analysed using appropriate statistical software (IBM SPSS Statistics: Release 25 IBM Corporation, Armonk NY, 2017 and GraphPad Prism: version 8.3.0, GraphPad Software, La Jolla CA, 2019). Significance for all inferential statistics was set at p<0.05.
Informant consultation
To add contextual detail to the quantitative results, we obtained feedback from DHDSS fieldworkers (n=2) and the community engagement officer through an interview [33]. Key findings were discussed and informants were encouraged to provide feedback, which was captured via notes. After the interview, the notes were distributed for confirmation.