Study design and study population
This is a cross-sectional analysis of data collected in the fifth examination (October 2011 – February 2015) of the Copenhagen City Heart Study (CCHS) (25). In total, 9215 individuals were invited of which 4543 participated (49.3%). Details about the source and study population, the invitation procedure, data collection and data processing are described elsewhere (25, 26).
The study participants filled out a questionnaire covering a wide range of domains including but not limited to socioeconomic status; general, physical and mental health; symptoms and diseases; smoking and alcohol consumption; diet; and medication use. We have made an overview of questions relevant for this study in Additional files, Table A1.
All participants were examined at the CCHS’s test centre at a public hospital in the Capital Region of Denmark by trained medical laboratory technicians, medical students and medical specialists.
The tests relevant for this study were measurements of blood pressure, WC, and LDL-C (i.e., our outcomes), and height, and weight (i.e., for descriptive purposes). WC was measured at the approximate midpoint between the lower margin of the last palpable rib and the top of the iliac crest. Three consecutive blood pressure measurements were taken on the participants’ non-dominant arm after five minutes of sitting with an automatic blood pressure monitor (OMRON M3, OMRON Healthcare, Hoofddorp, Netherlands). Venipunctures were taken according to standardised procedures and the level of LDL-C was determined directly (Sanofi Genzyme, Cambridge, Massachusetts, USA). Height was measured without shoes on a fixed scale to the nearest millimetre. Weight was measured with clothes, but without shoes, on a consultation scale (Seca, Hamburg, Germany) to the nearest 100 grams.
Accelerometer-based measurements of physical behaviours
All participants were invited to participate in a sub-study that involved wearing two tri-axial accelerometers (ActiGraph GT3X+; sampling frequency: 30 Hz; ActiGraph, Pensacola, Florida, USA) 24-hours per day for seven consecutive days to measure their daily physical behaviours. In total, 2335 participants gave consent to wear the accelerometers. The accelerometers were attached on i) the anterior aspect of the right thigh midway between the greater trochanter and patella oriented along the axis of the thigh, and ii) on the lateral aspect of the right iliac crest. They were attached directly to the skin using a double-sided medical tape (Hair-Set for hairpieces; 3M, Maplewood, Minnesota, USA) and wrapped with transparent adhesive film (OpSite Flexifix; Smith & Nephew, London, UK) to ensure a fixed position during the measurement period.
During the measurement period, the participants were asked to note their leisure time, working hours, time in bed, and periods of non-wear time in a diary. The participants were also asked to only remove the accelerometers in case of adverse skin reactions, discomfort or pain, affected sleep, and when going to a sauna. After the measurement period, the participants returned the accelerometers at the test centre or by mail using a pre-paid envelope. The measurements of physical behaviours have been described in detail in a previous publication (26).
Processing of raw accelerometer data
Detection of physical activity types and stationary behaviours
The MATLAB-software Acti4 (National Research Centre for the Working Environment, Copenhagen, Denmark) was used to detect and derive the time spent in the following physical activity types and stationary behaviours: lying, sitting, standing, moving (i.e., small movements without regular walking while in a standing posture), walking, climbing stairs (i.e., up and down), running, cycling and rowing. Acti4 detects the physical behaviours through an algorithm that uses inclinations and accelerations (27), with a high sensitivity and specificity (27, 28).
Quality control, time in bed and non-wear time
For each individual participant, we visually inspected the activity classification over time to identify and investigate any abnormalities in the data (e.g., high levels of rowing or lack of sitting). Time in bed was defined based on a combination of accelerometer and diary data (i.e., bedtime/get up time). Non-wear time was ‘operator-defined’ by diary information and visual inspection of the activity classification. In addition, Acti4 detects non-wear time automatically using a set of rules: 1) Periods <10 min without recorded movement were not regarded as non-wear time. 2) Periods between 10 and 90 min were classified as non-wear time if a) the vector sum of the standard deviation of acceleration was >0.5G for any second during a 5-second interval immediate before the period without recorded movement, and b) the accelerometer was placed in a horizontal position (±5°). 3) Periods >90 min without recorded movement were always considered as non-wear time (27). See previous publication for further details about the processing of the raw accelerometer data (26).
Our inclusion criteria were: 1) ≥5 days of measurements with ≥16 h of accelerometer recordings per 24-hours day, 2) not using antihypertensive, diuretics or cholesterol lowering medicine, and 3) no missing values in any of the outcome variables. All reported ‘sick days’ (i.e., diary information) were excluded.
Definition of variables
Physical behaviour composition
The physical behaviour composition consisted of time (min/24-h day) spent in sedentary behaviour (i.e., sum of lying and sitting), standing, moving, walking, HIPA (i.e., sum of climbing stairs [up/down], running, cycling and rowing), and time in bed. It hence reflects participants’ 24-hour time-use. Time spent in the physical behaviours was accumulated during waken hours only (i.e., except for time in bed). Time spent in each behaviour was represented by the individual’s daily mean time across the measurement period standardised to 24 hours.
Physical behaviours (i.e., compositional parts) consisting of zeros cannot be included in CoDA. Due to zero time spent climbing stairs, running, cycling, and rowing for some participants, we decided to merge these behaviours into the combined activity class HIPA.
We used SBP (mm Hg), WC (cm), and LDL-C (mmol/L) as outcome variables. WC was used rather than BMI or waist-hip ratio since it has been suggested to be a stronger predictor for CVD (29).
Covariates and variables for descriptive analyses
In addition to the physical behaviour composition, we used the following covariates in the analyses: sex, age, number of years of education, smoking status, average number of alcohol units/week and self-reported use of prescribed medication for cardiovascular disease, antidepressants or sedatives, asthma or bronchitis and diabetes.
For descriptive purposes, BMI was categorised according to WHO classification (30), blood pressure according to the classification used by the European Society of Hypertension and the European Society of Cardiology (31), and WC into >88 cm for women and >94 cm for men (29).
An overview including details about how we derived these variables can be found in the Additional files (Table A2).
We described the characteristics of the study population using frequencies and percentages (%) or medians and first and third quartiles (Q1-Q3) where appropriate. Medians were used instead of means due to skewed distributions of some of the continuous variables. We described the central tendency and dispersion of the physical behaviour composition with geometric means and a variation matrix, respectively.
Investigation of potential selection bias
The characteristics of the non-eligible participants (i.e., having accelerometer data but not fulfilling the eligibility criteria) were compared to the characteristics of the eligible participants. This was done using Mann-Whitney U test, Pearson’s Chi-squared test (i.e., p-values <0.05 were considered to indicate differences between groups) and assessing 95% confidence intervals (CI) of medians and proportions. We calculated CIs for medians and proportions using the normal approximation method and the Wilson’s score method, respectively (32).
Compositional data is bound to a sample space (i.e., the simplex) with a geometry that is incompatible with standard statistical methods. To allow the use of standard statistical methods, we transformed the physical behaviour composition with the isometric log-ratio (ilr) transformation based on a sequential binary partition process (19). This resulted in a set of pivot ilr-coordinates that represent the physical behaviour composition in a sample space (i.e., the real coordinate space) where standard statistical methods can be applied (20). Specifically, pivot ilr-coordinates were constructed, where the first coordinate (ilr1) represents the first part of the composition relative to the geometric mean of the remaining parts (33).
Modelling process and time reallocations
We investigated how sedentary behaviour, walking, and HIPA, expressed as ilr-coordinates, were associated with each outcome using linear regression models (i.e., crude and adjusted analyses). Due to the ilr-transformation, the model estimates of the ilr-coordinates are not directly interpretable. A solution to this challenge was to theoretically reallocate time between 1) sedentary behaviour and walking, and 2) sedentary behaviour and HIPA and thereby, quantify the measure of association in an understandable way (20). This was conducted in the following three steps.
i) For each outcome, we fitted a multiple linear regression model with the ilr-coordinates representing the physical behaviour composition and the previously mentioned covariates (i.e., only in adjusted analyses). The physical behaviour composition as a whole was associated with SBP, WC, and LDL-C in the crude and adjusted analyses (i.e., all p-values <0.001, data not shown). We tested for and found an interaction between the physical behaviour composition and age (i.e., p-value for interaction term in the SBP-, WC- and LDL-C-model: 0.006, <0.001, and <0.001, respectively). Subsequently, all analyses were stratified by age group (i.e., adults <65 years and older adults ≥65 years). We assessed the assumptions of the linear regression models by plotting residuals vs. continuous covariates, residuals vs. fitted values and by quantile-quantile (Q-Q) plots of the residuals (i.e., assumption of linearity, homogeneous variance of residuals, and assumption of normally distributed residuals).
ii) Since the beta-coefficients of the ilr-coordinates are not directly interpretable, we used the reallocation of time between the behaviours to quantify the measure of association in an understandable way. With the age group-specific geometric mean composition as the starting point (i.e., reference composition), we reallocated time according to our study objectives. The time reallocations were made pairwise. For example, if 10 minutes were reallocated from sedentary behaviour to walking in a theoretical reference composition consisting of 580 minutes sedentary behaviour, 190 minutes standing, 60 minutes moving, 90 minutes walking, 20 minutes HIPA and 500 minutes in bed (i.e., 24 hours), it would result in 570 minutes sedentary behaviour and 100 minutes walking, while all remaining physical behaviours were kept constant.
For reallocation 1), we reallocated time between sedentary behaviour and walking in 10-minute portions. That is, sedentary behaviour was decreased with 10 to 60 minutes with a corresponding increase in walking time. Similarly, walking was decreased alongside an increase in sedentary time, again from 10 to 60 minutes. For reallocation 2), we similarly reallocated time between sedentary behaviour and HIPA from 2 to 12 minutes in 2-minute portions.
iii) We estimated the outcome for the reference compositions and each reallocated composition using the fitted values from the regression models. Subsequently, we calculated the difference in outcome by subtracting the estimated outcome of the reference composition from the estimated outcome for each reallocated composition (20, 21).
We used the statistical software RStudio (version 1.1.463) (34) running R (version 3.5.3) for all analyses (35). Specifically, for the analyses involving CoDA, we used the following packages: compositions (36) and robCompositions (37).