Abstract
Objective
The health benefits of exercise are clear. In targeting interventions it would be valuable to know whether characteristic patterns of physical activity (PA) are associated with particular population subgroups. The present study used cluster analysis to identify characteristic hourly PA patterns measured by accelerometer.
Design
Cross-sectional design.
Setting
Objectively measured PA in Hong Kong adults.
Subjects
Four-day accelerometer data were collected during 2009 to 2011 for 1714 participants in Hong Kong (mean age 44·2 years, 45·9 % male).
Results
Two clusters were identified, one more active than the other. The ‘active cluster’ (n 480) was characterized by a routine PA pattern on weekdays and a more active and varied pattern on weekends; the other, the ‘less active cluster’ (n 1234), by a consistently low PA pattern on both weekdays and weekends with little variation from day to day. Demographic, lifestyle, PA level and health characteristics of the two clusters were compared. They differed in age, sex, smoking, income and level of PA required at work. The odds of having any chronic health conditions was lower for the active group (adjusted OR = 0·62, 95 % CI 0·46, 0·84) but the two groups did not differ in terms of specific chronic health conditions or obesity.
Conclusions
Implications are drawn for targeting exercise promotion programmes at the population level.
Keywords: Body composition, Chronic disease, Exercise, Motor activity, Sedentary lifestyle
The health benefits of regular physical activity (PA) are well established( 1 ), including positive associations with psychological well-being( 2 , 3 ) and an inverse relationship with various illnesses( 4 , 5 ). However, general efforts to promote exercise have met limited success; potentially more specifically targeted interventions may prove more effective. To plan these, we need information on population patterns of PA to examine whether there are identifiable groups for whom interventions should be tailored and also to better estimate the impact of such interventions. Several such studies have been conducted, including cohort studies of PA levels, patterns in children and adolescents( 6 – 9 ) and population-level comparisons between countries( 10 ).
Examination of PA patterns requires a reliable and valid measure of PA. While many studies have relied on self-report questionnaires( 11 , 12 ), their validity is questionable( 13 , 14 ). On the other hand, accelerometers offer an objective and reliable measurement( 15 ) and are therefore increasingly popular in research. These electronic devices are worn by the subject to record accelerations in movement to indicate the duration and intensity of PA. Most accelerometers count accelerations larger than a specified threshold, sometimes termed an ‘activity count’( 16 ). Most studies report average daily counts and/or total time spent on PA( 10 , 16 , 17 ), but this does not reveal the temporal pattern of PA, such as whether the activity is concentrated in certain periods of the day. Some researchers noted this as a limitation and they have recorded bouts of PA( 10 , 17 ). However this method is also has its limitation in not distinguishing the time patterns of PA. Only by looking at PA patterns can we compare the relative effect of the duration and intensity of PA; for example, comparing the benefits of brief but intensive PA with less-intensive activities of longer duration, or whether PA occurring in the morning and at night has similar effects on health. Furthermore, with a better understanding of PA patterns, one can tailor health programmes more precisely to the needs of people who exhibit specific characteristic PA patterns.
We present a population study of hourly PA patterns on weekdays and weekends in urban Hong Kong. To compress the large amount of data produced from the accelerometer readings, we applied a cluster analysis to group people with similar PA patterns( 18 ). This reveals distinct PA patterns and the clusters can be compared in terms of the people's characteristics in each cluster, including their health status. Cluster analysis has previously been used to provide daily summaries of PA, but only with questionnaire data( 19 , 20 ); ours appears to be the first study to have applied cluster analysis to accelerometer data. Whereas previous studies have provided daily summaries of PA patterns using accelerometer data( 9 , 21 ), we extend this by analysing hourly PA patterns.
Methods
Participants
The present study was part of the Hong Kong Jockey Club FAMILY Project Cohort Study, funded as an initiative to promote family health, happiness and harmony in Hong Kong. It includes families recruited during March 2009 to January 2011. Sampling was based on a random selection of residential addresses provided by the Hong Kong Census and Statistics Department. A family was eligible when all members aged 15 years or older, who lived in the same address and could understand Cantonese, agreed to participate. All eligible members were interviewed by trained interviewers who entered the data into tablet personal computers. Details of the interview have been described elsewhere( 22 ). Having completed the main survey (n 45 767), randomly chosen participants (n 32 530) were invited to take part in a sub-study by wearing an accelerometer for four consecutive days (including a weekend). Written consent was obtained from participants (parental consent was also obtained for participants under 18 years old) and the study was approved by the Institutional Review Board of the University of Hong Kong.
Measurements
Accelerometer
Previous studies have found the ActiGraph to be reliable and valid( 15 , 23 – 25 ). A recent study( 26 ) showed that 1 d of accelerometer data is adequate for estimating weekly moderate-to-vigorous PA (MVPA) in a representative US sample aged 20–85 years but we adopted a more stringent criterion because we aimed to examine hourly PA patterns. A total of 5898 participants agreed to participate and were instructed to wear an ActiGraph GT1 M uniaxial accelerometer (http://www.theactigraph.com). The ActiGraph was to be worn around the hip (right hip for right-handed and left hip for left-handed persons) for four consecutive days for all waking hours, removed only when bathing or sleeping. Instead of the conventional 7 d requirement, a 4 d measurement period was chosen to reduce user burden and encourage participation; two weekdays and two weekend days were selected on the assumption that there should be relatively little variation across weekdays. The choice of the start day (Thursday, Friday or Saturday) was up to the participant.
‘Non-wear’ time was defined by an interval of zero accelerometer counts for sixty consecutive minutes or more. ‘Wearing’ time in a day was computed by subtracting non-wear time from 24 h. A valid day had to include at least 12 h of wearing time. Only observations with four valid days of data were accepted for analysis( 26 ). While previous studies used a threshold of at least 10 h of registered time/d to report daily PA( 9 , 27 ), we again adopted a more stringent criterion of 12 h/d.
Counts were recorded using a 1 min epoch. PA level was assessed with counts per minute, i.e. total counts divided by (60 × number of wearing hours)( 10 , 16 ), and separated into time spent on MVPA and light PA (LPA)( 28 ). A 1 min period was classified as MVPA (≥3·00 metabolic equivalent tasks, MET) if the total counts within this period were greater than or equal to 1952( 29 ), and was classified as LPA if the total counts ranged between 101 and 1951 (inclusive). The time spent on MVPA and LPA was the total number of moderate-to-vigorous minutes and light minutes in a day. The ActiGraph firmware version 7·5·0 was used for data transformation.
The k-means cluster analysis( 18 , 30 ) was used to identify the number of distinct hourly PA patterns. Average counts per minute from 00.00 hours to 23.59 hours on both weekdays and weekends were used as cluster variables. This gave a total of forty-eight cluster variables (twenty-four for weekdays, twenty-four for weekends), allowing us to identify different PA patterns for weekdays and weekends. All cluster variables were standardized to a mean of 0 and a variance of 1 to equalize the importance of each variable. First, k (a pre-specified integer) cluster centres were randomly generated. Next, the Euclidean distance was computed between participant i and cluster centre j, which equals
 |
where d and e are the standardized average counts per minute in hour t on weekdays and weekends, respectively, for all centres j = 1 to k. Participant i was assigned to his or her nearest cluster (i.e. one with the shortest distance). After assigning all participants to their nearest cluster, the new cluster centres were recomputed using their mean, after which these steps were iterated until convergence was reached. The elbow method was used to determine the number of clusters and the pseudo R-square and Mann–Whitney U tests were used to assess the goodness-of-fit of the final cluster solution. To select an appropriate k, within-cluster sum of squares distances were computed for k = 2 to 10, and k was determined using the elbow method (i.e. a sudden drop of within-cluster sum of squares distances from
to k indicates k as the appropriate solution)(
31
). A dendrogram was then used to confirm the number of clusters obtained using the elbow method. Non-parametric (Kruskal–Wallis) tests were used because the forty-eight cluster variables were not normally distributed (P < 0·001 for Kolmogorov–Smirnov normality tests).
Chronic health conditions
Participants were asked whether a medical practitioner had told them that they had any of eight chronic health conditions chosen to represent relatively definitive diagnoses: cancer, diabetes mellitus, hypertension, high cholesterol, heart disease, stroke, asthma or chronic obstructive pulmonary disease. We also recorded self-reported medicine use for those who reported a chronic health condition, to permit some validation of the reported chronic conditions.
Body composition
Height (with SECA 214 stadiometer, http://www.seca.com), weight and body fat percentage (BFP; using the Omron fat analyser scale HBF-356, http://www.omron-healthcare.com.sg) were measured by trained interviewers following standard protocols. BMI was calculated as weight (kg) divided by the square of height (m2). Two definitions of obesity, based on BFP or BMI, were used in the present study. Obesity by BFP was defined as ≥25 % (males) and ≥35 % (females). Obesity by BMI was defined as ≥30 kg/m2.
Other statistical analyses
A comparison of self-reported chronic conditions and self-reported medication use indicated that 95 % of the people reporting a chronic condition were taking medication(s) that corresponded to that condition. Subsequent analyses comparing activity levels to chronic conditions were restricted to this sub-set of participants.
Outliers with ActiGraph counts per minute (≥ median+1·5 × interquartile range) were removed (n 617) before the cluster analysis( 32 ). Outliers did not differ in terms of demographic variables (all P > 0·05). For interval-scaled variables such as age and BMI, independent t tests were used to compare the differences between clusters and Cohen's d ( 33 ) was used to assess the effect size of these differences. Effect sizes of 0·2, 0·5 and 0·8 were classified as small, medium and large, respectively( 33 ). For categorical variables, Pearson's χ 2 test was used to compare the differences between clusters. For the prevalence of chronic health conditions, odds ratios, unadjusted and adjusted for age, sex, smoking, education and income, were used to compare the differences between clusters. All statistical analyses were performed using Predictive Analytics SoftWare (PASW 18·0, formerly known as SPSS).
Results
Cluster analysis
Analysable accelerometer data were obtained from a total of 1740 participants. The average of 1128 correlations for the forty-eight hourly variables was 0·09. Among those, only nine pairs (0·8 %) had a correlation larger than 0·5, ruling out multicollinearity among the forty-eight variables. A four-cluster solution was obtained using both the elbow method and the dendrogram (Fig. 1). It showed that, when the number of clusters increased from two to four, the within-cluster sums of squares were large. However, two clusters had a sample size of twenty or less. Upon further examination, one cluster (n 20) showed constant PA for the entire 24 h period, evidently an error; and the other included only six people. These twenty-six participants were removed, leaving 1714 participants, 480 (28 %) in cluster 1 (the ‘active’) and 1234 (72 %) in cluster 2 (the ‘less active’). We did not find groups of people who were active only on the weekend but not during the week (or vice versa), which if present would have formed other distinct clusters. The pseudo R-square and Mann–Whitney U test z-values for each of the forty-eight hourly variables ranged from 0·01 % to 19·92 % (mean 7·57 %) and from 0·26 to 17·54 (mean 8·71), respectively. Thirty-five out of the forty-eight Mann–Whitney U tests were significant at the 5 % level.
Fig. 1.
Dendrogram of the cluster analysis solution
Cluster physical activity profile
Figures 2 and 3 show weekday and weekend PA patterns, respectively, for the two clusters, as well as the average level. First, with respect to the overall PA patterns (solid lines in Figs 2 and 3) there was a slightly greater hourly variation on weekdays than on weekends. Slight increases in counts per minute were observed on weekdays during 08.00–08.59, 13.00–13.59 and 18.00–18.59 hours, most likely corresponding to times of commuting to and from work, and going to lunch. On weekends (Fig. 3, solid line), the overall PA trend was comparatively smooth between 11.00 and 18.59 hours.
Fig. 2.
Hourly physical activity pattern in the FAMILY Project Cohort Study, Hong Kong, 2009 (weekdays): ——, overall; – · · – · · –, cluster 1 (active); – – –, cluster 2 (less active)
Fig. 3.
Hourly physical activity pattern in the FAMILY Project Cohort Study, Hong Kong, 2009 (weekends): ——, overall; – · · – · · –, cluster 1 (active); – – –, cluster 2 (less active)
Second, on weekdays, both clusters showed similar temporal PA patterns, differing only in intensity. By contrast, the two groups showed different patterns on the weekend. For cluster 1, activity peaks were found at 10.00–10.59, 16.00–16.59 and 19.00–19.59 hours, suggesting periods of increased PA. For cluster 2, the hourly PA curve on weekends was low and smooth, suggesting no increased PA or sports activity. In sum, participants of cluster 1 were generally more active and showed a more varied PA pattern on weekends than those of cluster 2.
PA levels are summarized in Table 1. The overall average count per minute was 310·9. The participants spent 22·5 and 247·5 min/d on MVPA and on LPA, respectively. Participants were in general more active on weekdays than on weekends (323·6 v. 297·2 counts/min). They also spent slightly more time on MVPA on weekdays than on weekends (23·1 v. 20·9 min/d) and less time on LPA on weekdays than on weekends (242·1 v. 260·9 min/d). These differences were observed among participants in both clusters.
Table 1.
Summary of physical activity levels of the 1714 participants in the FAMILY Project Cohort Study, Hong Kong, 2009
| Overall (n 1714) | Cluster 1 (active; n 480) | Cluster 2 (less active; n 1234) | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Variable | Mean | sd | Mean | sd | Mean | sd | Effect size | ||||||
| PA counts/min (overall) | 309·2 | 125·1 | 456·4††† | 116·0 | 252·0 | 69·2 | 1·62 | ||||||
| PA counts/min (weekday) | 323·6*** | 147·8 | 470·4***,††† | 144·4 | 266·5*** | 103·1 | 1·37 | ||||||
| PA counts/min (weekend) | 297·2 | 140·1 | 443·2††† | 144·3 | 240·5 | 87·7 | 1·44 | ||||||
| Effect size | 0·21 | 0·23 | 0·38 | ||||||||||
| MVPA min/d (overall) | 22·5 | 14·0 | 35·4††† | 15·7 | 17·5 | 9·3 | 1·28 | ||||||
| MVPA min/d (weekday) | 23·1*** | 15·9 | 35·9*,††† | 18·7 | 18·1*** | 11·2 | 1·12 | ||||||
| MVPA min/d (weekend) | 20·9 | 15·3 | 33·9††† | 17·8 | 15·8 | 10·5 | 1·18 | ||||||
| Effect size | 0·16 | 0·13 | 0·25 | ||||||||||
PA, physical activity; MVPA, moderate-to-vigorous physical activity.
Mean values were significantly different from those at the weekend: *P < 0·05, ***P < 0·001.
Mean values were significantly different from those of cluster 2: †††P < 0·001.
There were large differences between the two clusters in terms of counts per minute (456·4 v. 252·0 counts/min, d = 1·62), daily MVPA minutes (35·4 v. 17·5 min/d, d = 1·28) and daily LPA minutes (319·0 v. 219·7 min/d, d = 0·82). An active participant (accumulating 150 min MVPA/week as recommended by the Centers for Disease Control and Prevention/American College of Sports Medicine)( 11 ) had 4·18 times the odds (95 % CI 2·75, 6·37) of being in the active cluster.
Demographic profiles of the clusters
Table 2 shows that people in the active cluster were heavier and had lower BFP than those in the less active cluster. After adjusting for sex, however, there was no difference in BFP between the two clusters. Compared with the less active group, males in the active group had lower BFP (22·3 % v. 23·5 % in the less active group, P < 0·05), were shorter (165·0 cm v. 167·3 cm in the less active group, P < 0·05) and younger (41·2 years v. 44·3 years in the less active group, P < 0·05), but these differences were not found among females.
Table 2.
Age and body composition measurements for the 1714 participants in the FAMILY Project Cohort Study, Hong Kong, 2009
| Overall (n 1714) | Cluster 1 (active; n 480) | Cluster 2 (less active; n 1234) | |||||
|---|---|---|---|---|---|---|---|
| Variable | Mean | sd | Mean | sd | Mean | sd | Effect size |
| Age (years) | 43·2 | 16·4 | 42·1 | 14·9 | 43·7 | 17·0 | −0·10 |
| Weight (kg) | 60·4 | 12·9 | 61·4† | 13·1 | 60·0 | 12·9 | 0·11 |
| Height (cm) | 160·5 | 9·3 | 160·8 | 9·8 | 160·5 | 9·1 | 0·03 |
| BMI (kg/m2) | 23·8 | 3·8 | 24·1 | 3·8 | 23·7 | 3·9 | 0·11 |
| BFP | 26·9 | 8·1 | 25·6† | 8·2 | 27·4 | 8·1 | 0·22 |
BFP, body fat percentage.
Mean values were significantly different from those of cluster 2: †P < 0·05.
The demographics of the participants (Table 3) were similar to those of the Hong Kong population (46·0 % male; 39·7 % had an income lower than $HK 10 000; 22·7 % attained only primary education, while 18·0 % had a bachelor's degree or above). Table 3 shows that the two clusters differed in age group, sex, smoking, income and PA required at work. The active cluster had a higher proportion of middle-aged participants (ages 35–54 years), while the proportions of adolescents and the elderly were higher in the less active cluster. There were more males (57·5 %) in the active cluster and more females (58·5 %) in the less active cluster. The proportion of smokers was higher in the active cluster (16·8 % v. 11·5 % in the less active cluster, OR = 1·55, P < 0·001). Classifying smokers by sex, it was found that the proportion of males who smoked was higher in the active cluster than in the less active cluster (26·4 % v. 19·0 %, OR = 1·53, P < 0·001), while for female smokers this pattern was reversed (3·1 % v. 4·4 %, OR = 0·70, P < 0·05). Monthly personal income groups (in Hong Kong dollars; $US 1 = $HK 7·8) were differently distributed in the two clusters, particularly in lower-income brackets. In the active cluster, 24·0 % and 30·0 % of the participants were in the $HK 1–5000 and $HK 5001–10 000 group, respectively, while in the less active cluster these numbers were 39·0 % and 18·2 %, respectively (OR = 0·74 and 1·93, both P < 0·001). Finally, more participants had physically demanding jobs in the active cluster than in the other cluster (P < 0·001).
Table 3.
Demographic and lifestyle characteristics of the 1714 participants in the FAMILY Project Cohort Study, Hong Kong, 2009
| Cluster 1 (active; n 480) | Cluster 2 (less active; n 1234) | |||||
|---|---|---|---|---|---|---|
| Variable | Category | Overall (n 1714) | n | Column % | n | Column % |
| Age group, years (χ 2 = 32·94†††) | 15–24 | 264 | 66 | 13·8 | 198 | 16·0 |
| 25–34 | 224 | 58 | 12·1 | 166 | 13·5 | |
| 35–44 | 352 | 115 | 24·0 | 237 | 19·2 | |
| 45–54 | 454 | 156 | 32·5 | 298 | 24·1 | |
| 55–64 | 258 | 63 | 13·1 | 195 | 15·8 | |
| ≥65 | 162 | 22 | 4·6 | 140 | 11·3 | |
| Sex (χ 2 = 35·82†††) | Male | 788 | 276 | 57·5 | 512 | 41·5 |
| Female | 926 | 204 | 42·5 | 722 | 58·5 | |
| Education level (χ 2 = 8·25) | Primary school | 282 | 83 | 20·6 | 199 | 19·7 |
| Secondary school | 788 | 233 | 58·0 | 555 | 54·8 | |
| Post-secondary school | 122 | 38 | 9·5 | 84 | 8·3 | |
| University level | 222 | 48 | 11·9 | 174 | 17·2 | |
| Missing§ | 300 | 78 | – | 222 | – | |
| Smoking (χ 2 = 7·63†) | Yes | 184 | 68 | 16·8 | 116 | 11·5 |
| Had quit | 60 | 18 | 4·5 | 42 | 4·1 | |
| No | 1173 | 318 | 78·7 | 855 | 84·4 | |
| Missing§ | 297 | 76 | 221 | |||
| Drinking (χ 2 = 1·66) | Yes | 437 | 137 | 34·0 | 300 | 29·7 |
| Had quit | 32 | 9 | 2·2 | 23 | 2·3 | |
| No | 944 | 257 | 63·8 | 687 | 68·0 | |
| Missing§ | 301 | 77 | – | 224 | – | |
| Monthly personal income, $HK‡ (χ 2 = 39·20†††) | 1–5000 | 472 | 93 | 24·0 | 379 | 39·0 |
| 5001–10 000 | 293 | 116 | 30·0 | 177 | 18·2 | |
| 10 001–15 000 | 230 | 76 | 19·6 | 154 | 15·8 | |
| 15 001–20 000 | 127 | 39 | 10·1 | 88 | 9·0 | |
| >20 000 | 238 | 63 | 16·3 | 175 | 18·0 | |
| Missing§ | 354 | 93 | – | 261 | – | |
| Physical activity required at work (χ 2 = 143·50†††) | No full-time job | 478 | 88 | 21·9 | 390 | 38·6 |
| Not physically demanding | 458 | 90 | 22·4 | 368 | 36·4 | |
| Somewhat physically demanding | 315 | 128 | 31·8 | 187 | 18·5 | |
| Physically demanding | 133 | 76 | 18·9 | 57 | 5·6 | |
| Very physically demanding | 28 | 20 | 5·0 | 8 | 0·8 | |
| Missing§ | 302 | 78 | – | 224 | – | |
Significant differences between the two clusters: †P < 0·05, †††P < 0·001.
‡$US 1 = $HK 7·8.
§Missing values are not included in column percentages.
Health profile of the clusters
The odds of having any of the eight chronic conditions in the less active cluster was 1·62 times (= 1/0·62) that in the active cluster (95 % CI 1·20, 2·20), adjusted for age, sex, smoking, education and income (Table 4). The odds of having been diagnosed with cancer, diabetes, heart disease, stroke or asthma for the less active cluster were more than twice those for the active, although statistical significance was observed only for asthma (P = 0·049). The prevalence of obesity was similar in both clusters regardless of the criterion used to define obesity (Table 4).
Table 4.
Prevalence (%) of chronic health conditions and obesity in the FAMILY Project Cohort Study, Hong Kong, 2009
| Prevalence (%) | ||||||
|---|---|---|---|---|---|---|
| Cluster 1 (active; n 480) | Cluster 2 (less active; n 1234) | OR‡ | 95 % CI | Adjusted OR‡,§ | 95 % CI | |
| No chronic health condition | 81·82 | 72·67 | 0·59 | 0·45, 0·78 | 0·62 | 0·46, 0·84 |
| Chronic health conditions | ||||||
| Cancer | 0·45 | 1·55 | 3·45 | 0·80, 14·94 | 3·37 | 0·71, 16·04 |
| Diabetes mellitus | 4·09 | 5·26 | 1·30 | 0·76, 2·23 | 1·35 | 0·77, 2·38 |
| High blood pressure | 10·00 | 14·14 | 1·48 | 1·04, 2·11 | 1·21 | 0·82, 1·79 |
| High cholesterol/TAG | 6·14 | 10·09 | 1·72 | 1·11, 2·65 | 1·52 | 0·96, 2·42 |
| Heart disease | 0·68 | 2·07 | 3·08 | 0·92, 10·27 | 2·54 | 0·71, 9·16 |
| Stroke | 0·23 | 1·03 | 4·59 | 0·60, 35·40 | 3·97 | 0·50, 31·71 |
| Asthma | 0·68 | 2·33 | 3·47 | 1·05, 11·50 | 4·48 | 1·02, 19·65 |
| COPD | 0·91 | 0·69 | 0·76 | 0·23, 2·53 | 1·29 | 0·32, 5·24 |
| Obesity∥ | ||||||
| BFP | 37·13 | 43·79 | 1·32 | 1·04, 1·67 | 1·12 | 0·87, 1·44 |
| BMI | 7·20 | 6·32 | 0·87 | 0·55, 1·37 | 0·89 | 0·55, 1·42 |
COPD, chronic obstructive pulmonary disease; BFP, body fat percentage.
‡Reference category for the odds ratio: cluster 2.
§Adjusted for age, sex, smoking, education and income.
∥Indicators of obesity: BFP ≥25 % (males), ≥35 % (females); BMI ≥ 30 kg/m2.
Discussion
We identified two temporal PA patterns in this Hong Kong population by using cluster analysis on hourly accelerometer data. The active cluster had higher overall PA levels than the other and a distinctive weekend PA pattern with clear peaks at certain hours. In contrast, the less active cluster had similar temporal patterns on weekdays and weekends. The two clusters differed in counts per minute by 204·4 (the equivalent of daily PA of an American woman aged 60 years or above)( 17 ).
In summary, middle-aged people, males, smokers, the middle-income group and those having physically demanding jobs were more likely to be in the active cluster, consisting of one-quarter of the sample. These clear differences in demographic characteristics between the two clusters are discussed as follows. First, middle-aged participants, aged 35–54 years, were prominently represented in the active cluster, whereas the less active cluster had more of the younger and older participants (ages 15–34 years and ≥55 years). The variable age distribution of the less active group suggests a need for different approaches in developing age-specific intervention programmes. A similar age difference was found in an accelerometer study of PA in the USA, in which those aged 40–49 years were the most active group( 34 ). This may reflect the pandemic of sedentary lifestyles (e.g. screen time) among adolescents worldwide. Second, and surprisingly, in our sample the proportion of male smokers was higher in the active than in the less active group, contrary to a previous study in China( 16 ). This is because smokers were more likely to have physically demanding jobs than non-smokers. In the present study, 18·5 % (= 34/184) of the smokers had a physically demanding job, compared with 10·1 % (= 118/1174) among non-smokers. Third, the income distribution was significantly different between the two clusters. In particular, the proportion of lower-income groups was higher in the less active cluster than in the active cluster. A possible explanation for this is that a large proportion of females (346/777 or 44·5 %) were in lower-income brackets, and they may be housewives with part-time jobs (169/346 or 48·8 % of them had full-time jobs) and less time for exercise or sports activities. In addition, it is possible that some males (19·2 % = 118/614) in the low-income group of $HK 5001–10 000 may have shorter workdays (hence a lower chance of working on weekends) that allows them to exercise on weekends. Fourth, the proportion of participants having a physically demanding job was significantly higher in the active cluster than in the less active cluster. This finding further supported the cluster solution as participants with a physically demanding job were in general more active than those with a less physically demanding job during weekdays (392·5 v. 289·2 counts/min, P < 0·001).
The two clusters did not differ regarding other demographic variables, some of which had previously been reported to be associated with PA levels. One such variable was education: lower education level was associated with lower PA level among US and Swedish males( 10 ) and Chinese adults( 16 ), but not in our sample. This was perhaps because PA in Hong Kong is mostly not in the form of choosing to exercise for personal fitness, but is more related to the virtually universal use of public transportation (which always involves some walking). Another such variable was BMI( 10 , 16 ). Other studies of PA pattern, mostly among children, found that higher BMI was associated with lower PA levels( 8 , 21 , 35 – 37 ). This association was not found in our study, again perhaps because much of the PA in Hong Kong relates to routine transportation that is undertaken regardless of body weight. The other inconsistency regarded alcohol usage. Alcohol consumption has been linked to higher PA level among adults( 38 ) and college students( 39 ) in the USA, but there was only a non-significant difference in our sample.
Strengths and limitations
The present study offered a large sample from a Westernized and urbanized Asian population that complements the picture obtained from existing studies in European and North American samples. It used cluster analysis to analyse accelerometer data, with which we could identify PA patterns that are specific to certain times of the day.
A limitation of the study concerned sample representativeness. While it began from a random population sample, the inclusion criteria required at least 12 h of recorded time daily for four consecutive days, which entailed some loss of respondents and a potential selection bias. It is likely that the PA patterns for those who provided four valid days of accelerometer data were different from those who do not. While the study provides useful descriptive data, its major limitation lies in the cross-sectional design. The association of PA with chronic health conditions cannot demonstrate causality, as chronic conditions could have limited the ability to exercise( 40 ). Nevertheless, the findings regarding the association of PA with chronic health problems are important, as they may raise the priority of promoting PA among those with chronic health conditions.
Conclusions
The present study has broadened the scope of research on PA patterns and shed light on the potential of using accelerometer data and PA patterns to classify individuals in more precise categories, e.g. active on weekends, sedentary in the morning, etc. It has also contributed to the understanding of PA patterns of Hong Kong Chinese, the most Westernized and urbanized city of China, by identifying two clusters, one more active than the other. Neither cluster showed signs of regularly increased PA (implying exercise or sports activity) on weekdays, most probably reflective of the intense pace in city life. Therefore, we suggest that potential interventions to promote PA in Hong Kong may be most effective in targeting those who are sedentary on weekends. Providing free weekend PA programmes may also help increase participation rates. Further research is needed to refine the classification of PA patterns and establish standard cut-off points. Also, future study is needed to examine the predictive power of different PA patterns on long-term health outcomes.
Acknowledgements
The present study was a part of the project ‘FAMILY: a Jockey Club Initiative for a Harmonious Society’ supported by The Hong Kong Jockey Club Charities Trust. The funding body played no part in the data analysis, drafting of the manuscript, or the decision to submit for publication. The authors declare that they have no conflict of interest. P.H.L. conducted the data analysis and drafted the manuscript. Y.-Y.Y. performed the data collection including quality control. I.M., G.M.L. and T.H.L. designed the study and planned the analysis strategy. All authors reviewed and approved the manuscript. The authors sincerely thank Dr Wilson W.S. Tam, Dr Ben K.K. Li and Mr Paul T.K. Wong (School of Public Health, The University of Hong Kong) for their role in the development of the survey instruments and the preparatory work for this research.
References
- 1. Janssen I & LeBlanc AG (2010) Systemic review of the health benefits of physical activity and fitness in school-aged children and youth. Int J Behav Nutr Phys Act 7, 40. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Buman MP, Hekler EB, Haskell WL et al. (2010) Objective light-intensity physical activity associations with rated health in older adults. Am J Epidemiol 172, 1155–1165. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Bize R, Johnson JA & Plotnikoff RC (2007) Physical activity level and health-related quality of life in the general adult population: a systematic review. Prev Med 45, 401–415. [DOI] [PubMed] [Google Scholar]
- 4. Roberts CK & Barnard RJ (2005) Effects of exercise and diet on chronic disease. J Appl Physiol 98, 3–30. [DOI] [PubMed] [Google Scholar]
- 5. Pedersen BK & Saltin B (2006) Evidence for prescribing exercise as therapy in chronic disease. Scand J Med Sci Sports 16, Suppl. 1, 3–63. [DOI] [PubMed] [Google Scholar]
- 6. Ortega FB, Ruiz JR, Labayen I et al. (2011) Sleep duration and activity levels in Estonian and Swedish children and adolescents. Eur J Appl Physiol 111, 2615–2623. [DOI] [PubMed] [Google Scholar]
- 7. Nader PR, Bradley RH, Houts RM et al. (2008) Moderate-to-vigorous physical activity from ages 9 to 15 years. JAMA 300, 295–305. [DOI] [PubMed] [Google Scholar]
- 8. Ness AR, Leary SD, Mattock C et al. (2007) Objectively measured physical activity and fat mass in a large cohort of children. PLoS Med 4, e97. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Riddoch CJ, Mattock C, Deere K et al. (2007) Objective measurement of levels and patterns of physical activity. Arch Dis Child 92, 963–969. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Hagströmer M, Trost SG, Sjöström M et al. (2010) Levels and patterns of objectively assessed physical activity – a comparison between Sweden and the United States. Am J Epidemiol 171, 1055–1064. [DOI] [PubMed] [Google Scholar]
- 11. Pate RR, Pratt M, Blair SN et al. (1995) Physical activity and public health. JAMA 273, 402–407. [DOI] [PubMed] [Google Scholar]
- 12. Guthold R, Ono T, Strong KL et al. (2008) Worldwide variability in physical inactivity: a 51-country survey. Am J Prev Med 34, 486–494. [DOI] [PubMed] [Google Scholar]
- 13. Lee PH, Macfarlane DJ, Lam TH et al. (2011) Validity of the International Physical Activity Questionnaire Short Form (IPAQ-SF): a systematic review. Int J Behav Nutr Phys Act 8, 115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. van Poppel MNM, Chinapaw MJM, Mokkink LB et al. (2010) Physical activity questionnaires for adults: a systematic review of measurement properties. Sports Med 40, 565–600. [DOI] [PubMed] [Google Scholar]
- 15. Ferrari P, Friedenreich C & Matthews CE (2007) The role of measurement error in estimating levels of physical activity. Am J Epidemiol 166, 832–840. [DOI] [PubMed] [Google Scholar]
- 16. Peters TM, Moore SC, Xiang YB et al. (2010) Accelerometer-measured physical activity in Chinese adults. Am J Prev Med 38, 583–591. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Troiano RP, Berrigan D, Dodd KW et al. (2008) Physical activity in the United States measured by accelerometer. Med Sci Sports Exerc 40, 181–188. [DOI] [PubMed] [Google Scholar]
- 18. Hastie T, Tibshirani R & Friedman J (2008) The Elements of Statistical Learning: Data Mining, Inference, and Prediction. New York: Springer-Verlag. [Google Scholar]
- 19. Monda KL & Popkin BM (2005) Cluster analysis methods help to clarify the activity-BMI relationship of Chinese youth. Obes Res 13, 1042–1051. [DOI] [PubMed] [Google Scholar]
- 20. Rovniak LS, Sallis JF, Saelens BE et al. (2010) Adults’ physical activity activity patterns across life domains: cluster analysis with replication. Health Psychol 29, 496–505. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Page A, Cooper AR, Stamatakis E et al. (2005) Physical activity patterns in nonobese and obese children asessed using minute-by-minute accelerometry. Int J Obes (Lond) 29, 1070–1076. [DOI] [PubMed] [Google Scholar]
- 22. Yu XN, Tam WWS, Wong PTK et al. (2012) The Patient Health Questionnaire-9 for measuring depressive symptoms among the general population in Hong Kong. Compre Psychiatry 53, 95–102. [DOI] [PubMed] [Google Scholar]
- 23. Reilly JJ, Penpraze V, Hislop J et al. (2008) Objective measurement of physical activity and sedentary behaviour: review with new data. Arch Dis Child 93, 614–619. [DOI] [PubMed] [Google Scholar]
- 24. Rothney MP, Apker GA, Song Y et al. (2008) Comparing the performance of three generations of ActiGraph accelerometers. J Appl Physiol 105, 1091–1097. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. de Vries SI, Bakker I, Hopman-Rock R et al. (2006) Clinimetric review of motion sensors in children and adolescents. J Clin Epidemiol 59, 670–680. [DOI] [PubMed] [Google Scholar]
- 26. McClain JJ, Dodd KW, Berrigan D et al. (2010) How many accelerometer days are needed for stable population and individual weekly activity estimates? Med Sci Sports Exerc 42, Suppl. 5, S117–S118. [Google Scholar]
- 27. Jakicic JM, Gregg EW, Knowler W et al. (2010) Activity patterns of obese adults with type 2 diabetes in the Look AHEAD Study. Med Sci Sports Exerc 42, 1995–2005. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Ward DS, Evenson KR, Vaughn A et al. (2005) Accelerometer use in physical activity: best practices and research recommendations. Med Sci Sports Exerc 37, 11 Suppl., S582–S588. [DOI] [PubMed] [Google Scholar]
- 29. Freedson P, Melanson E & Sirard J (1998) Calibration of the computer science and applications, inc. accelerometer. Med Sci Sports Exerc 30, 777–781. [DOI] [PubMed] [Google Scholar]
- 30. Lloyd SP (1982) Least squares quantization in PCM. IEEE Trans Inform Theory 28, 129–137. [Google Scholar]
- 31. Thorndike RL (1953) Who belong in the family? Psychometrika 18, 267–276. [Google Scholar]
- 32. Cho HJ, Kim YJ, Jung HJ et al. (2008) OutlierD: an R package for outlier detection using quantile regression on mass spectrometry data. Bioinformatics 24, 882–884. [DOI] [PubMed] [Google Scholar]
- 33. Cohen J (1988) Statistical Power Analysis for the Behavioral Sciences, 2nd ed. Hillsdale, NJ: Lawrence Erlbaum Associates. [Google Scholar]
- 34. Tucker JM, Welk GJ & Beyler NK (2011) Physical activity in US adults: compliance with the physical activity guidelines for Americans. Am J Prev Med 40, 454–461. [DOI] [PubMed] [Google Scholar]
- 35. Stevens J, Murray DM, Baggett CD et al. (2007) Objectively assessed associations between physical activity and body composition in middle-school girls. Am J Epidemiol 166, 1298–1305. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Basterfield L, Adamson AJ, Frary JK et al. (2011) Longitudinal study of physical activity and sedentary behavior in children. Pediatrics 127, e24–e30. [DOI] [PubMed] [Google Scholar]
- 37. Metcalf BS, Voss LD, Hosking J et al. (2008) Physical activity at the government-recommended level and obesity-related health outcomes: a longitudinal study (Early Bird 37). Arch Dis Child 93, 772–777. [DOI] [PubMed] [Google Scholar]
- 38. Gordon-Larsen P, Boone-Heinonen J, Sidney S et al. (2009) Active commuting and cardiovascular disease risk. Arch Intern Med 169, 1216–1223. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Musselman JRB & Rutledge PC (2010) The incongruous alcohol-activity association: Physical activity and alcohol consumption in college students. Psychol Sport Exerc 11, 609–618. [Google Scholar]
- 40. van den Berg-Emons RJ, Bussmann JB & Stam HJ (2010) Accelerometry-based activity spectrum in persons with chronic physical conditions. Arch Phys Med Rehabil 91, 1856–1861. [DOI] [PubMed] [Google Scholar]