Correlates of high-impact physical activity measured objectively in older British adults

Abstract Background Exposure to higher magnitude vertical impacts is thought to benefit bone health. The correlates of this high-impact physical activity (PA) in later life are unknown. Methods Participants were from the Cohort for Skeletal Health in Bristol and Avon, Hertfordshire Cohort Study and MRC National Survey of Health and Development. Associations of demographic, behavioural, physiological and psychological factors with vertical acceleration peaks ≥1.5 g (i.e. high-impact PA) from 7-day hip-worn accelerometer recordings were examined using linear regression. Results A total of 1187 participants (mean age = 72.7 years, 66.6% females) were included. Age, sex, education, active transport, self-reported higher impact PA, walking speed and self-rated health were independently associated with high-impact PA whereas BMI and sleep quality showed borderline independent associations. For example, differences in log-high-impact counts were 0.50 (P < 0.001) for men versus women and −0.56 (P < 0.001) for worst versus best self-rated health. Our final model explained 23% of between-participant variance in high impacts. Other correlates were not associated with high-impact activity after adjustment. Conclusions Besides age and sex, several factors were associated with higher impact PA in later life. Our findings help identify characteristics of older people that might benefit from interventions designed to promote osteogenic PA.


Background
The progressive age-related disorder of osteoporosis, characterized by loss of bone mass and strength leading to fragility fractures, has large associated societal costs that are expected to rise for future generations. 1 Physical activity (PA) produces wide ranging benefits for older adults that include increases in bone mineral density, 2 prevention of falls and fractures and the maintenance of independent living 3-5 and physical capability. 6,7 Importantly, it is thought that beneficial effects of PA on bone are mediated by deformations caused by higher impacts or loading forces, leading to new bone growth, which subsequently reduces risk of osteoporosis. [8][9][10][11] For example, we recently developed 12 and validated 13 an accelerometer-based method for characterizing PA according to vertical impact, and showed that positive associations with lower limb bone strength in postmenopausal women were explained by exposure to vertical impacts ≥1.5 g. 8 To underpin strategies to increase older adults' exposure to higher impact PA, greater understanding is needed of the determinants of high-impact PA.
Previous studies examining predictors of older people's PA have primarily relied on self-report with few using objective measures of PA. [14][15][16] The most consistent correlates identified include age, sex and health indicators like physical function, with insufficient evidence for most other factors. [14][15][16] Reviews of existing studies have identified a need for more research using representative samples of population-based older people, in addition to more objective assessments of PA. [14][15][16] In fact, studies using objective PA measures have identified similar correlates including age, sex and health status. 17,18 In previous descriptive analyses, we showed that walking speed and self-reported higher impact PA were related to accelerometer-measured higher impact PA among older adults from the general population, 19 and that older age and worse physical performance were related to lower levels of high-impact PA among older adults attending an aerobics class. 13 Further, in a recent qualitative study, we showed that older adults identified a fear of falling as a barrier to high-impact activities and that those with joint replacement reported being advised against high-impact PA by their surgeons. 20 However, no previous study has performed a detailed quantitative analysis of the range of factors associated with accelerometer-measured PA producing rare but highly osteogenic vertical impacts at older age. Further, examining differences in how factors relate to high-impact and overall PA may provide useful insights for intervention design.
The aim of this study was to examine the associations between demographic, behavioural, physiological, psychological and social factors and accelerometer-measured highimpact PA among a population-based sample of older adults. Given that many of these factors are likely to be interrelated, an important secondary aim was to identify which factors were independently associated with highimpact PA in later life. We also examined how these same factors relate to an objective estimate of overall PA.

Study population
Participants were from the Vertical Impacts on Bone in the Elderly (VIBE) study, a multicohort collaboration initially set up to investigate the health consequences of higher impact PA across three population-based cohorts of older people; the Cohort for Skeletal Health in Bristol and Avon (COSHIBA), Hertfordshire Cohort Study (HCS) and the Medical Research Council National Survey of Health and Development (MRC NSHD). 19 COSHIBA is a representative population-based cohort of 3200 women recruited through fifteen general practices in the Bristol and Avon area during 2007-09. 21 Only the 1286 COSHIBA participants who consented to be contacted about future research studies in 2014 and who remained resident in the Bristol and Avon area were eligible to participate in the VIBE study. NSHD is a nationally representative sample of 5362 singleton births from one week in March 1946. 22,23 Most participants (79%) included in the home visit phase of the NSHD 24th data collection (2015-16) 24 were invited to participate in the VIBE study. HCS comprises 3225 singleton births in Hertfordshire between 1931 and 1939 and who still lived in the area during 1998-2003. 25 Only the 443 HCS participants who were previously included in the UK arm of the European Project on Osteoarthritis (EPOSA) 26 were invited to participate in VIBE.

Accelerometer measurements and data processing
Participants who were invited and agreed to accelerometry monitoring, subject to availability of monitors, were provided with a GCDC X15-1c triaxial accelerometer (Gulf Coast Data Concepts, Waveland, Mississippi), custom designed size specific elasticated belt, a time log and a stamped addressed package along with written and, if seen in clinic (COSHIBA) or during a nurse home visit (NSHD), verbal instructions. Accelerometers were configured with standardized settings prior to participant use with a sampling frequency of 50 Hz, a deadband setting of 0.1 g (the threshold which must be exceeded before a recording is made) and a timeout setting of 10 s (a single sample every 10 s is forced even if the recording is <0.1 g). Participants were instructed to wear the accelerometer securely positioned in the belt over their right hip pointing toward the centre of their body for 7 continuous days, removing only for sleeping, washing and swimming. A time log was provided for participants to record when the monitor was put on in the morning and taken off at night for each monitoring day and to state if there was any reason why that day had not been reflective of their normal activity.
Following standardized cleaning and processing (to remove movement artefacts and non-wear time), described in detail elsewhere, 12 we derived a measure of high-impact PA based on vertical (i.e. Y-axis) accelerations peaks (i.e. accelerations higher than the preceding and subsequent readings 12 ) measuring ≥1.5 g. The ≥1.5 g cut-point was selected as very few acceleration peaks were observed within higher g bands. 8,12,19 To examine differences in how each factor relates to high-impact PA and total PA, we derived a measure of overall PA by summing the number of triaxial (i.e. X, Y and Z axes) accelerations peaks measuring ≥0.5 g (i.e. all movements producing both lower and higher magnitude impacts). Periods of inactivity were removed by excluding movements producing ≤0.5 g, and activity data were normalized for wear time based on 7 valid days (≥10 h recording time) of 14 h. 12 All g values represent g over and above 1 g from earth's gravitational force.

Hypothesized correlates of high-impact PA
The factors hypothesized to be associated with higher impact PA at old age (Table 1) were selected based on previous literature on correlates of PA in older adults. [13][14][15][16]19,20,[27][28][29] These factors were grouped into demographic (age, sex, educational level, occupational class, marital status), behavioural (regular active transport, self-reported time spent in moderate-highimpact PA, smoking and alcohol status), physiological (body mass index (BMI), walking speed, falls, walking restricted due to pain, joint replacement, mobility aid use, difficulty walking (limping), fractures since age 45) and psychological and social domains (self-rated health, fear of falling, mental wellbeing, sleep quality, and contact with relatives, friends and neighbours). A detailed description of each factor including any harmonization process performed for data analysis is provided in Table 1.

Statistical analyses
Means and standard deviations were used to summarize continuous measures and proportions to describe categorical measures. Differences in each measure between the three participating cohorts were investigated using ANOVA for continuous measures and chi-squared tests for categorical measures and where only two cohorts had relevant data, differences were examined using t-tests for continuous measures and Chi-squared tests for categorical measures. Accelerometer data were expressed as medians and interquartile ranges due to their skewed distributions and, differences between cohorts were examined using a nonparametric k-sample equality of medians test. Linear regression was subsequently used to examine associations between each selected factor and highimpact PA, and overall PA. Interaction terms were used to test cohort differences in associations of each factor with accelerometer outcomes and subsequent analyses were performed on all participants combined with cohortadjustment after little evidence of interactions was found. Interaction terms were also used to test sex-differences and subsequently men and women were combined with adjustment made for sex after no evidence of sex-interaction was found. For categorical factors, deviation from linearity was tested by comparing models with categorical exposures to models with same exposure entered as a continuous term, and where no evidence of deviation from linearity was found they were treated as continuous terms (with estimates representing a per category change).
First models were adjusted for age, sex and cohort. Second models included adjustment for all factors within each domain followed by final third models that were concurrently adjusted for all factors from all domains. Included in these second and third models were all those factors with a statistical significance of P ≤ 0.1 from tests of association with higher impact PA and/or overall PA in the first (age, sex and cohort-adjusted) models. Accelerometer outcomes were log-transformed due to their skewed distributions. To minimize the potential for bias due to missing data, we used multiple imputation by chained equations 32 to impute missing data for each factor thereby including all participants with valid accelerometer outcomes. Imputation models were run using 20 multiply imputed datasets which were combined using Rubin's combination rules. Results from imputed datasets were similar to results from complete case analysis and the former are presented. We also calculated adjusted R 2 of our final models to identify how much variance in high-impact PA and overall PA was explained by the selected factors, after accounting for important predictors and those which were not independently related to outcomes in final model.

Descriptive statistics
A total of 1187 participants aged between 69 and 88 years (mean = 72.4) (72.8% females) had valid measures of PA from accelerometers. Of these, 430 were from COSHIBA (100% females, mean age = 76.6), 649 from NSHD (50.2% females, mean age = 69) and 108 from HCS (42.1% females, mean age = 78.4). Men had greater levels of high-impact PA and overall PA than women (Table 2). Greater levels of both high-impact PA and overall PA were recorded in NSHD than in COSHIBA or HCS, reflecting their younger age and a higher proportion of males (Table 2). High-impact PA and overall PA were moderately correlated (Pearson correlation coefficient between log-high-impact PA and log-overall PA = 0.6). Distribution of each selected factor by cohort is provided in Table 2. These were used to derive a comparable measure of broken bones since age 45

Correlates of high-impact PA
In minimally adjusted models (Table 3), older age, female sex, lower education level and occupational class, lack of regular active transport, less self-reported time spent in higher impact PA, current smokers, higher BMI, slower walking speed, experiencing pain during walking, regular use of a mobility aid, presence of a noticeable limp, lower mental wellbeing, poorer self-rated health and a fear of falling were all associated with lower levels of high-impact PA. Conversely, marital status, alcohol drinking, recent falls, previous fracture and sleep quality were unrelated to high-impact PA whereas the association of previous joint replacement with high-impact PA was borderline (Table 3). Similarly, speaking with friends and relatives (COSHIBA), visiting or being visited by friends and relatives (NSHD) and the Lubben Social Network Scale (HCS) were not associated with high-impact PA (not shown). Following mutual adjustment for all factors showing initial associations, age, sex, education, active transport, selfreported high-impact PA, walking speed and self-rated health were all independently related to high-impact PA whereas occupational class, BMI, recent falls and sleep quality showed borderline associations with high-impact PA (Table 4). On the other hand, smoking status, pain during walking, mobility aid use, noticeable limp, mental wellbeing and fear of falling were no longer related to high-impact PA after adjustment (Table 4)

Correlates of overall PA
Except for marital status and previous fracture, all other factors examined were associated with overall PA in models with minimum adjustments (Table 3). As with high-impact PA, the social network measures were also unrelated to overall PA (not shown) though a weak association was observed in NSHD between regularly visiting friends and higher overall PA (sex-adjusted difference in log-overall PA for visiting friends less than once/week versus at least once/week was −0.11 (95% confidence intervals: −0.22, 0.00, P = 0.05).
After mutual adjustment for all factors showing initial associations with PA outcomes, age, sex, education, active transport, walking speed and self-rated health all independently predicted overall PA (Table 4). In addition, and contrary to findings for high-impact PA, smoking status, BMI and mobility aid use were also independently associated with overall PA whereas self-reported high-impact PA and sleep quality showed little evidence of associations with overall PA (Table 4). The demographic, behavioural, physiological, and psychological/social factors explained 20.7% (adjusted R 2 = 0.207, min = 0.204, max = 0.211), 31.6% (adjusted R 2 = 0.316, min = 0.312, max = 0.32), 42.5% (adjusted R 2 = 0.425, min = 0.422, max = 0.429) and 29.3% (adjusted R 2 = 0.293, In COSHIBA, height was measured to the nearest mm using a Harpenden stadiometer and weight to the nearest 50 g using Tanita weighing scales whereas in NSHD height was measured to the nearest mm using a Leicester stadiometer and weight to the nearest 100 g using Tanita weighing scales.
Ref: reference category.

Discussion
Main finding of this study We examined the associations of a wide range of demographic, behavioural, physiological, psychological and social factors with accelerometer-measured high-impact PA among participants aged in their late 60s, 70s and 80s recruited from three British population-based cohorts. Besides an older age and female sex, several factors, namely, lower education, lack of regular active transport, slower walking speed, less reported time in highimpact PA and poorer self-rated health were all independently associated with lower levels of high-impact PA in later life. On the other hand, smoking status, pain during walking, mobility aid use, noticeable limp, mental wellbeing and fear of falling were only related to high-impact PA prior to adjustment for other factors. All factors combined explained nearly a quarter of the variance in levels of high-impact PA between individuals. Moreover, while broadly similar findings were observed when examining correlates of overall PA (i.e. PA encompassing lower as well as higher impacts), there were important differences. Specifically, BMI, smoking status and mobility aid use were independently associated with overall but not high-impact PA whereas, based on qualitative assessment of differences in effect size, reported higher impact PA, sleep quality, education and self-rated health appeared more strongly related to high-impact PA. Walking speed was an important correlate of both highimpact PA and overall PA.

What is already known on this topic
This is the first study to examine the correlates of highimpact PA assessed by accelerometer at old age however, some of the factors identified as predictive of higher impact PA are similar to factors related to overall PA in previous studies of older adults. For example, regular PA reported by 8881 Australians aged 65+ years was independently associated with male sex, younger age, ability to travel independently, better physical functioning and lower psychological distress whereas no independent associations were found for employment status or fear of falling. Likewise, among a large sample of 48-83-year-old Swedish women, reported PA was lower with increasing age, BMI and in smokers. 29 PA counts per minute assessed via accelerometers in 850 70-77-yearold Norwegians were related to cardiorespiratory fitness and sex but not social support, 17 while among 560 British adults aged at least 65 years, independent predictors of average daily accelerometer step-counts included age, general health, disability, BMI and number of long walks. 18 What this study adds Our findings are important as they offer a first look at factors related to accelerometer-measured higher impact PA which is thought to be important for bone health in older populations. Consistent with and extending our previous unadjusted analyses 19 is that walking speed, a strong predictor of survival, 33 was related to both high-impact and overall PA even after adjustment, which may suggest an important role for underlying physical function. Likewise, self-rated health predicted both high-impact and overall PA, but appeared more strongly related to high-impact PA, which may reflect effects of underlying physical health. 34 In addition, reported high-impact PA was strongly predictive of accelerometer-measured high-impact PA, but not overall PA, including after adjustment, which indicates that our objective measure of high-impact PA is capturing time spent in highimpact activities.
That education was related to both high-impact and overall PA is consistent with studies showing lower self-reported

CORRELATES OF HIGH IMPACT PHYSICAL ACTIVITY
and objectively measured PA among older adults from lower socioeconomic backgrounds, 35 as well as with our qualitative findings that older adults identified greater knowledge of exercise benefits to be a facilitator of higher impact PA. 20 Fear of falling was initially related to high-impact PA, as previous reported in our qualitative study, 20 however, that this association, and that with other markers of functional status, was lost after adjustment may be because it was captured by other model covariates like self-rated health. Finally, that regular active transport was related to higher counts of both PA parameters might reflect effects of active lifestyles. 36

Limitations of this study
Strengths of this study include use of raw accelerometer recordings to derive objective measures of high-impact PA, comparison of findings with an objective measure of overall PA and inclusion of three population-based cohorts encompassing a broad age range of older individuals helps to increase power and generalizability of findings. Limitations of this study include its cross-sectional study design, which precludes inference regarding causality especially as reverse causation is possible. Allocation of the independent variables into domains was subjective which is another potential limitation of our approach, however this allowed for an organized   Walking restricted due to pain CORRELATES OF HIGH IMPACT PHYSICAL ACTIVITY sequential analysis. Additionally, as many correlated variables were simultaneously adjusted for in our final models these may be over-adjusted leading to an underestimation in effect sizes. It was also not possible to examine levels of PA based on conventional measures of energy expenditure as the GCDC accelerometers only recorded PA impact magnitude. Furthermore, we only had information on individual-level factors in VIBE, however, both perceived and observed environmental characteristics have been associated with PA 37,38 and thus future studies should investigate how they might relate to high-impact PA. Of further consideration is that VIBE participants tended to have lower BMI and higher educational level compared with others who did not participate in VIBE 19 and this selection bias may have led to underestimations of associations. Finally, measurement error in some of the factors studied might influence our findings, as could residual confounding due to unmeasured factors.

Implications and conclusions
Our findings suggest that maintaining physical function, wellbeing and health may be important for promoting osteogenic PA in later life. Further, certain groups of older people such as those with lower educational qualifications may benefit from supportive interventions to increase higher impact PA, whereas older women of any age and the oldest men and women may both be target populations for interventions. In conclusion, by using accelerometers calibrated to detect high magnitude vertical impacts from ground reaction forces, we showed that several factors were independently associated with osteogenic PA in older British men and women.