In order to address these limitations, we conducted a prospective cohort study using both self-reported and accelerometer-measured PA to examine whether the association of PA with all-cause mortality and the incidence of four relevant long-term conditions was modified by the concentrations of multiple air pollutants. Based on the Danish studies, we hypothesised that air pollution does not modify the association between PA and health outcomes.
A prospective cohort study was conducted using UK Biobank data. A total of 502,366 participants were recruited between 2006 and 2010. The ages of the participants ranged from 37 to 73 years old. Participants attended one of 22 assessment centres across England, Scotland, and Wales. At the assessment centre, all participants underwent face-to-face interviews, completed a self-administered touchscreen questionnaire, and underwent a physical examination (including height, weight, and blood pressure) conducted by trained staff. Participants’ ethnicity, smoking status, and alcohol intake were all self-reported. Comorbidities and medical history were obtained from self-report of physician diagnoses and verified by staff during face-to-face interviews. This study was conducted under project 71,392.
The outcomes of the study were all-cause mortality and incident long-term conditions: major adverse cardiovascular disease (MACE), chronic obstructive pulmonary disease (COPD), and type 2 diabetes (T2D). Death certificates were obtained from the National Health Service Information Centre (England and Wales) and the Scottish National Health Service Central Register (Scotland). Hospital admission records were obtained from the Health Episode Statistics (England and Wales) and Scottish Morbidity Records (Scotland). All linkages were undertaken by the UK Biobank and detailed procedures can be found at http://content.digital.nhs.uk/services. Mortality data were available up to November 2022 at the time of analysis and hospital admission data were available up to October 2022 in England, August 2022 in Scotland, and May 2022 in Wales. Therefore, analysis of all-cause mortality and disease-specific outcomes were censored respectively at these dates or date of death if this occurred earlier. International Classification of Diseases Tenth Revision (ICD-10) codes were used to ascertain incidence of long-term conditions (hospitalisations or deaths) in this study. MACE was defined as incident myocardial infarction, heart failure, or stroke (ICD-10 codes: I11.0, I21, I42.0, I42.6-42.7, I42.9, I50, I60-64); COPD using ICD-10 codes J41-44 and J47; T2D using ICD-10 code E11; and cancer using ICD-10 codes C00-C97. Cancer registry was not used for this study because the data was outdated when the analysis was conducted.
Self-reported physical activity (PA) data were collected by UK Biobank via a touchscreen questionnaire, using the International Physical Activity Questionnaire (IPAQ) Short Form [19] to provide information on the frequency, intensity and duration of walking, moderate and vigorous activity. Data processing followed the guidelines published by IPAQ [19]. The volume of PA was calculated by applying weights to the time spent on activities of different intensities to produce total physical activity in MET (metabolic equivalent of task)-hour/week.
Accelerometer-measured PA was obtained using Axivity AX3 wrist-worn triaxial accelerometers worn by 103,686 UK Biobank participants between 2013 and 2015. Participants who provided an email address to UK Biobank were invited at random [20]. The dominant wrist of each individual was used over a period of 7 days at 100 Hz, as has been described elsewhere [20]. The 7161 participants with insufficient wear time (< 72 h), missing data, or poor device calibration were excluded. A two-step approach based on random forest and raw acceleration was used to categorise time spent in light, moderate, and vigorous PA based on a method described previously [21, 22]. In this study, the time spent in intensity-specific PA were combined to produce total MET-hour/week as 1.5 * LPA + 4 * MPA + 8 * VPA, where LPA, MPA, and VPA were weekly hour spent in light, moderate, and vigorous intensity PA.
Based on the European Study of Cohorts for Air Pollution Effects (ESCAPE) protocol, annual outdoor air pollutant concentrations at home addresses of cohort participants in 2010 were estimated using a European land use regression model, including PM ≤ 10 µm (PM), PM ≤ 2.5 µm (PM), PM 2.5-10 µm (PM), PM filter absorbance, nitrogen oxides (NO), and nitrogen dioxide (NO) [23, 24]. PM filter absorbance was calculated using the measured filter reflectance using the ISO (International Standardization Organization) 9835 (1993) formula. Absorbance has previously been found to be highly correlated with elemental carbon [25]. Land use regression models are a statistical method that uses multiple linear regression to analyse the association between pollutant concentrations measured at multiple monitoring points and predictors such as traffic, land use, and topography [24]. Separate models had previously been conducted for PM and NO. The PM model included measurements taken between 2008 and 2011 from 20 places in Europe with 20-40 monitoring sites each [26]. These included two areas (39 sites) in the UK: Manchester and London/Oxford. Nitrogen oxides (NO and NO) were measured, between 2008 and 2011, in 20 European study regions, and NO was measured in 36 regions, including three areas (119 sites) in the UK: Manchester, London/Oxford, and Bradford [24]. The measurements were validated with acceptable leave-one-out-cross-validation root mean squared error, e.g. < 1.4 μg/m for PM [23] and median) was used as the reference in the binary analyses to ensure correct estimation of additive interaction [34].
Cox proportional hazard regression models were used to examine the associations of PA and air pollution with health outcomes. Hazard ratios (HRs) and 95% confidence intervals (CI) were used to estimate the effect sizes. Follow-up started at the UK Biobank baseline assessment visit for self-reported PA and at the final date of accelerometer wear for objectively measured PA. Proportional hazard assumptions were checked via tests of the Schoenfeld residuals. None of the tests showed a significant violation of the assumptions.
Both multiplicative and additive interactions (relative excess risk due to additive interaction [RERI]) were calculated for the binary PA and air pollution variables. The above analyses also were replicated treating PA and air pollution variables as numeric variables (scaled to their sample interquartile range [IQR]) since even though dichotomised variables could be easier to interpret they could lead to imprecise and underpowered estimation.
The inferential analyses described above were conducted in the 10 imputed datasets using full conditional specification multiple imputation using chained equations to address missing data in covariates [35]. Rubin’s rule was used to pool the findings from the 10 imputed data. Ten imputations were selected based on the 16% total missingness in the data [36]. PA and air pollution variables were dichotomised as above or below the sample median. Complete case analysis including all people with non-missing data was conducted as a sensitivity analysis. Analysis of self-reported PA was replicated among people with accelerometer data only. Statistical analyses were performed using R statistical software 4.3.0 with the packages mice, survival, and interactionR.
