Study design
The Allied-Dunbar National Fitness Survey (ADNFS) was conducted in a representative sample of English adults between February and November 1990 (http://discover.ukdataservice.ac.uk/catalogue?sn=3303 [14]). Probability sampling procedures randomly selected 30 English parliamentary constituencies out of 523. Within each constituency, 200 addresses were randomly chosen from the electoral register and one adult per household chosen at random [15]. Out of the 5,698 men and women aged 16 and over approached for survey, 4,316 participated in ADNFS; a 76 % response rate. Due to the sampling procedure which focused on the adult population in households, non-responders tended to be younger and from lower social classes. However, differences were small and ADNFS participants were representative of the age and sex distribution of the English population at that time [12, 13]. Interviewers from the Social Survey Division of the Office of Population Censuses and Surveys (OPCS) conducted structured interviews in participants’ homes. Information on socio-demographic characteristics, physical activity, health and lifestyle were collected by questionnaire at baseline interview [14]. The survey protocol was approved by the local Research Ethical Committees of each Health District involved [16].
Outcome measurement
The main outcome measure was all-cause mortality. All ADNFS participants were tagged for mortality and migration at the Office of National Statistics (ONS) from their survey date in 1990 to 14th May 2014 (n = 1,175 deaths and n = 145). Deaths were coded into four categories (cardiovascular, cancer, suicide/violence/accidental, and other) based on the classification of the underlying cause of death against the International Classification of Diseases, tenth edition (ICD-10). Classifications for CVD deaths were defined by ICD codes in the range I00–I99, cancer deaths by codes in the range C00–D48, and suicide/violence/accidental deaths by codes in the range V01–Y98. This classification was independently done by an assessor masked to exposure data. A 5 % sample was randomly selected and independently classified by a second researcher, with 100 % agreement.
Explanatory variables and covariates
The primary exposure was the number of occasions of self-reported 20-min episodes of moderate to vigorous physical activity in the past month (activity bouts). The ADNFS questionnaire was designed to capture the frequency (number of times in past month), duration (length of all activity engaged in lasting at least 1 min) and intensity (scored according to published energy costs [17–21]) of all activity engaged in and has been validated against walking speed and stair climbing [22]. At the time of the ADNFS survey (1990), three episodes of at least moderate activity of 20 min duration per week were recommended for maintaining/improving cardio-respiratory fitness and provide the rationale for producing a summary of current activity based on the number of occasions of moderate to vigorous activity of at least 20 min duration for each main activity type [14]. Information on bouts activity <20 min in length was not available to us. Participants were classified according to the range, frequency and intensity of self-reported physical activity bouts lasting at least 20 min over the 4 weeks prior to interview. Habitual activities comprised all sports and recreation, transportation, home activities and occupation, and was summarised into three energy bands; vigorous: ≥7.5 kcal/min (approximately ≥6.5 METs), moderate: 5–7.49 kcal/min (4–6.49 METs), and light: 2–4.9 kcal/min (1–3.99 METs) [14]. A habitual physical activity variable was derived based on the number of 20-min bouts of moderate to vigorous activity (>5 kcal/min; approximately >4 METs) in the past month, referred to here as physical activity bout. Current guidelines recommend at 150 min of at least moderate activity per week [9] and as the reference period used in the ADNFS study to assess current activity was past 4 weeks, recommended levels equate to 600 min of at least moderate activity per month. Thus, a categorical habitual activity measure was derived based on the number of 20-min physical activity bouts achieved, where the inactives reported 0 bouts, low actives reported 1–14, moderate actives reported 15–29, and actives—those meeting physical activity guidelines, here 30 bouts of 20 min—reported 30+ bouts per month, respectively.
A lifetime physical activity variable was collected at baseline and classified participants according to the proportion of their life spent regularly active in sports and exercise (participating in sports/recreation at least once a week, for at least 2 months of the year) since 14 years [23]. A lifetime participation proportion was calculated for every sports and exercise activity as previously published [23], by dividing the number of years of regular participation since age 14, by the current age minus 14 years. The decision to use the 14 year cut off, to restrict PA to only sports and recreational activities and to define regular lifetime activity of once a week for at least 2 months a year was based on early evaluation work by the ADNFS study team. They found inconsistencies in the reporting of childhood (<14 years), school curriculum linked activities and the frequency of lifetime activity (see ADNFS technical report [14].
Interviewers collected information on date of interview, socio-demographic characteristics (age, sex, occupation and marital status), regional health authority (RHA; NHS administrative units between 1974 and 1996) other lifestyle habits (smoking status and alcohol consumption) and prevalent disease at time of interview (stroke/MI, cancer). Socio-economic categories were assigned on the basis of occupation, according to the 1980 Registrar-General’s OPCS classification and comprised: (I) professional, (II) intermediate, (III) skilled, (IV) partly skilled, (V) unskilled and (VI) unclassified. Smoking (smokers, ex-smokers and non-smokers) and alcohol consumption (lights, moderate, heavy, none) were self-reported. BMI was recorded for a sub-sample of participants (n = 2,708/3,918) using a calibrated digital weighing scale and a metal stadiometer. Overweight was defined as 25 kg/m2 ≤ BMI < 30 kg/m2 and obesity as BMI ≥ 30 kg/m2, according to WHO criteria [24].
Statistical analyses
Baseline characteristics were summarised separately according to survival status using means (SD) and percentages, and differences were examined using logistic regression. Individuals with missing data for an exposure of interest were included in all analyses not involving that specific exposure. To assess the nature of the relationship between activity and mortality, tests for departure from linear trend comprised a model including both categorical and log-linear terms for physical activity, followed by a Wald test for joint effect of categorical terms. A Cox regression model for the log hazard of death as a function of a restricted cubic spline for bouts of activity was fit to the data.
Cox proportional hazards regression was used to estimate the hazard ratios and corresponding 95 % confidence intervals (CI) for the association between meeting minimum activity guidelines [9] (here 30 or more 20-min bouts per month), lifetime physical activity, and all-cause mortality. Age is a strong determinant of mortality risk [25] and was used as the underlying time-scale for all models. Person-time for each participant was calculated from age at ADNFS interview to age at death or the study censor date (14th May 2014), whichever came first. A step-wise forward regression approach assessed the strength of the association between each variable and mortality, and overall model fit. Only those variables improving model fit were included in final models. Model 1 adjusted for age, sex and smoking status; known strong risk factors for mortality. Model 2 additionally adjusted for social class, geographical area, anxiety/depression and season of interview. Likelihood ratio tests (LRT) compared models with and without potential predictor variables. Interactions between physical activity and sex, social class and smoking status were examined via LRT.
To ascertain whether physical activity mediates its effects on mortality through BMI, Model 2 was repeated with and without BMI as a covariate and the percentage change in HR associated with mortality risk for physical activity variables was assessed. To assess bias from antecedent disease, sensitivity analyses were conducted omitting [1] those with a self-report of stroke, heart attack (MI) or cancer at baseline (n = 258), [2] deaths occurring within 5 years after interview (n = 31) and [3] those who were underweight (BMI < 18.5 kg/m2, n = 43). To assess the influence of migration on effect estimates, sensitivity analyses considering person-time before emigration were carried out, which censored individuals at date of first emigration [26], where available. Proportional hazards assumption was formally tested using the Schoenfeld and scaled Schoenfeld residuals. To estimate how much premature mortality could be prevented if all inactive individuals became low active, moderately active or active, the population attributable fraction (PAF) was calculated [27], by subtracting the marginal mean between–scenario hazard ratio and its confidence limits from 1 and adjusting for all known measured confounders (Model 2). The PAF for all-cause mortality associated with incremental increases in activity bouts were also calculated for the population as a whole, as well as stratified by sex (Supplementary Table 2). Statistical significance was set at a level of p < 0.05. Data was analysed using STATA version 13.0 (Stata, College Station, TX, USA).