Study design
We analyzed data from the 1958 National Child Development Survey (NCDS). The dataset is freely available upon request at the UK Data Service[15] (https://www.ukdataservice.ac.uk/deposit-data/). Full description of sampling design and methods can be found elsewhere[16]. Briefly, this birth cohort collected data from approximately 94% (N=17,415) of all births between 3rd and 9th March 1958 in England, Wale, and Scotland. Sociodemographic and behavioral information from both parents besides data regarding to pregnancy and the child were assessed. Further, any child born in the Great Britain in that specific week were identified by school registers and added to cohort sample during second through fourth sweeps (1965, 1969, and 1974). Flow diagram about sample composition is shown in Figure S1 (supplementary material).
Fifty-five years after baseline assessments, this cohort remains largely representative of the sample that it was drawn. In the latest sweep (2013) 9,137 participants were interviewed and cumulative deaths through cohort summed 1,548 (response rate: 61.3%) [16]. After baseline measurements, new sweeps occurred in 1965 (cohort age: 7 years), 1969 (11 years), 1974 (16 years), 1981 (23 years), 1991 (33 years), 2000 (42 years), 2002 (44 years), 2004 (46 years), 2008 (50 years), and the latest in 2013 (55 years). The 2002 and 2004 sweeps were not included in the present analysis due to methodological distinctions compared to the others (i.e. self-reported and telephone-based interview, respectively). The next sweep was programed to occur in 2020 and 2021; however, fieldwork was paused in light of COVID-19.
Outcome
We considered multimorbidity as the co-occurrence of more than one of the following morbidities at age 55: obesity, hypertension, diabetes, depression, asthma, cancer, visual and hearing impairment. Then, based on this classification, a dichotomic variable was created.
Exposure
LTPA was assessed from 1965 to 2013. Data collected in 2002 (age 44 sweep) and 2004 (age 46 sweep) regarding to LTPA were not used in this study because they were assessed by different interview methods (self-reported and telephone-based interview, respectively), which may decrease significantly both reliability and comparability of measurements[17]. All included LTPA data were assessed by face-to-face interview-administered paper-based questionnaire. Table S1 (electronic supplemental material) describes how it was measured in each sweep and how we operationalized it for this study. Briefly, participants were classified as physically active when performed physical activity regularly (age 7, 11, and 16) and at least once per week (from age 23 to the latest), as in previous work[18].
Confounding variables
All multivariate analyses were adjusted for the following variables: gender, marital status, education level, income, country of birth, ethnicity, body mass index (BMI), smoking, alcohol intake, hours of sleep, and LTPA. We considered sex, country of birth, and ethnicity collected at birth sweep. We used LTPA as a possible confounder in analysis where the main exposure was a time-specific sweep and it was adjusted for previous and future LTPA practice. All remaining used variables were assessed in the latest sweep (2013).
Statistical analysis
Descriptive analysis is reported as absolute and relative frequencies. Difference between groups was verified using chi-square tests. To evaluate the PA effect during earlier life, we stratified the variable in childhood (age 7 and 11 sweeps), teenage (age 16 sweep), young (age 23 sweep), middle (age 33 sweep), and middle-to-old age (age 50 and 55 sweeps) adults.
Logistic regression was performed as crude analysis (model 1) and using hierarchical model adjusting for gender, marital status, education level, income, country of birth, ethnicity (model 2), as well as BMI, smoking, alcohol intake, and hours of sleep. For regression analyses, PA was categorized as reported in Table S1: inactive and active.
Then, a structured modeling approach developed by Mishra et al.[19] was used to select the most appropriate life course model for multimorbidity at age 55. Four different hypothesized life course model were examined: saturated, critical, sensitive, and accumulation. Saturated model included all possible exposure combinations and interactions and describes all possible trajectories of LTPA throughout life course (childhood, teenage, young, middle, and middle-to-older adults).
Accumulation model was tested in two versions. First, a strict model (continuous) was assessed by adding the number of times an individual reported being physically active across their life course to form an overall score, which was then used as the exposure. This model assumes that the effect of physical activity at each period is the same. Second, a relaxed model (categorical) was examined in which all time periods are contributing to multimorbidity at age 55 but not necessarily in an equal way.
Critical period model assumes that only physical activity in a certain age influences multimorbidity at age 55 regardless of any other time period. Similarly, sensitive model was tested to allow the examination of the varied effect of LTPA across the life course, which can be modelled by simultaneously including all physical activity variables in the model. Finally, a null model was tested with only our outcome at the model[19].
To identify the most appropriate life course model to explain multimorbidity at age 55, likelihood ratio test was conducted comparing each life course model to the saturated model. When nested life course models (critical period, sensitive period, and accumulation) provided similar fit to the fully saturated model (p>0.05), the one with the lowest Akaike’s information criterion (AIC) was selected. When more than one model presented p-value higher than 0.05 and there is not a large difference in p-values, the simpler model was selected[19].
To minimize data loss, missing data were imputed using multiple imputation chained equations as recommended by the NCDS user guide[20]. We ran imputation models with all variables from our logistic models across 20 imputed datasets. When imputed results were similar to those obtained using observed values, the latest was presented. All statistical analyzes were carried out using STATA 13.1 software[21]. A p value of <0.05 was accepted as statistically significant.