Data sources and searches
The protocol for this this systematic review and meta-analysis was registered in the PROSPERO prospective register of systematic reviews (CRD42021277514). The review was conducted according to PRISMA and MOOSE guidelines [17, 18] (Electronic Supplementary Materials 1–2) (Fig. 1). We searched MEDLINE and Embase (using OvidSP interface) from inception to 15 September 2021 with no restrictions on language. We planned to translate all non-English language studies using translators or the “Google Translate” service. The computer-based searches used a combination of keywords or terms relating to the exposure (“physical activity”, “exercise”, “aerobic training”) and outcome (“pneumonia”, “lower respiratory tract infection”). The full search strategy is presented in Electronic Supplementary Material 3. One author (SKK) initially screened the titles and abstracts of the retrieved references to assess the potential for inclusion in the review. Screening was conducted using Rayyan, a free online bibliographic tool that helps expedite the initial screening of abstracts and titles using a process of semi-automation while incorporating a high level of usability. This was then followed by full-text evaluation of the selected titles and abstracts. This was independently conducted by two authors (SKK and SS), with disagreements resolved in consultation with a third author (JAL). To identify potential articles missed by the search strategy, the reference lists of relevant studies and review articles were manually scanned and citing references were also checked in Web of Science.
We included all observational population-based observational cohort (retrospective or prospective) studies that had evaluated the relationship of physical activity with the risk of pneumonia in adult general populations and had at least 1 year of follow-up. We excluded the following studies: (i) case–control study designs because of lack of temporality, (ii) those involving elite athletes and/or evaluated competitive or endurance sports, and (iii) those evaluating the associations between measures of fitness (e.g., cardiorespiratory fitness, physical fitness, exercise capacity) and risk of pneumonia.
Data extraction and risk of bias assessment
Using a standardised data collection form, the lead author (SKK) initially extracted relevant data from eligible studies and a second author (SS) independently checked the data using the original articles. A third author (JAL) was involved to help resolve any disagreements. Data on the following study characteristics were extracted: first author surname and year of publication, geographical location, year of enrolment/data collection, study design, demographic characteristics (age and percentage of males), sample size, duration of follow-up, assessment of physical activity, risk comparisons, number of outcome events, the most fully adjusted risk ratios of outcomes (and corresponding 95% confidence interval [CIs]), and list of covariates adjusted for. The level of adjustment was defined as ‘ + ’ minimally adjusted analysis, i.e. age and/or sex; ‘ + + ’ as adjustment for established risk factors without inflammation, i.e. age and/or sex plus body mass index (BMI), socioeconomic status, alcohol consumption, smoking, and comorbidities; and ‘ + + + ” as adjustment for established risk factors including inflammation. We also extracted data on minimally adjusted estimates (unadjusted, age, or age and sex) to be used for sensitivity analysis. When there were multiple publications of studies using data of the same cohorts, the study selection was based on a single set of most comprehensive results to avoid double counting of a cohort in the pooled analysis. The most up-to-date comprehensive study was used (i.e. the one with the most extended follow-up or analysis covering the largest number of participants and events). The Cochrane Risk of Bias in Non-randomised Studies–of Interventions (ROBINS-I) tool was used to assess the risk of bias within individual observational studies . This tool assesses risk of bias for the following domains: confounding, participant selection, classification of interventions, deviations from intended interventions, missing data, outcome measurements, and selective reporting. The risk is quantified in each domain as low risk, moderate risk, serious risk, or critical risk; then, an overall judgement of the risk of bias is provided for each study. Finally, to grade the quality of evidence across outcomes, we used the Grading of Recommendations Assessment, Development and Evaluation (GRADE) tool, a widely adopted reproducible and transparent framework for grading certainty in evidence and used in clinical decision making . GRADE considers the following criteria: study limitations, inconsistency of effect, imprecision, indirectness, and publication bias, and has four levels of evidence: very low, low, moderate, and high.
Data synthesis and analysis
Summary measures of association were reported as relative risks (RRs) with 95% CIs. All studies categorised physical activity exposure (e.g., leisure-time physical activity, total or any physical activity) into two or more groups. There was varied reporting of risk comparisons; hence, the need to provide some consistency to enhance interpretation of the findings. Since the risk estimates could not be transformed into consistent comparisons, the extreme groups (i.e., top versus bottom or maximum versus the minimal amount of physical activity) reported for each study were used for the analyses. This approach, which we have utilised in previous similar meta-analyses [22,23,24,25], is considered reliable as it has been shown that pooled estimates from transformed and untransformed data are qualitatively similar . When a study assessed specific types of physical activity in addition to total or any physical activity, we only used risk estimates for total or any physical activity in the pooled analysis. Standard chi-square tests and the I2 statistic were used to quantify the extent of statistical heterogeneity across studies [27, 28]. Given the absence of substantial heterogeneity among contributing studies, RRs were pooled using a fixed effects model. We explored for the effect of interactions on the association using study-level characteristics such as geographical location (Europe vs. North America vs. Asia), sex (men vs. women), the average age at baseline (≥ 55 vs. < 55 years), the average duration of follow-up (< 10 vs. ≥ 10 years) based on the distribution of the data, type of pneumonia outcome (incident pneumonia vs. pneumonia-related mortality), number of events (≥ 350 vs. < 350), and degree of adjustment (+ vs. + +), which was conducted using stratified analysis and random effects meta-regression . To test the robustness of the observed association, we conducted a sensitivity analysis by investigating the influence of omitting each study in turn on the overall result (stata module metaninf). As a sensitivity analysis, we also pooled studies that reported minimally adjusted estimates (age and/or sex adjusted). To explore for small study effects, we visually inspected constructed Begg’s funnel plots  and performed Egger’s regression symmetry test . We employed Stata version MP 17 (Stata Corp, College Station, Texas) for all statistical analyses.