Physical Activity and Longevity: How to Move Closer to Causal Inference
Total Page:16
File Type:pdf, Size:1020Kb
BJSM Online First, published on March 15, 2018 as 10.1136/bjsports-2017-098995 Editorial Br J Sports Med: first published as 10.1136/bjsports-2017-098995 on 15 March 2018. Downloaded from rather than self-reported, retrospec- Physical activity and longevity: how to tive PA in a large sample may provide a more precise predictor of mortality.7 move closer to causal inference Furthermore, inadequate measurement, limited knowledge or poor adjustment for Kaitlin H Wade, Rebecca C Richmond, George Davey Smith confounding variables, such as smoking status in the setting of physical activity and mortality, can severely bias observed Kujala provides an insightful review confounding by baseline adiposity biased associations. contesting epidemiological findings that findings that bus conductors had lower As presented by Kujala, RCTs, the increased physical activity (PA) lengthens risk of coronary heart disease than their gold standard in epidemiology for infer- 1 the life span, arguing that intervention less-active driver counterparts (although ring causality, have failed to provide (randomised controlled trial (RCT) and this issue was acknowledged by Morris conclusive evidence in this context (eg, experimental) studies do not support PA who performed analysis stratified by the Lifestyle Interventions and Independence causing a reduced risk of death and high- busmen’s uniform size to account for for Elders,8 Look Action for Health in 3 lighting several limitations in previous potential confounding by adiposity). Diabetes,9 Heart Failure: A Controlled observational studies that may have led to The possibility of reverse causation Trial Investigating Outcomes of Exer- spurious conclusions. (whereby the ‘outcome’ is responsible for cise Training10 and other large-scale The review coincides with the publi- variation in the ‘exposure’, rather than meta-analyses).11 12 In the absence of cation of findings from the large-scale the direction of interrogation) may also long-term trials, the focus moves to other Prospective Urban Rural Epidemiologic lead to misinterpretation of observed approaches for strengthening causal infer- (PURE) study (n=130 843), which iden- associations. For example, the notion that ence. Some such methods are discussed by tified a graded lower rate of mortality reducing PA increases the risk of becoming Kujala and are outlined in table 1. among those individuals achieving overweight/obese is as plausible as the One approach acknowledged is the moderate and high levels of PA compared reverse, where being overweight/obese comparison of associations between recre- 4 with those with low PA (HR 0.80; renders PA difficult. Studies of older ational leisure time and obligatory occu- 95% CI 0.74 to 0.87 and 0.65; 95% CI adults or those with many comorbidi- pational PA, the latter of which has not 2 0.60 to 0.70; P for trend <0.0001). ties are particularly vulnerable to reverse consistently been associated with a reduced While this study is undeniably an impres- causation. For example, in reference to risk of death.13 Physiologically, there are sive endeavour, collecting prospective Kujala’s ‘healthy exerciser bias’, aged indi- no compelling reasons why recreational data on participants from 17 countries, viduals who are healthy enough to partic- and occupational PA should have system- the conclusion to support increased forms ipate in PA due to a lack of chronic illness atically different effects on mortality and of PA levels for all individuals (irrespec- will seemingly have a reduced risk of so, if activity were truly causal, effect tive of age, gender or country of origin) death compared with their less-fit peers. sizes should be similar between these two has major public health implications. The Furthermore, comparing estimates of risk contexts. One explanation for this poten- findings are, as so often, qualified by the for physically demanding versus sedentary tial discordance is confounding by socio- study, being unable to fully assert a causal occupations may suffer reverse causation, economic position. For example, earlier http://bjsm.bmj.com/ (rather than correlational) role for PA particularly when high fitness and good studies of the association between occu- levels in reducing mortality. health are criteria for recruitment into pation and PA, at a time when there may Kujala emphasises how epidemiological such physically demanding occupations. have been a positive social class gradient study designs are vulnerable to limitations Related to this, in the setting of evalu- for cardiovascular disease (CVD), tended that may skew or distort observational ating potential causes of mortality, both to show that doing more occupational PA associations and impede the distinction selection and survival biases,5 which influ- was related to lower CVD.14 However, between correlation and causation. Such ence participation rates in epidemiolog- with a change in social class gradient on September 28, 2021 by guest. Protected copyright. distortions of observed relationships may ical studies, can also lead to distortion of over time, more recent studies have typi- arise due to confounding by measured/ associations among respondents. In these cally failed to demonstrate consistent and unmeasured lifestyle, behavioural and cases, the population under study (and protective effects of occupational PA.13 biological factors (such as higher fitness, therefore the observed associations) may While the recent PURE study found that lower body mass index (BMI), genetic differ from the population not selected or occupational PA was protective against variation and socioeconomic factors) who were unable/unwilling to participate mortality risk across countries at different correlated with both the exposure (here, (due to morbidity or lack of interest in economic levels, it is important to high- PA) and outcome (here, longevity). If not surveys relating to health).6 light that definitions of occupational PA appropriately accounted for, confounding Kujala also highlights the limitation of included travel to work, which may be factors make the ascertainment of under- measurement error, which can bias esti- strongly influenced by health-related lying causal mechanisms and path- mates within epidemiological studies, selection.15 Interestingly, the PURE study ways exceptionally complex. Such was particularly those relying on self-report does not seem to be as supportive for the illustrated by Jerry Morris’ London or questionnaire-based information. role of recreational activity on reducing busmen study revisited by Kujala, where Recent developments have highlighted the mortality risk, where differences in under- trade-off between sample size and measure- lying confounding structures between Medical Research Council Integrative Epidemiology Unit, Population Health Sciences, Bristol Medical School, ment precision in obtaining adequate varying income countries investigated University of Bristol, Bristol, UK statistical power with minimum measure- may explain the heterogeneity in effects ment error. For example, measuring observed. Correspondence to Dr Kaitlin H Wade, Integrative Epidemiology Unit, University of Bristol, Bristol BS8 maximal oxygen consumption in a formal A further causal inference approach not 2BN, UK; Kaitlin. Wade@ bristol. ac. uk fitness test, VO2max, in a smaller sample directly considered by Kujala is that of the Wade KH, et al. Br J Sports Med Month 2018 Vol 0 No 0 1 Copyright Article author (or their employer) 2018. Produced by BMJ Publishing Group Ltd under licence. Editorial Br J Sports Med: first published as 10.1136/bjsports-2017-098995 on 15 March 2018. Downloaded from Table 1 Methods for strengthening causal inference in physical activity epidemiology studies Method Example Strength Limitation Lifecourse approach Use of prospective studies to investigate PA Useful for assessing temporal associations; ability to Logistically demanding as it requires levels at different ages and how they might adjust for the respective outcome measures at baseline repeat assessments; residual differently affect lifespan (where possible) makes it possible to disentangle confounding; measurement error in prospective associations from tracking effects exposure, outcome and covariables; selection bias Cross-context comparison Comparison of associations between Exploring residual confounding; reliable findings if Assumptions about different confounding voluntary leisure-time PA and compulsory estimates are similar across different contexts (where the structures may not be correct; variables in occupational PA; PA across different cultures confounding structure in these settings is likely to differ) different studies might be measured with or dissimilar countries varying accuracy and generalisability Sibling comparison MZ or DZ twin comparisons among siblings Using MZ best controls for familial background and Assumes a stable family environment; discordant for PA genetic confounding, compared with DZ (or siblings), confounding by factors not perfectly where 50% of genetic information is shared shared by siblings; reverse causation still possible Mendelian randomisation The use of genetic variants associated with Genetic instruments are not subject to confounding from Low power; lack of instruments; exercise and fitness, incorporated into a environmental or lifestyle factors, are not influenced by pleiotropy and linkage disequlibrium; Mendelian randomisation analysis, whereby the outcome, do not change over time and are measured population stratification;