The Association between Leisure Time Physical Activity and Pancreatic Cancer Risk in Adults: A Systematic Review and Meta-analysis

Pancreatic cancer is the 12th most common cancer globally with 337,872 new cases and 330,391 deaths worldwide in 2012 (1). Pancreatic cancer has one of the lowest 1-year survival rates among cancer sites at 26%, and an overall survival rate of only 6% (2, 3). This heavy mortality burden is largely attributed to ambiguous symptom presentation, lack of available screening methods for early detection and intervention, and a very rapid natural history of disease progression (4).

Evidence suggests that several factors, including tobacco smoking, obesity, diabetes mellitus, a family history of pancreatic cancer, pancreatitis, particularly familial pancreatitis, previously identified genes, and heavy alcohol consumption (5–9), could contribute to increased risk of pancreatic cancer. Furthermore, Helicobacter pylori and/or hepatitis B infection, exposure to chemicals such as petroleum compounds and solvents, and physical inactivity have also been implicated as potential risk factors (10–14); however, results are inconclusive (4).

Physical activity as a protective factor has been related to reductions in postmenopausal breast (evidence is not consistent in premenopausal women), colorectal, and endometrial cancers (15–17). The body of literature examining the association between physical activity and other cancer sites, including pancreatic cancer (15, 18), continues to expand. Previous meta-analyses by Bao and colleagues (19) and O'Rorke and colleagues (20) suggested a possible protective effect of occupational physical activity on pancreatic cancer incidence [relative risk (RR), 0.75; 95% confidence interval (CI), 0.58–0.96 and RR, 0.75; 95% CI, 0.59–0.96, respectively]. However, they reported lack of evidence to support leisure time physical activity (LTPA) in pancreatic cancer risk reduction. Since the time of publication, 6 additional studies, including 4 cohorts (refs. 21–24; 3 with <50 cases and one with >400 cases) and 2 case–control studies (refs. 15, 25; 1 hospital-based and 1 population-based) have been published examining the effects of LTPA in pancreatic cancer development. More recently, Behrens-G and colleagues (26) updated the previous meta-analyses, using a broach approach to examine all types of physical activity on pancreatic cancer risk. Building on this previous work, we have conducted a meta-analysis of these associations to conduct detailed subgroup analyses and examine potential sources of heterogeneity while focusing on LTPA. We focused on LTPA, as this has been more consistently measured in population-based studies and with less variation in assessment in comparison to other types of physical activity (occupational, household, and transportation physical activity). Furthermore, LTPA is a modifiable exposure and a more readily targetable behavior change from a public health perspective.

We conducted a search of the PubMed and MEDLINE databases up until February 1, 2015, with the keywords and medical subject heading terms of: “physical activity,” “motor activity,” “exercise,” “cancer,” “neoplasm,” “carcinoma,” “tumor,” “pancreas,” “pancreatic,” “risk factor,” “risk factors,” and “risk.” Detailed search terms and additional online databases searched are provided in Supplementary Appendix SI. We applied no date, language, or geographical restrictions. Abstracts and unpublished results were not included. Reference lists of relevant articles were examined for additional studies.

Two independent reviewers (M.S. Farris and A.A. McFadden) completed an initial screening process using predefined inclusion criteria. Articles were included for further review if they were original studies with primary data, human subjects, and related to the topic of interest (physical activity and/or lifestyle factors and incidence of pancreatic cancer). Search results were reviewed independently by the 2 reviewers and then compared for discrepancies. A third reviewer (D.R. Brenner) was involved where any discrepancy between the two reviewers occurred. In the case of multiple publications from the same study population, only the article with the largest sample size and the study assessing incidence rather than mortality was included.

Following the Meta-analyses of Observational Studies in Epidemiology (MOOSE) criteria and PRISMA checklist (27, 28), study information was collected using a standardized abstraction template including information on study design, number of cases and controls, participant and study characteristics, assessment of LTPA and pancreatic cancer, effect estimates, and adjustments for confounding (9). Studies were placed into multiple categories based on participant and study design characteristics, including study design (cohort or case–control), median age (<50, 50–60, or >60 years), gender (combined, male, or female), location (United States, Canada, Europe, or Asia), and median year of data collection calculated by taking the earliest study recruitment year subtracted from the last follow-up or analysis/publication year (before the year 2000 or after the year 2000). The dichotomization for median year of data collection was derived to examine temporal trends in the relation potentially related to methodologic limitations of older studies or changes in reporting of LTPA over time. To assess adjustment for potential confounding, studies were categorized into 4 subgroups on the basis of the variables that were included in the statistical models used: no adjustment, basic model (including adjustment for age, sex, and site), basic model with smoking, and other (including a combination of the basic model, smoking and/or lifestyle/heath factors). Assessment of LTPA involved the reporting period of LTPA separated into 3 subgroups; lifetime LTPA [assessing LTPA over the participants lifetime or several decades (>30 years) prior to study recruitment], past year LTPA, and 2–10 past years LTPA as well as the type/intensity of activity. The type/intensity was separated into 5 subgroups including effects corresponding to an estimated prevention guideline from the WHO Global Recommendations for Physical Activity and Health (>150 minutes moderate physical activity per week or >75 minutes of vigorous physical activity per week; ref. 29), quartiles/quintiles representing multiple levels of LTPA, low versus high, frequency (times), and sports participation. The incidence of pancreatic cancer was assessed on the basis of the method used to confirm pancreatic cancer diagnosis. In studies, pancreatic cancer diagnosis was collected through either pathology reports, International Classification of Disease (ICD) codes, cancer registry, a combination of methods or subjective measures such as death certificates. Furthermore, studies reporting incidence and mortality were included, with preference given to measures of incidence. As previously mentioned, the 5-year survival rate for pancreatic cancer is extremely low (2, 3) and, therefore, within a small lag-time window, pancreatic cancer mortality roughly approximates incidence in most populations.

HRs, ORs, and/or RRs and their 95% CIs assessing LTPA and pancreatic cancer were extracted. We included the estimate from each study characterizing risk associated with the largest comparison of low-to-high LTPA. For example, where quartiles were presented, the lowest compared with the highest quartiles were used. For studies comparing the highest versus the lowest LTPA, that is, those studies estimating the risk associated with a deficit of activity, the reciprocal of the study estimate was derived. This approach enabled distinctive contrasts between sedentary behavior and physically active behavior to a relatively comparable extent across studies in an attempt to minimize heterogeneity in LTPA measures. LTPA was assessed in all included studies through subjective measures such as self-reported questionnaires or in-person interviews. While measures may be subjective, in this study through a quality assessment, objective definitions of physical activity were determined for inclusion, as shown in Table 1. If there were several study estimates in each article, the multivariate estimate adjustment with the most covariates was extracted. In studies with multiple follow-up time points, the study estimate with the longest follow-up time was presented.

Meta-analytic methods were used to pool estimates from included studies using the DerSimonian and Laird random-effects model (30). As stated above, the primary analyses used pooled effect estimates for the lowest versus highest comparison of LTPA and the weighted summary and 95% CI. Analyses were then performed separately on cohort and case–control studies to assess differential magnitude of effects.

To assess sources of heterogeneity the Cochran Q statistic, the I² statistic were estimated across studies. Meta-regression analyses were also conducted using the variables of interest from data abstraction. Heterogeneity across subgroups was evaluated by comparing stratified estimates including: study design, median age at baseline/recruitment, gender, location, median year of data collection, confounding adjustment, reporting period of LTPA, type/intensity of LTPA, assessment of outcome, and incidence versus mortality. Stratum-specific analyses and meta-regression were then conducted separately among case–control and cohort studies using all previously stated subgroups to assess heterogeneity within each study design. In the analysis of case–control studies, differences in effects across types of controls (population-based or hospital-based) were also examined.

To evaluate for potential publication bias, the Begg test and a visual interpretation of funnel plots were used (31). All statistical analyses were performed in the statistical packages STATA version 13, R version 9.3 (32) and assessed with a significance level of 0.05.

The search initially identified 1,431 citations; of these, 833 were excluded as either not meeting inclusion criteria or because they were duplicates. There were 598 abstracts reviewed and 39 full text reviewed, leaving 27 eligible articles for systematic review (6–9, 11, 15, 18, 21–25, 33–47) and 26 studies eligible for meta-analysis (refs. 6–9, 11, 15, 18, 21–25, 33, 34, 36–47; Fig. 1). Of these 26 studies, 7 reported men and women as separate populations (6, 18, 22, 33, 41–43); therefore, 33 risk estimates were accounted for in the overall analysis.

Study characteristics from the 27 studies (6–9, 11, 15, 18, 21–25, 33–47) included in the systematic review are presented in Table 1. Twenty-one (6–8, 11, 21–24, 35–47) of the 27 included studies were cohort studies. Of the 6 case–control studies (9, 15, 18, 25, 33, 34) included in the analyses, 4 had population controls (9, 15, 18, 34) and 5 were frequency-matched (9, 15, 18, 33, 34). Twenty-one studies were conducted among participants with a median age of 50+ years (6–8, 11, 15, 18, 21–24, 34–37, 40–47), whereas 6 studies included participants with a median age <50 years (9, 25, 33, 38, 39).

Assessment of study quality for inclusion into the meta-analysis is presented in Table 1. The majority (n = 24) of studies (7–11, 15, 18, 26–28, 30–40, 42, 44, 45) had sufficient follow-up time of greater than 5 years, objective definitions and collection methods of LTPA data, and quality case confirmation methods. Two studies did not have objective definitions of physical inactivity (23, 24), however, did satisfy the other quality assessment measures. After the quality assessment, one study, conducted by Paffenbarger and colleagues (35), was excluded from the meta-analysis because did it not present CIs or adequate data to calculate SEs.

Overall, 33 risk estimates were included in this meta-analysis. Seven studies had derived separate estimates for men and women of the association between the lowest versus the highest groups of LTPA and pancreatic cancer risk. Of these 33 risk estimates, 23 suggested a protective effect (6–9, 11, 15, 18, 21, 22, 25, 33, 34, 36–47). The overall combined summary estimate also suggested a significant protective effect (RR, 0.89; 95% CI, 0.82–0.96; Fig. 2). There was some evidence of low-to-moderate heterogeneity across the studies (I² = 22.1%, P_{heterogeneity} = 0.130). Subsequent removal of those studies (6, 8, 25) most contributing to heterogeneity did not significantly alter the results in sensitivity analyses (RR, 0.90; 95% CI, 0.84–0.96). Furthermore, study design and median age of participants contributed to a majority of the heterogeneity in the subgroup analyses, as discussed below.

Subgroup analyses are presented in Table 2. Effects were significantly different by study design with stronger effects observed among the 8 case–control risk estimates (RR, 0.69; 95% CI, 0.59–0.81; refs. 9, 15, 18, 25, 33, 34) than in the 25 cohort risk estimates (RR, 0.96; 95% CI, 0.91–1.02; refs. 7, 8, 21, 23, 24, 36–42, 44–47) reporting a nonsignificant association (P < 0.05). Differential effects were observed by median age of the study population. The risk estimates from studies that had a median age less than 50 years were statistically significant (RR, 0.61; 95% CI, 0.50–0.75; refs. 18, 23, 25, 37, 39), 16 risk estimates with a median age between 50 and 60 years were nonsignificant (RR, 0.93; 95% CI, 0.87–1.01; refs. 8, 9, 11, 15, 24, 33, 34, 36, 38, 40–42, 46), and 11 risk estimates with a median age greater than 60 years did not support any effect (RR, 1.00; 95% CI, 0.89–1.12; refs. 6, 21, 22, 43–45, 47). Effects also seemed to differ by gender slightly with a significant risk reduction observed in 10 risk estimates from combined populations (RR, 0.79; 95% CI, 0.68–0.91; refs. 7, 9, 11, 21, 23, 25, 34, 38, 39, 44) but not among male-specific risk estimates (RR, 0.95; 95% CI, 0.80–1.13; refs. 6, 15, 18, 22, 24, 33, 36, 40–43, 46, 47) and female-specific risk estimates (RR, 0.92; 95% CI, 0.82–1.03; refs. 6, 8, 18, 22, 33, 37, 41–43, 45).

The intensity of activity appeared to influence the estimates such that the strength of the effect became attenuated for lower activity levels. For example, there was a protective, suggestive association (RR, 0.85; 95% CI, 0.69–1.05) in the 9 studies (8, 18, 21, 22, 25, 41) that reported risk estimates for LTPA corresponding to the WHO recommendations for intensity of physical activity (>150 min/wk of moderate or >75 min/wk of vigorous LTPA; ref. 29). Risk estimates reporting LTPA in quartiles/quintiles showed evidence of risk reduction (RR, 0.80; 95% CI, 0.60–1.06; refs. 6, 9, 11, 37), whereas low versus high (RR, 0.92; 95% CI, 0.82–1.03; refs. 33, 36, 38–40, 43, 45–47) and frequency measures (sessions per week; RR, 0.92; 95% CI, 0.92–1.03; refs. 7, 15, 34, 42) subgroups produced limited evidence of a protective effect. In addition, the sports participation subgroup reported no effect (RR, 0.99; 95% CI, 0.69–1.44; refs. 23, 24, 44). With respect to the duration of activity, lifetime (RR, 0.89; 95% CI, 0.80–0.98; refs. 7, 11, 15, 23–25, 34, 36, 37, 39–42, 44–47) and studies that assessed 2–10 years of activity (RR, 0.62; 95% CI, 0.46–0.84; refs. 9, 18) supported a protective effect; however, past year reporting of LTPA revealed no association (6, 8, 21, 22, 33, 38, 43).

Both study design and median age of the study population were statistically significant sources of heterogeneity across studies (P = 0.001) according to meta-regression. Other study characteristics accounting for heterogeneity across studies were the confounding variables used for adjustment (P = 0.064) and gender (P = 0.054). Effects were also heterogeneous within and across LTPA level characterizations, for example, heterogeneity among risk estimates reporting under the WHO recommendations was moderately high (I² = 57.4%, P_{heterogeneity} = 0.016). This heterogeneity may be accounted for by either Stevens and colleagues (8) or Brenner and colleagues (25), as when these studies were removed, the I² statistic decreased by 18% and 25%, respectively, and became non-statistically significant. No other studies in the subgroup analysis based on the recommended levels of LTPA as defined by the WHO made any significant difference in the heterogeneity. In addition, for 4 studies that reported risk estimates for LTPA as quartiles or quintiles, there was moderate but non-statistically significant heterogeneity (I² = 42.8%, P_{heterogeneity} = 0.137; refs. 6, 9, 11, 37). When Nothlings (6) was removed, the heterogeneity decreased by 42.8%. When Stevens and colleagues (8), Brenner and colleagues (25), and Nothlings (6) were excluded, heterogeneity is substantially reduced in the overall results (I² = 0.0%, P_{heterogeneity} = 0.467). Furthermore, within cohort studies and case–control studies, there was very good agreement overall, thus heterogeneity may be indicative of a difference in magnitude of effect between study designs. In addition, 25 studies reporting effects from models with more detailed covariate adjustment supported a statistically significant protective effect (RR, 0.90; 95% CI, 0.82–0.98; refs. 6–9, 18, 21, 22, 24, 25, 36–43, 45, 47), in comparison to other confounder subgroups.

In cohort studies alone (Table 3), age of the population was also a determinant of differences in effect estimates. Studies with median age of <50 years suggested a significant effect (RR, 0.60; 95% CI, 0.41–0.88; refs. 23, 37, 39) but no effect was found in the 50- to 60-year-old (RR, 0.96; 95% CI, 0.89–1.04; refs. 8, 11, 24, 36, 38, 40–42, 46) and >60 years' subgroups (RR, 1.00; 95% CI, 0.89–1.12; refs. 6, 7, 21, 22, 43–45, 47). We also observed large differences in the effects between the one risk estimate with no covariate adjustment (RR, 1.20; 95% CI, 0.63–2.27; ref. 44) in comparison to the 3 studies with basic covariate adjustment (RR, 0.82; 95% CI, 0.64–1.05; refs. 11, 23, 46) and the 21 studies with more comprehensive covariate adjustment (RR, 0.97; 95% CI, 0.91–1.03; refs. 6–8, 21, 22, 24, 36–43, 45, 47).

In case–control studies (Table 4), there was sufficient agreement between all subgroup analyses including study characteristics, assessment of LTPA, and pancreatic cancer, accounting for little to no difference in effects.

To determine whether publication bias was present, a funnel plot was produced and visually assessed for asymmetry and verified by the Begg test. On the funnel plot, there were “missing” small studies (large value for 1/se) showing a harmful effect (positive β). This visual finding was not supported by the Begg test (P = 0.889); therefore, publication bias is not present in this meta-analysis.

The results of this meta-analysis support a protective association between LTPA and pancreatic cancer risk, with a risk reduction of 9% in cohort studies, 31% in case–control studies, and despite heterogeneity, 11% overall. Although the magnitude of the association is not as high as for other physical activity and cancer site associations (15–17), this result may be attributable to several limitations of the present study, including heterogeneity between included studies, inconsistent measurements of LTPA, insufficient adjustment for confounding, and other methodologic limitations related to the conduct of studies regarding pancreatic cancer.

LTPA is associated with reduced risk of several cancers, the observational epidemiologic evidence is classified as “convincing” for colon and breast cancer, “probable” for prostate cancer, and “possible” for endometrial and lung cancers (48, 49). There are several hypothesized biologic mechanisms whereby this reduction of risk may occur, including lowering of insulin levels, reduction of abdominal fat mass, elevated tolerance to oxidative stress through induction of antioxidant gene expression, and increased levels of various antitumor defenses (34, 49, 50). Specifically important for pancreatic cancer, LTPA improves insulin resistance through lowering fasting insulin and C-peptide levels and increasing insulin stimulated synthesis of glycogen in muscles (51, 52). As insulin resistance, hyperinsulinemia, and hyperglycemia have been hypothesized to play an important etiologic role in pancreatic carcinogenesis (53, 54), it is plausible that LTPA may exert protective effects through modulation of these pathways.

Age is recognized a nonmodifiable risk factor for pancreatic cancer (55). Estimates from studies in this meta-analysis differed significantly in terms of median age of the study population. A significant protective effect was found in studies including participants of a younger median age (<50) in both case–control and cohort studies, in comparison to a null finding in studies with a median age of >60 years. This finding is particularly of interest in the cohort studies, in which across subgroups, the strongest significant protective effects were observed among younger populations. This result suggests the impact of LTPA with pancreatic cancer risk reduction may be age-dependent and will dilute as age increases. In a rat model of pancreatic cancer, when compared with sedentary controls, aerobic LTPA inhibited pancreatic carcinogenesis when LTPA was implemented at 6 weeks of age but promoted pancreatic carcinogenesis when implemented at 13 weeks of age, suggesting that the effect of LTPA on pancreatic carcinogenesis may vary with age (56). There is also evidence that increased duration of some exposures, such as adiposity (57) and smoking (58), may result in increased pancreatic cancer risk. This risk accumulation may diminish the apparent protective effect of LTPA in older populations.

Noncausal explanations for the protective association between LTPA and pancreatic cancer risk should also be considered. Adjustment for confounding variables was limited; other modifiable factors were not accounted for in all of these studies and therefore, residual confounding may be present. A significant difference in adjustment for confounders was found between studies. Smoking was often adjusted for, but other unhealthy behaviors associated with smoking, such as alcohol consumption, poor diet, family history of cancer, or presence of other chronic conditions, were not consistently included as confounders, which have been shown to be probable risk factors for pancreatic cancer (26, 57, 58). It is possible that those who participate in LTPA are also likely to make other healthy lifestyle choices and vice versa. The observed protective effect, therefore, may be attributable to these factors, as opposed to LTPA itself. It is also possible that confounding factors may mask the “real” association, therefore leading to an underestimation of the protective effects of LTPA. In support of this possibility, those studies adjusting for a higher number of potential confounders observed significant results. Future studies should consistently involve a comprehensive list of covariate information to isolate this relation further and reduce the chance of spurious or attenuated associations. Through careful methodologic considerations, the issues of residual confounding, reverse causation, and selection bias may subside. However, because of the nature of the studies included in this review, these concerns are still present. The issue of reverse causality where the potential for subclinical disease prior to study recruitment may exist resulting in study participants not feeling well and therefore, not exercising, should be considered. More definitive methods for identifying pancreatic cancer cases through screening programs and early diagnosis would reduce the concern of reverse causality. Moreover, because of the nature of case–control studies, selection bias may be present. Controls who choose to participate in research studies tend to be more healthy, have a different distribution of LTPA and, therefore, may cause an overinflation in the risk estimate.

This meta-analysis was strengthened by the large number of studies included, allowing for extensive subgroup analysis to characterize various subgroups of interest and examine potential sources of heterogeneity. Furthermore, the subgroup analysis by characterization of LTPA was a considerable strength in this study; the WHO recommendations may be incorporated in the development of future studies assessing this relation. The implications of these results are strengthened regarding LTPA as an easily targeted modifiable behavior change from a public health perspective in comparison to overall physical activity and other types of physical activity examined in past meta-analyses (19, 20, 26). Furthermore, from a methodologic perspective, many studies do not include any or adequate measures of occupational, household- and transportation-related activity, therefore, there is an increased potential for publication bias or gaps in this literature when examining total physical activity in comparison to LTPA.

In the analysis stratified by study design, we observed significant differences in results between case–control and cohort studies. The magnitude of effect in the majority of case–control studies was generally lower in comparison to cohort studies. This likely is related to the more detailed assessments of LTPA that is often done and possible in case–control studies and would have decreased some measurement error associated with these assessments. All included studies measured LTPA through subjective measures and, therefore, may be subject to misclassification of LTPA. Misclassification may be exacerbated by differential recall, as case–control studies are more prone to recall bias (59). Consequently, risk estimates obtained from self-reported questionnaires may be attenuated (60). LTPA is inherently difficult to measure because it includes many unstructured activities that occur in different contexts. To examine the relation and reduce measurement bias in LTPA reporting more accurately, we performed subgroup analyses identifying characteristics of LTPA by intensity. However, we were limited in this analysis by the LTPA assessments used in these studies and, consequently, a need exists in future studies to include more accurate reports of LTPA through the use of objective measurements such as with activity monitors or supervised LTPA.

A second concern in our meta-analysis is that the primary analysis compared lowest versus highest LTPA to assess the risk of pancreatic cancer. In making this comparison, the assumption was that the relationship between LTPA and pancreatic cancer risk is relatively uniform and appropriate for this meta-analysis. However, some studies' highest measure of LTPA may be equivalent to other studies' middle measure of LTPA, in which case an underestimate of the association may occur. This limitation underscores the importance for future studies to conform to a uniform method of reporting LTPA, for example, through use of the guidelines made by the WHO Global Recommendations for Physical Activity and Health (29). In so doing, comparability between studies would be improved that would permit better assessments across populations. Although we found a non-statistically significant risk reduction among the 9 studies reporting LTPA corresponding to WHO recommendations (8, 18, 21, 22, 25, 41), this finding may be attributable to the moderately high heterogeneity among these studies. An important risk reduction may be present in these studies that could be more fully assessed with a pooled analysis that would apply a common definition for LTPA across studies. Such an analysis would still be limited by the original methods used for assessing LTPA and the heterogeneity across studies could not be fully resolved. A pooled analysis could also address some of the issues that arise when comparing populations that have underlying differences in levels of LTPA. To address the concerns, in part, we used random-effects models to combine the risk estimates. This approach made the assumption that the true effect is being sampled from a distribution of effects. Therefore, this method accounts statistically, to some extent, for differences in measurements and definitions of LTPA.

In conclusion, the current data suggest a statistically significant association between LTPA and pancreatic cancer risk. While the effects are comparatively small (11% risk reduction), this finding may be attributable to several limiting factors of the present study and to the fact that the protective effect of LTPA may depend on age. There is substantial evidence in other cancer sites for the beneficial effect of LTPA on cancer risk, and biological plausibility also exists for an effect of LTPA in pancreatic cancer etiology. To incorporate the benefits of LTPA in potentially reducing the risk of pancreatic cancer, future research should focus on more accurate measures of LTPA, such as supervised LTPA and activity monitors.

No potential conflicts of interest were disclosed.

The authors thank Marcus Vaska for his support in generating a search strategy. The authors would like to thank Dr. Doreen Rabi and Dr. Derek Roberts for their contributions to the development of the study protocol and guidance with the drafting of the manuscript.

C.M. Friedenreich was supported by Alberta Innovates Health Solutions Health Senior Scholar Award and Alberta Cancer Foundation Weekend to End Women's Cancers Breast Cancer Chair.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.