Studies and participants
All the studies participating in the StoP consortium (23) were conducted in accordance with applicable laws, regulations and guidelines for protection of human subjects, and the StoP Project received ethical approval from the University of Milan Review Board (reference no. 19/15 of 01/04/2015). Overall, ten out of over 30 studies included in the latest release (number 2.1) of the StoP dataset collected data on leisure-time PA (24–33). However, five studies were not considered for the present investigation, mainly because available PA variables were characterized by a large amount of missing data or were qualitative in their nature (29–33). Studies with large percentage of missing values on PA were not included (34). Five studies were ultimately included in the pooled analysis: Italy (28), Canada (27), Russia (26), USA (25), and Spain (24).
Exposure assessment and data standardization
The questionnaires administered to study subjects usually include demographic and lifestyle data on PA, cigarette smoking, alcohol use, dietary habits, and family history of cancer. Additional data from the study cases were obtained from cancer registries or hospital medical records. All data were collected and standardized according to a pre-specified format at the data-pooling center.
The main characteristics of PA variables are outlined in Table 1. Studies reported (25–28) the duration (number of hours) of leisure-time PA over a certain interval of time (a week period). In particular, Italian and US centers (25,28) had pre-specified criteria with cut-off points in their questionnaires, while Canada, Russia and Spain (24,26,27) reported continuous values. Only the Canadian center incorporated data on the PA intensity (moderate or vigorous) (27). The period for study-specific PA variables ranged from one year to 5 years.
Since the variables related to leisure-time PA as well as period of referral showed certain levels of variation across the centers, we designed PA exposure category based on study-specific tertiles that were created using control populations. Study subjects were assigned to one of three PA categories, either none/low, intermediate or high (most commonly expressed as number of hours engaging in recreational PA across a week period). None and low levels of exposure were combined together due to the nature of individual study questionnaires and were considered as a reference category.
Statistical analysis
Descriptive analyses were conducted to describe the study population in terms of demographic characteristics, selected lifestyle habits, GC characteristics (subsite and histotype) and H. pylori infection. The relationship between PA and GC was evaluated using a two-stage approach (35). Firstly, multivariable logistic regression models were applied to obtain study-specific odds ratios (ORs) and the corresponding 95% Confidence Intervals (CIs). These models were adjusted for sex, age, H. Pylori, smoking, alcohol consumption, BMI, social class, occupational PA, cancer history, and dietary habits (Additional file 1). In the second phase, a random-effect model was applied in order to estimate summary (pooled) effect measures. Heterogeneity across studies was assessed with the Q and I2 statistics measures (36).
In order to investigate the effects of leisure-time PA across strata of selected covariates, we performed stratified analyses according to: sex, age (≤55, 56-65, >65), BMI (normal weight, overweight, obese), social class (study-specific low, intermediate, high), smoking status (never, former, current smoker), alcohol drinking status (never, ever), vegetables and fruit intake (study-specific low, intermediate, high), occupational PA (study-specific low, intermediate, high), cancer history among first degree relatives (yes, no), GC subsite (cardia, non-cardia), GC histotype (intestinal, diffuse, undifferentiated) and H. Pylori status (positive, negative). Heterogeneity tests were performed across all the strata estimates. Lastly, we performed a test for linear trend across the three levels of PA variable (37).
All statistical tests were two-sides and a p-value < 0.05 was considered as statistically significant. Statistical analyses were carried out using STATA software, version 12.