The rate ratio of injurious falls was the outcome of interest. If the rate ratio was not presented in the article, we calculated it from the ratio of the total number of injurious falls divided by the total length of time falls were monitored (person years) in the two comparison groups. In cases where data were available only for people who had completed the study, or where the trial authors had stated there were no losses to follow-up, we assumed that these participants had been followed up for the maximum possible period. We estimated the standard error of the rate ratio by using the formula given in the Cochrane handbook.17
We used the generic inverse variance method in Review Manager (RevMan 5.1) to group the trial results and we compiled forest plots for each category of injurious falls. To allow for variability among the participants, type of exercise intervention, and outcome definitions we used a random effect model. We report the pooled rate ratios for each injurious fall outcome, along with 95% confidence intervals. We assessed heterogeneity with the Q test and the I² statistic.18
We conducted a prespecified subgroup analysis based on fall risk at enrolment (a priori); that is, trials with participants selected for inclusion based on fall history or other specific risk factors for falling (at higher risk) compared with non-selected participants. We also explored the possible impact of risk of bias on statistically significant pooled estimates of exercise effect by removing studies of poorer quality—that is, those for which the risk of bias was unclear for at least three of the quality components considered, or the risk was at least unclear for one category and high for another. To explore the possibility of publication bias we constructed funnel plots of effect estimates against their standard errors for analyses that contained at least 10 data points.