Article type
Year
Abstract
Many treatments are more effective in some patients than others, which means that the effectiveness of treatment varies with some factors, usually called 'interaction'. There are several methods available for investigating interaction in aggregate-data meta-analyses. Suppose we want to examine the potential interaction between sex and treatment, using relative risk (RR) as the outcome measure. The best method ('interaction term approach') would be to calculate an interaction term based on RR-male and RR-female within each trial, and then combine the interaction terms across trials to obtain a summary interaction term.
However, the majority of existing meta-analyses do not use this method. Instead, they combined the treatment effects in male subgroups across trials first (RR-maletotal), and then combined the treatment effects in female subgroups (RR-femaletotal), and then test for difference between the two subgroup-specific summary estimates ('subgroup approach'). This method can be easily implemented with commonly available software for meta-analysis such as RevMan.
However, the subgroup approach is problematic. First, it calculates the subgroup-specific summary estimates first, thus ignoring the within-trial correlations between male and female in the trials with data on both subgroups. This would lead to reduced efficiency and inappropriate estimation of standard errors. In addition, to compare the two summary estimates of male and female subgroups, the heterogeneity in the meta-analyses to obtain the two summary estimates has to be ignored. For these reasons and based on the conceptual comparisons, researchers have recommended avoiding using the subgroup approach.
In this study, we searched PubMed and randomly selected 100 empirical meta-analyses to re-evaluate the interactions, which were originally assessed by the subgroup approach, by using the interaction term approach. The results on interactions by the two methods were compared both qualitatively and quantitatively. Study characteristics associated with differences in the results were described. Recommendations were made for future meta-analyses. Details will be presented at the meeting.
However, the majority of existing meta-analyses do not use this method. Instead, they combined the treatment effects in male subgroups across trials first (RR-maletotal), and then combined the treatment effects in female subgroups (RR-femaletotal), and then test for difference between the two subgroup-specific summary estimates ('subgroup approach'). This method can be easily implemented with commonly available software for meta-analysis such as RevMan.
However, the subgroup approach is problematic. First, it calculates the subgroup-specific summary estimates first, thus ignoring the within-trial correlations between male and female in the trials with data on both subgroups. This would lead to reduced efficiency and inappropriate estimation of standard errors. In addition, to compare the two summary estimates of male and female subgroups, the heterogeneity in the meta-analyses to obtain the two summary estimates has to be ignored. For these reasons and based on the conceptual comparisons, researchers have recommended avoiding using the subgroup approach.
In this study, we searched PubMed and randomly selected 100 empirical meta-analyses to re-evaluate the interactions, which were originally assessed by the subgroup approach, by using the interaction term approach. The results on interactions by the two methods were compared both qualitatively and quantitatively. Study characteristics associated with differences in the results were described. Recommendations were made for future meta-analyses. Details will be presented at the meeting.