Article type
Year
Abstract
Background: Treatments may be more effective in specific patient subgroups, but trials may lack power to detect treatment-subgroup interactions. Although meta-analysis may increase power, such interactions are best assessed using individual participant data (IPD) which is often not feasible. Since meta-regression of treatment effects by summaries of patient characteristics is prone to bias, an alternative is needed for meta-analysis of summary data.
Objective: To demonstrate how interactions between treatment effects and patient subgroups based may be assessed using summary data.
Methods: We conducted a systematic review in advanced colorectal cancer where hazard ratios (HRs) for survival were reported (or could be estimated) for various patient subgroups. To assess subgroup interactions within trials for binary subgroup variables (e.g. sex), we used the ratio of subgroup HRs. For ordered subgroup variables of >2 levels (e.g. cancer stage), we used weighted linear regression to estimate the change in treatment effect per subgroup level. These within-trial interactions were then combined across trials using standard meta-analysis methodology to obtain a pooled interaction effect. We compared our results to those from meta-regression.
Results: Despite difficulties obtaining subgroup data for all included trials and inconsistencies in the classification of some subgroup categories, we were able to assess the interaction between treatment and KRAS genotype, age, sex, performance status and number and site of metastases. Results and comparison with meta-regression will be presented, as will similar approaches for assessing treatment by subgroup interactions for dichotomous and continuous outcomes. We shall also discuss how such analyses may be presented graphically.
Conclusions: Identifying treatment by patient subgroup interactions can enable therapies to be targeted at those who most benefit. We encourage systematic reviewers to use appropriate methods to assess these interactions where possible; a process that will be facilitated if trialists report treatment effects by patient subgroups.
Objective: To demonstrate how interactions between treatment effects and patient subgroups based may be assessed using summary data.
Methods: We conducted a systematic review in advanced colorectal cancer where hazard ratios (HRs) for survival were reported (or could be estimated) for various patient subgroups. To assess subgroup interactions within trials for binary subgroup variables (e.g. sex), we used the ratio of subgroup HRs. For ordered subgroup variables of >2 levels (e.g. cancer stage), we used weighted linear regression to estimate the change in treatment effect per subgroup level. These within-trial interactions were then combined across trials using standard meta-analysis methodology to obtain a pooled interaction effect. We compared our results to those from meta-regression.
Results: Despite difficulties obtaining subgroup data for all included trials and inconsistencies in the classification of some subgroup categories, we were able to assess the interaction between treatment and KRAS genotype, age, sex, performance status and number and site of metastases. Results and comparison with meta-regression will be presented, as will similar approaches for assessing treatment by subgroup interactions for dichotomous and continuous outcomes. We shall also discuss how such analyses may be presented graphically.
Conclusions: Identifying treatment by patient subgroup interactions can enable therapies to be targeted at those who most benefit. We encourage systematic reviewers to use appropriate methods to assess these interactions where possible; a process that will be facilitated if trialists report treatment effects by patient subgroups.