Pooling of Individual Patient Data from Clinical Trials. Improvement of analyses of subgroups

Abstract

Meta-analyses that use individual patient data (IPD), that is, the raw data of individual trials, rather than simply the overall results of each trial have been proposed as a major improvement in subgroup analyses. Since IPD meta-analyses often include more detailed data, they usually have greater statistical power to carry out informative subgroup analyses. By using the individual patient data the flexibility of subgroup analyses may be enhanced. Consequently, the estimated subgroup effects may be less influenced by misclassification and ecological bias. IPD meta-analyses, therefore, allow a more thorough assessment as to whether differences in treatment effects between subgroups are spurious or not. The merits of the IPD meta-analytic approach to study subgroup effects have, however, not yet been studied properly. Thus far, there is no consensus on the best methodology of analysing subgroup effects in IPD meta-analyses. The general objective of this thesis, therefore, is to study the accuracy, flexibility, and validity of IPD meta-analyses in analysing subgroup effects. The results of an extensive literature study, in which we tried to identify all published IPD meta-analyses and their related conventional meta-analyses, were used to study the methods used to study subgroups in IPD meta-analyses and to compare subgroup effect estimates of conventional and IPD meta-analyses. We showed that a wide variety of analytical methods was used in both conventional and IPD meta-analyses. We also showed that IPD meta-analyses appear to provide more relevant clinical information, since they more frequently reported subgroups based on individual patient and disease characteristics. The data of six trials (n=1,643) on the effectiveness of antibiotics in children with acute otitis media were used for several (methodological) studies. We compared the subgroup effects of 1) conventional meta-analyses using summary statistics derived from published data, 2) two-stage approach to IPD meta-analyses where summary statistics derived from IPD are used, and 3) one-stage approach to IPD meta-analyses where the IPD is pooled into a single dataset. From this study we concluded that conventional meta-analysis do not allow proper subgroup analyses, whereas IPD meta-analyses produce more accurate subgroup effects. We showed an empirical example of an IPD meta-analysis, on the (subgrouping) effects of antibiotics in preventing the development of asymptomatic middle ear effusion (MEE) in subgroups of children. We studied the influence of various methods of handling missing data (complete case analyses, single imputation within and over trials, and multiple imputations within and over trials) on the subgroup effect estimates of IPD meta-analyses. From this study we concluded that imputation within trials appears to be the most appropriate approach of handling missing data in IPD meta-analyses. We finally discussed the findings reported in this thesis and provided recommendations on how and when treatment effects in subgroups should best be analysed and reported in IPD meta-analyses. We showed that IPD meta-analyses are indeed more accurate, flexible, and valid in analysing subgroup effects, and should therefore be performed. Future studies should adhere to our recommendations regarding the analyses and report of subgroup effects in IPD meta-analyses.