From: Meta-analysis and The Cochrane Collaboration: 20 years of the Cochrane Statistical Methods Group

Topic | Comments |
---|---|

Continuous outcome measures | More work is needed on methods for combining results from trials with continuous outcome measures, especially with regard to the choice of effect measure, distribution of the data, baseline assessments and missing data. Of particular interest, perhaps, is the issue of combining results from trials where continuous outcome data have been categorised using varying cut-points and number of groups, or where some trials have presented results grouped and others as continuous. |

Combining parallel group and cross-over trials | Work is needed to evaluate strategies for combining information from parallel group and cross-over trials, and to consider the information that needs to be supplied in reports of cross-over trials to enable them to be used in this way. |

Heterogeneity | Given the lack of consensus at the workshop, it would be valuable to gain more insight into the merits of various strategies for combining trials when statistical heterogeneity is found. More empirical research into the effects of variation in methodology on heterogeneity would be valuable |

Combining trials with different endpoints | It would be useful to consider what, if anything, can be done to combine trials that use different endpoints (This is a generalisation of the issue of combining data from trials where the same endpoint is assessed in different ways). |

Summary statistics | More work is needed examining the relative merits of odds ratios, relative risks, NNTs and so on for meta-analysis of RCTs. |

Extraction of summary statistics from published trials where the endpoint is survival time | Methods to combine results from survival studies where individual data are not available need to be investigated. Guidelines for minimum standards for reporting such trials should be developed. |

Ignoring survival times | A study could be carried out of the loss of information (and power?) arising from classifying patients by presence or absence of the event of interest rather than by time to event. The effect of the varying length of follow-up should be examined. |

Presentation of meta-analyses - numerical | Studies should be encouraged examining the merits of odds ratios, relative risks, NNT and other approaches for producing numerical summaries of the results of meta-analyses. |

Presentation of meta-analyses - graphical | Various aspects of graphical presentation vary among published meta-analyses. The relative merits of these should be systematically reviewed. If possible, recommendations should be developed for standard graphical presentation. Aspects to consider include ordering of trials, symbols used, whether or not log scale used for treatment effect, and whether any additional graphs might usefully supplement the standard type. |

Random effects models | It would be useful to have clarification of the properties of the various strategies for fitting random effects models. |

Combining results from observational studies | Methods for combining results from observational studies, such as case-control studies, need to be investigated. In particular, methods are needed to pool estimated regression coefficients. |