18.4.2  General approach

Most IPD meta-analyses to date have used a two-stage approach to analysis. In the first stage, each individual study is analysed in the same way, as set out in the meta-analysis protocol or analysis plan.  In the second step, the results, or summary statistics, of each of these individual study analyses are combined to provide a pooled estimate of effect in the same way as for a conventional systematic review (Simmonds 2005). More complex approaches using multilevel modelling have been described for binary data (Turner 2000), continuous data (Higgins 2001), ordinal data (Whitehead 2001) and time-to-event data (Tudor Smith 2005b) but, currently, their application is less common. When there is no heterogeneity between trials, a stratified log-rank two-stage approach for time-to-event data may be best avoided for estimating larger intervention effects (Tudor Smith 2005a)