Two-Stage Meta-Analysis

An individual patient data meta-analysis first analysing each study separately, then combining the resulting study-level estimates using standard methods.

Source: Stewart & Parmar 1993

Two-stage meta-analysis of individual patient data works in two steps: first each study's raw records are analysed separately to produce a summary estimate for that study, then those estimates are combined using standard meta-analysis methods, exactly as aggregate results would be. Keeping the studies separate at the first step makes the analysis simpler and more transparent than modelling all patients at once, while still drawing on the underlying data. Analyse each, then pool, is its sequence. Riley and colleagues (2010) describe this approach.

Source: Riley et al. 2010

Two-stage meta-analysis works by, in the first stage, analysing each study's patient-level data separately to produce a summary estimate, such as a treatment effect, with its precision; and, in the second stage, combining these study-level estimates using standard meta-analysis, weighting by precision under a fixed-effect or random-effects model. This mirrors conventional meta-analysis but derives the study estimates from the raw data. So two-stage meta-analysis works by first summarising each study from its patient-level data and then pooling those summaries with established methods, separating the within-study analysis from the across-study combination into two distinct steps.

Source: Stewart & Parmar 1993

Two-stage meta-analysis analyses each study's patient-level data separately and then combines the resulting summary estimates, while one-stage meta-analysis analyses the combined patient-level data from all studies together in a single model that accounts for the studies. The two-stage approach is often simpler and mirrors conventional meta-analysis, while the one-stage approach can be more flexible for modelling patient-level interactions and may be more efficient in some settings. So the two differ in whether the studies are summarised first and then pooled or modelled jointly in one step, both being approaches to individual patient data meta-analysis that often give similar results.

Source: Stewart & Parmar 1993

The advantages of two-stage meta-analysis include its relative simplicity and transparency, since the second stage uses familiar meta-analysis methods and produces familiar outputs such as forest plots; its clear separation of within-study and between-study analysis, which makes heterogeneity easy to examine; and its avoidance of some complex modelling choices required by the one-stage approach. So two-stage meta-analysis is advantageous where a straightforward, transparent synthesis is wanted, giving results comparable to conventional meta-analysis while using patient-level data within each study, which is why it is a common and accessible way to conduct an individual patient data meta-analysis.

Source: Stewart & Parmar 1993

Two-stage meta-analysis is appropriate when patient-level data are available and a transparent, conventional-style synthesis is wanted, and when each study has enough data to be analysed reliably on its own to produce a stable summary estimate. Where studies are very small or the interest lies in modelling patient-level interactions in detail, the one-stage approach may be preferred. So two-stage meta-analysis is appropriate for many individual patient data syntheses, particularly where simplicity and familiar outputs are valued and the studies are individually analysable, with the choice between two-stage and one-stage depending on the data and the questions being addressed.

Source: Stewart & Parmar 1993