Optimising cohort data in Europe

y y One-stage approach can simultaneously model an outcome value at multiple time points using all the data. The two-stage approach does not automatically account for the correlation between multiple points, the second stage treats each point as independent, ignoring the correlation. y y Treatment-covariate interactions by the two-stage approach is an important methodological approach, which is used in personalized medicine, and guidance for using such analysis has been extensively reviewed (Donegan et al., 2015). Sharing and pooling data from different cohort studies is not always possible due to ELSI (Ethical, Legal, and Societal Issues) considerations such as the absence of a legal basis for data processing (e.g. informed consent of the research participant). Federated analysis can be a solution, because it usually only shares metadata (i.e. not affected by informed consent requirements) instead of the individual raw data (Banerjee et al., 2022). Therefore, the owners of each study can pool the metadata to form a common scheme of harmonisation between all parties. On the other hand, web-based network technologies and new database management systems are presently opening up a solution to avoid restrictions on sharing individual data. It is now possible that different databases previously harmonised with common metadata can be interconnected through database federation systems. These systems make it possible to ensure that analyses are secure while retaining individual data within the institutions that own the participating studies.

Differential features of the three types of analytical approaches are shown in Table 1.

Table 1. Relevant characteristics and conclusions of the three types of analytical approaches

Characteristics

Pooled analysis Meta-analysis

Federated analysis

Data at individual level should be shared

Yes

No

No

Harmonised data are required

Yes

No

Yes

How heterogeneity is checked across study cohorts

Tests to assess between-study heterogeneity

Models with fixed or random effects

The same as previous approaches

Higher statistical power and better modelling of the combined data

It explicitly accounts for between-study heterogeneity

It overcomes technical, ethical and legal issues in sharing data

Main advantage