Table 2.

Guidelines for performing subgroup analyses that build upon those provided by Yusuf et al. (1991)

Design

State plausible subgroup hypotheses and note which are defined by post-randomization features

Rank hypotheses in order of plausibility

Calculate power. Consider adjusting the design if necessary

State methods for analysis and be specific (i.e., include functional form of all relevant variables—continuous or categorical and if categorical, specify the cut-offs)

State conclusions that can be drawn from the analysis plan and any resulting decisions that may occur as a consequence of the findings

Analysis

Use tests of interaction to formally assess heterogeneity of effects

Distinguish between a priori and data-driven hypotheses. Do not present p values for data-driven hypotheses

Adjust for multiplicity for a priori subgroup analyses

If post-randomization:

• Consider causal inference tools

• Consider method for incorporating time into model

• Consider sensitivity analyses, where several models are fit and results compared across models

Interpretation and reporting

Report findings corresponding to primary hypothesis

Report the number of a priori hypotheses tested

Report the number of data-driven hypotheses examined

Interpret findings in the context of previous studies and/or similar data from other trials, and based on biological plausibility

Consider pooling findings for subgroup analyses with other studies

Consider external data set where methods can be applied to replicate findings and/or provide code for fitting model, particularly if a post-randomization analysis was conducted so that other investigators can more easily replicate