Subgroup analysis

  • Separating patients into subgroups and then performing analysis within the subgroup is how observational studies are performed.
  • RCTs, however, run into trouble with subgroup analysis.
    • There should not be too many subgroups.
    • Subgroups should be identifiable by features apparent before randomisation.
    • There should be some test of interaction, rather than within-subgroup analysis.
    • each subgroup may be underpowered to detect the desired effect size
    • If enough subgroup analyses are performed, eventually one will be positive by chance alone.
    Thus, even though a well-designed clinical trial may have validity, its subgroup analysis may not.

How do I make my subgroup analysis appear more credible?

  • Use within-study comparison rather than within-subgroup comparison
  • Specify the subgroups in the study design, rather than post-hoc
  • It helps if the effect size is large and if there is compelling biological rationale.