Assessing subgroup effects with binary data: can the use of different effect measures lead to different conclusions?

Area of Science:

• Biostatistics
• Clinical Trials
• Epidemiology

Background:

• Understanding subgroup effects in randomized trials is vital for result applicability.
• Statistical tests for interaction guide decisions on whether benefits/harms differ across subgroups.
• Different effect measures (e.g., risk ratios, odds ratios) produce varying interaction test results for binary outcomes.

Purpose of the Study:

• To investigate why different effect measures yield different statistical interaction tests for binary outcomes.
• To clarify the implications of these differences for interpreting randomized trial results.
• To provide guidance on selecting appropriate interaction tests in clinical research.

Main Methods:

• Analysis of statistical interaction tests across different effect measures.
• Examination of the UK Hip trial data as a case study.
• Development of a graphical technique to illustrate differences in interaction tests.

Main Results:

• Interaction tests on different effect measures assess distinct null hypotheses.
• Differences in interaction test results are pronounced when subgroup risks vary considerably.
• The choice of effect measure significantly influences the outcome of interaction tests (e.g., P=0.14 vs. P<0.001 in the UK Hip trial).

Conclusions:

• The interaction test serves as a crucial check on the generalizability of trial findings to all subgroups.
• The interaction test should be applied to the effect measure least likely to show a priori interaction.
• Pre-specification and clinical knowledge should guide the choice of interaction test, particularly when outcome risks differ widely between subgroups.